paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1309.7524 | 3 | 1309 | 2019-07-09T20:06:48 | Meme and Variations: A Computer Model of Cultural Evolution | [
"cs.MA",
"cs.NE"
] | Holland's (1975) genetic algorithm is a minimal computer model of natural selection that made it possible to investigate the effect of manipulating specific parameters on the evolutionary process. If culture is, like biology, a form of evolution, it should be possible to similarly abstract the underlying skeleton of the process and develop a minimal model of it. Meme and Variations, or MAV, is a computational model, inspired by the genetic algorithm, of how ideas evolve in a society of interacting individuals (Gabora 1995). The name is a pun on the classical music form 'theme and variations', because it is based on the premise that novel ideas are variations of old ones; they result from tweaking or combining existing ideas in new ways (Holland et al. 1981). MAV explores the impact of biological phenomena such as over-dominance and epistasis as well as cognitive and social phenomena such as the ability to learn generalizations or imitate others on the fitness and diversity of cultural transmissible actions. | cs.MA | cs | Gabora, L. (1995). Meme and Variations: A comp uter model of cultural evolution. In L. Nadel & D. L.
Stein (Eds.), 1993 Lectures in complex systems (pp. 471-486). Boston: Addison-Wesley.
Meme and Variations:
A Computer Model of Cultural Evolution
Liane Gabora
Holland's (1975) genetic algorithm is a minimal computer model of natural selection that made it
possible to investigate the effect of manipulating specific parameters on the evolutionary process.
If culture is, like biology, a form of evolution, it should be possible to similarly abstract the
underlying skeleton of the process and develop a minimal model of it. Meme and Variations, or
MAV, is a computational model, inspired by the genetic algorithm, of how ideas evolve in a
society of interacting individuals (Gabora 1995). The name is a pun on the classical music form
'theme and variations', because it is based on the premise that novel ideas are variations of old
ones; they result from tweaking or combining existing ideas in new ways (Holland et al. 1981).
MAV explores the impact of several phenomena that are unique to culture. These are introduced
briefly here, and the technical details of how they are implemented will be presented shortly.
The first is knowledge-based operators. Brains detect regularity and build schemas with which
they adapt the mental equivalents of mutation and recombination to the constraints of their world,
and the situation at hand. Thus they generate novelty strategically, on the basis of past
experience. Knowledge-based operators are a crude attempt to incorporate this into the model.
The second crudely implemented cultural phenomenon is imitation. Ideas for how to perform
actions spread when agents copy what one another is doing. This enables them to share solutions
to the problems they face.
The third is mental simulation; agents can 'imagine', or guess, how successful an idea would be if
it were implemented before they actually commit to implementing it. This provides them with a
rudimentary form of selection before the phenotypic expression of an idea. Once again, the way
this phenomenon is implemented is not terribly life-like, but the goal here was to abstract the
essence of the phenomenon and see how it affects the evolution of culture.
Every iteration, each neural-network based agent in an artificial society has the opportunity to
acquire a new idea, either through 1) innovation, by changing a previously learned idea, or 2)
imitation, by copying what a neighbor is doing. Thus, variation occurs through mutation of pre-
existing ideas, selection occurs through choice of which pre-existing idea to mutate, and how to
mutate it, and ideas spread through imitation.
1. THE MODEL
Since the model is artificially limited with respect to the number and kinds of features an idea can
have, it does not hurt in this case to adopt the terminology of biology. Thus the features or
components of an idea are referred to as loci, and alternative forms of a locus are referred to as
1
alleles. The processes that generate variation are referred to as mutation operators. Forward
mutation is mutation away from the original (or, as biologists refer to it, the wild type) allele, and
backmutation is mutation from an alternative form back to the original. An individual is referred
to as an agent, and the set of all agents is referred to as the society.
1.1 The Domain
Donald (1991) has provided substantial evidence that the earliest culture took the form of
physical actions, such as displays of aggression or submission. The ideas in MAV can be thought
of as mating displays. An idea is a pattern consisting of six loci that dictate the degree of
movement for six body parts: left arm, right arm, left leg, right leg, head, and tail. Each locus has
a floating point activation between -0.5 and 0.5 which determines the amount of movement
(angle of rotation from rest position) of the corresponding body part when the idea is
implemented. A value of 0.0 corresponds to rest position; values above 0.0 correspond to upward
movement, and values below 0.0 correspond to downward movement. Floating point loci
activations produce graded limb movement. However, for the purpose of mutation, loci are
treated as if there are only three possible alleles at each locus: stationary, up, and down. Six loci
with three possible alleles each gives a total of 729 possible ideas.
1.2 The Neural Network
The neural network is an autoassociator; it learns the identity function between input and output
patterns. It has six input/output units numbered 1 through 6, corresponding to the six body parts.
It has six hidden units numbered 7 through 12, corresponding to the general concepts, "arms",
"legs", "left", "right", "movement", and "symmetry" (Figure 1).
Figure 1. The neural network. Arrows represent connections with positive weights. For clarity,
negative connections and connections to the symmetry unit are not shown.
Hidden units are linked with positive weights to input/output units that are positive instances of
the concepts they represent, and linked with negative weights to input/output units that represent
2
negative instances of the ideas they represent (thus "left arm" is positively linked to "left" and
negatively linked to "right"). Hidden units that represent opposite concepts have negative
connections between them. The hidden units enable the network to encode the semantic structure
of an idea, and their activations are used to bias the generation of variation.
The neural network starts with small random weights between input/output nodes. Weights
between hidden nodes, and weights between hidden nodes and input/output nodes, are fixed at +/-
1.0. Patterns (representing ideas) are learned by training for 50 iterations using the generalized
delta rule (Rumelhart et. al. 1986) with a sigmoid activation function (Hinton 1981). The basic
idea is that patterns are repeatedly presented to the network, and each time, the network's actual
output is compared to the desired output. Since this neural network is an autoassociator, we
desire simply that the output be identical to the input. The relevant variables are:
aj = activation of unit j
tj = the jth component of the target pattern (the external input)
ωij = weight on line from unit i to unit j
β = 0.15
θ = 0.5
(1)
For the movement node, we use the absolute value of ai (since negative movement is not
possible; the least you can move is to not move at all).
The comparison between input and output involves computing an error term, which is used to
modify the pattern of connectivity in the network such that its responses become more correct.
The error signal, δj, is calculated such that the more intermediate (and thus 'uncommitted's) the
activation of the unit, the more it is affected. For input/output units it is computed as follows:
(2)
For hidden units we do the same thing by determining how much each unit is contributing to the
error as follows:
(3)
1.3 The Embodiment
The embodiment is a six-digit array specifying the behavior of the six body parts. While the
output of the neural network represents what the agent is thinking about, the embodiment
represents what it is actually doing. An idea cannot be observed and imitated by other agents
until it has been copied from the neural network to the embodiment and is implemented as an
action.
1.4 The Fitness Function
An optimal action is one where all body parts except the head are moving, and limb movement is
anti-symmetrical. (Thus if the left arm is moving up, the right arm is moving down, and vice
versa.) This is implemented as follows:
3
F = fitness
µ = 2.5
am = activation of movement hidden node
as = activation of symmetry hidden node
ah = activation of head node
i = 1 if ah = 0.0, otherwise i = 0
F = µam + 2µas + µi (4)
This fitness function corresponds to a relatively realistic display, but it also has some
interesting properties. An agent that develops the general rule "movement improves fitness" risks
overgeneralization since head stability contributes as much to fitness as movement at every other
limb. This creates a situation that is the cultural analog of overdominance in genetics; the optimal
value of this locus lies midway between the two extremes. We also have a situation analogous to
bidirectional selection or underdominance; the optimal value of the tail locus lies at either of the
two extremes. (The desire to investigate underdominance was the reason for giving the agents
tails). There is a cultural analog of epistasis-where the fitness at one locus depends on which
allele is present at another locus. Epistasis is present because the value of what one limb is doing
depends on what its counterpart is doing; for example, if the left leg is moving backward the right
should be moving forward, and vice versa. Finally, since there is one optimal allele for the head,
two optimal alleles for the tail, two optimal arm combinations, and two optimal leg combinations,
we have a total of eight different optimal actions. This enables us to perform a comparative
analysis of diversity under different ratios of creation to imitation.
1.5 Using Experience to Bias the Generation of Novelty
The idea here is to translate knowledge acquired during evaluation of an action into educated
guesses about what increases fitness. Each locus starts out with the allele for no movement, and
with an equal probability of mutating to each of the other two alleles (those for upward and
downward movement). A new action is not learned unless it is fitter than the currently-
implemented action, so we use the difference between the two to bias the direction of mutation.
Two rules of thumb are used. The first rule is: if the fitter action codes for more movement,
increase the probability of forward mutation and decrease the probability of back mutation. Do
the opposite if the fitter action codes for less movement. This rule of thumb is based on the
assumption that movement in general (regardless of which particular body part is moving) can be
beneficial or detrimental. This seems like a useful generalization since movement of any body
part uses energy and increases the likelihood of being detected. It is implemented as follows:
am1 = activation of movement unit for currently-implemented meme
am2 = activation of movement unit for new meme
p(fmut)i = probability of forward mutation at allele i (increased movement)
p(bmut)i = probability of backward mutation at allele i (decreased movement)
IF (am2 > am1)
THEN p(fmut)i = MAX(1.0, p(fmut)i + 0.1)
ELSE IF (am2 < am1)
THEN p(fmut)i= MIN(0.0, p(fmut)i - 0.1)
p(bmut)i = 1 - p(fmut)i
4
The second rule of thumb biases the agent either toward or away from symmetrical limb
movement. It has two parts. First, if in the fitter action both members of one pair of limbs are
moving either up or down, increase the probability that you will do the same with the other pair
of limbs. Second, if in the fitter action, one member of a pair of limbs is moving in one direction
and its counterpart is moving in the opposite direction, increase the probability that you will do
the same with the other pair of limbs. This generalization is also biologically useful, since many
beneficial behaviors (walking, etc.) entail movement of limbs in opposite directions, while others
(galloping, etc.) entail movement of limbs in the same direction. The implementation of this rule
is analogous to that of the first rule.
In summary, each action is associated with a measure of its effectiveness, and
generalizations about what seems to work and what does not are translated into guidelines that
specify the behavior of the cultural algorithm.
2. PROTOCOL
Agents are in a two-dimensional wrap-around 10x10 grid-cell world, one agent per cell. Each
iteration, every agent has the opportunity to (1) acquire an idea for a new action, either by
imitating a neighbor, or creating one anew (2) update the mutation operator, and (3) implement
the new action.
Agents have an equal probability of creating and imitating. To invent or create a new idea,
the cultural algorithm is applied to the idea currently represented on the input/output layer of the
neural network. For each locus, the agent decides whether mutation will take place. The
probability of mutation is specified globally at the beginning of a run. If it decides to mutate, the
direction of mutation is stochastically determined. If the new idea has a higher fitness than the
currently-implemented idea, the agent learns and implements the action specified by that idea.
To acquire an idea through imitation, an agent randomly chooses one of its eight neighbors
and evaluates the fitness of the action the neighbor is implementing. If its own action is fitter than
that of the neighbor, it chooses another neighbor, until it has either observed all eight neighbors,
or found one with a fitter action. If no fitter action is found, the agent does nothing. Otherwise,
the neighbor's action is copied to the input/output layer, and it is learned and implemented.
Since in both creation and imitation, a new idea is not acquired unless it is fitter than the
currently implemented action, the new idea provides information that is used by the cultural
algorithm. For example, since we arbitrarily chose a fitness function in which movement is
generally beneficial, if the new action does code for more movement than the old one, the
probability of forward mutation will almost always increase.
No matter how the new idea has been acquired, it gets implemented as an action when it is
copied from the neural network to the embodiment. In the 'no mental simulation' condition,
whether the new idea was acquired through creation or imitation, it must be implemented as an
action for at least one iteration before its fitness can be assessed. In this case, mutation operators
are updated the following iteration.
3. RESULTS
5
The following experiments were conducted using a mutation rate of 0.17 per locus, a 1:1 creation
to imitation ratio, and all cultural evolution strategies operative, unless otherwise indicated.
3.1 Outline of a Run: Culture Evolves
Initially all agents were immobile, thus the number of different actions implemented was zero, as
shown in Figure 2. The immobility idea quickly mutated to a new idea that coded for movement
of a body part. This new idea had a higher fitness and was preferentially implemented. As ideas
continued to be created, get implemented as actions, and spread through imitation, the society
evolved toward increasingly fit actions.
Figure 2. Effect of varying p(create) to p(imitate) ratio on number of different ideas.
3.2 Trade-off Between Diversity and Global Optimization
Figure 2 shows how diversity peaked when the first maximally fit idea was found, and decreased
as the society converged on maximally-fit ideas. As in a genetic algorithm, increasing the
frequency of variation-inducing operations-in this case, the creativity to imitation ratio-increased
diversity. This was true both as the society was evolving, and when it finally stabilized. An
interesting result can be seen if one looks more closely at how diversity varied with the creation
to imitation ratio. Diversity ranged from 1-2 actions when p(create) = 0.25, to 10-11 actions when
p(create) = 1.0. When p(create) = 0.75, the society converged on 7-8 actions. Thus it found all (or
nearly all) of the fittest actions. A nice balance was thereby struck between the diversifying effect
of creation and the converging effect of imitation.
6
3.3 Frequency of Change Must be Intermediate
Agents could vary with respect to not only just the frequency with which they invented, but with
respect to how much change they introduced when they did. As in a genetic algorithm, evolution
did not occur in the complete absence of mutation. The best performance was obtained with
between 0.07 and 0.22 mutations per locus (Figure 3). In biological terms this would constitute a
very high mutation rate, however because there are so few loci it ends up being approximately
one mutation per innovation event, which is very reasonable. There are however two reasons why
it can afford to be high. The first is that mental simulation ensures that unfit ideas are not
implemented as actions. The second is that fit actions are imitated by others, thus never lost from
the society as a whole.
Figure 3. Optimization time decreases sharply, and then increases slowly, as mutation rate
increases. This trend holds true for both the mean fitness of all ideas and the fittest idea that has
appeared in a given iteration.
Note that this is similar to what happens in a genetic algorithm when the mutation rate is much
above the minimum necessary for evolution. In both MAV and the GA, the frequency of change
must be intermediate between the two extremes.
3.4 Epistasis Decreases Rate of Fitness Increase
As in biology, fitness increased more slowly, and stabilization took longer, for epistatically
7
linked loci than for either over- or underdominant loci Figures 4 and 5 show this at two different
mutation rates; in fact, it was observed at every mutation rate tested. In figure 4, we see that
whereas the over-dominant locus had stabilized by the 100th iteraction, and the under-dominant
by the 150th, the two epistatic loci took 200 iterations to stabilize. In figure 5 whereas the
overdominant locus stabilizes by the 70th iteration, and the under-dominant by approximately
iteration 130, the two epistatic loci never do manage to stabilize. The phenomenon can be
attributed to the fact that for epistatically linked loci there are more constraints to be met; what
one arm should be doing, for example, depends on what the other arm is doing. It is also
interesting to note that the epistatic loci to some extent mirror one another, such that if the mean
activation for one locus increases, the mean activation for the epistatically linked locus decreases,
and vice versa.
3.5 Cultural Drift
Since we have eight optimal ideas, there are many stable configurations for the distribution of
ideas. Figures 4 and 5 reveal amongst equally-fit alleles the presence of drift -- the term
biologists use to refer to changes in the relative frequencies of different alleles due to random
Figure 4. Mean loci activations with a mutation rate of 0.01. Over-dominant locus stabilizes most
quickly, followed by under-dominant locus, and then epistatic loci.
8
sampling processes in a finite population (Wright, 1969). Drift is indicated by the fact that since
we are looking at mean activation values across the entire society, if the activation value is very
high or very low that means that almost all agents had stabilized on the same value at a particular
locus. This is the case for both the over-dominant and epistatic loci. (For the under-dominant
locus it is not possible to distinguish between all individuals stabilized on the intermediate value,
or as many stabilized above this value as below.) This is in accord with Cavalli-Sforza and
Feldman's (1981) mathematical model of culture.
Figure 5. Mean loci activations with a mutation rate of 0.67. Over-dominant locus stabilizes most
quickly. Under-dominant locus takes longer. After 200 iterations, epistatic loci still have not
stabilized.
3.6 Effect of Knowledge-based Operators, Imitation, and Mental Simulation
The three cultural evolution strategies-knowledge-based operators, imitation, and mental
simulation-were made inoperative one at a time to determine their contribution to optimization
speed and peak mean fitness. These experiments were performed separately, and also
cumulatively, adding each strategy one at a time. Since, the results were comparable for the
separate and cumulative experiments, only the cumulative results are presented here (Figure 6).
9
Figure 6. Cumulative improvement with successively applied cultural strategies. MS refers to
mental simulation. 'Imit' refers to imitation. 'Kmut' refers to knowledge based operators.
All three cultural evolution strategies increased the rate of optimization. Mental simulation and
imitation also increased peak mean fitness.
3.7 Fittest Society with Creation to Imitation Ratio of 2:1
The highest mean fitness was achieved when both creation and imitation were employed, as
illustrated in Figure 7. The best performance was observed when the creativity to imitation ratio
was either 1:1 or 3:1. The society with a 3:1 ratio improved most quickly until the 21st iteration,
at a mean fitness of 8.6. The society with a 1:1 ratio then overtook it and converged entirely on
optimal actions slightly earlier than the 3:1 society (32 iterations as opposed to 37). Thus it can
be said that overall the optimal creativity to imitation ratio is most likely midway between these
values, or approximately 2:1. The 1:3 ratio society took longer to converge, though it did
eventually after 47 iterations. The society that just created and never imitated had the opposite
problem. It did as well as the 1:1 and 3:1 ratio societies for the first ten iterations or so, at which a
mean fitness of 6.2 was attained. However, after that, its performance dropped compared to the
others, such that even after 100 generations it never did converge on optimal solutions. As might
be expected, when there was no creativity, just imitation, then there was no evolution at all; the
agents simply remained immobile.
10
Figure 7. Effect of varying the creation-to-imitation ratio. The optimum seems to be between 0.5
and 0.75.
However, it is interesting to note that the fitness of the fittest idea (Figure 8) increased as a
function of the ratio of creation to imitation. Since the agents with the fittest ideas gain nothing
by imitating others, there is a trade-off between average action fitness, and fitness of the fittest
action. (Of course, this result should not be taken too seriously as indicating that smart people
don't need to imitate, since the agents in MAV only had one problem to solve. Thus, those who
happened to be lucky simply got a head start.)
11
Figure 8. Effect of varying p(create) to p(imitate) ratio on the fitness of the fittest idea.
4. COMPARISON WITH OTHER APPROACHES
The computational approach taken here allows us to look for patterns that arise over time when
the cultural activities of inventing and imitating are carried out in parallel in a society of
interacting individuals. To my knowledge, MAV is the first computational model of the evolution
of culture as opposed to the effect of culture on biological evolution. For example, Hutchins and
Hazelhurst (1992) used a computer model to explore the relationship between environment,
internal representation of the environment, and cultural artifacts that mediate the transmission of
knowledge about environmental regularity vertically across generations. In MAV, on the other
hand, we stick to one generation so that the effects of cultural evolution can be disentangled from
biological evolution. Similarly, computer models of the evolution of creativity (Sims, 1991; Todd
& Latham, 1992) and cooperation (Axelrod, 1985), although they explore a cultural process, they
use a genetic algorithm-a model of biological evolution. Axelrod's work has inspired others, who
have taken a more culturally realistic approach (e.g., Conte & Paolucci forthcoming; Hales
forthcoming), but in these studies the space of possible cultural entities is too small to evolve, so
the study is really a study of how particular cooperative strategies diffuse across a population of
agents. Note, however, that although MAV is modeled after cultural evolution, even it is too
simple to explore many cultural phenomena. The space of possible ideas is fixed and small, and
12
(unlike real life) the fitness function is predetermined and static. It would be particularly
interesting to explore the effects of a dynamically changing fitness function given recent findings
that maintaining diversity in a population is, over the long term, more important than global
fitness in the short term (Hutter, 2001). Another shortcoming of MAV, imitation and innovation
are not as discrete in real life as MAV would suggest. Despite these shortcomings, however,
MAV demonstrates the feasibility of computationally modeling the processes by which creative
ideas spread through a society giving rise to observable patterns of cultural diversity.
5. SUMMARY
If culture, like biology, is a form of evolution, it should be possible to abstract the underlying
skeleton of the process and develop a minimal model of it analogous to the genetic algorithm.
MAV is a minimal computer model of the process by which culture evolves. It consists of an
artificial society of neural network-based agents that don't have genomes, and neither die nor
have offspring, but they invent, imitate, and implement ideas, and thereby their actions gradually
become more fit. Every iteration, each neural-network based agent in an artificial society has the
opportunity to acquire a new idea, either through 1) innovation, by mutating a previously learned
idea, or 2) imitation, by copying an action performed by a neighbor.
The program exhibits features observed in biology such as: (1) drift, (2) epistasis increases
time to reach equilibrium, (3) increasing the frequency of innovative or variation-generating
operations increases diversity, and (4) although in the absence of variation-generating operations,
culture does not evolve, increasing innovation much beyond the minimum necessary for
evolution causes average fitness to decrease.
The model also addresses the evolutionary consequences of phenomena specific to culture.
Imitation, mental simulation (ability to assess the relative fitness of an action before actually
implementing it), and strategic (as opposed to random) generation of variation all increase the
rate at which fitter actions evolve. The higher the ratio of innovation to imitation, the greater the
diversity, and the higher the fitness of the fittest action. Interestingly, however, for the society as
a whole, the optimal innovation-to-imitation ratio was approximately 2:1 (but diversity is then
compromised). For the agent with the fittest behavior, the less it imitated (i.e. the more
computational effort reserved for innovation), the better.
REFERENCES
Axelrod, R. Modeling the Evolution of Norms. Paper presented at the American Political Science
Association Meeting, New Orleans, 1985.
Bonner, J. T. The Evolution of Culture in Animals. Princeton: Princeton University Press, 1980.
Boyd, R. and P. J. Richerson. Culture and the Evolutionary Process. Chicago: The University of
Chicago Press, 1985.
Braitenberg, V. Vehicles. Cambridge: MIT Press, 1984.
Cavalli-Sforza, L. L. and M. W. Feldman. Cultural Transmission and Evolution: A Quantitative
Approach. Princeton: Princeton University Press, 1981.
Dawkins, R. The Selfish Gene. Oxford: Oxford University Press, 1976.
13
Donald, M. Origins of the Modern Mind. Cambridge: Harvard University Press, 1991.
Hinton, G. E. Implementing Semantic Networks in Parallel Hardware. In Parallel Models of
Associative Memory. Hillsdale NJ: Erlbaum Press, 1981.
Holland, J. K. Adaptation in Natural and Artificial Systems. Ann Arbor: University of Michigan
Holland, J. H., K. J. Holyoak, R. E. Nisbett, and P. R. Thagard. Induction. Cambridge: MIT
Press, 1975.
Press, 1986.
1981.
1969.
Hutchins, E. and B. Hazelhurst. Learning in the Cultural Process. In Artificial Life II. Redwood
City, CA: Addison-Wesley, 1992.
Lumsden, C. and E. O. Wilson. Genes, Mind, and Culture. Cambridge: Harvard University Press,
Robert, M. Observational Learning in Fish, Birds, and Mammals: A Classified Bibliography
Spanning over 100 Years of Research. Psych Record, 40 (1990): 289-311.
Smith, W. J. The Behavior of Communicating. Cambridge: Harvard University Press, 1977.
Wright, S. Evolution and the Genetics of Populations. Chicago: University of Chicago Press,
14
|
1604.04359 | 2 | 1604 | 2017-07-13T03:30:09 | Complexity of Manipulation with Partial Information in Voting | [
"cs.MA",
"cs.AI",
"cs.CC",
"cs.CY",
"cs.DS"
] | The Coalitional Manipulation problem has been studied extensively in the literature for many voting rules. However, most studies have focused on the complete information setting, wherein the manipulators know the votes of the non-manipulators. While this assumption is reasonable for purposes of showing intractability, it is unrealistic for algorithmic considerations. In most real-world scenarios, it is impractical for the manipulators to have accurate knowledge of all the other votes. In this paper, we investigate manipulation with incomplete information. In our framework, the manipulators know a partial order for each voter that is consistent with the true preference of that voter. In this setting, we formulate three natural computational notions of manipulation, namely weak, opportunistic, and strong manipulation. We say that an extension of a partial order is if there exists a manipulative vote for that extension.
1. Weak Manipulation (WM): the manipulators seek to vote in a way that makes their preferred candidate win in at least one extension of the partial votes of the non-manipulators.
2. Opportunistic Manipulation (OM): the manipulators seek to vote in a way that makes their preferred candidate win in every viable extension of the partial votes of the non-manipulators.
3. Strong Manipulation (SM): the manipulators seek to vote in a way that makes their preferred candidate win in every extension of the partial votes of the non-manipulators.
We consider several scenarios for which the traditional manipulation problems are easy (for instance, Borda with a single manipulator). For many of them, the corresponding manipulative questions that we propose turn out to be computationally intractable. Our hardness results often hold even when very little information is missing, or in other words, even when the instances are quite close to the complete information setting. | cs.MA | cs |
Complexity of Manipulation with Partial Information in
Voting
Palash Dey
Tata Institute of Fundamental Research, Mumbai
[email protected]
Neeldhara Misra
Y. Narahari
Indian Institute of Technology, Gandhinagar
Indian Institute of Science, Bangalore
[email protected]
[email protected]
October 11, 2018
Abstract
The Coalitional Manipulation problem has been studied extensively in the literature for many
voting rules. However, most studies have focused on the complete information setting, wherein
the manipulators know the votes of the non-manipulators. While this assumption is reason-
able for purposes of showing intractability, it is unrealistic for algorithmic considerations. In
most real-world scenarios, it is impractical to assume that the manipulators to have accurate
knowledge of all the other votes. In this work, we investigate manipulation with incomplete
information. In our framework, the manipulators know a partial order for each voter that is
consistent with the true preference of that voter.
In this setting, we formulate three natural
computational notions of manipulation, namely weak, opportunistic, and strong manipulation.
We say that an extension of a partial order is viable if there exists a manipulative vote for that ex-
tension. We propose the following notions of manipulation when manipulators have incomplete
information about the votes of other voters.
1. WEAK MANIPULATION: the manipulators seek to vote in a way that makes their preferred
candidate win in at least one extension of the partial votes of the non-manipulators.
2. OPPORTUNISTIC MANIPULATION:
the manipulators seek to vote in a way that makes
their preferred candidate win in every viable extension of the partial votes of the non-
manipulators.
3. STRONG MANIPULATION: the manipulators seek to vote in a way that makes their preferred
candidate win in every extension of the partial votes of the non-manipulators.
We consider several scenarios for which the traditional manipulation problems are easy (for
instance, Borda with a single manipulator). For many of them, the corresponding manipulative
questions that we propose turn out to be computationally intractable. Our hardness results often
hold even when very little information is missing, or in other words, even when the instances are
very close to the complete information setting. Our results show that the impact of paucity of
information on the computational complexity of manipulation crucially depends on the notion
of manipulation under consideration. Our overall conclusion is that computational hardness
continues to be a valid obstruction to manipulation, in the context of a more realistic model.
Keywords: voting, manipulation, incomplete information, algorithm, computational complexity
1
1 Introduction
In many real life and AI related applications, agents often need to agree upon a common decision
although they have different preferences over the available alternatives. A natural tool used in these
situations is voting. Some classic examples of the use of voting rules in the context of multiagent
systems include Clarke tax [ER91], collaborative filtering [PHG00], and similarity search [FKS03],
etc. In a typical voting scenario, we have a set of candidates and a set of voters reporting their
rankings of the candidates called their preferences or votes. A voting rule selects one candidate as
the winner once all voters provide their votes. A set of votes over a set of candidates along with
a voting rule is called an election. A central issue in voting is the possibility of manipulation. For
many voting rules, it turns out that even a single vote, if cast differently, can alter the outcome.
In particular, a voter manipulates an election if, by misrepresenting her preference, she obtains an
outcome that she prefers over the "honest" outcome. In a cornerstone impossibility result, Gibbard
and Satterthwaite [Gib73, Sat75] show that every unanimous and non-dictatorial voting rule with
three candidates or more is manipulable. We refer to [BCE+15] for an excellent introduction to
various strategic issues in computational social choice theory.
Considering that voting rules are indeed susceptible to manipulation, it is natural to seek ways
by which elections can be protected from manipulations. The works of Bartholdi et al. [BITT89,
BIO91] approach the problem from the perspective of computational intractability. They exploit
the possibility that voting rules, despite being vulnerable to manipulation in theory, may be hard to
manipulate in practice. Indeed, a manipulator is faced with the following decision problem: given
a collection of votes P and a distinguished candidate c, does there exist a vote v that, when tallied
with P, makes c win for a (fixed) voting rule r? The manipulation problem has subsequently been
generalized to the problem of COALITIONAL MANIPULATION by Conitzer et al. [CSL07], where one
or more manipulators collude together and try to make a distinguished candidate win the election.
The manipulation problem, fortunately, turns out to be NP-hard in several settings. This established
the success of the approach of demonstrating a computational barrier to manipulation.
However, despite having set out to demonstrate the hardness of manipulation, the initial results
in [BITT89] were to the contrary, indicating that many voting rules are in fact easy to manipulate.
Moreover, even with multiple manipulators involved, popular voting rules like plurality, veto, k-
approval, Bucklin, and Fallback continue to be easy to manipulate [XZP+09]. While we know that
the computational intractability may not provide a strong barrier [PR06, PR07, FKN08, XC08a,
XC08b, FHH10, Wal10, Wal11, IKM12, Dey15, DMN15b, DMN16, DMN15a, DN14, DN15] even for
rules for which the coalitional manipulation problem turns out to be NP-hard, in all other cases the
possibility of manipulation is a much more serious concern.
1.1 Motivation and Problem Formulation
In our work, we propose to extend the argument of computational intractability to address the
cases where the approach appears to fail. We note that most incarnations of the manipulation
problem studied so far are in the complete information setting, where the manipulators have com-
plete knowledge of the preferences of the truthful voters. While these assumptions are indeed the
best possible for the computationally negative results, we note that they are not reflective of typ-
ical real-world scenarios. Indeed, concerns regarding privacy of information, and in other cases,
the sheer volume of information, would be significant hurdles for manipulators to obtain complete
information. Motivated by this, we consider the manipulation problem in a natural partial infor-
2
mation setting. In particular, we model the partial information of the manipulators about the votes
of the non-manipulators as partial orders over the set of candidates. A partial order over the set
of candidates will be called a partial vote. Our results show that several of the voting rules that
are easy to manipulate in the complete information setting become intractable when the manipu-
lators know only partial votes. Indeed, for many voting rules, we show that even if the ordering
of a small number of pairs of candidates is missing from the profile, manipulation becomes an in-
tractable problem. Our results therefore strengthen the view that manipulation may not be practical
if we limit the information the manipulators have at their disposal about the votes of other voters
[CWX11].
We introduce three new computational problems that, in a natural way, extend the question of
manipulation to the partial information setting. In these problems, the input is a set of partial votes
P corresponding to the votes of the non-manipulators, a non-empty set of manipulators M, and a
preferred candidate c. The task in the WEAK MANIPULATION (WM) problem is to determine if there
is a way to cast the manipulators' votes such that c wins the election for at least one extension of
the partial votes in P. On the other hand, in the STRONG MANIPULATION (SM) problem, we would
like to know if there is a way of casting the manipulators' votes such that c wins the election in
every extension of the partial votes in P.
We also introduce the problem of OPPORTUNISTIC MANIPULATION (OM), which is an "interme-
diate" notion of manipulation. Let us call an extension of a partial profile viable if it is possible
for the manipulators to vote in such a way that the manipulators' desired candidate wins in that
extension. In other words, a viable extension is a YES-instance of the standard COALITIONAL MANIP-
ULATION problem. We have an opportunistic manipulation when it is possible for the manipulators
to cast a vote which makes c win the election in all viable extensions. Note that any YES-instance
of STRONG MANIPULATION is also an YES-instance of OPPORTUNISTIC MANIPULATION, but this may
not be true in the reverse direction. As a particularly extreme example, consider a partial profile
where there are no viable extensions: this would be a NO-instance for STRONG MANIPULATION, but
a (vacuous) YES-instance of OPPORTUNISTIC MANIPULATION. The OPPORTUNISTIC MANIPULATION
problem allows us to explore a more relaxed notion of manipulation: one where the manipula-
tors are obliged to be successful only in extensions where it is possible to be successful. Note that
the goal with STRONG MANIPULATION is to be successful in all extensions, and therefore the only
interesting instances are the ones where all extensions are viable.
It is easy to see that YES instance of STRONG MANIPULATION is also a YES instance of OPPOR-
TUNISTIC MANIPULATION and WEAK MANIPULATION. Beyond this, we remark that all the three
problems are questions with different goals, and neither of them render the other redundant. We
refer the reader to Figure 1 for a simple example distinguishing these scenarios.
All the problems above generalize COALITIONAL MANIPULATION, and hence any computa-
tional intractability result for COALITIONAL MANIPULATION immediately yields a corresponding
intractability result for WEAK MANIPULATION, STRONG MANIPULATION, and OPPORTUNISTIC MA-
NIPULATION under the same setting. For example, it is known that the COALITIONAL MANIPULA-
TION problem is intractable for the maximin voting rule when we have at least two manipula-
tors [XZP+09]. Hence, the WEAK MANIPULATION, STRONG MANIPULATION, and OPPORTUNISTIC
MANIPULATION problems are intractable for the maximin voting rule when we have at least two
manipulators.
3
Figure 1: An example of a partial profile. Consider the plurality voting rule with one manipulator.
If the favorite candidate is A, then the manipulator simply has to place A on the top of his vote
to make A win in any extension. If the favorite candidate is B, there is no vote that makes B win
in any extension. Finally, if the favorite candidate is C, then with a vote that places C on top, the
manipulator can make C win in the only viable extension (Extension 2).
1.2 Related Work
A notion of manipulation under partial
information has been considered by Conitzer et
al. [CWX11]. They focus on whether or not there exists a dominating manipulation and show
that this problem is NP-hard for many common voting rules. Given some partial votes, a dom-
inating manipulation is a non-truthful vote that the manipulator can cast which makes the win-
ner at least as preferable (and sometimes more preferable) as the winner when the manipulator
votes truthfully. The dominating manipulation problem and the WEAK MANIPULATION, OPPOR-
TUNISTIC MANIPULATION, and STRONG MANIPULATION problems do not seem to have any apparent
complexity-theoretic connection. For example, the dominating manipulation problem is NP-hard
for all the common voting rules except plurality and veto, whereas, the STRONG MANIPULATION
problem is easy for most of the cases (see Table 1). However, the results in [CWX11] establish the
fact that it is indeed possible to make manipulation intractable by restricting the amount of infor-
mation the manipulators possess about the votes of the other voters. Elkind and Erd´elyi [EE12]
study manipulation under voting rule uncertainty. However, in our work, the voting rule is fixed
and known to the manipulators.
Two closely related problems that have been extensively studied in the context of incomplete
votes are POSSIBLE WINNER and NECESSARY WINNER [KL05]. In the POSSIBLE WINNER problem,
we are given a set of partial votes P and a candidate c, and the question is whether there ex-
ists an extension of P where c wins, while in the NECESSARY WINNER problem, the question
is whether c is a winner in every extension of P. Following the work in [KL05], a number
of special cases and variants of the POSSIBLE WINNER problem have been studied in the litera-
ture [CLMM10, BBF10, BRR11, BRR+12, GNNW14, XC11, DL13, NW14, BFLR12, ML15]. The
flavor of the WEAK MANIPULATION problem is clearly similar to POSSIBLE WINNER. However, we
emphasize that there are subtle distinctions between the two problems. A more elaborate compar-
ison is made in the next section.
4
1.3 Our Contribution
Our primary contribution in this work is to propose and study three natural and realistic general-
izations of the computational problem of manipulation in the incomplete information setting. We
summarize the complexity results in this work in Table 1. Our results provide the following inter-
esting insights on the impact of lack of information on the computational difficulty of manipulation.
We note that the number of undetermined pairs of candidates per vote are small constants in all
our hardness results.
(cid:66) We observe that the computational problem of manipulation for the plurality and veto vot-
ing rules remains polynomial time solvable even with lack of information, irrespective of the
notion of manipulation under consideration [Proposition 1, Theorem 11 and 15, and Obser-
vation 4]. We note that the plurality and veto voting rule also remain vulnerable under the
notion of dominating manipulation [CWX11].
(cid:66) The impact of absence of information on the computational complexity of manipulation is
more dynamic for the k-approval, k-veto, Bucklin, Borda, and maximin voting rules. Only
the WEAK MANIPULATION and OPPORTUNISTIC MANIPULATION problems are computationally
intractable for the k-approval [Theorem 1 and 5], k-veto [Theorem 2 and 6], Bucklin [Theo-
rem 3 and 10], Borda [Observation 3 and Theorem 7], and maximin [Observation 3 and The-
orem 8] voting rules, whereas the STRONG MANIPULATION problem remains computationally
tractable [Theorem 11 to 14].
(cid:66) Table 1 shows an interesting behavior of the Fallback voting rule. The Fallback voting rule is
the only voting rule among the voting rules we study here for which the WEAK MANIPULATION
problem is NP-hard [Theorem 3] but both the OPPORTUNISTIC MANIPULATION and STRONG
MANIPULATION problems are polynomial time solvable [Theorem 13 and Observation 4]. This
is because the OPPORTUNISTIC MANIPULATION problem can be solved for the Fallback voting
rule by simply making manipulators vote for their desired candidate.
(cid:66) Our results show that absence of information makes all the three notions of manipulations
intractable for the Copelandα voting rule for every rational α ∈ [0, 1] \ {0.5} for the WEAK
MANIPULATION problem [Observation 3] and for every α ∈ [0, 1] for the OPPORTUNISTIC
MANIPULATION and STRONG MANIPULATION problems [Theorem 4 and 9].
Our results (see Table 1) show that whether lack of information makes the manipulation prob-
lems harder, crucially depends on the notion of manipulation applicable to the situation under
consideration. All the three notions of manipulations are, in our view, natural extension of manip-
ulation to the incomplete information setting and tries to capture different behaviors of manipula-
tors. For example, the WEAK MANIPULATION problem maybe applicable to an optimistic manipula-
tor whereas for an pessimistic manipulator, the STRONG MANIPULATION problem may make more
sense.
Organization of the paper: We define the problems and introduce the basic terminology in the
next section. We present our hardness results in Section 3. In Section 4, we present our polynomi-
ally solvable algorithms. Finally we conclude with future directions of research in Section 5.
5
WM, (cid:96) = 1 WM OM, (cid:96) = 1 OM SM, (cid:96) = 1
SM
Plurality
Veto
k-Approval
k-Veto
Bucklin
Fallback
Borda
maximin
Copelandα
P
P
NP-complete
coNP-hard
P
NP-complete
coNP-hard
P
P
NP-hard
coNP-hard
Table 1: Summary of Results ((cid:96) denotes the number of manipulators). The results in white follow
immediately from the literature (Observation 1 to 3). Our results for the Copelandα voting rule
hold for every rational α ∈ [0, 1]\{0.5} for the WEAK MANIPULATION problem and for every α ∈ [0, 1]
for the OPPORTUNISTIC MANIPULATION and STRONG MANIPULATION problems.
2 Preliminaries
In this section, we begin by providing the technical definitions and notations that we will need in
the subsequent sections. We then formulate the problems that capture our notions of manipulation
when the votes are given as partial orders, and finally draw comparisons with related problems
that are already studied in the literature of computational social choice theory.
2.1 Notations and Definitions
Let V = {v1, . . . , vn} be the set of all voters and C = {c1, . . . , cm} the set of all candidates. If not
specified explicitly, n and m denote the total number of voters and the total number of candidates
respectively. Each voter vi's vote is a preference (cid:31)i over the candidates which is a linear order over
C. For example, for two candidates a and b, a (cid:31)i b means that the voter vi prefers a to b. We
denote the set of all linear orders over C by L(C). Hence, L(C)n denotes the set of all n-voters'
preference profile ((cid:31)1, . . . ,(cid:31)n). A map r : ∪n,C∈N+L(C)n −→ 2C \ {∅} is called a voting rule. For
some preference profile (cid:31)∈ L(C)n, if r((cid:31)) = {w}, then we say w wins uniquely and we write
r((cid:31)) = w. From here on, whenever we say some candidate w wins, we mean that the candidate w
wins uniquely. For simplicity, we restrict ourselves to the unique winner case in this paper. All our
proofs can be easily extended in the co-winner case.
A more general setting is an election where the votes are only partial orders over candidates.
A partial order is a relation that is reflexive, antisymmetric, and transitive. A partial vote can be
extended to possibly more than one linear vote depending on how we fix the order of the unspeci-
fied pairs of candidates. For example, in an election with the set of candidates C = {a, b, c}, a valid
partial vote can be a (cid:31) b. This partial vote can be extended to three linear votes namely, a (cid:31) b (cid:31) c,
6
a (cid:31) c (cid:31) b, c (cid:31) a (cid:31) b. In this paper, we often define a partial vote like (cid:31) \A, where (cid:31)∈ L(C)
and A ⊂ C × C, by which we mean the partial vote obtained by removing the order among the pair
of candidates in A from (cid:31). Also, whenever we do not specify the order among a set of candidates
while describing a complete vote, the statement/proof is correct in whichever way we fix the order
among them. We now give examples of some common voting rules.
(cid:66) Positional scoring rules: An m-dimensional vector (cid:126)α = (α1, α2, . . . , αm) ∈ Rm with α1 (cid:62)
α2 (cid:62) . . . (cid:62) αm and α1 > αm naturally defines a voting rule -- a candidate gets score αi
from a vote if it is placed at the ith position, and the score of a candidate is the sum of the
scores it receives from all the votes. The winners are the candidates with maximum score.
Scoring rules remain unchanged if we multiply every αi by any constant λ > 0 and/or add
any constant µ. Hence, we assume without loss of generality that for any score vector (cid:126)α, there
exists a j such that αj − αj+1 = 1 and αk = 0 for all k > j. We call such a (cid:126)α a normalized score
vector. For (cid:126)α = (m − 1, m − 2, . . . , 1, 0), we get the Borda voting rule. With αi = 1 ∀i (cid:54) k
and 0 else, the voting rule we get is known as k-approval. For the k-veto voting rule, we have
αi = 0 ∀i (cid:54) m − k and −1 else. Plurality is 1-approval and veto is 1-veto.
(cid:66) Bucklin and simplified Bucklin: Let (cid:96) be the minimum integer such that at least one can-
didate gets majority within top (cid:96) positions of the votes. The winners under the simplified
Bucklin voting rule are the candidates having more than n/2 votes within top (cid:96) positions. The
winners under the Bucklin voting rule are the candidates appearing within top (cid:96) positions of
the votes highest number of times.
(cid:66) Fallback and simplified Fallback: For these voting rules, each voter v ranks a subset Xv ⊂ C
of candidates and disapproves the rest of the candidates [BS09]. Now for the Fallback and
simplified Fallback voting rules, we apply the Bucklin and simplified Bucklin voting rules re-
spectively to define winners. If there is no integer (cid:96) for which at least one candidate gets more
than n/2 votes, both the Fallback and simplified Fallback voting rules output the candidates
with most approvals as winners. We assume, for simplicity, that the number of candidates
each partial vote approves is known.
(cid:66) Maximin: For any two candidates x and y, let D(x, y) be N(x, y) − N(y, x), where N(x, y)
(respectively N(y, x)) is the number of voters who prefer x to y (respectively y to x). The
election we get by restricting all the votes to x and y only is called the pairwise election
between x and y. The maximin score of a candidate x is miny(cid:54)=x D(x, y). The winners are the
candidates with maximum maximin score.
(cid:66) Copelandα. The Copelandα score of a candidate x is {y (cid:54)= x : DE(x, y) > 0} + α{y (cid:54)= x :
DE(x, y) = 0}, where α ∈ [0, 1]. That is, the Copelandα of a candidate x is the number of
other candidates it defeats in pairwise election plus α times the number of other candidates it
ties with in pairwise elections. The winners are the candidates with the maximum Copelandα
score.
2.2 Problem Definitions
We now formally define the three problems that we consider in this work, namely WEAK MANIPU-
LATION, OPPORTUNISTIC MANIPULATION, and STRONG MANIPULATION. Let r be a fixed voting rule.
We first introduce the WEAK MANIPULATION problem.
7
Definition 1. r-WEAK MANIPULATION
Given a set of partial votes P over a set of candidates C, a positive integer (cid:96) (> 0) denoting the number
of manipulators, and a candidate c, do there exist votes (cid:31)1, . . . ,(cid:31)(cid:96) ∈ L(C) such that there exists an
extension (cid:31)∈ L(C)P of P with r((cid:31),(cid:31)1, . . . ,(cid:31)(cid:96)) = c?
To define the OPPORTUNISTIC MANIPULATION problem, we first introduce the notion of an (r, c)-
opportunistic voting profile, where r is a voting rule and c is any particular candidate.
Definition 2. (r, c)-Opportunistic Voting Profile
Let (cid:96) be the number of manipulators and P a set of partial votes. An (cid:96)-voter profile ((cid:31)i)i∈[(cid:96)] ∈ L(C)(cid:96) is
called an (r, c)-opportunistic voting profile if for each extension P of P for which there exists an (cid:96)-vote
profile ((cid:31)(cid:48)
(cid:16)
P ∪ ((cid:31)i)i∈[(cid:96)]
i)i∈[(cid:96)] ∈ L(C)(cid:96) with r
(cid:16)
P ∪(cid:0)(cid:31)(cid:48)
(cid:17)
i∈[(cid:96)]
i
= c, we have r
(cid:1)
(cid:17)
= c.
In other words, an (cid:96)-vote profile is (r, c)-opportunistic with respect to a partial profile if, when
put together with the truthful votes of any extension, c wins if the extension is viable to begin with.
We are now ready to define the OPPORTUNISTIC MANIPULATION problem.
Definition 3. r-OPPORTUNISTIC MANIPULATION
Given a set of partial votes P over a set of candidates C, a positive integer (cid:96) (> 0) denoting the number
of manipulators, and a candidate c, does there exist an (r, c)-opportunistic (cid:96)-vote profile?
We finally define the STRONG MANIPULATION problem.
Definition 4. r-STRONG MANIPULATION
Given a set of partial votes P over a set of candidates C, a positive integer (cid:96) (> 0) denoting the number
of manipulators, and a candidate c, do there exist votes ((cid:31)i)i∈(cid:96) ∈ L(C)(cid:96) such that for every extension
(cid:31)∈ L(C)P of P, we have r((cid:31), ((cid:31)i)i∈[(cid:96)]) = c?
We use (P, (cid:96), c) to denote instances of WEAK MANIPULATION, OPPORTUNISTIC MANIPULATION,
and STRONG MANIPULATION, where P denotes a profile of partial votes, (cid:96) denotes the number of
manipulators, and c denotes the desired winner.
For the sake of completeness, we provide the definitions of the COALITIONAL MANIPULATION
and POSSIBLE WINNER problems below.
Definition 5. r-COALITIONAL MANIPULATION
Given a set of complete votes (cid:31) over a set of candidates C, a positive integer (cid:96) (> 0) denoting the number
of manipulators, and a candidate c, do there exist votes ((cid:31)i)i∈(cid:96) ∈ L(C)(cid:96) such that r
=
c?
(cid:16)(cid:31), ((cid:31)i)i∈[(cid:96)]
(cid:17)
Definition 6. r-POSSIBLE WINNER
Given a set of partial votes P and a candidate c, does there exist an extension (cid:31) of the partial votes in
P to linear votes such that r((cid:31)) = c?
2.3 Comparison with Possible Winner and Coalitional Manipulation
For any fixed voting rule, the WEAK MANIPULATION problem with (cid:96) manipulators reduces to the
POSSIBLE WINNER problem. This is achieved by simply using the same set as truthful votes and
introducing (cid:96) empty votes. We summarize this in the observation below.
8
Observation 1. The WEAK MANIPULATION problem many-to-one reduces to the POSSIBLE WINNER
problem for every voting rule.
Proof. Let (P, (cid:96), c) be an instance of WEAK MANIPULATION. Let Q be the set consisting of (cid:96) many
copies of partial votes {∅}. Clearly the WEAK MANIPULATION instance (P, (cid:96), c) is equivalent to the
POSSIBLE WINNER instance (P ∪ Q, c).
However, whether the POSSIBLE WINNER problem reduces to the WEAK MANIPULATION problem
or not is not clear since in any WEAK MANIPULATION problem instance, there must exist at least
one manipulator and a POSSIBLE WINNER instance may have no empty vote. From a technical
point of view, the difference between the WEAK MANIPULATION and POSSIBLE WINNER problems
may look marginal; however we believe that the WEAK MANIPULATION problem is a very natural
generalization of the COALITIONAL MANIPULATION problem in the partial information setting and
thus worth studying. Similarly, it is easy to show, that the COALITIONAL MANIPULATION problem
with (cid:96) manipulators reduces to WEAK MANIPULATION, OPPORTUNISTIC MANIPULATION, and STRONG
MANIPULATION problems with (cid:96) manipulators, since the former is a special case of the latter ones.
Observation 2. The COALITIONAL MANIPULATION problem with (cid:96) manipulators many-to-one reduces
to WEAK MANIPULATION, OPPORTUNISTIC MANIPULATION, and STRONG MANIPULATION problems
with (cid:96) manipulators for all voting rules and for all positive integers (cid:96).
Proof. Follows from the fact that every instance of the COALITIONAL MANIPULATION problem is also
an equivalent instance of the WEAK MANIPULATION, OPPORTUNISTIC MANIPULATION, and STRONG
MANIPULATION problems.
Finally, we note that the COALITIONAL MANIPULATION problem with (cid:96) manipulators can be
reduced to the WEAK MANIPULATION problem with just one manipulator, by introducing (cid:96)−1 empty
votes. These votes can be used to witness a good extension in the forward direction. In the reverse
direction, given an extension where the manipulator is successful, the extension can be used as the
manipulator's votes. This argument leads to the following observation.
Observation 3. The COALITIONAL MANIPULATION problem with (cid:96) manipulators many-to-one reduces
to the WEAK MANIPULATION problem with one manipulator for every voting rule and for every positive
integer (cid:96).
Proof. Let (P, (cid:96), c) be an instance of COALITIONAL MANIPULATION. Let Q be the set of consisting of
(cid:96)−1 many copies of partial vote {c (cid:31) others}. Clearly the WEAK MANIPULATION instance (P∪Q, 1, c)
is equivalent to the COALITIONAL MANIPULATION instance (P, (cid:96), 1).
This observation can be used to derive the hardness of WEAK MANIPULATION even for one ma-
nipulator whenever the hardness for COALITIONAL MANIPULATION is known for any fixed number
of manipulators (for instance, this is the case for the voting rules such as Borda, maximin and
Copeland). However, determining the complexity of WEAK MANIPULATION with one manipulator
requires further work for voting rules where COALITIONAL MANIPULATION is polynomially solvable
for any number of manipulators (such as k-approval, Plurality, Bucklin, and so on).
9
3 Hardness Results
In this section, we present our hardness results. While some of our reductions are from the POS-
SIBLE WINNER problem, the other reductions in this section are from the EXACT COVER BY 3-SETS
problem, also referred to as X3C. This is a well-known NP-complete [GJ79] problem, and is defined
as follows.
Definition 7 (Exact Cover by 3-Sets (X3C)). Given a set U and a collection S = {S1, S2, . . . , St} of t
subsets of U with Si = 3 ∀i = 1, . . . , t, does there exist a T ⊂ S with T =
3 such that ∪X∈TX = U?
We use X3C to refer to the complement of X3C, which is to say that an instance of X3C is a YES
instance if and only if it is a NO instance of X3C. The rest of this section is organized according to
the problems being addressed.
U
3.1 Weak Manipulation
To begin with, recall that the COALITIONAL MANIPULATION problem is NP-complete for the
Borda [DKNW11, BNW11], maximin [XZP+09], and Copelandα [FHS08, FHHR09, FHS10] vot-
ing rules for every rational α ∈ [0, 1] \ {0.5}, when we have two manipulators. Therefore, it follows
from Observation 3 that the WEAK MANIPULATION problem is NP-complete for the Borda, maximin,
and Copelandα voting rules for every rational α ∈ [0, 1] \ {0.5}, even with one manipulator.
For the k-approval and k-veto voting rules, we reduce from the corresponding POSSIBLE WIN-
NER problems. While it is natural to start from the same voting profile, the main challenge is in
undoing the advantage that the favorite candidate receives from the manipulator's vote, in the
reverse direction.
We begin with proving that the WEAK MANIPULATION problem is NP-complete for the k-approval
voting rule even with one manipulator and at most 4 undetermined pairs per vote.
Theorem 1. The WEAK MANIPULATION problem is NP-complete for the k-approval voting rule even
with one manipulator for any constant k > 1, even when the number of undetermined pairs in each
vote is no more than 4.
Proof. For simplicity of presentation, we prove the theorem for 2-approval. We reduce from the
POSSIBLE WINNER problem for 2-approval which is NP-complete [XC11], even when the number of
undetermined pairs in each vote is no more than 4. Let P be the set of partial votes in a POSSIBLE
WINNER instance, and let C = {c1, . . . , cm, c} be the set of candidates, where the goal is to check if
there is an extension of P that makes c win. For developing the instance of WEAK MANIPULATION,
we need to "reverse" any advantage that the candidate c obtains from the vote of the manipulator.
Notice that the most that the manipulator can do is to increase the score of c by one. Therefore,
in our construction, we "artificially" increase the score of all the other candidates by one, so that
despite of the manipulator's vote, c will win the new election if and only if it was a possible winner
in the POSSIBLE WINNER instance. To this end, we introduce (m + 1) many dummy candidates
d1, . . . , dm+1 and the complete votes:
wi = ci (cid:31) di (cid:31) others, for every i ∈ {1, . . . , m}
Further, we extend the given partial votes of the POSSIBLE WINNER instance to force the dummy
candidates to be preferred least over the rest - by defining, for every vi ∈ P, the corresponding
10
partial vote v(cid:48)
i as follows.
i = vi ∪ {C (cid:31) {d1, . . . , dm+1}}.
v(cid:48)
This ensures that all the dummy candidates do not receive any score from the modified partial votes
corresponding to the partial votes of the POSSIBLE WINNER instance. Notice that since the number
of undetermined pairs in vi is no more than 4, the number of undetermined pairs in v(cid:48)
i is also no
more than 4. Let (C(cid:48), Q, c) denote this constructed WEAK MANIPULATION instance. We claim that
the two instances are equivalent.
vi ∈ P}. Notice that this is well-defined since vi and v(cid:48)
In the forward direction, suppose c is a possible winner with respect to P, and let P be an
extension where c wins. Then it is easy to see that the manipulator can make c win in some
extension by placing c and dm+1 in the first two positions of her vote (note that the partial score
of dm+1 is zero in Q). Indeed, consider the extension of Q obtained by mimicking the extension P
on the "common" partial votes, {v(cid:48)
i have
exactly the same set of incomparable pairs. In this extension, the score of c is strictly greater than
the scores of all the other candidates, since the scores of all candidates in C is exactly one more
than their scores in P, and all the dummy candidates have a score of at most one.
In the reverse direction, notice that the manipulator puts the candidates c and dm+1 in the top
two positions without loss of generality. Now suppose the manipulator's vote c (cid:31) dm+1 (cid:31) others
makes c win the election for an extension Q of Q. Then consider the extension P obtained by
restricting Q to C. Notice that the score of each candidate in C in this extension is one less than
their scores in Q. Therefore, the candidate c wins this election as well, concluding the proof.
i
The above proof can be imitated for any other constant values of k by reducing it from the
POSSIBLE WINNER problem for k-approval and introducing (m + 1)(k − 1) dummy candidates.
We will use Lemma 1 in subsequent proofs which has been used before [BRR11, DMN15b,
DMN16].
Lemma 1. Let C = {c1, . . . , cm} (cid:93) D, (D > 0) be a set of candidates, and (cid:126)α a normalized score vector
of length C. Then, for any given X = (X1, . . . , Xm) ∈ Zm, there exists λ ∈ R and a voting profile such
that the (cid:126)α-score of ci is λ + Xi for all 1 (cid:54) i (cid:54) m, and the score of candidates d ∈ D is less than λ.
Moreover, the number of votes is O(poly(C ·
(cid:80)
m
i=1
Xi, λ)).
Note that the number of votes used in Lemma 1 is polynomial in m if λ and Xi is polynomial
in m for every i ∈ [m], which indeed is the case in all the proofs that use Lemma 1. We next show
that the WM problem is NP-complete for the k-veto voting rule.
Theorem 2. The WEAK MANIPULATION problem for the k-veto voting rule is NP-complete even with
one manipulator for any constant k > 1.
Proof. We reduce from the POSSIBLE WINNER problem for the k-veto voting rule which is known to
be NP-complete [BD09]. Let P be the set of partial votes in a POSSIBLE WINNER problem instance,
and let C = {c1, . . . , cm, c} be the set of candidates, where the goal is to check if there is an extension
that makes c win with respect to k-veto. We assume without loss of generality that c's position is
fixed in all the partial votes (if not, then we fix the position of c as high as possible in every vote).
We introduce k + 1 many dummy candidates d1, . . . , dk, d. The role of the first k dummy can-
didates is to ensure that the manipulator is forced to place them at the "bottom k" positions of her
vote, so that all the original candidates get the same score from the additional vote of the manip-
ulator. The most natural way of achieving this is to ensure that the dummy candidates have the
11
same score as c in any extension (note that we know the score of c since c's position is fixed in
all the partial votes). This would force the manipulator to place these k candidates in the last k
positions. Indeed, doing anything else will cause these candidates to tie with c, even when there is
an extension of P that makes c win.
To this end, we begin by placing the dummy candidates in the top k positions in all the partial
votes. Formally, we modify every partial vote as follows:
w = di (cid:31) others, for every i ∈ {1, . . . , k}
At this point, we know the scores of c and di, for every i ∈ {1, . . . , k}. Using Lemma 1, we add
complete votes such that the final score of c is the same with the score of every di and the score of
c is strictly more than the score of d. The relative score of every other candidate remains the same.
This completes the description of the construction. We denote the augmented set of partial votes
by P.
We now argue the correctness. In the forward direction, if there is an extension of the votes that
makes c win, then we repeat this extension, and the vote of the manipulator puts the candidate di
at the position m + i + 2; and all the other candidates in an arbitrary fashion. Formally, we let the
manipulator's vote be:
v = c (cid:31) c1 (cid:31) ··· (cid:31) cm (cid:31) d (cid:31) d1 (cid:31) ··· (cid:31) dk.
By construction c wins the election in this particular setup.
In the reverse direction, consider a
vote of the manipulator and an extension Q of P in which c wins. Note that the manipulator's vote
necessarily places the candidates di in the bottom k positions -- indeed, if not, then c cannot win
the election by construction. We extend a partial vote w ∈ P by mimicking the extension of the
corresponding partial vote w(cid:48) ∈ P, that is, we simply project the extension of w(cid:48) on the original
set of candidates C. Let Q denote this proposed extension of P. We claim that c wins the election
given by Q. Indeed, suppose not. Let ci be a candidate whose score is at least the score of c in the
extension Q. Note that the scores of ci and c in the extension Q are exactly the same as their scores
in Q, except for a constant offset -- importantly, their scores are offset by the same amount. This
implies that the score of ci is at least the score of c in Q as well, which is a contradiction. Hence,
the two instances are equivalent.
We next prove, by a reduction from X3C, that the WEAK MANIPULATION problem for the Bucklin
and simplified Bucklin voting rules is NP-complete even with one manipulator and at most 16
undetermined pairs per vote.
Theorem 3. The WEAK MANIPULATION problem is NP-complete for Bucklin, simplified Bucklin, Fall-
back, and simplified Fallback voting rules, even when we have only one manipulator and the number
of undetermined pairs in each vote is no more than 16.
Proof. We reduce the X3C problem to WEAK MANIPULATION for simplified Bucklin. Let (U =
{u1, . . . , um}, S := {S1, S2, . . . , St}) be an instance of X3C, where each Si is a subset of U of size
three. We construct a WEAK MANIPULATION instance based on (U, S) as follows.
Candidate set: C = W ∪ X ∪ D ∪ U ∪ {c, w, a, b}, where W = m − 3, X = 4, D = m + 1
We first introduce the following partial votes P in correspondence with the sets in the family as
follows.
W (cid:31) X (cid:31) Si (cid:31) c (cid:31) (U \ Si) (cid:31) D \ ({X × ({c} ∪ Si)}) ,∀i (cid:54) t
12
Notice that the number of undetermined pairs in every vote in P is 16. We introduce the following
additional complete votes Q:
(cid:66) t copies of U (cid:31) c (cid:31) others
(cid:66) m/3 − 1 copies of U (cid:31) a (cid:31) c (cid:31) others
(cid:66) m/3 + 1 copies of D (cid:31) b (cid:31) others
The total number of voters, including the manipulator, is 2t + 2m/3 + 1. Now we show equivalence
of the two instances.
In the forward direction, suppose we have an exact set cover T ⊂ S. Let the vote of the
manipulator v be c (cid:31) D (cid:31) others. We consider the following extension P of P.
On the other hand, if Si ∈ S \ T, then we have:
W (cid:31) Si (cid:31) c (cid:31) X (cid:31) (U \ Si) (cid:31) D
W (cid:31) X (cid:31) Si (cid:31) c (cid:31) (U \ Si) (cid:31) D
We claim that c is the unique simplified Bucklin winner in the profile (P, W, v). Notice that the
simplified Bucklin score of c is m + 1 in this extension, since it appears in the top m + 1 positions
in the m/3 votes corresponding to the set cover, t votes from the complete profile Q and one vote
v of the manipulator. For any other candidate ui ∈ U, ui appears in the top m + 1 positions once
in P and t + m
3 − 1 times in Q. Thus, ui does not get majority in top m + 1 positions making its
simplified Bucklin score at least m + 2. Hence, c is the unique simplified Bucklin winner in the
profile (P, W, v). Similarly, the candidate w1 appears only t times in the top m + 1 positions. The
same can be argued for the remaining candidates in D, W, and w.
In the reverse direction, suppose the WEAK MANIPULATION is a YES instance. We may assume
without loss of generality that the manipulator's vote v is c (cid:31) D (cid:31) others, since the simplified
Bucklin score of the candidates in D is at least 2m. Let P be the extension of P such that c is the
unique winner in the profile (P, Q, v). As every candidate in W is ranked within top m + 2 positions
in t + m
3 votes in P. In those votes, all the
candidates in Si are also within top m + 2 positions. Now if any candidate in U is within top m + 1
positions in P more than once, then c will not be the unique winner. Hence, the Si's corresponding
to the votes where c (cid:31) X in P form an exact set cover.
3 + 1 votes in Q, for c to win, c (cid:31) X must hold in at least m
The reduction above also works for the Bucklin voting rule. Specifically, the argument for the
forward direction is exactly the same as the simplified Bucklin above and the argument for the
reverse direction is as follows. Every candidate in W is ranked within top m + 2 positions in
3 + 1 votes in Q and c is never placed within top m + 2 positions in any vote in Q. Hence, for
t + m
c to win, c (cid:31) X must hold in at least m
3 votes in P. In those votes, all the candidates in Si are also
within top m positions. Notice that c never gets placed within top m positions in any vote in (P, Q).
Now if any candidate x ∈ U is within top m positions in P more than once, then x gets majority
within top m positions and thus c cannot win.
The result for the Fallback and simplified Fallback voting rules follow from the corresponding
results for the Bucklin and simplified Bucklin voting rules respectively since every Bucklin and
simplified Bucklin election is also a Fallback and simplified Fallback election respectively.
13
3.2 Strong Manipulation
We know that the COALITIONAL MANIPULATION problem is NP-complete for the Borda, maximin,
and Copelandα voting rules for every rational α ∈ [0, 1] \ {0.5}, when we have two manipulators.
Thus, it follows from Observation 2 that STRONG MANIPULATION is NP-hard for Borda, maximin,
and Copelandα voting rules for every rational α ∈ [0, 1] \ {0.5} for at least two manipulators.
For the case of one manipulator, STRONG MANIPULATION turns out to be polynomial-time solv-
able for most other voting rules. For Copelandα, however, we show that the problem is co-NP-hard
for every α ∈ [0, 1] for a single manipulator, even when the number of undetermined pairs in each
vote is bounded by a constant. This is achieved by a careful reduction from X3C. The following
lemma has been used before [McG53].
Lemma 2. For any function f : C × C −→ Z, such that
1. ∀a, b ∈ C, f(a, b) = −f(b, a).
2. ∀a, b, c, d ∈ C, f(a, b) + f(c, d) is even,
there exists a n-voters' profile such that for all a, b ∈ C, a defeats b with a margin of f(a, b). Moreover,
(cid:88)
{a,b}∈C×C
n is even and n = O
f(a, b)
We have following intractability result for the STRONG MANIPULATION problem for the
Copelandα rule with one manipulator and at most 10 undetermined pairs per vote.
Theorem 4. STRONG MANIPULATION is co-NP-hard for Copelandα voting rule for every α ∈ [0, 1]
even when we have only one manipulator and the number of undetermined pairs in each vote is no
more than 10.
Proof. We reduce X3C to STRONG MANIPULATION for Copelandα rule. Let (U = {u1, . . . , um}, S =
{S1, S2, . . . , St}) is an X3C instance. We assume, without loss of generality, t to be an even integer
(if not, replicate any set from S). We construct a corresponding WEAK MANIPULATION instance for
Copelandα as follows.
Candidate set C = U ∪ {c, w, z, d}
Partial votes P:
∀i (cid:54) t, (U \ Si) (cid:31) c (cid:31) z (cid:31) d (cid:31) Si (cid:31) w \ {{z, c} × (Si ∪ {d, w})}
Notice that the number of undetermined pairs in every vote in P is 10. Now we add a set Q of
complete votes with Q even and Q = poly(m, t) using Lemma 2 to achieve the following margin
of victories in pairwise elections. Figure 2 shows the weighted majority graph of the resulting
election.
(cid:66) DQ(d, z) = DQ(z, c) = DQ(c, d) = DQ(w, z) = 4t
(cid:66) DQ(ui, d) = DQ(c, ui) = 4t ∀ui ∈ U
(cid:66) DQ(z, ui) = t ∀ui ∈ U
14
3 − 2
(cid:66) DQ(c, w) = t − 2m
(cid:66) DQ(ui, ui+1 (mod ∗)m) = 4t ∀ui ∈ U
(cid:66) DQ(a, b) = 0 for every a, b ∈ C, a (cid:54)= b, not mentioned above
4t
d
c
4t
4t
4t
t − 2m
3 − 2
4t
uj
4t
uj+1
w
t
4t
z
Figure 2: Weighted majority graph of the reduced instance in Theorem 4. The weight of all the
edges not shown in the figure are 0. For simplicity, we do not show edges among {u1, . . . , um}.
We have only one manipulator who tries to make c winner. Notice that the number of votes in
the STRONG MANIPULATION instance (P ∪ Q, 1, c) including the manipulator's vote is odd (since P
and Q are even integers). Therefore, DP∗∪Q∪{v∗}(a, b) is never zero for every a, b ∈ C, a (cid:54)= b in
every extension P∗ of P and manipulators vote v∗ and consequently the particular value of α does
not play any role in this reduction. Hence, we assume, without loss of generality, α to be zero from
here on and simply use the term Copeland instead of Copelandα.
Now we show that the X3C instance (U, S) is a YES instance if and only if the STRONG MANIPU-
LATION instance (P∪ Q, 1, c) is a NO instance (a STRONG MANIPULATION instance is a NO instance if
there does not exist a vote of the manipulator which makes c the unique winner in every extension
of the partial votes). We can assume without loss of generality that manipulator puts c at first
position and z at last position in her vote v.
Assume that the X3C instance is a YES instance. Suppose (by renaming) that S1, . . . , S m
forms
an exact set cover. We claim that the following extension P of P makes both z and c Copeland
co-winners.
3
Extension P of P:
, (U \ Si) (cid:31) c (cid:31) z (cid:31) d (cid:31) Si (cid:31) w
i (cid:54) m
3
+ 1, (U \ Si) (cid:31) d (cid:31) Si (cid:31) w (cid:31) c (cid:31) z
i (cid:62) m
3
We have summarize the pairwise margins between z and c and the rest of the candidates from
the profile (P ∪ Q ∪ v) in Table 2. The candidates z and c are the co-winners with Copeland score
(m + 1).
15
C \ {z} DP∪Q∪v(z,·)
c
w, d
ui ∈ U
(cid:62) 3t
(cid:54) −3t
1
C \ {c}
z, ui ∈ U
DP∪Q∪v(c,·)
(cid:62) 3t
w
d
−1
(cid:54) −3t
Table 2: DP∪Q∪v(z,·) and DP∪Q∪v(c,·)
For the other direction, notice that Copeland score of c is at least m + 1 since c defeats d and
every candidate in U in every extension of P. Also notice that the Copeland score of z can be at
most m + 1 since z loses to w and d in every extension of P. Hence the only way c cannot be the
unique winner is that z defeats all candidates in U and w defeats c.
3 extensions of P. We claim that the sets Si in the remaining of
the extensions where c (cid:31) w forms an exact set cover for (U, S). Indeed, otherwise some candidate
ui ∈ U is not covered. Then, notice that ui (cid:31) z in all t votes, making D(z, ui) = −1.
This requires w (cid:31) c in at least t− m
3.3 Opportunistic Manipulation
All our reductions for the co-NP-hardness for OPPORTUNISTIC MANIPULATION start from X3C. We
note that all our hardness results hold even when there is only one manipulator. Our overall
approach is the following. We engineer a set of partial votes in such a way that the manipulator is
forced to vote in a limited number of ways to have any hope of making her favorite candidate win.
For each such vote, we demonstrate a viable extension where the vote fails to make the candidate
a winner, leading to a NO instance of OPPORTUNISTIC MANIPULATION. These extensions rely on
the existence of an exact cover. On the other hand, we show that if there is no exact set cover,
then there is no viable extension, thereby leading to an instance that is vacuously a YES instance of
OPPORTUNISTIC MANIPULATION. Our first result on OPPORTUNISTIC MANIPULATION shows that the
OPPORTUNISTIC MANIPULATION problem is co-NP-hard for the k-approval voting rule for constant
k (cid:62) 3 even when the number of manipulators is one and the number of undetermined pairs in each
vote is no more than 15.
Theorem 5. The OPPORTUNISTIC MANIPULATION problem is co-NP-hard for the k-approval voting
rule for constant k (cid:62) 3 even when the number of manipulators is one and the number of undetermined
pairs in each vote is no more than 15.
Proof. We reduce X3C to OPPORTUNISTIC MANIPULATION for k-approval rule.
Let (U =
{u1, . . . , um}, S = {S1, S2, . . . , St}) is an X3C instance. We construct a corresponding OPPORTUNISTIC
MANIPULATION instance for k-approval voting rule as follows. We begin by introducing a candidate
for every element of the universe, along with k − 3 dummy candidates (denoted by W), and special
candidates {c, z1, z2, d, x, y}. Formally, we have:
Candidate set C = U ∪ {c, z1, z2, d, x, y} ∪ W.
Now, for every set Si in the universe, we define the following total order on the candidate set,
which we denote by P(cid:48)
i:
W (cid:31) Si (cid:31) y (cid:31) z1 (cid:31) z2 (cid:31) x (cid:31) (U \ Si) (cid:31) c (cid:31) d
16
Using P(cid:48)
i, we define the partial vote Pi as follows:
Pi = P(cid:48)
i
\ ({{y, x, z1, z2} × Si} ∪ {(z1, z2), (x, z1), (x, z2)}).
We denote the set of partial votes {Pi : i ∈ [t]} by P and {P(cid:48)
i : i ∈ [t]} by P(cid:48). We remark that the
number of undetermined pairs in each partial vote Pi is 15.
We now invoke Lemma 1 from [DMN16], which allows to achieve any pre-defined scores on the
candidates using only polynomially many additional votes. Using this, we add a set Q of complete
votes with Q = poly(m, t) to ensure the following scores, where we denote the k-approval score of
a candidate from a set of votes V by sV(·): sQ(z1) = sQ(z2) = sQ(y) = sQ(c) − m/3; sQ(d), sQ(w) (cid:54)
sQ(c) − 2t ∀w ∈ W; sQ(x) = sQ(c) − 1; sP(cid:48)∪Q(uj) = sQ(c) + 1 ∀j ∈ [m].
Our reduced instance is (P∪ Q, 1, c). The reasoning for this score configuration will be apparent
as we argue the equivalence. We first argue that if we had a YES instance of X3C (in other words,
there is no exact cover), then we have a YES instance of OPPORTUNISTIC MANIPULATION. It turns
out that this will follow from the fact that there are no viable extensions, because, as we will show
next, a viable extension implies the existence of an exact set cover.
To this end, first observe that the partial votes are constructed in such a way that c gets no
additional score from any extension. Assuming that the manipulator approves c (without loss of
generality), the final score of c in any extension is going to be sQ(c) + 1. Now, in any viable
extension, every candidate uj has to be "pushed out" of the top k positions at least once. Observe
that whenever this happens, y is forced into the top k positions. Since y is behind the score of c by
only m/3 votes, Si's can be pushed out of place in only m/3 votes. For every uj to lose one point,
these votes must correspond to an exact cover. Therefore, if there is no exact cover, then there is
no viable extension, showing one direction of the reduction.
On the other hand, suppose we have a NO instance of X3C -- that is, there is an exact cover. Let
{Si : i ∈ [m/3]} forms an exact cover of U. We will now use the exact cover to come up with two
viable extensions, both of which require the manipulator to vote in different ways to make c win.
Therefore, there is no single manipulative vote that accounts for both extensions, leading us to a
NO instance of OPPORTUNISTIC MANIPULATION.
First, consider this completion of the partial votes:
i = 1, W (cid:31) y (cid:31) x (cid:31) z1 (cid:31) z2 (cid:31) Si (cid:31) (U \ Si) (cid:31) c (cid:31) d
2 (cid:54) i (cid:54) m/3, W (cid:31) y (cid:31) z1 (cid:31) z2 (cid:31) x (cid:31) Si (cid:31) (U \ Si) (cid:31) c (cid:31) d
m/3 + 1 (cid:54) i (cid:54) t, W (cid:31) Si (cid:31) y (cid:31) z1 (cid:31) z2 (cid:31) x (cid:31) (U \ Si) (cid:31) c (cid:31) d
Notice that in this completion, once accounted for along with the votes in Q, the score of c is tied
with the scores of all uj's, z1, x and y, while the score of z2 is one less than the score of c. Therefore,
the only k candidates that the manipulator can afford to approve are W, the candidates c, d and z2.
However, consider the extension that is identical to the above except with the first vote changed to:
W (cid:31) y (cid:31) x (cid:31) z2 (cid:31) z1 (cid:31) Si (cid:31) (U \ Si) (cid:31) c (cid:31) d
Here, on the other hand, the only way for c to be an unique winner is if the manipulator ap-
proves W, c, d and z1. Therefore, it is clear that there is no way for the manipulator to provide
a consolidated vote for both these profiles. Therefore, we have a NO instance of OPPORTUNISTIC
MANIPULATION.
17
We next move on to the k-veto voting rule and show that the OPPORTUNISTIC MANIPULATION
problem for the k-veto is co-NP-hard for every constant k (cid:62) 4 even when the number of manipula-
tors is one and the number of undetermined pairs in each vote is no more than 15.
Theorem 6. The OPPORTUNISTIC MANIPULATION problem is co-NP-hard for the k-veto voting rule for
every constant k (cid:62) 4 even when the number of manipulators is one and the number of undetermined
pairs in each vote is no more than 15.
Proof. We reduce X3C to OPPORTUNISTIC MANIPULATION for k-veto rule. Let (U = {u1, . . . , um}, S =
{S1, S2, . . . , St}) is an X3C instance. We construct a corresponding OPPORTUNISTIC MANIPULATION
instance for k-veto voting rule as follows.
Candidate set C = U ∪ {c, z1, z2, d, x, y} ∪ A ∪ W, where A = {a1, a2, a3}, W = k − 4
For every i ∈ [t], we define P(cid:48)
i as follows:
∀i (cid:54) t, c (cid:31) A (cid:31) (U \ Si) (cid:31) d (cid:31) Si (cid:31) y (cid:31) x (cid:31) z1 (cid:31) z2 (cid:31) W
i, we define partial vote Pi = P(cid:48)
\ ({{y, x, z1, z2} × Si} ∪ {(z1, z2), (x, z1), (x, z2)}) for every
Using P(cid:48)
i ∈ [t]. We denote the set of partial votes {Pi : i ∈ [t]} by P and {P(cid:48)
i : i ∈ [t]} by P(cid:48). We note that
the number of undetermined pairs in each partial vote Pi is 15. Using Lemma 1, we add a set Q
of complete votes with Q = poly(m, t) to ensure the following. We denote the k-veto score of a
candidate from a set of votes W by sW(·).
i
(cid:66) sP(cid:48)∪Q(z1) = sP(cid:48)∪Q(z2) = sP(cid:48)∪Q(c) − m/3
(cid:66) sP(cid:48)∪Q(ai) = sP(cid:48)∪Q(uj) = sP(cid:48)∪Q(w) = sP(cid:48)∪Q(c) ∀ai ∈ A, uj ∈ U, w ∈ W
(cid:66) sP(cid:48)∪Q(y) = sP(cid:48)∪Q(c) − m/3 − 1
(cid:66) sP(cid:48)∪Q(x) = sP(cid:48)∪Q(c) − 2
We have only one manipulator who tries to make c winner. Now we show that the X3C instance
(U, S) is a YES instance if and only if the OPPORTUNISTIC MANIPULATION instance (P ∪ Q, 1, c) is a
NO instance.
In the forward direction, let us now assume that the X3C instance is a YES instance. Suppose
(by renaming) that S1, . . . , Sm/3 forms an exact set cover. Let us assume that the manipulator's vote
v disapproves every candidate in W ∪ A since otherwise c can never win uniquely. We now show
that if v does not disapprove z1 then, v is not a c-optimal vote. Suppose v does not disapprove z1.
Then we consider the following extension P of P.
i = 1, c (cid:31) A (cid:31) (U \ Si) (cid:31) d (cid:31) y (cid:31) z1 (cid:31) x (cid:31) z2 (cid:31) Si (cid:31) W
2 (cid:54) i (cid:54) m/3, c (cid:31) A (cid:31) (U \ Si) (cid:31) d (cid:31) y (cid:31) z1 (cid:31) z2 (cid:31) x (cid:31) Si (cid:31) W
m/3 + 1 (cid:54) i (cid:54) t, c (cid:31) A (cid:31) (U \ Si) (cid:31) d (cid:31) Si (cid:31) y (cid:31) x (cid:31) z1 (cid:31) z2 (cid:31) W
We have the following scores sP∪Q(c) = sP∪Q(z1) = sP∪Q(z2) + 1 = sP∪Q(x) + 1 = sP∪Q(uj) +
1 ∀uj ∈ U. Hence, both c and z1 win for the votes P∪Q∪{v}. However, the vote v(cid:48) which disapproves
a1, a2, a3, z1 makes c a unique winner for the votes P ∪ Q ∪ {v(cid:48)}. Hence, v is not a c-optimal vote.
Similarly, we can show that if the manipulator's vote does not disapprove z2 then, the vote is not
18
c-optimal. Hence, there does not exist any c-optimal vote and the OPPORTUNISTIC MANIPULATION
instance is a NO instance.
In the reverse direction, we show that if the X3C instance is a NO instance, then there does not
exist a vote v of the manipulator and an extension P of P such that c is the unique winner for the
votes P∪ Q∪ {v(cid:48)} thereby proving that the OPPORTUNISTIC MANIPULATION instance is vacuously YES
(and thus every vote is c-optimal). Notice that, there must be at least m/3 votes P1 in P where the
corresponding Si gets pushed in bottom k positions since sP(cid:48)∪Q(uj) = sP(cid:48)∪Q(c) ∀ai ∈ A, uj ∈ U.
However, in each vote in P1, y is placed within top m − k many position and thus we have P1
is exactly m/3 since sP(cid:48)∪Q(y) = sP(cid:48)∪Q(c) − m/3 − 1. Now notice that there must be at least one
candidate u ∈ U which is not covered by the sets Sis corresponding to the votes P1 because
the X3C instance is a NO instance. Hence, c cannot win the election uniquely irrespective of the
manipulator's vote. Thus every vote is c-optimal and the OPPORTUNISTIC MANIPULATION instance
is a YES instance.
We show next similar intractability result for the Borda voting rule too with only at most 7
undetermined pairs per vote.
Theorem 7. The OPPORTUNISTIC MANIPULATION problem is co-NP-hard for the Borda voting rule
even when the number of manipulators is one and the number of undetermined pairs in every vote is
no more than 7.
Proof. We reduce X3C to OPPORTUNISTIC MANIPULATION for the Borda rule.
Let (U =
{u1, . . . , um}, S = {S1, S2, . . . , St}) is an X3C instance. Without loss of generality we assume that
m is not divisible by 6 (if not, then we add three new elements b1, b2, b3 to U and a set {b1, b2, b3}
to S). We construct a corresponding OPPORTUNISTIC MANIPULATION instance for the Borda voting
rule as follows.
Candidate set C = U ∪ {c, z1, z2, d, y}
For every i ∈ [t], we define P(cid:48)
i as follows:
∀i (cid:54) t, y (cid:31) Si (cid:31) z1 (cid:31) z2 (cid:31) (U \ Si) (cid:31) d (cid:31) c
Using P(cid:48)
i
i, we define partial vote Pi = P(cid:48)
\ ({({y} ∪ Si) × {z1, z2}} ∪ {(z1, z2)}) for every i ∈ [t]. We
denote the set of partial votes {Pi : i ∈ [t]} by P and {P(cid:48)
i : i ∈ [t]} by P(cid:48). We note that the number of
undetermined pairs in each partial vote Pi is 7. Using Lemma 1, we add a set Q of complete votes
with Q = poly(m, t) to ensure the following. We denote the Borda score of a candidate from a set
of votes W by sW(·).
(cid:66) sP(cid:48)∪Q(y) = sP(cid:48)∪Q(c) + m + m/3 + 3
(cid:66) sP(cid:48)∪Q(z1) = sP(cid:48)∪Q(c) − 3(cid:98)m/6(cid:99) − 2
(cid:66) sP(cid:48)∪Q(z2) = sP(cid:48)∪Q(c) − 5(cid:98)m/6(cid:99) − 3
(cid:66) sP(cid:48)∪Q(ui) = sP(cid:48)∪Q(c) + m + 5 − i ∀i ∈ [m]
(cid:66) sP(cid:48)∪Q(d) (cid:54) sP(cid:48)∪Q(c) − 5m
19
We have only one manipulator who tries to make c winner. Now we show that the X3C instance
(U, S) is a YES instance if and only if the OPPORTUNISTIC MANIPULATION instance (P∪Q, 1, c) is a NO
instance. Notice that we can assume without loss of generality that the manipulator places c at the
first position, d at the second position, the candidate ui at (m + 5 − i)th position for every i ∈ [m],
and y at the last position, since otherwise c can never win uniquely irrespective of the extension of
P (that it, the manipulator's vote looks like c (cid:31) d (cid:31) {z1, z2} (cid:31) um (cid:31) um−1 (cid:31) ··· (cid:31) u1 (cid:31) y).
In the forward direction, let us now assume that the X3C instance is a YES instance. Suppose
(by renaming) that S1, . . . , Sm/3 forms an exact set cover. Let the manipulator's vote v be c (cid:31) d (cid:31)
z1 (cid:31) z2 (cid:31) um (cid:31) ··· (cid:31) u1 (cid:31) y. We now argue that v is not a c-optimal vote. The other case where
the manipulator's vote v(cid:48) be c (cid:31) d (cid:31) z2 (cid:31) z1 (cid:31) um (cid:31) ··· (cid:31) u1 (cid:31) y can be argued similarly. We
consider the following extension P of P.
1 (cid:54) i (cid:54) (cid:98)m/6(cid:99), z2 (cid:31) y (cid:31) Si (cid:31) z1 (cid:31) (U \ Si) (cid:31) d (cid:31) c
(cid:100)m/6(cid:101) (cid:54) i (cid:54) m/3, z1 (cid:31) y (cid:31) Si (cid:31) z2 (cid:31) (U \ Si) (cid:31) d (cid:31) c
m/3 + 1 (cid:54) i (cid:54) t, y (cid:31) Si (cid:31) z1 (cid:31) z2 (cid:31) (U \ Si) (cid:31) d (cid:31) c
We have the following Borda scores sP∪Q∪{v}(c) = sP∪Q∪{v}(y) + 1 = sP∪Q∪{v}(z2) + 6 =
sP∪Q∪{v}(z1) = sP∪Q∪{v}(ui) + 1 ∀i ∈ [m]. Hence, c does not win uniquely for the votes P∪ Q∪ {v}.
However, c is the unique winner for the votes P∪ Q∪ {v(cid:48)}. Hence, there does not exist any c-optimal
vote and the OPPORTUNISTIC MANIPULATION instance is a NO instance.
In the reverse direction, we show that if the X3C instance is a NO instance, then there does not
exist a vote v of the manipulator and an extension P of P such that c is the unique winner for the
votes P ∪ Q ∪ {v(cid:48)} thereby proving that the OPPORTUNISTIC MANIPULATION instance is vacuously
YES (and thus every vote is c-optimal). Notice that the score of y must decrease by at least m/3 for
c to win uniquely. However, in every vote v where the score of y decreases by at least one in any
extension P of P, at least one of z1 or z2 must be placed at top position of the vote v. However,
the candidates z1 and z2 can be placed at top positions of the votes in P at most m/3 many times
while ensuring c does not lose the election. Also, even after manipulator places the candidate ui at
(m + 5 − i)th position for every i ∈ [m], for c to win uniquely, the score of every ui must decrease
by at least one. Hence, altogether, there will be exactly m/3 votes (denoted by the set P1) in any
extension of P where y is placed at the second position. However, since the X3C instance is a NO
instance, the Sis corresponding to the votes in P1 does not form a set cover. Let u ∈ U be an
element not covered by the Sis corresponding to the votes in P1. Notice that the score of u does
not decrease in the extension P and thus c does not win uniquely irrespective of the manipulator's
vote. Thus every vote is c-optimal and thus the OPPORTUNISTIC MANIPULATION instance is a YES
instance. Thus every vote is c-optimal and the OPPORTUNISTIC MANIPULATION instance is a YES
instance.
For the maximin voting rule, we show intractability of OPPORTUNISTIC MANIPULATION with one
manipulator even when the number of undetermined pairs in every vote is at most 8.
Theorem 8. The OPPORTUNISTIC MANIPULATION problem is co-NP-hard for the maximin voting rule
even when the number of manipulators is one and the number of undetermined pairs in every vote is
no more than 8.
20
Proof. We reduce X3C to OPPORTUNISTIC MANIPULATION for the maximin rule.
Let (U =
{u1, . . . , um}, S = {S1, S2, . . . , St}) is an X3C instance. We construct a corresponding OPPORTUNISTIC
MANIPULATION instance for the maximin voting rule as follows.
Candidate set C = U ∪ {c, z1, z2, z3, d, x, y}
For every i ∈ [t], we define P(cid:48)
i as follows:
∀i (cid:54) t, Si (cid:31) x (cid:31) d (cid:31) y (cid:31) (U \ Si) (cid:31) z1 (cid:31) z2 (cid:31) z3
i, we define partial vote Pi = P(cid:48)
i
Using P(cid:48)
the set of partial votes {Pi : i ∈ [t]} by P and {P(cid:48)
undetermined pairs in each partial vote Pi is 8. We define another partial vote p as follows.
\ ({({x} ∪ Si) × {d, y}}) for every i ∈ [t]. We denote
i : i ∈ [t]} by P(cid:48). We note that the number of
p = (z1 (cid:31) z2 (cid:31) z3 (cid:31) others ) \ {(z1, z2), (z2, z3), (z1, z3)}
Using Lemma 2, we add a set Q of complete votes with Q = poly(m, t) to ensure the following
pairwise margins (notice that the pairwise margins among z1, z2, and z3 does not include the partial
vote p). Figure 3 shows the weighted majority graph of the resulting election.
(cid:66) DP(cid:48)∪Q∪{p}(d, c) = 4t + 1
(cid:66) DP(cid:48)∪Q∪{p}(x, d) = 4t + 2m/3 + 1
(cid:66) DP(cid:48)∪Q∪{p}(y, x) = 4t − 2m/3 + 1
(cid:66) DP(cid:48)∪Q∪{p}(d, uj) = 4t − 1 ∀uj ∈ U
(cid:66) DP(cid:48)∪Q(z1, z2) = DP(cid:48)∪Q(z2, z3) = DP(cid:48)∪Q(z3, z1) = 4t + 2
(cid:66) DP(cid:48)∪Q∪{p}(a, b) (cid:54) 1 for every a, b ∈ C not defined above.
We have only one manipulator who tries to make c winner. Now we show that the X3C instance
(U, S) is a YES instance if and only if the OPPORTUNISTIC MANIPULATION instance (P ∪ Q ∪ {p}, 1, c)
is a NO instance. Notice that we can assume without loss of generality that the manipulator's vote
prefers c to every other candidate, y to x, x to d, and d to uj for every uj ∈ U.
In the forward direction, let us now assume that the X3C instance is a YES instance. Suppose
(by renaming) that S1, . . . , Sm/3 forms an exact set cover. Notice that the manipulator's vote must
prefer either z2 to z1 or z1 to z3 or z3 to z2. We show that if the manipulator's vote v prefers z2
to z1, then v is not a c-optimal vote. The other two cases are symmetrical. Consider the following
extension P of P and p of p.
1 (cid:54) i (cid:54) m/3, d (cid:31) y (cid:31) Si (cid:31) x (cid:31) (U \ Si) (cid:31) z1 (cid:31) z2 (cid:31) z3
m/3 + 1 (cid:54) i (cid:54) t, Si (cid:31) x (cid:31) d (cid:31) y (cid:31) (U \ Si) (cid:31) z1 (cid:31) z2 (cid:31) z3
p = z2 (cid:31) z3 (cid:31) z1 (cid:31) others
From the votes in P ∪ Q ∪ {v, p}, the maximin score of c is −4t, of d, x, uj ∀uj ∈ U are −4t − 2,
of z1, z3 are at most than −4t − 2, and of z2 is −4t. Hence, c is not the unique maximn winner.
21
4t + 1
d
c
4t − 1
uj
z1
4t + 2
z3
4t + 2
4t + 2
4t + 2m/3 + 1
z2
y
x
4t − 2m/3 + 1
Figure 3: Weighted majority graph of the reduced instance in Theorem 8. Solid line and dashed
line represent pairwise margins in P(cid:48)∪Q∪{p} and P(cid:48)∪Q respectively. The weight of all the edges not
shown in the figure are within −1 to 1. For simplicity, we do not show edges among {u1, . . . , um}.
However, the manipulator's vote c (cid:31) z1 (cid:31) z2 (cid:31) z3 (cid:31) other makes c the unique maximin winner.
Hence, v is not a c-optimal vote.
For the reverse direction, we show that if the X3C instance is a NO instance, then there does
not exist a vote v of the manipulator and an extension P of P such that c is the unique winner for
the votes P∪ Q∪ {v(cid:48)} thereby proving that the OPPORTUNISTIC MANIPULATION instance is vacuously
YES (and thus every vote is c-optimal). Consider any extension P of P. Notice that, for c to win
uniquely, y (cid:31) x must be at least m/3 of the votes in P; call these set of votes P1. However, d (cid:31) x in
every vote in P1 and d (cid:31) x can be in at most m/3 votes in P for c to win uniquely. Hence, we have
P1 = m/3. Also for c to win, each d (cid:31) uj must be at least one vote of P and d (cid:31) uj is possible
only in the votes in P1. However, the sets Sis corresponding to the votes in P1 does not form a set
cover since the X3C instance is a NO instance. Hence, there must exist a uj ∈ U for which uj (cid:31) d
in every vote in P and thus c cannot win uniquely irrespective of the vote of the manipulator. Thus
every vote is c-optimal and the OPPORTUNISTIC MANIPULATION instance is a YES instance.
Our next result proves that the OPPORTUNISTIC MANIPULATION problem is co-NP-hard for the
Copelandα voting rule too for every α ∈ [0, 1] even with one manipulator and at most 8 undeter-
mined pairs per vote.
Theorem 9. The OPPORTUNISTIC MANIPULATION problem is co-NP-hard for the Copelandα voting
rule for every α ∈ [0, 1] even when the number of manipulators is one and the number of undetermined
pairs in each vote is no more than 8.
Proof. We reduce X3C to OPPORTUNISTIC MANIPULATION for the Copelandα voting rule. Let (U =
{u1, . . . , um}, S = {S1, S2, . . . , St}) is an X3C instance. We construct a corresponding OPPORTUNISTIC
MANIPULATION instance for the Copelandα voting rule as follows.
Candidate set C = U ∪ {c, z1, z2, z3, d1, d2, d3, x, y}
For every i ∈ [t], we define P(cid:48)
i as follows:
22
Using P(cid:48)
i, we define partial vote Pi = P(cid:48)
∀i (cid:54) t, Si (cid:31) x (cid:31) y (cid:31) c (cid:31) others
of partial votes {Pi : i ∈ [t]} by P and {P(cid:48)
pairs in each partial vote Pi is 8. We define another partial vote p as follows.
p = (z1 (cid:31) z2 (cid:31) z3 (cid:31) others ) \ {(z1, z2), (z2, z3), (z1, z3)}
i
\({({x}∪Si)×{c, y}}) for every i ∈ [t]. We denote the set
i : i ∈ [t]} by P(cid:48). We note that the number of undetermined
Using Lemma 2, we add a set Q of complete votes with Q = poly(m, t) to ensure the following
pairwise margins (notice that the pairwise margins among z1, z2, and z3 does not include the partial
vote p). Figure 4 shows the weighted majority graph of the resulting election.
(cid:66) DP(cid:48)∪Q∪{p}(uj, c) = 2 ∀uj ∈ U
(cid:66) DP(cid:48)∪Q∪{p}(x, y) = 2m/3
(cid:66) DP(cid:48)∪Q∪{p}(c, y) = DP(cid:48)∪Q∪{p}(x, c) = DP(cid:48)∪Q∪{p}(di, c) = DP(cid:48)∪Q∪{p}(zk, c) =
DP(cid:48)∪Q∪{p}(uj, x) = DP(cid:48)∪Q∪{p}(x, zk) = DP(cid:48)∪Q∪{p}(di, x) = DP(cid:48)∪Q∪{p}(y, uj) =
DP(cid:48)∪Q∪{p}(di, y) = DP(cid:48)∪Q∪{p}(y, zk) = DP(cid:48)∪Q∪{p}(zk, uj) = DP(cid:48)∪Q∪{p}(uj, di) =
DP(cid:48)∪Q∪{p}(zk, d1) = DP(cid:48)∪Q∪{p}(zk, d2) = DP(cid:48)∪Q∪{p}(d3, zk) = 4t ∀i, k ∈ [3], j ∈ [m]
(cid:66) DP(cid:48)∪Q∪{p}(uj, u(cid:96)) = −4t for at least m/3 many u(cid:96) ∈ U
(cid:66) DP(cid:48)∪Q(z1, z2) = DP(cid:48)∪Q(z2, z3) = DP(cid:48)∪Q(z3, z1) = 1
(cid:66) DP(cid:48)∪Q∪{p}(a, b) (cid:54) 1 for every a, b ∈ C not defined above.
We have only one manipulator who tries to make c winner. Now we show that the X3C instance
(U, S) is a YES instance if and only if the OPPORTUNISTIC MANIPULATION instance (P ∪ Q ∪ {p}, 1, c)
is a NO instance. Since the number of voters is odd, α does not play any role in the reduction and
thus from here on we simply omit α. Notice that we can assume without loss of generality that the
manipulator's vote prefers c to every other candidate and x to y.
In the forward direction, let us now assume that the X3C instance is a YES instance. Suppose
(by renaming) that S1, . . . , Sm/3 forms an exact set cover. Suppose the manipulator's vote v order
z1, z2, and z3 as z1 (cid:31) z2 (cid:31) z3. We will show that v is not a c-optimal vote. Symmetrically, we
can show that the manipulator's vote ordering z1, z2, and z3 in any other order is not c-optimal.
Consider the following extension P of P and p of p.
1 (cid:54) i (cid:54) m/3, y (cid:31) c (cid:31) Si (cid:31) x (cid:31) others
m/3 + 1 (cid:54) i (cid:54) t, Si (cid:31) x (cid:31) y (cid:31) c (cid:31) others
p = z1 (cid:31) z2 (cid:31) z3 (cid:31) others
From the votes in P∪ Q∪ {v, p}, the Copeland score of c is m + 4 (defeating y, zk, uj ∀k ∈ [3], j ∈
[m]), of y is m+3 (defeating zk, uj ∀k ∈ [3], j ∈ [m]), of uj is at most 2m/3+4 (defeating x, di ∀i ∈ [3]
and at most 2m/3 many u(cid:96) ∈ U), of x is 5 (defeating c, y, zk ∀l ∈ [3]), of d1, d2 is 2 (defeating y
and c), of d3 is 5 (defeating y, c, zk ∀k ∈ [3]). of z3 is m + 3 (defeating di, uj∀i ∈ [3], j ∈ [m])
for every k ∈ [3], of z3 is m + 2 (defeating d1, d2, uji ∈ [3], j ∈ [m]), z2 is m + 3 (defeating
23
uj
2
c
y
2m/3
x
z1
1
1
z2
1
z3
d1
d2
d3
Figure 4: Weighted majority graph of the reduced instance in Theorem 9. Solid line and dashed
line represent pairwise margins in P(cid:48) ∪ Q ∪ {p} and P(cid:48) ∪ Q respectively. The weight of all the edges
not shown in the figure are within −1 to 1. The weight of all unlabeled edges are 4t. For simplicity,
we do not show edges among {u1, . . . , um}.
24
d1, d2, z3, uji ∈ [3], j ∈ [m]), z1 is m + 4 (defeating d1, d2, z2, z3, uji ∈ [3], j ∈ [m]). Hence, c
co-wins with z1 with Copeland score m + 4. However, the manipulator's vote c (cid:31) z3 (cid:31) z2 (cid:31) z1
makes c win uniquely. Hence, v is not a c-optimal vote and thus the OPPORTUNISTIC MANIPULATION
instance is a NO instance.
For the reverse direction, we show that if the X3C instance is a NO instance, then there does
not exist a vote v of the manipulator and an extension P of P such that c is the unique winner for
the votes P∪ Q∪ {v(cid:48)} thereby proving that the OPPORTUNISTIC MANIPULATION instance is vacuously
YES (and thus every vote is c-optimal). Consider any extension P of P. Notice that, for c to win
uniquely, c must defeat each uj ∈ U and thus c is preferred over uj in at least one vote in P; we
call these votes P1. However, in every vote in P1, y is preferred over x and thus P1 (cid:54) m/3 because
x must defeat y for c to win uniquely. Since the X3C instance is a NO instance, there must be a
candidate u ∈ U which is not covered by the sets corresponding to the votes in P1 and thus u is
preferred over c in every vote in P. Hence, c cannot win uniquely irrespective of the vote of the
manipulator. Thus every vote is c-optimal and the OPPORTUNISTIC MANIPULATION instance is a YES
instance.
For the Bucklin and simplified Bucklin voting rules, we show intractability of the OPPORTUNIS-
TIC MANIPULATION problem with at most 15 undetermined pairs per vote and only one manipulator.
Theorem 10. The OPPORTUNISTIC MANIPULATION problem is co-NP-hard for the Bucklin and simpli-
fied Bucklin voting rules even when the number of manipulators is one and the number of undetermined
pairs in each vote is no more than 15.
Proof. We reduce X3C to OPPORTUNISTIC MANIPULATION for the Bucklin and simplified Bucklin
voting rules. Let (U = {u1, . . . , um}, S = {S1, S2, . . . , St}) is an X3C instance. We assume without
loss of generality that m is not divisible by 6 (if not, we introduce three elements in U and a set
containing them in S) and t is an even integer (if not, we duplicate any set in S). We construct
a corresponding OPPORTUNISTIC MANIPULATION instance for the Bucklin and simplified Bucklin
voting rules as follows.
Candidate set C = U ∪ {c, z1, z2, x1, x2, d} ∪ W, where W = m − 3
For every i ∈ [t], we define P(cid:48)
i as follows:
Using P(cid:48)
i
∀i (cid:54) t, (U \ Si) (cid:31) Si (cid:31) d (cid:31) x1 (cid:31) x2 (cid:31) z1 (cid:31) z2 (cid:31) others
i, we define partial vote Pi = P(cid:48)
\ ({({d} ∪ Si) × {x1, x2, z1, z2}} ∪ {(z1, z2)}) for every
i ∈ [t]. We denote the set of partial votes {Pi : i ∈ [t]} by P and {P(cid:48)
i : i ∈ [t]} by P(cid:48). We note that the
number of undetermined pairs in each partial vote Pi is 15. We introduce the following additional
complete votes Q:
(cid:66) t/2 − (cid:98)m/6(cid:99) − 1 copies of W (cid:31) z1 (cid:31) z2 (cid:31) x1 (cid:31) c (cid:31) others
(cid:66) t/2 − (cid:98)m/6(cid:99) − 1 copies of W (cid:31) z1 (cid:31) z2 (cid:31) x2 (cid:31) c (cid:31) others
(cid:66) 2(cid:100)m/6(cid:101) copies of W (cid:31) z1 (cid:31) z2 (cid:31) d (cid:31) c (cid:31) others
(cid:66) (cid:98)m/6(cid:99) copies of W (cid:31) z1 (cid:31) d (cid:31) x1 (cid:31) c (cid:31) others
(cid:66) (cid:98)m/6(cid:99) copies of W (cid:31) z1 (cid:31) d (cid:31) x2 (cid:31) c (cid:31) others
25
(cid:66) 2(cid:100)m/6(cid:101) − 1 copies of U (cid:31) x1 (cid:31) others
(cid:66) One U (cid:31) c (cid:31) others
We have only one manipulator who tries to make c winner. Now we show that the X3C instance
(U, S) is a YES instance if and only if the OPPORTUNISTIC MANIPULATION instance (P∪Q, 1, c) is a NO
instance. The total number of voters in the OPPORTUNISTIC MANIPULATION instance is 2t + 2m/3 + 1.
We notice that within top m + 1 positions of the votes in P(cid:48) ∪ Q, c appears t + m/3 times, z1 and z2
appear t + (cid:98)m/6(cid:99) times, x1 appears t/2 + m/3 − 1 times, x2 appears t/2 − 1 times, every candidate in
W appears t + m/3 − 1 times, every candidate in U appears t + m/3 + 1 times. Also every candidate
in U appears t + m/3 + 1 times within top m positions of the votes in P∪ Q. Hence, for both Bucklin
and simplified Bucklin voting rules, we can assume without loss of generality that the manipulator
puts c, every candidate in W, x1, x2, and exactly one of z1 and z2.
In the forward direction, let us now assume that the X3C instance is a YES instance. Suppose
(by renaming) that S1, . . . , Sm/3 forms an exact set cover. Suppose the manipulator's vote v puts
c, every candidate in W, x1, x2, and z1 within top m + 1 positions. We will show that v is not
c-optimal. The other case where the manipulator's vote v(cid:48) puts c, every candidate in W, x1, x2, and
z2 within top m + 1 positions is symmetrical. Consider the following extension P of P:
1 (cid:54) i (cid:54) (cid:98)m/6(cid:99), (U \ Si)d (cid:31) x1 (cid:31) x2 (cid:31) z2 (cid:31) Si (cid:31)(cid:31) z1 (cid:31) others
(cid:100)m/6(cid:101) (cid:54) i (cid:54) m/3, (U \ Si)d (cid:31) x1 (cid:31) x2 (cid:31) z1 (cid:31) Si (cid:31)(cid:31) z2 (cid:31) others
m/3 + 1 (cid:54) i (cid:54) t, (U \ Si) (cid:31) Si (cid:31) d (cid:31) x1 (cid:31) x2 (cid:31) z1 (cid:31) z2 (cid:31) others
For both Bucklin and simplified Bucklin voting rules, c co-wins with z1 for the votes in P ∪ Q ∪ {v}.
However, c wins uniquely for the votes in P∪ Q∪ {v(cid:48)}. Hence, v is not a c-optimal vote and thus the
OPPORTUNISTIC MANIPULATION instance is a NO instance.
For the reverse direction, we show that if the X3C instance is a NO instance, then there does
not exist a vote v of the manipulator and an extension P of P such that c is the unique winner for
the votes P∪ Q∪ {v(cid:48)} thereby proving that the OPPORTUNISTIC MANIPULATION instance is vacuously
YES (and thus every vote is c-optimal). Consider any extension P of P. Notice that, for c to win
uniquely, every candidate must be pushed out of top m + 1 positions in at least one vote in P; we
call these set of votes P1. Notice that, P1 (cid:62) m/3. However, in every vote in P1, at least one of
z1 and z2 appears within top m + 1 many positions. Since, the manipulator has to put at least
one of z1 and z2 within its top m + 1 positions and z1 and z2 appear t + (cid:98)m/6(cid:99) times in the votes
in P(cid:48) ∪ Q, we must have P1 (cid:54) m/3 and thus P1 = m/3, for c to win uniquely. However, there
exists a candidate u ∈ U not covered by the Sis corresponding to the votes in P1. Notice that u gets
majority within top m positions of the votes and c can never get majority within top m + 1 positions
of the votes. Hence, c cannot win uniquely irrespective of the vote of the manipulator. Thus every
vote is c-optimal and the OPPORTUNISTIC MANIPULATION instance is a YES instance.
4 Polynomial Time Algorithms
We now turn to the polynomial time cases depicted in Table 1. This section is organized in three
parts, one for each problem considered.
26
4.1 Weak Manipulation
Since the POSSIBLE WINNER problem is in P for the plurality and the veto voting rules [BD09], it
follows from Observation 1 that the WEAK MANIPULATION problem is in P for the plurality and veto
voting rules for any number of manipulators.
Proposition 1. The WEAK MANIPULATION problem is in P for the plurality and veto voting rules for
any number of manipulators.
Proof. The POSSIBLE WINNER problem is in P for the plurality and the veto voting rules [BD09].
Hence, the result follows from Observation 1.
4.2 Strong Manipulation
We now discuss our algorithms for the STRONG MANIPULATION problem. The common flavor in all
our algorithms is the following: we try to devise an extension that is as adversarial as possible for
the favorite candidate c, and if we can make c win in such an extension, then roughly speaking,
such a strategy should work for other extensions as well (where the situation only improves for c).
However, it is challenging to come up with an extension that is globally dominant over all the others
in the sense that we just described. So what we do instead is we consider every potential nemesis
w who might win instead of c, and we build profiles that are "as good as possible" for w and "as
bad as possible" for c. Each such profile leads us to constraints on how much the manipulators can
afford to favor w (in terms of which positions among the manipulative votes are safe for w). We
then typically show that we can determine whether there exists a set of votes that respects these
constraints, either by using a greedy strategy or by an appropriate reduction to a flow problem.
We note that the overall spirit here is similar to the approaches commonly used for solving the
NECESSARY WINNER problem, but as we will see, there are non-trivial differences in the details. We
begin with the k-approval and k-veto voting rules.
Theorem 11. The STRONG MANIPULATION problem is in P for the k-approval and k-veto voting rules,
for any k and any number of manipulators.
Proof. For the time being, we just concentrate on non-manipulators' votes. For each candidate
c(cid:48) ∈ C \ {c}, calculate the maximum possible value of smax
NM (c, c(cid:48)) = sNM(c(cid:48)) − sNM(c) from non-
manipulators' votes, where sNM(a) is the score that candidate a receives from the votes of the
non-manipulators. This can be done by checking all 4 possible score combinations that c and c(cid:48) can
get in each vote v and choosing the one which maximizes sv(c(cid:48)) − sv(c) from that vote. We now fix
the position of c at the top position for the manipulators' votes and we check if it is possible to place
NM (c, c(cid:48))+ sM(c(cid:48))− sM(c)
other candidates in the manipulators' votes such that the final value of smax
is negative which can be solved easily by reducing it to the max flow problem which is polynomial
time solvable.
We now prove that the STRONG MANIPULATION problem for scoring rules is in P for one manip-
ulator.
Theorem 12. The STRONG MANIPULATION problem is in P for any scoring rule when we have only
one manipulator.
27
Proof. For each candidate c(cid:48) ∈ C \ {c}, calculate smax
NM (c, c(cid:48)) using same technique described in
the proof of Theorem 11. We now put c at the top position of the manipulator's vote. For each
candidate c(cid:48) ∈ C \ {c}, c(cid:48) can be placed at positions i ∈ {2, . . . , m} in the manipulator's vote which
NM (c, c(cid:48)) + αi − α1 negative. Using this, construct a bipartite graph with C \ {c} on left and
makes smax
{2, . . . , m} on right and there is an edge between c(cid:48) and i iff the candidate c(cid:48) can be placed at i in
the manipulator's vote according to the above criteria. Now solve the problem by finding existence
of perfect matching in this graph.
Our next result proves that the STRONG MANIPULATION problem for the Bucklin, simplified
Bucklin, Fallback, and simplified Fallback voting rules are in P.
Theorem 13. The STRONG MANIPULATION problem is in P for the Bucklin, simplified Bucklin, Fall-
back, and simplified Fallback voting rules, for any number of manipulators.
Proof. Let (C, P, M, c) be an instance of STRONG MANIPULATION for simplified Bucklin, and let m
denote the total number of candidates in this instance. Recall that the manipulators have to cast
their votes so as to ensure that the candidate c wins in every possible extension of P. We use
Q to denote the set of manipulating votes that we will construct. To begin with, without loss of
generality, the manipulators place c in the top position of all their votes. We now have to organize
the positioning of the remaining candidates across the votes of the manipulators to ensure that c is
a necessary winner of the profile (P, Q).
To this end, we would like to develop a system of constraints indicating the overall number of
times that we are free to place a candidate w ∈ C \ {c} among the top (cid:96) positions in the profile Q.
In particular, let us fix w ∈ C \ {c} and 2 (cid:54) (cid:96) (cid:54) m. Let ηw,(cid:96) be the maximum number of votes of Q
in which w can appear in the top (cid:96) positions. Our first step is to compute necessary conditions for
ηw,(cid:96).
We use Pw,(cid:96) to denote a set of complete votes that we will construct based on the given partial
votes. Intuitively, these votes will represent the "worst" possible extensions from the point of view
of c when pitted against w. These votes are engineered to ensure that the manipulators can make
c win the elections Pw,(cid:96) for all w ∈ C \ {c} and (cid:96) ∈ {2, . . . , m}, if, and only if, they can strongly
manipulate in favor of c. More formally, there exists a voting profile Q of the manipulators so that
c wins the election Pw,(cid:96) ∪ Q, for all w ∈ C \ {c} and (cid:96) ∈ {2, . . . , m} if and only if c wins in every
extension of the profile P ∪ Q.
We now describe the profile Pw,(cid:96). The construction is based on the following case analysis,
where our goal is to ensure that, to the extent possible, we position c out of the top (cid:96) − 1 positions,
and incorporate w among the top (cid:96) positions.
(cid:66) Let v ∈ P be such that either c and w are incomparable or w (cid:31) c. We add the complete vote
v(cid:48) to Pw,(cid:96), where v(cid:48) is obtained from v by placing w at the highest possible position and c at
the lowest possible position, and extending the remaining vote arbitrarily.
(cid:66) Let v ∈ P be such that c (cid:31) w, but there are at least (cid:96) candidates that are preferred over w in
v. We add the complete vote v(cid:48) to Pw,(cid:96), where v(cid:48) is obtained from v by placing c at the lowest
possible position, and extending the remaining vote arbitrarily.
(cid:66) Let v ∈ P be such that c is forced to be within the top (cid:96) − 1 positions, then we add the
complete vote v(cid:48) to Pw,(cid:96), where v(cid:48) is obtained from v by first placing w at the highest possible
position followed by placing c at the lowest possible position, and extending the remaining
vote arbitrarily.
28
(cid:66) In the remaining votes, notice that whenever w is in the top (cid:96) positions, c is also in the top
(cid:96) − 1 positions. Let P∗
w,(cid:96) denote this set of votes, and let t be the number of votes in P∗
w,(cid:96).
We now consider two cases. Let d(cid:96)(c) be the number of times c is placed in the top (cid:96) − 1 positions
in the profile Pw,(cid:96) ∪ Q, and let d(cid:96)(w) be the number of times w is placed in the top (cid:96) positions in
the profile Pw,(cid:96). Let us now formulate the requirement that in Pw,(cid:96) ∪ Q, the candidate c does not
have a majority in the top (cid:96) − 1 positions and w does have a majority in the top (cid:96) positions. Note
that if this requirement holds for any w and (cid:96), then strong manipulation is not possible. Therefore,
to strongly manipulate in favor of c, we must ensure that for every choice of w and (cid:96), we are able
to negate the conditions that we derive.
The first condition from above simply translates to d(cid:96)(c) (cid:54) n/2. The second condition amounts
to requiring first, that there are at least n/2 votes where w appears in the top (cid:96) positions, that is,
d(cid:96)(w) + ηw,(cid:96) + t > n/2. Further, note that the gap between d(cid:96)(w) + ηw,(cid:96) and majority will be filled
by using votes from P∗
w,(cid:96) to "push" w forward. However, these votes contribute equally to w and
c being in the top (cid:96) and (cid:96) − 1 positions, respectively. Therefore, the difference between d(cid:96)(w) +
ηw,(cid:96) and n/2 must be less than the difference between d(cid:96)(c) and n/2. Summarizing, the following
conditions, which we collectively denote by ((cid:63)), are sufficient to defeat c in some extension: d(cid:96)(c) (cid:54)
n/2, d(cid:96)(w) + ηw,(cid:96) + t > n/2, n/2 − d(cid:96)(w) + ηw,(cid:96) < n/2 − d(cid:96)(c).
From the manipulator's point of view, the above provides a set of constraints to be satisfied as
they place the remaining candidates across their votes. Whenever d(cid:96)(c) > n/2, the manipulators
place any of the other candidates among the top (cid:96) positions freely, because c already has majority.
On the other hand, if d(cid:96)(c) (cid:54) n/2, then the manipulators must respect at least one of the following
constraints: ηw,(cid:96) (cid:54) n/2 − t − d(cid:96)(w) and ηw,(cid:96) (cid:54) d(cid:96)(c) − d(cid:96)(w).
Extending the votes of the manipulator while respecting these constraints (or concluding that
this is impossible to do) can be achieved by a natural greedy strategy -- construct the manipulators'
votes by moving positionally from left to right. For each position, consider each manipulator and
populate her vote for that position with any available candidate. We output the profile if the process
terminates by completing all the votes, otherwise, we say NO.
We now argue the proof of correctness. Suppose the algorithm returns NO. This implies that
there exists a choice of w ∈ C \ {c} and (cid:96) ∈ {2, . . . , m} such that for any voting profile Q of the
manipulators, the conditions in ((cid:63)) are satisfied. (Indeed, if there exists a voting profile that violated
at least one of these conditions, then the greedy algorithm would have discovered it.) Therefore,
no matter how the manipulators cast their vote, there exists an extension where c is defeated. In
particular, for the votes in P \ P∗
w,(cid:96), this extension is given by Pw,(cid:96). Further, we choose n/2 − ηw,(cid:96) −
d(cid:96)(w) votes among the votes in P∗
w,(cid:96) and extend them by placing w in the top (cid:96) positions (and
extending the rest of the profile arbitrary). We extend the remaining votes in P∗
w,(cid:96) by positioning
w outside the top (cid:96) positions. Clearly, in this extension, c fails to achieve majority in the top (cid:96) − 1
positions while w does achieve majority in the top (cid:96) positions.
On the other hand, if the algorithm returns YES, then consider the voting profile of the manipu-
lators. We claim that c wins in every extension of P ∪ Q. Suppose, to the contrary, that there exists
an extension R and a candidate w such that the simplified Bucklin score of c is no more than the
simplified Bucklin score of w in R. In this extension, therefore, there exists (cid:96) ∈ {2, . . . , m} for which
w attains majority in the top (cid:96) positions and c fails to attain majority in the top (cid:96) − 1 positions.
However, note that this is already impossible in any extension of the profile Pw,l ∪ P∗
w,(cid:96), because
of the design of the constraints. By construction, the number of votes in which c appears in the top
(cid:96) − 1 positions in R is only greater than the number of times c appears in the top (cid:96) − 1 positions in
29
any extension of Pw,l ∪ P∗
w,(cid:96) (and similarly for w). This leads us to the desired contradiction.
for some w ∈ C \ {c} and (cid:96) < m, then we make ηw,(cid:96) =∞. The proof of correctness for the Bucklin
For the Bucklin voting rule, we do the following modifications to the algorithm. If d(cid:96)(c) > d(cid:96)(w)
voting rule is similar to the proof of correctness for the simplified Bucklin voting rule above.
For Fallback and simplified Fallback voting rules, we consider the number of candidates each
voter approves while computing ηw,(cid:96). We output YES if and only if ηw,(cid:96) (cid:62) 0 for every w ∈ C \ {c}
and every (cid:96) (cid:54) m, since we can assume, without loss of generality, that the manipulator approves
the candidate c only. Again the proof of correctness is along similar lines to the proof of correctness
for the simplified Bucklin voting rule.
We next show that the STRONG MANIPULATION problem for the maximin voting rule is
polynomial-time solvable when we have only one manipulator.
Theorem 14. The STRONG MANIPULATION problem for the maximin voting rules are in P, when we
have only one manipulator.
Proof. For the time being, just concentrate on non-manipulators' votes. Using the algorithm for
NW for maximin in [XC11], we compute for all pairs w, w(cid:48) ∈ C, N(w,w(cid:48))(w, d) and N(w,w(cid:48))(c, w(cid:48))
for all d ∈ C \ {c}. This can be computed in polynomial time. Now we place c at the top position
in the manipulator's vote and increase all N(w,w(cid:48))(c, w(cid:48)) by one. Now we place a candidate w at
the second position if for all w(cid:48) ∈ C, N(cid:48)
(w,w(cid:48))(w, d) < N(w,w(cid:48))(c, w(cid:48)) for all d ∈ C \ {c}, where
N(cid:48)
(w,w(cid:48))(w, d) = N(w,w(cid:48))(w, d) of the candidate d has already been assigned some position in the
manipulator's vote, and N(cid:48)
(w,w(cid:48))(w, d) = N(w,w(cid:48))(w, d) + 1 else. The correctness argument is in the
similar lines of the classical greedy manipulation algorithm of [BITT89].
4.3 Opportunistic Manipulation
For the plurality, Fallback, and simplified Fallback voting rules, it turns out that the voting profile
where all the manipulators approve only c is a c-opportunistic voting profile, and therefore it is
easy to devise a manipulative vote.
Observation 4. The OPPORTUNISTIC MANIPULATION problem is in P for the plurality and Fallback
voting rules for a any number of manipulators.
For the veto voting rule, however, a more intricate argument is needed, that requires building
a system of constraints and a reduction to a suitable instance of the maximum flow problem in a
network, to show polynomial time tractability of OPPORTUNISTIC MANIPULATION.
Theorem 15. The OPPORTUNISTIC MANIPULATION problem is in P for the veto voting rule for a
constant number of manipulators.
Proof. Let (P, (cid:96), c) be an input instance of OPPORTUNISTIC MANIPULATION. We may assume without
loss of generality that the manipulators approve c. We view the voting profile of the manipulators
as a tuple (na)a∈C\{c} ∈ (N ∪ {0})m−1 with
a∈C\{c} na = (cid:96), where the na many manipulators
disapprove a. We denote the set of such tuples as T and we have T = O((2m)(cid:96)) which is poly-
nomial in m since (cid:96) is a constant. A tuple (na)a∈C\{c} ∈ T is not c-optimal if there exists another
a)a∈C\{c} ∈ T and an extension P of P with the following properties. We denote the veto
tuple (n(cid:48)
score of a candidate from P by s(·). For every candidate a ∈ C \ {c}, we define two quantities w(a)
and d(a) as follows.
(cid:80)
30
(cid:66) s(c) > s(a) for every a ∈ C \ {c} with na = n(cid:48)
(cid:66) s(c) > s(a)−n(cid:48)
a for every a ∈ C\{c} with na (cid:62) n(cid:48)
a = 0 and we define w(a) = s(c) − 1, d(a) = 0
a and we define w(a) = s(c)−n(cid:48)
a −1, d(a) =
0
(cid:66) s(a) − na (cid:62) s(c) > s(a) − n(cid:48)
a for every a ∈ C \ {c} with na < n(cid:48)
a and we define w(a) =
s(c) − n(cid:48)
a, d(a) = s(a) − na
We guess the value of s(c). Given a value of s(c), we check the above two conditions by reducing
this to a max flow problem instance as follows. We have a source vertex s and a sink t. We have a
vertex for every a ∈ C (call this set of vertices Y) and a vertex for every vote v ∈ P (call this set of
vertices X). We add an edge from s to each in X of capacity one. We add an edge of capacity one
from a vertex x ∈ X to a vertex y ∈ Y if the candidate corresponding to the vertex y can be placed
at the last position in an extension of the partial vote corresponding to the vertex x. We add an
edge from a vertex y to t of capacity w(a), where a is the voter corresponding to the vertex y. We
also set the demand of every vertex y d(a) (that is the total amount of flow coming into vertex y
must be at least d(a)), where a is the voter corresponding to the vertex y. Clearly, the above three
conditions are met if and only if there is a feasible P amount of flow in the above flow graph. Since
s(c) can have only P + 1 possible values (from 0 to P) and T = O((2m)(cid:96)), we can iterate over
all possible pairs of tuples in T and all possible values of s(c) and find a c-optimal voting profile if
there exists a one.
5 Conclusion
We revisited many settings where the complexity barrier for manipulation was non-existent, and
studied the problem under an incomplete information setting. Our results present a fresh perspec-
tive on the use of computational complexity as a barrier to manipulation, particularly in cases that
were thought to be dead-ends (because the traditional manipulation problem was polynomially
solvable). To resurrect the argument of computational hardness, we have to relax the model of
complete information, but we propose that the incomplete information setting is more realistic,
and many of our hardness results work even with very limited incompleteness of information.
Our work is likely to be the starting point for further explorations. To begin with, we leave
open the problem of completely establishing the complexity of strong, opportunistic, and weak
manipulations for all the scoring rules. Other fundamental forms of manipulation and control do
exist in voting, such as destructive manipulation and control by adding candidates. It would be
interesting to investigate the complexity of these problems in a partial information setting.
Another exciting direction is the study of average case complexity, as opposed to the worst
case results that we have pursued. These studies have already been carried out in the setting of
complete information [PR06, FP10, Wal10]. Studying the problems that we propose in the average-
case model would reveal further insights on the robustness of the incomplete information setting
as captured by our model involving partial orders.
Our results showed that the impact of paucity of information on the computational complexity
of manipulation crucially depends on the notion of manipulation under consideration. We also ar-
gued that different notions of manipulation may be applicable to different situations, maybe based
of how optimistic (or pessimistic) the manipulators are. One important direction of future research
31
is to run extensive experimentations on real and synthetic data to know how people manipulate in
the absence of complete information.
Acknowledgement
Palash Dey wishes to gratefully acknowledge support from Google India for providing him with a
special fellowship for carrying out his doctoral work. Neeldhara Misra acknowledges support by
the INSPIRE Faculty Scheme, DST India (project IFA12-ENG-31).
References
[BBF10]
Yoram Bachrach, Nadja Betzler, and Piotr Faliszewski. Probabilistic possible winner
determination. In International Conference on Artificial Intelligence (AAAI), volume 10,
pages 697 -- 702, 2010.
[BCE+15] Felix Brandt, Vincent Conitzer, Ulle Endriss, J´erome Lang, and Ariel D Procaccia. Hand-
book of computational social choice, 2015.
[BD09]
Nadja Betzler and Britta Dorn. Towards a dichotomy of finding possible winners in
In Mathematical Foundations of Computer Science
elections based on scoring rules.
(MFCS), pages 124 -- 136. Springer, 2009.
[BFLR12] Dorothea Baumeister, Piotr Faliszewski, J´erome Lang, and Jorg Rothe. Campaigns for
lazy voters: truncated ballots. In International Conference on Autonomous Agents and
Multiagent Systems, AAMAS 2012, Valencia, Spain, June 4-8, 2012 (3 Volumes), pages
577 -- 584, 2012.
[BIO91]
John Bartholdi III and James B. Orlin. Single transferable vote resists strategic voting.
Soc. Choice Welf., 8(4):341 -- 354, 1991.
[BITT89]
John Bartholdi III, C.A. Tovey, and M.A. Trick. The computational difficulty of manip-
ulating an election. Soc. Choice Welf., 6(3):227 -- 241, 1989.
[BNW11] Nadja Betzler, Rolf Niedermeier, and Gerhard J Woeginger. Unweighted coalitional
In IJCAI, volume 11, pages 55 -- 60,
manipulation under the borda rule is NP-hard.
2011.
[BRR11]
Dorothea Baumeister, Magnus Roos, and Jorg Rothe. Computational complexity of
two variants of the possible winner problem. In The 10th International Conference on
Autonomous Agents and Multiagent Systems (AAMAS), pages 853 -- 860, 2011.
[BRR+12] Dorothea Baumeister, Magnus Roos, Jorg Rothe, Lena Schend, and Lirong Xia. The
possible winner problem with uncertain weights. In ECAI, pages 133 -- 138, 2012.
[BS09]
Steven J Brams and M Remzi Sanver. Voting systems that combine approval and pref-
erence. In The mathematics of preference, choice and order, pages 215 -- 237. Springer,
2009.
32
[CLMM10] Yann Chevaleyre, J´erome Lang, Nicolas Maudet, and J´erome Monnot. Possible winners
In Proc. International
when new candidates are added: The case of scoring rules.
Conference on Artificial Intelligence (AAAI), 2010.
[CSL07]
Vincent Conitzer, Tuomas Sandholm, and J´erome Lang. When are elections with few
candidates hard to manipulate? J. ACM, 54(3):14, 2007.
[CWX11]
[Dey15]
Vincent Conitzer, Toby Walsh, and Lirong Xia. Dominating manipulations in voting
with partial information. In International Conference on Artificial Intelligence (AAAI),
volume 11, pages 638 -- 643, 2011.
Palash Dey. Computational complexity of fundamental problems in social choice theory.
In Proc. 2015 International Conference on Autonomous Agents and Multiagent Systems,
pages 1973 -- 1974. International Foundation for Autonomous Agents and Multiagent
Systems, 2015.
[DKNW11] Jessica Davies, George Katsirelos, Nina Narodytska, and Toby Walsh. Complexity of
and algorithms for borda manipulation. In Proc. International Conference on Artificial
Intelligence (AAAI), pages 657 -- 662, 2011.
[DL13]
Ning Ding and Fangzhen Lin. Voting with partial information: what questions to ask?
In Proc. 12th International Conference on Autonomous Agents and Multi-agent Systems
(AAMAS), pages 1237 -- 1238. International Foundation for Autonomous Agents and
Multiagent Systems, 2013.
[DMN15a] Palash Dey, Neeldhara Misra, and Y. Narahari. Detecting possible manipulators in
elections. In Proc. 2015 International Conference on Autonomous Agents and Multiagent
Systems, AAMAS 2015, Istanbul, Turkey, May 4-8, 2015, pages 1441 -- 1450, 2015.
[DMN15b] Palash Dey, Neeldhara Misra, and Y. Narahari. Kernelization complexity of possi-
In Proc. 2015 Interna-
ble winner and coalitional manipulation problems in voting.
tional Conference on Autonomous Agents and Multiagent Systems, AAMAS 2015, Istan-
bul, Turkey, May 4-8, 2015, pages 87 -- 96, 2015.
[DMN16]
Palash Dey, Neeldhara Misra, and Y. Narahari. Kernelization complexity of possible
winner and coalitional manipulation problems in voting. Theor. Comput. Sci., 616:111 --
125, 2016.
[DN14]
[DN15]
[EE12]
Palash Dey and Y Narahari. Asymptotic collusion-proofness of voting rules: the case
of large number of candidates. In Proc. 13th International Conference on Autonomous
Agents and Multiagent Systems (AAMAS), pages 1419 -- 1420. International Foundation
for Autonomous Agents and Multiagent Systems, 2014.
Palash Dey and Y Narahari. Asymptotic collusion-proofness of voting rules: The case
of large number of candidates. Studies in Microeconomics, 3(2):120 -- 139, 2015.
Edith Elkind and G´abor Erd´elyi. Manipulation under voting rule uncertainty. In Proc.
11th International Conference on Autonomous Agents and Multiagent Systems (AAMAS),
pages 627 -- 634. International Foundation for Autonomous Agents and Multiagent Sys-
tems, 2012.
33
[ER91]
Eithan Ephrati and Jeffrey S Rosenschein. The Clarke tax as a consensus mechanism
among automated agents. In Proc. Ninth International Conference on Artificial Intelli-
gence (AAAI), pages 173 -- 178, 1991.
[FHH10]
Piotr Faliszewski, Edith Hemaspaandra, and Lane A Hemaspaandra. Using complexity
to protect elections. Commun ACM, 53(11):74 -- 82, 2010.
[FHHR09] Piotr Faliszewski, Edith Hemaspaandra, Lane A. Hemaspaandra, and Jorg Rothe. Llull
and copeland voting computationally resist bribery and constructive control. J. Artif.
Intell. Res., 35:275 -- 341, 2009.
[FHS08]
[FHS10]
[FKN08]
[FKS03]
[FP10]
[Gib73]
[GJ79]
Piotr Faliszewski, Edith Hemaspaandra, and Henning Schnoor. Copeland voting: Ties
In Proc. 7th International Conference on Autonomous Agents and Multiagent
matter.
Systems (AAMAS), pages 983 -- 990. International Foundation for Autonomous Agents
and Multiagent Systems, 2008.
Piotr Faliszewski, Edith Hemaspaandra, and Henning Schnoor. Manipulation of
In Proc. 9th International Conference on Autonomous Agents
copeland elections.
and Multiagent Systems (AAMAS), pages 367 -- 374. International Foundation for Au-
tonomous Agents and Multiagent Systems, 2010.
Ehud Friedgut, Gil Kalai, and Noam Nisan. Elections can be manipulated often.
In
IEEE 49th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages
243 -- 249. IEEE, 2008.
Ronald Fagin, Ravi Kumar, and D. Sivakumar. Efficient similarity search and classifi-
cation via rank aggregation. In Proc. 2003 ACM SIGMOD International Conference on
Management of Data, SIGMOD '03, pages 301 -- 312, New York, NY, USA, 2003. ACM.
Piotr Faliszewski and Ariel D Procaccia. Ai's war on manipulation: Are we winning?
AI Magazine, 31(4):53 -- 64, 2010.
Allan Gibbard. Manipulation of voting schemes: a general result. Econometrica, pages
587 -- 601, 1973.
Michael R Garey and David S Johnson. Computers and Intractability, volume 174.
freeman New York, 1979.
[GNNW14] Serge Gaspers, Victor Naroditskiy, Nina Narodytska, and Toby Walsh. Possible and
In Proc. 13th International Conference on
necessary winner problem in social polls.
Autonomous Agents and Multiagent Systems (AAMAS), pages 613 -- 620. International
Foundation for Autonomous Agents and Multiagent Systems, 2014.
[IKM12] M. Isaksson, G. Kindler, and E. Mossel. The geometry of manipulation - a quantitative
proof of the gibbard-satterthwaite theorem. Combinatorica, 32(2):221 -- 250, 2012.
[KL05]
Kathrin Konczak and J´erome Lang. Voting procedures with incomplete preferences. In
Proc. International Joint Conference on Artificial Intelligence-05 Multidisciplinary Work-
shop on Advances in Preference Handling, volume 20, 2005.
34
[McG53]
David C McGarvey. A theorem on the construction of voting paradoxes. Econometrica,
pages 608 -- 610, 1953.
[ML15]
[NW14]
[PHG00]
[PR06]
[PR07]
[Sat75]
[Wal10]
[Wal11]
[XC08a]
[XC08b]
[XC11]
Vijay Menon and Kate Larson. Complexity of manipulation in elections with partial
votes. CoRR, abs/1505.05900, 2015.
Nina Narodytska and Toby Walsh. The computational impact of partial votes on strate-
In Proc. 21st European Conference on Artificial Intelligence, 18-22 August
gic voting.
2014, Prague, Czech Republic - Including Prestigious Applications of Intelligent Systems
(PAIS 2014), pages 657 -- 662, 2014.
David M. Pennock, Eric Horvitz, and C. Lee Giles. Social choice theory and recom-
mender systems: Analysis of the axiomatic foundations of collaborative filtering.
In
Proc. Seventeenth National Conference on Artificial Intelligence and Twelfth Conference
on on Innovative Applications of Artificial Intelligence, July 30 - August 3, 2000, Austin,
Texas, USA., pages 729 -- 734, 2000.
Ariel D Procaccia and Jeffrey S Rosenschein. Junta distributions and the average-
case complexity of manipulating elections. In Proc. Fifth International Conference on
Autonomous Agents and Multiagent Systems (AAMAS), pages 497 -- 504. ACM, 2006.
Ariel D. Procaccia and Jeffrey S. Rosenschein. Average-case tractability of manipulation
in voting via the fraction of manipulators. In Proc. 6th International Joint Conference
on Autonomous Agents and Multiagent Systems (AAMAS 2007), Honolulu, Hawaii, USA,
May 14-18, 2007, page 105, 2007.
Mark Allen Satterthwaite. Strategy-proofness and Arrow's conditions: Existence and
correspondence theorems for voting procedures and social welfare functions. J. Econ.
Theory, 10(2):187 -- 217, 1975.
Toby Walsh. An empirical study of the manipulability of single transferable voting. In
Proc. 19th European Conference on Artificial Intelligence (ECAI), pages 257 -- 262, 2010.
Toby Walsh. Where are the hard manipulation problems? J. Artif. Intell. Res., pages
1 -- 29, 2011.
Lirong Xia and Vincent Conitzer. Generalized scoring rules and the frequency of coali-
tional manipulability. In Proc. 9th ACM conference on Electronic Commerce (EC), pages
109 -- 118. ACM, 2008.
Lirong Xia and Vincent Conitzer. A sufficient condition for voting rules to be frequently
manipulable. In Proc. 9th ACM conference on Electronic Commerce (EC), pages 99 -- 108.
ACM, 2008.
Lirong Xia and Vincent Conitzer. Determining possible and necessary winners under
common voting rules given partial orders. volume 41, pages 25 -- 67. AI Access Founda-
tion, 2011.
35
[XZP+09] Lirong Xia, Michael Zuckerman, Ariel D Procaccia, Vincent Conitzer, and Jeffrey S
Rosenschein. Complexity of unweighted coalitional manipulation under some common
voting rules. In Proc. 21st International Joint Conference on Artificial Intelligence (IJCAI),
volume 9, pages 348 -- 352, 2009.
36
|
1911.07290 | 1 | 1911 | 2019-10-31T02:30:45 | Dynamic Conflict Resolution Using Justification Based Reasoning | [
"cs.MA",
"cs.LO"
] | We study conflict situations that dynamically arise in traffic scenarios, where different agents try to achieve their set of goals and have to decide on what to do based on their local perception. We distinguish several types of conflicts for this setting. In order to enable modelling of conflict situations and the reasons for conflicts, we present a logical framework that adopts concepts from epistemic and modal logic, justification and temporal logic. Using this framework, we illustrate how conflicts can be identified and how we derive a chain of justifications leading to this conflict. We discuss how conflict resolution can be done when a vehicle has local, incomplete information, vehicle to vehicle communication (V2V) and partially ordered goals. | cs.MA | cs | Dynamic Conflict Resolution Using Justification Based
Reasoning∗
Werner Damm Martin Fränzle Willem Hagemann
Paul Kröger Astrid Rakow
Department of Computing Science, University of Oldenburg, Germany
{werner.damm, martin.fraenzle, willem.hagemann, paul.kroeger, a.rakow}@uol.de
We study conflict situations that dynamically arise in traffic scenarios, where different agents try
to achieve their set of goals and have to decide on what to do based on their local perception. We
distinguish several types of conflicts for this setting. In order to enable modelling of conflict situations
and the reasons for conflicts, we present a logical framework that adopts concepts from epistemic and
modal logic, justification and temporal logic. Using this framework, we illustrate how conflicts can
be identified and how we derive a chain of justifications leading to this conflict. We discuss how
conflict resolution can be done when a vehicle has local, incomplete information, vehicle to vehicle
communication (V2V) and partially ordered goals.
1 Introduction
As humans are replaced by autonomous systems, such systems must be able to interact with each other
and resolve dynamically arising conflicts. Examples of such conflicts arise when a car wants to enter the
highway in dense traffic or simply when a car wants to drive faster than the preceding. Such "conflicts"
are pervasive in road traffic and although traffic rules define a jurisdictional frame, the decision, e.g.,
to give way, is not uniquely determined but influenced by a list of prioritised goals of each system and
the personal preferences of its user. If it is impossible to achieve all goals simultaneously, autonomous
driving systems (ADSs) have to decide "who" will "sacrifice" what goal in order to decide on their
manoeuvres. Matters get even more complicated when we take into account that the ADS has only partial
information. It perceives the world via sensors of limited reach and precision. Moreover, measurements
can be contradicting. An ADS might use V2V to retrieve more information about the world, but it
inevitably has a confined insight to other traffic participants and its environment. Nevertheless, for the
acceptance of ADSs, it is imperative to implement conflict resolution mechanisms that take into account
the high dimensionality of decision making. These decisions have to be explained and in case of an
incident, the system's decisions have to be accountable.
In this paper we study conflict situations as dynamically occurring in road traffic and develop a
formal notion of conflict between two agents. We distinguish several types of conflicts and propose a
conflict resolution process where the different kinds of conflicts are resolved in an incremental fashion.
This process successively increases the required cooperation and decreases the privacy of the agents,
finally negotiating which goals of the two agents have to be sacrificed. We present a logical framework
enabling the analysis of conflicts. This framework borrows from epistemic and modal logic in order to
accommodate the bookkeeping of evidences used during a decision process. The framework in particular
∗This work is partly supported by the German Research Council (DFG) as part of the PIRE SD-SSCPS project (Science
of Design of Societal Scale CPS, grant no. DA 206/11-1, FR 2715/4-1) and the Research Training Group SCARE (System
Correctness under Adverse Conditions, grant no. DFG GRK 1765).
G. Caltais and J. Krivine (eds.): CREST 2019
EPTCS 308, 2019, pp. 47 -- 65, doi:10.4204/EPTCS.308.4
c(cid:13) W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
This work is licensed under the
Creative Commons Attribution License.
48
Dynamic Conflict Resolution
provides a mean to summarise consistent evidences and keep them apart from inconsistent evidences.
We hence can, e.g., fuse compatible perceptions into a belief b about the world and fuse another set of
compatible perceptions to a belief b(cid:48) and model decisions that take into account that b might contradict
b(cid:48). Using the framework we illustrate how conflicts can be explained and algorithmically analysed as
required for our conflict resolution process. Finally we report on a small case study using a prototype
implementation (employing the Yices SMT solver [14]) of the conflict resolution algorithm. We discuss
related work in Sect. 5. In particular we discuss work regarding the notion of traffic conflict and relate
our works with work on the perimeter in game theory [10] and strategy synthesis for levels of cooperation
like [11, 7].
Outline
In Sect. 2 we introduce the types of conflict on a running example and develop a formal notion
of conflict between two agents. We elaborate on the logical foundations for modelling and analysing
conflicts and the logical framework itself in Sect. 3. We sketch our case study on conflict analysis
in Sect. 4 and outline in Sect. 4.2 an algorithm for analysing conflict situations as requested by our
resolution protocol and for deriving explanation of the conflict for the resolution. Before drawing the
conclusions in Sect. 6, we discuss related work in Sect. 5.
2 Conflict
Already in 1969 in the paper "Violence, Peace and Peace Research" [18] J. Galtung presents his theory
of the Conflict Triangle, a framework used in the study of peace and conflict. Following this theory a
conflict comprises three aspects: opposing actions, incompatible goals, inconsistent beliefs (regarding
the reasons of the conflict, knowledge of the conflict parties,. . . ).
We focus on conflicts that arise dynamically between two agents in road traffic. We develop a char-
acterisation of conflict as a situation where one agent can accomplish its goals with the help of the other,
but both agents cannot accomplish all their goals simultaneously and the agents have to decide what to
do based on their local beliefs. In Sect. 2.1 we formalise our notion of conflict. For two agents with
complete information, we may characterise a conflict as: Agents A and B are in conflict, if 1. A would
accomplish its set of goals ΦA, if B will do what A requests, while 2. B would accomplish its set of
goals ΦB, if A will do what B requests, and 3. it is impossible to accomplish the set of goals ΦA ∪ ΦB.
A situation where A and B both compete to consume the same resource is thus an example of a conflict
situation. Since we study conflicts from the view-point of an agent's beliefs, we also consider believed
conflicts, which can be resolved by sharing information regarding the others observations, strategies or
goals. To resolve a conflict we propose a sequence of steps that require an increasing level of cooperation
and decreasing level of privacy -- the steps require to reveal information or to constrain acting options.
Our resolution process defines the following steps:
(C1) Shared situational awareness
(C2) Sharing strategies
(C3) Sharing goals
(C4) Agreeing on which goals to sacrifice and which strategy to follow
Corresponding to (C1) to (C4), we introduce different kinds of conflicts on a running example -- a two
lane highway, where one car, A, is heading towards an obstacle at its lane and at the lane to its left a fast
car, B, is approaching from behind (cf. Fig. 1). An agent has a prioritised list of goals (like 1. "collision-
freedom", 2."changing lane" and 3. "driving fast"). We assume that an agent's goals are achievable.
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
49
Figure 1: Car A wants to circumvent the
obstacle (grey box). Car B is approaching
from behind.
An agent A has a set of actions actA and exists within a world. At a time the world has a certain
state. The world "evolves" (changes state) as determined by the chosen actions of the agents within the
world and events determined by the environment within the world. The agent perceives the world only
via a set of observation predicates, that are predicates whose valuation is determined by an observation
of the agent. Without an observation the agent has no (direct) evidence for the valuation of the respective
observation predicate.
Example 1. Let car A want to change lane. It perceives that it is on a two lane highway, the way ahead
is free for the next 500 m and B is approaching. Let A perceive B's speed via radar. That is A makes the
observation car B is fast justified by the evidence radar. We annotate this briefly as radar:car B
is fast. Further let A derive from lidar data that B is slow -- lidar:car B is slow.
In this situation we say agent A has contradicting evidences. Certain evidences can be combined
without contradiction and others not. We assume that an agent organises its evidences in maximal con-
sistent sets (i.e., justification graphs of Sect. 3), where each represents a set of possible worlds:
Example 2. There are possible worlds of A where it is on a two lane highway, the way ahead is free for
the next 500 m and B is slowly approaching. Analogously A considers possible worlds where B is fast.
The state of the world outside of its sensors' reach is unconstrained.
Observing the world (for some time), an agent A assesses what it can do to achieve its goals in all
possible worlds. That is, A tries to find a strategy that guarantees to achieve its goals in all its possible
worlds. A strategy determines at each state the action of the agent -- the agent decides for an action based
on its believed past. If there is one such strategy for A to accomplish its goals ΦA, then A has a (believed)
winning strategy for ΦA. This strategy might not be winning in the "real" world though, e.g., due to
misperceptions.
Example 3. Let A want to drive slowly and comfortably. A wants to avoid collisions and it assumes that
also B wants to avoid collisions. Although A has contradicting evidences on the speed of B and hence
believes that it is possible that "B is fast" and also that "B is slow", it can follow the strategy to stay at
its lane and wait until B has passed. This strategy is winning in all of A's possible worlds.
Even when A has no believed winning strategy, it can have a winning strategy for a subset of possible
worlds. Additional information on the state of world might resolve the conflict by eliminating possible
worlds. We call such conflicts observation-resolvable conflicts.
Example 4. Let A want to change lane to circumvent the obstacle. It is happy to change directly after B
but only if B is fast. If B is slow, it prefers to change before B passed. Further let A have contradicting
evidences on the speed of B. A considers a conflict with B possible in some world and hence has no
believed winning strategy. Now it has to resolve its inconsistent beliefs. Let B tell A, it is fast, and A trust
B more than its own sensors, then A might update its beliefs by dismissing all worlds where B is slow.
Then "changing after B passed" becomes a believed winning strategy.
In case of inconsistent evidences, as above, A has to decide how to update its beliefs. The decision
how to update its beliefs will be based on the analysis of justifications (cf. Sect. 3) of (contradicting)
evidences. The lidar contradicts the radar and B reports on its speed. Facing the contradiction of
evidences justified by lidar and radar A trusts the evidence justified by B.
AB 50
Dynamic Conflict Resolution
Let the agents already have exchanged observations and A still have no believed winning strategy. A
conflict might be resolved by communicating part of the other agent's (future) strategy:
Example 5. Let A want to change lane. It prefers to change directly after B, if B passes A fast. Otherwise,
A wants to change in front of B. Let B so far away that B might decelerate, in which case it might slow
down so heavily that A would like to change in front of B even if B currently is fast.
Let A believe "B is fast". Now A has no believed winning strategy, as B might decelerate. According
to (C2), information about parts of the agent's strategies are now communicated. A asks B whether it
plans to decelerate. Let B be cooperative and tell A that it will not decelerate. Then A can dismiss all
worlds where B slows down and "changing after B passed" becomes a believed winning strategy for A.
Let the two agents have performed steps (C1) and (C2), i.e., they exchanged missing observations
and strategy parts, and still A has no winning strategy for all possible worlds.
Example 6. Let now, in contrast to Ex. 5, B not tell A whether it will decelerate. Then step (C3) is
performed. So A asks B to respect A's goals. Since A prefers B to be fast and B agrees to adopt A's goal
as its own, A can again dismiss all worlds where B slows down.
Here the conflict is resolved by communicating goals and the agreement to adopt the other's goals.
So an agent's strategy might change in order to support the other agent. We call this kind of conflicts
goal-disclosure-resolvable conflicts.
The above considered conflicts can be resolved by some kind of information exchange between the
two agents, so that the sets of an agent's possible worlds is adapted and in the end all goals ΦA of A and
ΦB of B are achievable in all remaining possible worlds. The price to pay for conflict resolution is that
the agents will have to reveal information. Still there are cases where simply not all goals are (believed to
be) achievable. In this case A and B have to negotiate which goals ΦAB ⊆ ΦA∪ΦB shall be accomplished.
While some goals may be compatible, other goals are conflicting. We hence consider goal subsets ΦAB
for which a combined winning strategy for A and B exists. We assume that there is a weight assignment
function w that assigns a value to a given goal combination 2ΦA∪ΦB → N based on which decision for
a certain goal combination is taken. This weighting of goals reflects the relative value of goals for the
individual agents. Such a function will have to reflect, e.g., moral, ethics and jurisdiction.
Example 7. Let A's and B's highest priority goal be collision-freedom, reflected in goals ϕA,col and
ϕB,col. Further let A want to go fast ϕA, f ast and change lane immediately ϕA,lc. Let also B want to go fast
ϕB, f ast, so that A cannot change immediately. Now in step (C4) A and B negotiate what goals shall be
accomplished. In our scenario collision-freedom is valued most, and B's goals get priority over A's, since
B is on the fast lane. Hence our resolution is to agree on a strategy accomplishing {ϕA,col,ϕB,col,ϕB, f ast},
which is the set of goals having the highest value among all those for which a combined winning strategy
exists.
Note that additional agents are captured as part of the environment here. At each step an agent can
also decide to negotiate with some other agent than B in order to resolve its conflict.
2.1 Formal Notions
In the following we introduce basic notions to define a conflict. Conflicts, as introduced above, arise in a
wide variety of system models, but we consider in this paper only a propositional setting.
Let f1 : X → Y1,. . . , fn : X → Yn, and f : X → Y1× . . .×Yn be functions. We will write f = ( f1, . . . , fn)
if and only if f (x) = ( f1(x), . . . , fn(x)) for all x ∈ X. Note that for any given f as above the decomposition
into its components fi is uniquely determined by the projections of f onto the corresponding codomain.
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
51
Each agent A has a set of actions AA. The sets of actions of two agents are disjoint. To formally
define a (possible) world model of an agent A, let S be a set of states and V be a set of propositional
variables. A believes at a state s ∈ S that a subset V of V is true and V\V is false. A (possible) world
model M for an agent A is a transition system over S with designated initial state and current state, all
states are labelled with the belief propositions that hold at that state and transitions labeled with actions
(cid:104)actA,actB,actEnv(cid:105) where actA ∈ AA is an action of A, actB ∈ AB an action of B and actEnv ∈ AEnv an
action of the environment. The set of actions of an agent includes send and receive actions via which
information can be exchanged, the environment guarantees to transmit a send message to the respective
receiver. Formally a possible world is MA = (S,T,λ ,π,s∗,sc) with T ⊆ S× S, λ : T → AA × AB × AEnv,
π : S → 2V and s∗,sc ∈ S representing the initial state and the current state respectively. We use MA
to encode the current believes on the present, past and about the possible futures. If an agent, let us
say A wlog, follows a strategy, it decides for an action based on its believed past, i.e., a strategy for
A is a function δA : (2V)∗ → AA. Given strategies δA : (2V)∗ → AA for A, δB : (2V)∗ → AB for B,
δ = (δA,δB) : (2V)∗ → AA × AB is a common strategy of A and B, which chooses the actions of A
according to δA and the actions of B according to δB.
A (believed) run r = (s0,s1, . . .) is an infinite sequence of states starting with the initial state s0 = s∗
and (si,si+1) ∈ T for all i ∈ N. A run r = (s0,s1, . . .) results from a strategy δA in A's world MA, denoted as
r ∈ r(δA,MA), if and only if λ (si,si+1) = δA(π(s0),π(s1), . . .π(si)) for all i ∈ N. Given a set of possible
worlds M(A) for A, we use r(δA, M(A)) to denote the set of runs that result from δA in a MA ∈ M(A),
r(δA, M(A)) =(cid:83)
MA∈M(A) r(δA,MA).
write r = Φ as shorthand for r =(cid:86)
We use linear-time temporal logic (LTL) to specify goals. For a run r and a goal (or a conjunction of
goals) ϕ we write r = ϕ, if the valuation of propositions along r satisfies ϕ1. We say δ is a (believed)
winning strategy for ϕ in MA, if for all r ∈ r(δ ,MA) it holds that r = ϕ. An agent A has a set of goals
Φ and a weight assignment function wA : 2Φ → N that assigns values to a given goal combination. We
ϕ∈Φ ϕ. We say subgoal Φ(cid:48) ⊆ Φ is maximal for r if r = Φ(cid:48) and
w(Φ(cid:48)) ≥ w(Φ(cid:48)(cid:48)) for all Φ(cid:48)(cid:48) ⊆ Φ with r = Φ(cid:48)(cid:48). true is the empty subgoal. We say Φ(cid:48) ⊆ Φ is maximal for
a set R of runs if for all r ∈ R r = Φ(cid:48) and for all Φ(cid:48)(cid:48) ⊆ Φ with r = Φ(cid:48)(cid:48) for all r ∈ R, w(Φ(cid:48)) ≥ w(Φ(cid:48)(cid:48)).
There is one "special" world model that represents the ground truth, i.e., it reflects how the reality
evolves. We refer the interested reader to [12] for a more elaborate presentation of our concept of reality
and associated beliefs. Agent A considers several worlds possible at a time. At each state s of the real
world A has a set of possible worlds MA(s) and for each world MA ∈ M(A) a believed current state
and beliefs on the goals of B, ΦB(MA) and the goal weight assignment function of B, wB. A possible
world is labeled with the set of evidences that justifies that the world is regarded as possible. The real
world changes states according to the actions of A, B and E. The set of possible worlds MA(s) changes
to MA(s(cid:48)) due to the believed passing of time and due to belief updates triggered by e.g. observations.
For the scope of this paper though, we do not consider the actual passing of time, but study the conflict
analysis at a single state of the real world.
strategy for ΦA in all possible worlds MA ∈ MA(sc).
Definition 1 (Believed Possible Conflict). Let Φmax
which a believed winning strategy (δA,δ(cid:48)
(δA,δ(cid:48)
Agent A believes at state s it is in a possible conflict with B, if for each of its winning strategies
B) : (2VA)∗ → AA × AB for a maximal subgoal ΦA ∈ Φmax
We say that A has a (believed) winning strategy δA for ΦA at the real world state sc if δA is a winning
be the set of maximal subgoals of A at state s for
A
B) : (2VA)∗ → AA × AB in MA exists.
,
A
1cf. Def. 9
52
Dynamic Conflict Resolution
A,δB) : (2VA)∗ → AA × AB and a possible world M ∈ MA such that (δ(cid:48)
winning strategy in M for ΦB, a believed maximal subgoal of the believed goals of B in M.
• there is a strategy (δ(cid:48)
• but (δA,δB) is not a winning strategy for ΦA ∪ ΦB in MA.
A,δB) is a
A
In Def. 1 Φmax
is the set of maximal subgoals that A can achieve in all possible worlds with the help
of B. A believes that B might decide for a strategy to accomplish some of its maximal subgoals and B
takes this decision wrt A's possible worlds. Note that A assuming the goals of B being true means that
A has to deal with arbitrary behaviour of B. Also note that B always has a winning strategy in every M
since ΦB is maximal wrt M. If A cannot find one winning strategy that fits all possible choices of B then
A believes that it is in a conflict with B. Note that A analyses the conflict within its possible worlds MA(s)
and in particular it beliefs that B believes that in one of its possible wolds. That A has got beliefs about
(deviations of) the beliefs of B is an interesting future extension.
3 Epistemic Logic, Justifications and Justification Graph
Conflict analysis demands to know who believes to be in conflict with whom and what pieces of infor-
mation made him belief that he is in conflict. To this end we introduce the logic of justification graphs
that allows to keep track of external information and extends purely propositional formulae by so called
belief atoms (cf. p. 54), which are used to label the sources of information. In Sect. 2 we already used
such formulae, e.g., "radar:car B is fast". Our logic provides several atomic accessibility relations
representing justified beliefs of various sources, as required for our examples of Sect. 2. It provides jus-
tification graphs as a mean to identify belief entities which compose different justifications to consistent
information even when the information base contains contradicting information of different sources, as
required for analysing conflict situations.
First, this section provides a short overview on epistemic modal logics and multi-modal extensions
thereof. Such logics use modal operators to expressing knowledge and belief stemming from different
sources. Often we will refer to this knowledge and belief as information, especially when focusing on the
sources or of the information. Thereupon the basic principles of justifications logics are shortly reviewed.
Justification logics are widely seen as interesting variants to epistemic logics as they allow to trace back
intra-logical and external justifications of derived information. In the following discussion it turns out
that tracing back external justifications follows the same principles as the distribution of information over
different sources.
Consequently, the concept of information source and external justification are then unified in our
variant of an epistemic modal logic. This logic of justification graphs extends the modal logic by a
justification graph. The nodes of a justification graph are called belief entities and represent groups
of consistent information. The leaf nodes of a justification graphs are called belief atoms, which are
information source and external justifications at the same time, as they are the least constituents of
external information. We provide a complete axiomatisation with respect to the semantics of the logic of
justification graphs.
Justification Graphs
3.1
Modal Logic and Epistemic Logic Modal logic extends the classical logic by modal operators ex-
pressing necessity and possibility. The formula (cid:50)φ is read as "φ is necessary" and (cid:51)φ is read as "φ is
possible". The notions of possibility and necessity are dual to each other, (cid:51)φ can be defined as ¬(cid:50)¬φ.
The weakest modal logic K extends propositional logic by the axiom K(cid:50) and the necessitation rule Nec(cid:50)
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
53
as follows
(K(cid:50))
(cid:96)(cid:50)(φ → ψ) → ((cid:50)φ →(cid:50)ψ),
(Nec(cid:50))
The axiom K(cid:50) ensures that whenever φ → ψ and φ necessarily hold, then also ψ necessarily has to hold.
The necessitation rule Nec(cid:50) allows to infer the necessity of φ from any proof of φ and, hence, pushes
any derivable logical truth into the range of the modal operator(cid:50). This principle is also known as logical
awareness. Various modalities like belief or knowledge can be described by adding additional axioms
encoding the characteristic properties of the respective modal operator. The following two axioms are
useful to model knowledge and belief:
from (cid:96) φ conclude (cid:96)(cid:50)φ.
(cid:96)(cid:50)φ → φ,
(cid:96)(cid:50)φ →(cid:51)φ.
(T(cid:50))
it postulates the weaker property that everything which is necessary is also possible. Under both axioms
(D(cid:50))
The axiom T(cid:50) and D(cid:50) relate necessity with the factual world. While the truth axiom T(cid:50) characterises
knowledge as it postulates that everything which is necessary is also factual, D(cid:50) characterises belief as
(cid:96)(cid:50)⊥ → ⊥ holds, i.e. a necessary contradiction yields also a factual contradiction.
Multi-modal logics are easily obtained by adding several modal operators with possibly different
properties and can be used to express the information of more than one agent. E.g., the formula ei:φ
expresses that the piece of information φ belongs to the modality ei. Modal operators can also be used
to represent modalities referring to time. E.g., in the formula Xφ the temporal modality X expresses that
φ will hold in the next time step. An important representative of a temporal extension is linear temporal
logic (LTL).
In multi-agent logics the notions of common information and distributed information play an im-
portant role. While common knowledge captures the information which is known to every agent ei, we
are mainly interested in information that is distributed within a group of agents E = {e1, . . . ,en}. The
distributed information within a group E contains any piece of information that at least one of the agent
e1, . . . , en has. Consequently, we introduce a set-like notion for groups, where an agent e is identified
with the singleton group {e} and the expression {e1, . . . ,en}:φ is used to denote that φ is distributed
information within the group E. The distribution of information is axiomatised by
(cid:96) E:φ → F:φ, where E is a subgroup of F.
(DistE,F)
Note that groups may not be empty. The modal logic for distributed information contains for every group
E at least the axiom KE, the necessitation rule NecE, and the axiom DistE,F for any group F with E ⊆ F.
Justification Logics
Justification logics [5] are variants of epistemic modal logics where the modal
operators of knowledge and belief are unfolded into justification terms. Hence, justification logics allow
a complete realisation of Plato's characterisation of knowledge as justified true belief. A typical formula
of justification logic has the form s:φ, where s is a justification term built from justification constants,
and it is read as "φ is justified by s". The basic justification logic J0 results from extending propositional
logic by the application axiom and the sum axioms
(cid:96) s:(φ → ψ) → (t:φ → [s·t]:ψ),
(Sum)
where s, t, [s·t], [s +t], and [t +s] are justification terms which are assembled from justification constants
using the operators + and · according to the axioms. Justification logics tie the epistemic tradition
(cid:96) s:φ → [s +t]:φ,
(cid:96) s:φ → [t + s]:φ,
(Appl)
54
Dynamic Conflict Resolution
together with proof theory. Justification terms are reasonable abstractions for constructions of proofs. If
s is a proof of φ → ψ and t is a proof of φ then the application axiom postulates that there is a common
proof, namely s·t, for ψ. Moreover, if we have a proof s for φ and some proof t then the concatenations
of both proofs, s + t and t + s, are still proofs for φ. In our framework we were not able to derive any
meaningful example using the sum axiom of justification logic. Therefore this axiom is omitted in the
following discussion.
Discussion All instances of classical logical tautologies, like A∨¬A and s:A∨¬s:A, are provable in
justification logics. But in contrast to modal logics, justification logics do not have a necessitation rule.
The lack of the necessitation rule allows justification logics to break the principle of logical awareness,
as s:(A∨¬A) is not necessarily provable for an arbitrary justification term s. Certainly, restricting the
principle of logical awareness is attractive to provide a realistic model of restricted logical resources.
Since we are mainly interested in revealing and resolving conflicts, the principle of logical awareness is
indispensable in our approach.
Nevertheless, justification logic can simulate unrestricted logical awareness by adding proper axiom
internalisation rules (cid:96) e:φ for all axioms φ and justification constants e. In such systems a weak variant
of the necessitation rule of modal logic holds: for any derivation (cid:96) φ there exists a justification term t
such that (cid:96) t:φ holds. Since φ was derived using axioms and rules only, also the justification term t is
exclusively built from justification constants dedicated to the involved axioms. Beyond that, t is hardly
informative as it does not help to reveal external causes of a conflict. Hence, we omit the axiom internal-
isation rule and add the modal axiom Kt and the modal necessitation rule Nect for any justification term t
to obtain a justification logic where each justification term is closed under unrestricted logical awareness.
An important consequence of the proposed system is that · becomes virtually idempotent and com-
mutative.2 These insights allows us to argue merely about justification groups instead of justification
terms. It turns out that a proper reformulation of Appl with regard to justification groups is equivalent to
DistE,F, finally yielding the same axiomatisation for distributed information and compound justifications.
Belief Atoms, Belief Groups, and Belief Entities So far, we argued that assembling distributed in-
formation and compound justifications follow the same principle. In the following we even provide a
unified concept for the building blocks of both notions. A belief atom e is the least constituent of ex-
ternal information in our logic. To each e we assign the modal operator e:. Hence, for any formula φ
also e:φ is a formula saying "e has information φ". Belief atoms play different roles in our setting. A
belief atom may represent a sensor collecting information about the state of the world, or it may represent
certain operational rules as well as a certain goal of the system. The characteristic property of a belief
atom is that the information of a belief atom has to be accepted or rejected as a whole. Due to its external
and indivisible nature, e is the only source of evidence for its information. The only justification for
information of e is e itself. Consequently, e:φ can also be read as "e is the justification for φ". This is
what belief atoms and justifications have in common: either we trust a justification or not.
The information of a system is distributed among its belief atoms. The modal logic for distributed
information allows us to consider the information which is distributed over a belief group. While belief
groups can be built arbitrarily from belief atoms, we also introduce the concept of belief entities. A belief
entity is either a belief atom, or a distinguished group of belief entities. Belief entities are dynamically
2For any instance (cid:96) s:(φ → ψ) → (s:φ → [s· s]:ψ) of Appl there is an instance (cid:96) s:(φ → ψ) → (s:φ → s:ψ) of Ks in the
proposed system. Moreover, it is an easy exercise to show that any instance of (cid:96) s:(φ → ψ) → (t:φ → [t · s]:ψ) is derivable in
the proposed system.
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
55
distinguished by a justification graph. In contrast to belief groups, belief entities and belief atoms are not
allowed to have inconsistent information. Hence a justification graph allows us to restrict the awareness
of extra-logical evidences -- so we can distinctively integrate logical resources that have to be consistent.
Justification Graphs Let V be a set of propositional variables and let E be the set of belief entities.
The designated subset EA of E denotes the set of belief atoms.
Definition 2 (Language of Justification Graphs). A formula φ is in the language of justification graphs
if and only if φ is built according to the following BNF, where A ∈ V and /0 (cid:54)= E ⊆ E:
φ ::= ⊥ A (φ → φ ) E:(φ ) X(φ ) P(φ ) (φ )U(φ ) (φ )S(φ ).
Using the descending sequence of operator precedences (:, ¬, ∨, ∧, →, ↔), we can define the well-
known logical connectives ¬, ∨, ∧ and ↔ from → and ⊥. Often, we omit brackets if the formula is still
uniquely readable. We define → to be right associative. For singleton sets {e} ⊆ E we also write e:φ
instead of {e}:φ. The language allows the usage of temporal operators for next time (X), previous time
(P), until (U), and since (S). Operators like always in the future (G) or always in the past (H) can be
defined from the given ones.
Definition 3 (Justification Graph). A justification graph is a directed acyclic graph G whose nodes are
belief entities of E. An edge e (cid:55)→G f denotes that the belief entity e has the component f . The set of all
direct components of an entity e is defined as G(e) := { f e (cid:55)→G f}.
holds e ∈ EA if and only if G(e) = /0.
Definition 4 (Axioms of a Justification Graph). Let G be a justification graph. The logic of a justification
graph has the following axioms and rules.
(i) As an extension of propositional logic the rule of modus ponens MP has to hold: from (cid:96) φ and
The leaf nodes of a justification graph are populated by belief atoms, i.e. for any belief entity e it
(cid:96) φ → ψ conclude (cid:96) ψ. Any substitution instance of a propositional tautology φ is an axiom.
(ii) Belief groups are closed under logical consequence and follow the principle of logical awareness.
Information is freely distributed along the subgroup-relation. For any belief group E the axiom KE
and the necessitation rule NecE hold. For groups E and F with E ⊆ F the axiom DistE,F holds.
(iii) Belief entities are not allowed to have inconsistent information. Non-atomic belief entities inherit
all information of their components. For any belief entity e the axiom De holds. If E is a subgroup
of the components of e, then the axiom DistE,e holds.
Past-LTL (LTL with past operator). A comprehensive list of axioms can be found in [24].
(iv) In order to express temporal relation the logic for the justification graph includes the axioms of
(v) Information of a belief entity e ∈ E and time are related. The axiom (PRE) : (cid:96) e:Pφ ↔ Pe:φ
ensures that every belief entity e correctly remembers its prior beliefs and establishes a principle
which is also known as perfect recall (e.g., see [17]).
Definition 5 (Proof). Let G be a justification graph. A proof (derivation) of φ in G is a sequence of
formulae φ1, . . . ,φn with φn = φ such that each φi is either an axiom of the justification graph or φi is
obtained by applying a rule to previous members φ j1, . . . ,φ jk with j1, . . . , jk < i. We will write (cid:96)G φ if
and only if such a sequence exists.
Definition 6 (Proof from a set of formulae). Let G be a justification graph and Σ be a set of formulae.
The relation Σ (cid:96)G φ holds if and only if (cid:96)G (σ1 ∧···∧ σk) → φ for some finite subset {σ1, . . . ,σk} ⊆ Σ
with k ≥ 0.
56
Dynamic Conflict Resolution
Definition 7 (Consistency with respect to a justification graph). Let G be a justification graph.
(i) A set Σ of formulae is G-inconsistent if and only if Σ (cid:96)G ⊥. Otherwise, Σ is G-consistent. A formula
(ii) A set Σ of formulae is maximally G-consistent if and only if Σ is G-consistent and for all φ (cid:54)∈ Σ the
φ is G-inconsistent if and only if {φ} is G-inconsistent. Otherwise, φ is G-consistent.
set Σ∪{φ} is G-inconsistent.
Semantics Let S be the state space, that is the set of all possible states of the world. An interpretation
π over S is a mapping that maps each state s to a truth assignment over s, i.e. π(s) ⊆ V is the subset of
all propositional variables which are true in the state s. In Sect. 2 we introduced world models and runs
on world models. There a world model captured the evolution of a states in time.
In this section we focus an the epistemic notions of knowledge and belief and therefore our main
concern is the accessibility relation of information. We hence presume that the set of runs of a possible
world of Sect. 2 is given, that then defines the evolution in time. Formally a run over S is a function r
from the natural numbers (the time domain) to S. The set of all runs is denoted by R.
Definition 8. Let G be a justification graph. A Kripke structure M for G is a tuple M = (S, R,π, ((cid:55)→e)e∈E)
where
(i) S is a state space,
(ii) R is the set of all runs over S,
(iii) π is an interpretation over S,
(iv) each (cid:55)→e in ((cid:55)→e)e∈E is an individual accessibility relation (cid:55)→e⊆ S× S for a belief entity e in E.
Definition 9 (Model for a Justification Graph). Let M = (S, R,π, ((cid:55)→e)e∈E) be a Kripke structure for the
justification graph G, where
(i) (cid:55)→e is a serial relation for any belief entity e ∈ E,
e∈E (cid:55)→e for any belief group E ⊆ E,
(iii) (cid:55)→e ⊆ (cid:55)→E holds for all non-atomic belief entities e ∈ E\ EA and any subgroup E ⊆ G(e).
We recursively define the model relation (M,r(t)) =G φ as follows:
(ii) (cid:55)→E is defined as (cid:55)→E =(cid:84)
(M,r(t)) (cid:54)=G ⊥.
(M,r(t)) =G Q
(M,r(t)) =G φ → ψ :⇐⇒ (M,r(t)) =G φ implies (M,r(t)) =G ψ.
(M,r(t)) =G E:φ
(M,r(t)) =G Xφ
(M,r(t)) =G Pφ
(M,r(t)) =G φUψ
(M,r(t)) =G φSψ
:⇐⇒ Q ∈ π(r(t)).
:⇐⇒ (M,r(cid:48)(t)) =G φ for all r(cid:48) with r(t(cid:48)) (cid:55)→E r(cid:48)(t(cid:48)) for all t(cid:48) ≤ t.
:⇐⇒ (M,r(t + 1)) =G φ.
:⇐⇒ (M,r(t(cid:48))) =G φ for some t(cid:48) with t(cid:48) + 1 = t.
:⇐⇒ (M,r(t(cid:48))) =G ψ for some t(cid:48) ≥ t and
(M,r(t(cid:48)(cid:48))) =G φ for all t(cid:48)(cid:48) with t ≤ t(cid:48)(cid:48) < t(cid:48).
:⇐⇒ (M,r(t(cid:48))) =G ψ for some 0 ≤ t(cid:48) ≤ t and
(M,r(t(cid:48)(cid:48))) =G φ for all t(cid:48)(cid:48) with t(cid:48) < t(cid:48)(cid:48) ≤ t.
When (M,r(t)) =G φ holds, we call (M,r(t)) a pointed model of φ for G. If (M,r(0)) is a pointed
model of φ for G, then we write (M,r) =G φ and say that the run r satisfies φ. Finally, we say that φ is
satisfiable for G, denoted by =G φ if and only if there exists a model M and a run r such that (M,r) =G φ
holds.
Proposition 1 (Soundness and Completeness). The logic of a justification graph G is a sound and com-
plete axiomatisation with respect to the model relation =G. That is, a formula φ is G-consistent if and
only if φ is satisfiable for G.
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
57
While the soundness proof is straightforward, a self-contained completeness proof involve lengthy
sequences of various model constructions and is far beyond the page limit. However, it is well-known,
(e.g., [19]), that KD
n , the n-agent extension of K with distributive information is a sound and complete
axiomatisation with respect to the class of Kripke structures having n arbitrary accessibility relations,
where the additional accessibility relations for groups are given as the intersection of the participating
agents, analogously to Def. 9.(ii). Also the additional extension KDD
n with DE for any belief group E is
sound and complete with respect to Kripke structures having serial accessibility relations, analogously
to Def. 9.(i). The axioms of justification graph are between these two systems. Def. 9.(iii) explicitly
allows belief entities to have more information than its components. Various completeness proofs for
combining LTL and epistemic logics are given e.g., in [17].
Extracting Justifications Let Σ = {σ1, . . . ,σn} be a finite set of formulae logically describing the
situation which is object of our investigation. Each formula σi ∈ Σ encodes information of belief atoms
(σi ≡ ei:φi with ei ∈ EA), facts (σi ≡ φi where φi does not contain any epistemic modal operator), or is an
arbitrary Boolean combinations thereof. Further, let G be a justification graph such that Σ is G-consistent
and e be a non-atomic belief entity of G. For any formula φ we may now ask whether φ is part of
the information of e. If there is a proof Σ (cid:96)G e:φ, then φ is included in e's information. To extract a
justification for e:φ we use that Σ∪{¬e:φ} is G-inconsistent and accordingly unsatisfiable for G. If we
succeed in extracting a minimal unsatisfiable core Σ(cid:48) ⊆ Σ∪{¬e:φ} a minimal inconsistency proof can
be recovered, from which finally the used justifications are extracted.
The following proposition allows to use SAT/SMT-solvers for a restricted setting and has been used
in our case study.
Proposition 2 (SAT Reduction). Let Σ = {σ1, . . . ,σn} be a set of formulae such that each element σi is
of the form ei:φi with ei ∈ EA and φi does not contain any epistemic modal operators. Further, let e be
an arbitrary belief entity that does not occur in Σ. Then G = {e (cid:55)→G eiei occurs in Σ} is a justification
graph for Σ if and only if Φ = {φ1, . . .φn} is satisfiable over the non-epistemic fragment of the logic of
justification graphs.
In order to proof the proposition one shows that any model of Φ in the non-epistemic fragment can
be extended to a model of Σ for the given G by adding trivial accessibility relations. On the other hand,
for any model M and run r with (M,r) =G Σ there exists a run r(cid:48) which is accessible from (cid:55)→e such
that (M,r(cid:48)) =G Φ. Since Φ does not contain any epistemic modal operators, dropping the accessibility
relations from M still yields a model of Φ. A more detailed version of this proof can be found in [12].
4 Identifying and Analysing Conflicts
In this section we first present an abstract algorithm for the conflict resolution of Sect. 2 that starts at
level (C1) and proceeds resolution stepwise up to level (C4). We then sketch our small case study where
we applied an implementation of the abstract algorithm.
4.1 Analysing Conflicts
For the analysis of conflicts we employ SMT solvers. Prop. 2 reduces the satisfiability of a justification
graph to a SAT problem. To employ SMT solving for conflict analysis, we encode the (real and possible)
worlds of Sect. 2 via logic formulae as introduced in Sect. 3. Each state si is represented as a conjunction
of literals, si ≡(cid:86)v∧(cid:86)¬v(cid:48). Introducing a dedicated propositional variable vt for each v ∈ V and time step
10:
11:
12:
13:
14:
15:
16:
17:
for i ∈ [1,2,3,4] do
(cid:48),Φmax
B
if ∆A = /0 then
Σ(cid:48),Φmax
A
if (Σ(cid:48) (cid:54)= Σ)∨ (Φmax
A )∨ (Φmax
∆A ← FINDSTRATEGY(Σ(cid:48),Φmax
if ∆A (cid:54)= /0 then
(cid:48) (cid:54)= Φmax
B
A
A
break
return ∆A
(cid:48) ← FIXCONFLICT(C, (Ci),Σ,Φmax
A ,Φmax
B )
(cid:48) (cid:54)= Φmax
(cid:48),Φmax
B
B ) then
(cid:48), AA, AB)
58
Dynamic Conflict Resolution
B , AA, AB)
A ,Φmax
ΣM ← POSSIBLEWORLDS(Σ, AA, AB)
∆A ← STRATA(AA, AB,ΣM,Φmax
A )
C ← /0
B) ∈ ∆A with r((δ A,δ(cid:48)
for all (δ A,δ(cid:48)
Algorithm 1 Determining winning strategy based on observations, goals, and possible actions.
1: function FINDSTRATEGY(Σ,Φmax
2:
3:
4:
5:
6:
7:
8:
9:
(cid:46) construct(cid:8)(δ A,δ(cid:48)
B),ΣM) = ΦA ∈ Φmax
B),ΣM,ΦA,Φmax
(cid:46) (δ A,δ(cid:48)
E ← TESTIFNOTWINNING((δ A,δ(cid:48)
if E (cid:54)= /0 then
B) is not winning for all ΣM ∈ ΣM, i.e. r((δ A,δ(cid:48)
C ← C∪{E}
∆A = ∆A \{(δ A,δ(cid:48)
A do
B , AA, AB)
B) r((δ A,δ(cid:48)
B)}
(cid:46) construct set of possible worlds
B),ΣM) = ΦA with ΦA ∈ Φmax
(cid:46) set of conflict causes
A
(cid:46) cf. Alg. 2
B),ΣM) (cid:54)= ΦA
(cid:46) memorize justifications E
(cid:9)
(cid:46) A is in conflict with B
(cid:46) traverse resolution levels
(cid:46) cf. Alg. 3
(cid:46) new information generated
(cid:46) new attempt with new information
(cid:46) new attempt was successful, stop and return
(cid:46) select (δ A,δ(cid:48)
B) ∈ ∆A to reach some goal in Φmax
A
t allows us to obtain a formula describing a finite run on M. A predicate of the form(cid:86)
t+1)
encodes the transition relation T. The effect of performing an action at at state s is captured by a formula
of the form at → (st → s(cid:48)
t+1). Using this we can encode a strategy δ in a formula ψδ such that its
valuations represent runs of M according to δ . All runs according to δ achieve goals Φ if and only if
ψδ ∧¬Φ is unsatisfiable. These logical encodings are the main ingredients for using a SAT solver for
our conflict analysis. Since there are only finitely many possible strategies, we examine for each strategy
which goals can be (maximally) achieved in a world M or in a set of worlds M. Likewise we check
whether A has a winning strategy that is compatible with the strategies A believes B might choose.
(s,s(cid:48))∈T(st → s(cid:48)
Since we iterate over all possible worlds for our conflict analysis, we are interested in summarising
possible worlds. We are usually not interested in all vt -- e.g. the speed of B may at times t be irrelevant.
We are hence free to ignore differences in vt in different possible worlds and are even free to consider
all valuations of vt, even if A does not consider them possible. This insight leads us to a symbolic
representation of the possible worlds, collecting the relevant constraints. Now the justification graph
groups the constraints that are relevant, with other words, e:φ and e(cid:48):¬φ will not be components of the
same justification graph if the valuation of φ is relevant. In the following we hence consider the maximal
consistent set of possible worlds, meaning encodings of possible worlds that are uncontradictory wrt. the
relevant propositions, which are specified via the justification graph.
4.2 Algorithmic approach
In this section, we sketch an abstract algorithm for the conflict resolution at levels (C1) to (C4) as in
Sect. 2. Note that we do not aim with Alg. 1 for efficiency or optimal solutions but aim to illustrate how
satisfiability checks can be employed to analyse our conflicts.
The following algorithms describe how we deal with logic formulae encoding sets of possible worlds,
sets of runs on them, etc. to analyse conflicts (cf. Def. 1, p. 51) via SMT solving. We use ΣM to refer to a
formula that encodes a maximal consistent set of possible worlds (cf. Sect. 4.1), i.e., that corresponds to a
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
59
Figure 2: Abstract resolution process with information base, possible worlds and strategies.
justification graph. We use ΣM to refer to a set of formulas ΣM ∈ ΣM that encode the set of possible worlds
M structured into sets of possible worlds via justification graphs. We use M and ΣM synonymously. Also
we often do not distinguish between ΣM and M -- neglecting that ΣM represents a set of worlds that are
like M wrt to the relevant constraints.
Fig. 2 provides an overview of the relation between the initial information base ΣI of agent A, its
set ΣM of possible worlds ΣM, winning strategies, resolution, and stepwise update of the information ΣR
during our conflict resolution process. The initial information base defines the set of possible worlds
M, which is organised in sets of maximal consistent worlds ΣM. Based on ΣM, A's set of strategies ∆A
is checked whether it comprises a winning strategies in presence of an agent B that tries to achieve its
own goals. If no such winning strategy exists, A believes to be in conflict with B. At each level (Ci) the
resolution procedure tries to determine information ΣR of level (Ci) to resolve the conflict. If the possible
worlds are enriched by this information, the considered conflict vanishes. The new information is added
to the existing information base and the over-all process is re-started again until either winning strategies
are found or ΣR is empty.
How to find a believed winning strategy Alg. 1 finds a winning strategy of agent A for a goal ΦA in
ΣM tolerating that B follows an arbitrary winning strategy in ΣM ∈ ΣM for its goals, i.e. it finds a strategy
that satisfies ΦA in all possible worlds ΣM where ΦA is maximal for ΣM and in each possible world
ΣM ∈ ΣM agent B may also follow a winning strategy for one of its maximal goals ΦB. If such a strategy
cannot be found, A believes to be in conflict with B (cf. Def. 1).
Input for the algorithm is (i) a set Σ of formulae describing the current belief of A, e.g. its current
observations and its history of beliefs -- we call it the information base in the sequel -- , (ii) a set of goals
of B that is maximal for a ΣM ∈ ΣM,
Φmax
(iv) a set of possible actions AA for A and (v) a set of believed possible actions AB for B.
of A that is maximal in M, (iii) a set of believed goals Φmax
B
A
First (L. 2 of Alg. 1) is to construct sets of maximal consistent sets of possible worlds that together
represent M. In L. 3 the set ∆A is determined, which is the set of strategies accomplishing a maximal goal
B) that satisfy ΦA ∈ Φmax
combination for A assuming B agrees to help, i.e., all winning strategies (δ A,δ(cid:48)
in all possible worlds ΣM ∈ ΣM, where ΦA is maximal for M.
B follows its strategy to achieve one of its maximal goals ΦB ∈ Φmax
In lines 5 ff. we examine whether one of A's strategies (where B is willing to help) works even when
B) ∈ ∆A (L. 6). The
To this end TESTIFNOTWINNING is called for all of A's winning strategies (δ A,δ(cid:48)
function TESTIFNOTWINNING performs this test iteratively for one maximal consistent set of worlds ΣM
(Alg. 2 L. 2). Let ∆B be the set of joint strategies achieving a goal ΦB ∈ Φmax
that is maximal in ΣM.
A,δ B) ∈ ∆B (Alg. 2 L. 3). A strategy of A
B) to every (δ(cid:48)
We check the compatibility of A's strategy (δ A,δ(cid:48)
(δ A,δ(cid:48)
B) is compatible to all of B's strategies (δ(cid:48)
A,δ B) if all joint strategies (δ A,δ B) achieve the maximal
goals for A and B (Alg. 2 L. 6).3 If the joint strategy (δ A,δ B) is not a winning strategy for the joint goal
3Note that according to Sect. 2.1, we have ΦB = true if B cannot achieve any goal. This reflects that A cannot make any
in ΣM.
B
B
A
assumption about B's behaviour in such a situation.
initialjustifiedinformationbaseΣIΣ∪ΣMΣMΣMΣMΣM∆⊆∆Awinning?resolutionatlevelCi∆A∆constraints[no]justifications[yes]requiredinformationΣRCi←Ci+1 60
Dynamic Conflict Resolution
B , AA, AB)
B),ΣM,ΦA,Φmax
for all ΣM ∈ ΣM do
Algorithm 2 Test if a strategy is winning in all possible worlds.
1: function TESTIFNOTWINNING((δ A,δ(cid:48)
2:
∆B ← STRATB(AA, AB,ΣM,Φmax
3:
B )
for all (δ(cid:48)
4:
A,δ B),ΣM) = ΦB and ΦB is maximal in ΣM do
B with r((δ(cid:48)
5:
6:
7:
8:
9:
A,δ B) ∈ ∆B do
if r((δ A,δ B),ΣM) (cid:54)= ΦA ∪ ΦB then
return GETJUSTIFICATIONS((δ A,δ B) (cid:54)= ΦA ∪ ΦB)
(cid:46) construct(cid:8)(δ(cid:48)
A,δ B) r(cid:0)(δ(cid:48)
for all ΦB ∈ Φmax
(cid:46) (δ A,δ(cid:48)
A,δ B),ΣM
return /0
else
B) is not winning for all M and all (δ(cid:48)
A,δ B)
(cid:1) = ΦB with ΦB ∈ Φmax
(cid:9)
B
Algorithm 3 Try to fix a conflict by resolving contradictions.
1: function FIXCONFLICT(C, (Ci),Σ,Φmax
2:
3:
4:
C ← C\{E}
A ,Φmax
Σ,Φmax
for E ∈ C do
A ,Φmax
B )
A ,Φmax
B ← RESOLVE(Σ,E,Φmax
A ,Φmax
return Σ,Φmax
B , (Ci))
5:
B
(cid:46) try resolution according to level (Ci)
ΦA∪ΦB (Alg. 2 L. 6), the function GETJUSTIFICATIONS extracts the set of justifications for this conflict
situation (Alg. 2 L. 7). The set of justifications is added to the set of conflict causes C (Alg. 1 L. 8.).
Since strategy (δ A,δ(cid:48)
B) is not compatible to all of B's strategies, it is hence not further considered as a
possible conflict-free strategy for A (Alg. 1 L. 9).
A strategy that remains in ∆A at Alg. 1 L. 10 is a winning strategy for one of A's goals in all possible
worlds ΣM regardless of what maximal goals B tries to achieve in ΣM. However, if ∆A is empty at
Alg. 1 L. 10 , A is in a (believed) conflict with B (Def. 1). In this case, conflict resolution is attempted
(cf. lines 10 ff. in Alg. 1). Function FIXCONFLICT from Alg. 3 is called with the set of conflict causes,
the current conflict resolution level, and the current information base and goals. For each conflict cause,
an attempt of resolution is made by function RESOLVE. The conflict is analysed to identify whether
adding/updating information of the current resolution level helps to resolve the conflict. If there are
several ways to resolve a conflict, justifications can be used to decide which resolution should be chosen.
Note that conflict resolution hence means updating of the information base Σ or goal sets Φmax
and Φmax
B .
Line 13 of Alg. 1 checks if some new information was obtained from the resolution procedure. If not,
resolution will be restarted at the next resolution level. If new information was obtained, FINDSTRATEGY
is called with the updated information.
If the result is a non-empty set of strategies, the algorithm
terminates by returning them as (believed) winning strategies for A. However, if the result is the empty
set, resolution is restarted at the next resolution level. If ∆A is empty at level (C4), the conflict cannot be
resolved and the algorithm terminates.
A
Termination Alg. 1 eventually terminates under the following assumptions. The first assumption is
that the set of variables V and hence the information base Σ is finite. In this case, the construction of
maximal consistent possible worlds ΣM terminates since there is a finite number of possible consistent
combinations of formulae and the time horizon for the unrolling of a possible world ΣM is bounded.
Together with finite sets Φmax
B , AA, and AB, the construction of strategies, i.e. functions
STRATA and STRATA, terminates since there are only finite numbers of combinations of input histo-
A , Φmax
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
61
ries and output action and there is only a finite number of goals to satisfy. All loops in algorithms 1, 2,
and 3 hence iterate over finite sets.
The extraction of justifications terminates since runs are finite and consequently the number of ac-
tions involved in the run, too. Furthermore, for each state in a run, there are only a finite number of
propositions that apply. Together with a finite number of propositions representing the goal, GETJUSTI-
FICATIONS can simply return the (not necessarily minimal) set of justifications from all these finite many
formulae as a naive approach.
B
A , and Φmax
Alg. 1 terminates if a non-empty set ∆A is derived by testing and/or resolution, or if a fixed point re-
garding Σ, Φmax
is reached. Since all other loops and functions terminate, the only open aspect
is the fixed point whose achievement depends on RESOLVE. We assume that Alg. 1 is executed at a fixed
time instance s.t. A's perception of the environment does not change during execution. Thus, Σ contains
only a finite number of pieces of information to share. If we assume that that sharing information leads
only to dismissing possible worlds rather then considering more worlds possible, then this a monotonic
process never removing any information.4 Furthermore, we assume that the partial order of goals leads,
if necessary, to a monotonic process of goal negotiation which itself can repeated finite many times until
no further goals can be sacrificed or adopted from B. Thus, if ∆A remains to be the empty set in line 15,
the fixed point will eventually be reached.
Furthermore, we do not consider any kind of race conditions occurring from concurrency, e.g. dead-
lock situations where A can't serve B's request because it does not know what its strategy will be since A
wait's for B's respond, and vice versa.
So in summary, the algorithm terminates under certain artificial assumptions but cannot determine
a resolution in case without an outside arbiter. In practice such a conflict resolution process has to be
equipped with time bounds and monitors. We consider these aspects as future work.
4.3 Case study
We implemented the algorithm sketched above employing Yices [14] to determine contradictions and
analysed variations of a toy example to evaluate and illustrate our approach. More details on the
case study can be found in [12]. The implementation is available under https://uol.de/en/hs/
downloads/.
We modelled a system of two agents on a two lane highway. Each agent is represented by its position
and its lane. Each agent has a set of actions: it can change lane and drive forward with different speeds.
We captured this via a discrete transition relation where agents hop from position to position. The
progress of time is encoded via unrolling, that is we have for each point in time a corresponding copy
of a variable to hold the value of the respective attribute at that time. Accordingly the transition relation
then refers to these copies.
Since we analyse believed conflicts of an agent, we consider several worlds. In other words, we
consider several variations of a Yices model. Each variation represents a justification graph summarising
the maximal consistent set of evidences and thereby representing a set of worlds which is justified by
this set of evidences.
We modify the Yices file by adding additional constraints according to the algorithm Sect. 4.2. For
the steps (C1) to (C4) we add constraint predicates, e.g., that encode that information about certain ob-
servations have been communicated by say B to A, constraints that specify that B tells A it will decelerate
at step 4 and constraints that encode goal combinations.
4Otherwise the set of already examined worlds can be used to define a fixed point.
62
Dynamic Conflict Resolution
We employed Yices to determine whether there is conflict. The key observation is: If Yices deter-
mines that it holds that ¬ϕ is satisfiable in our system model, then there is the possibility that the goal is
not achieved -- otherwise each evolution satisfies ϕ and there is a winning strategy for the model.
5 Related work
Studying Traffic Conflicts According to Tiwari in his 1998 paper [26] studying traffic conflicts in
India, one of the earliest studies concerned with traffic conflicts is the 1963 paper [21] of Perkins and
Harris. It aims to predict crashes in road traffic and to obtain a better insight to causal factors. The term
traffic conflict is commonly used according to [26] as "an observable situation in which two or more
road users approach each other in space and time to such an extent that a collision is imminent if their
movements remain unchanged" [1]. In this paper we are interested in a more general and formal notion
of conflict. We are not only interested in collisions-avoidance but more generally in situations where
traffic participants have to cooperate with each other in order to achieve their goals -- which might be
collision-freedom. Moreover, we aim to provide a formal framework that allows to explain real world
observations as provided by, e.g., the studies of [26, 21].
Tiwari also states in [26] that it is necessary to develop a better understanding of conflicts and con-
jectures that illusion of control [23] and optimism bias theories like in [13] might explain fatal crashes.
In this paper we develop a formal framework that allows us to analyse conflicts based on beliefs of the
involved agents, -- although supported by our framework -- we here do not compare the real world evolu-
tion with the evolution that an agent considers possible. Instead we analyse believed conflicts, that are
conflicts which an agent expects to occur based on its beliefs. Such conflicts will have to be identified
and analysed by prediction components of the autonomous vehicles architecture, especially in settings
where misperception and, hence, wrong beliefs are possible.
In [15] Sameh et al. present their approach to modelling conflict resolution as done by humans in or-
der to generate realistic traffic simulations. The trade-off between anticipation and reactivity for conflict
resolution is analysed in [25] in order to determine trajectories for vehicles at an intersection. Both works
[15, 25] focus on conflicts leading to accidents. Regarding the suggested resolution approaches, our res-
olution process suggests cooperation steps with increasing level cooperation. This resolution process is
tailored for autonomous vehicles that remain autonomous during the negation process.
Strategies and Games For strategy synthesis Finkbeiner and Damm [10] determined the right perime-
ter of a world model. The approach aims to determine the right level of granularity of a world model
allowing to find a remorse-free dominant strategy. In order to find a winning (or remorse-free dominant)
strategy, the information of some aspects of the world is necessary to make a decision. We accommodated
this as an early step in our resolution protocol. Moreover in contrast to [10], we determine information
that agent A then want requests from agent B in order to resolve a conflict with B -- there may still be
no winning (or remorse-free dominant) strategy for all goals of A. In [11] Finkbeiner et. al.presented
an approach to synthesise a cooperative strategy among several processes, where the lower prioritised
process sacrifices its goals when a process of higher priority achieves its goals. In contrast to [11] we do
not enforce a priority of agents but leave it open how a conflict is resolved in case not all their goals are
achievable. Our resolution process aims to identify the different kinds of conflict as introduced in Sect. 2
that arise when local information and beliefs are taken into account and which not necessarily imply that
actually goals have to be sacrificed.
We characterize our conflict notion in a game theoretic setting by considering the environment of
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
63
agents A and B as adversarial and compare two scenarios where (i) the agent B is cooperative (angelic)
with the scenario where (ii) B is not cooperative and also not antagonistic but reasonable in following
a strategy to achieve its own goals. As Brenguier et al. in [7] remark, a fully adversarial environment
(including B) is usually a bold abstraction. By assuming in (ii) that B maximises its own goals -- we
assume that B follows a winning strategy for its maximal accomplishable goals. So we are in a similar
mind set than at assume-guarantee [9] and assume-admissible [7] synthesis. Basically we consider the
type of strategy (winning/admissible/dominant) as exchangeable, the key aspect of our definition is that
goals are not achievable but can be achieved with the help of the other.
Logics
Justification logic was introduced in [5, 6] as an epistemic logic incorporating knowledge and
belief modalities into justification terms and extends classical modal logic by Plato's characterisation of
knowledge as justified true belief. However, even this extension might be epistemologically insufficient
as Gettier already pointed out in 1963 [20]. In [4] a combination of justification logics and epistemic
logic is considered with respect to common knowledge. The knowledge modality Ki of any agent i
inherits all information that are justified by some justification term t, i.e. t:φ → Kiφ. In such a setting any
justified information is part of common knowledge. Moreover, justified common knowledge is obtained
by collapsing all justification terms into one modality J and can be regarded as a special constructive
sort of common knowledge. While our approach neglects the notion of common information, we use a
similar inheritance principle where a belief entity inherits information of its components, cf. Def. 4.(iii).
A comparison of the strength of this approach with different notions of common knowledge can be
found in [2]. While justification logic and related approaches [16, 3], aim to restrict the principle of
logical awareness and the related notion of logical omniscience, we argue in Sec. 3 that the principle of
logical awareness as provided by modal logic is indispensable in our approach. A temporal (LTL-based)
extension of justification logic has been sketched in [8]. This preliminary work differs from our approach
wrt. the axiom systems used for the temporal logic part and the justification / modal logic part, cf. the
logic of justification graphs axiomatised in Section 3. Our logic and its axiomatisation incorporates
a partial order on the set of beliefs that underlies their prioritization during conflict resolution, which
contrasts with the probabilistic extension of justification logic outlined in [22].
6 Conclusion
Considering local and incomplete information, we presented a new notion of conflict that captures sit-
uations where an agent believes it has to cooperate with another agent. We proposed steps for conflict
resolution with increasing level of cooperation. Key for conflict resolution is the analysis of a conflict,
tracing and identifying contradictory evidences. To this end we presented a formal logical framework
unifying justifications with modal logic. Alas, to the authors' best knowledge there are no efficient satis-
fiability solvers addressing distributed information so far. However, we exemplified the applicability of
our framework in a restricted but non-trivial setting. On the one hand, we plan to extend this framework
by efficient implementations of adapted satisfiability solvers, on the other hand by integrating richer log-
ics addressing decidable fragments of first order logic, like linear arithmetic, and probabilistic reasoning.
References
[1] F. H. Amundsen & C. Hyden (1977): Proc. of the first Workshop on Traffic Conflicts, Oslo, Norway. 1st
Workshop on Traffic Conflicts, LTH Lund.
64
Dynamic Conflict Resolution
[2] E. Antonakos (2007): Justified and common knowledge: Limited conservativity. In: International Symposium
on Logical Foundations of Computer Science, Springer, pp. 1 -- 11, doi:10.1007/978-3-540-72734-7_1.
[3] S. Artemov & R. Kuznets (2009): Logical omniscience as a computational complexity problem.
In:
Proc. of the 12th Conference on Theoretical Aspects of Rationality and Knowledge, ACM, pp. 14 -- 23,
doi:10.1145/1562814.1562821.
[4] S. Artemov & E. Nogina (2005): Introducing justification into epistemic logic. Journal of Logic and Com-
putation 15(6), pp. 1059 -- 1073, doi:10.1093/logcom/exi053.
Justified common knowledge.
[5] S. N. Artemov (2006):
doi:10.1016/j.tcs.2006.03.009.
[6] S. N. Artemov (2008):
The Logic of Justification.
doi:10.1017/S1755020308090060.
Theor. Comput. Sci. 357(1-3), pp. 4 -- 22,
Rew. Symb. Logic 1(4), pp. 477 -- 513,
[7] R Brenguier, J.-F. Raskin & O. Sankur (2017): Assume-admissible synthesis. Acta Informatica 54(1), pp.
41 -- 83, doi:10.1007/s00236-016-0273-2.
[8] S. Bucheli, M. Ghari & T. Studer (2017): Temporal Justification Logic.
In: Proc. of the 9th Work-
shop on Methods for Modalities, January 2017, EPTCS 243, Open Publishing Association, pp. 59 -- 74,
doi:10.4204/EPTCS.243.5.
[9] K. Chatterjee & T. A. Henzinger (2007): Assume-Guarantee Synthesis. In: Tools and Algorithms for the
Construction and Analysis of Systems, Springer, pp. 261 -- 275, doi:10.1007/978-3-540-71209-1_21.
[10] W. Damm & B. Finkbeiner (2011): Does It Pay to Extend the Perimeter of a World Model? In: FM 2011:
Formal Methods, Springer, pp. 12 -- 26, doi:10.1007/978-3-642-21437-0_4.
[11] W. Damm, B. Finkbeiner & A. Rakow (2016): What You Really Need To Know About Your Neighbor. In:
Proc. Fifth Workshop on Synthesis, SYNT@CAV 2016, EPTCS 229, pp. 21 -- 34, doi:10.4204/EPTCS.229.4.
[12] W. Damm, M Fränzle, W. Hagemann, P. Kröger & A. Rakow (2019): Justification Based Reasoning for
Dynamic Conflict Resolution. arXiv e-prints:arxiv:1905.11764. Available at http://arxiv.org/abs/
1905.11764.
[13] D. M. DeJoy (1989): The optimism bias and traffic accident risk perception. Accident Analysis & Prevention
21(4), pp. 333 -- 340, doi:10.1016/0001-4575(89)90024-9.
[14] B. Dutertre (2014): Yices 2.2. In: Computer Aided Verification, Springer, pp. 737 -- 744, doi:10.1007/978-3-
319-08867-9_49.
[15] S. El hadouaj, A. Drogoul & S. Espié (2001): How to Combine Reactivity and Anticipation: The Case of
Conflicts Resolution in a Simulated Road Traffic. In: Multi-Agent-Based Simulation, Springer, pp. 82 -- 96,
doi:10.1007/3-540-44561-7_6.
[16] R. Fagin & J. Y. Halpern (1987): Belief, awareness, and limited reasoning. Artificial intelligence 34(1), pp.
39 -- 76, doi:10.1016/0004-3702(87)90003-8.
[17] R. Fagin, J. Y. Halpern, Y. Moses & Moshe Y. Vardi (2003): Reasoning About Knowledge. MIT Press.
[18] J. Galtung (1969): Violence, Peace, and Peace Research. Journal of Peace Research 6(3), pp. 167 -- 191,
doi:10.1177/002234336900600301.
[19] J. Gerbrandy (1998): Distributed knowledge.
Dialogue, 98, pp. 111 -- 124.
In: Twendial 1998: Formal Semantics and Pragmatics of
[20] E. L. Gettier
Is
doi:10.1093/analys/23.6.121.
(1963):
justified true belief knowledge?
Analysis 23(6), pp. 121 -- 123,
[21] J. I. Harris & S. R. Perkins (1967): Traffic conflict characteristics: accident potential at intersections. High-
way Research Board 225, pp. 35 -- 43.
[22] I. Kokkinis, Z. Ognjanovic & T. Studer (2016): Probabilistic Justification Logic. In: Logical Foundations
of Computer Science - International Symposium, LFCS 2016. Proc., LNCS 9537, Springer, pp. 174 -- 186,
doi:10.1007/978-3-319-27683-0_13.
W. Damm, M. Fränzle, W. Hagemann, P. Kröger and A. Rakow
65
[23] J. E. Langer (1975): The Illusion of Control. Journal of Personality and Social Psychology 32, pp. 311 -- 328,
doi:10.1037/0022-3514.32.2.311.
[24] O. Lichtenstein, A. Pnueli & L. Zuck (1985): The glory of the past. In: Workshop on Logic of Programs,
Springer, pp. 196 -- 218, doi:10.1007/3-540-15648-8_16.
[25] N. Murgovski, G. R. de Campos & J. Sjöberg (2015): Convex modeling of conflict resolution at traf-
In: 2015 54th IEEE Conference on Decision and Control (CDC), pp. 4708 -- 4713,
fic intersections.
doi:10.1109/CDC.2015.7402953.
[26] G. Tiwari, D. Mohan & J. Fazio (1998): Conflict analysis for prediction of fatal crash locations in mixed
traffic streams. Accident Analysis & Prevention 30(2), pp. 207 -- 215, doi:10.1016/S0001-4575(97)00082-1.
|
cs/0407021 | 2 | 0407 | 2004-07-14T06:28:48 | Multi-agent coordination using nearest neighbor rules: revisiting the Vicsek model | [
"cs.MA",
"cs.AI"
] | Recently, Jadbabaie, Lin, and Morse (IEEE TAC, 48(6)2003:988-1001) offered a mathematical analysis of the discrete time model of groups of mobile autonomous agents raised by Vicsek et al. in 1995. In their paper, Jadbabaie et al. showed that all agents shall move in the same heading, provided that these agents are periodically linked together. This paper sharpens this result by showing that coordination will be reached under a very weak condition that requires all agents are finally linked together. This condition is also strictly weaker than the one Jadbabaie et al. desired. | cs.MA | cs |
Multi-agent coordination using nearest-neighbor
rules: revisiting the Vicsek model ∗
Sanjiang Li †
State Key Laboratory of Intelligent Technology and Systems
Department of Computer Science and Technology
Tsinghua University, Beijing 100084, China
Huaiqing Wang
Department of Information Systems
City University of Hong Kong, Hong Kong, China
October 28, 2018
Abstract
Recently, Jadbabaie, Lin, and Morse (2003) offered a mathematical
analysis of the discrete time model of groups of mobile autonomous agents
raised by Vicsek et al.
in 1995. In their paper, Jadbabaie et al. showed
that all agents shall move in the same heading, provided that these agents
are periodically linked together. This paper sharpens this result by showing
that coordination will be reached under a very weak condition that requires
all agents are finally linked together. This condition is also strictly weaker
than the one Jadbabaie et al. desired.
Index Terms -- Decentralized control, multi-agent coordination, switched
systems.
1
Introduction
Coordination of groups of mobile autonomous agents (particles [9], or boids [5])
has attracted researchers in a surprisingly wide variety of disciplines ranging
from physics [9, 7, 8], to the biological sciences [10, 1], to computer science and
engineering [5, 3, 6, 4].
This paper is mainly concerned with one particular discrete time model of
groups of mobile autonomous agents, viz., the one proposed by Vicsek et al. [9]
∗This work was partly supported by the National Foundation of Natural Science of China
(60305005, 60321002, 60496321) and by a Hong Kong CERG (Grant No. CityU 1234/03E).
†Email: [email protected] (S. Li), [email protected] (H. Wang)
1
in 1995. In this model, a group of autonomous agents is moving in the plane with
all agents moving at the same speed but with different headings. Each agent's
motion is updated using a local rule based on the average of its own heading and
the heading of its "neighbors." This is known as the nearest-neighbor rule in [3].
Agent i's neighbors at time t are those agents that are either in or on a circle of
pre-specified radius r centered at agent i's current position. Known as the Vicsek
model, this can be viewed as a special version of a model proposed by Reynolds [5]
for simulating animal aggregation for the computer animation industry. Although
the Vicsek model is very simple, simulation results in [9] show that, using the
local update rule, all agents shall eventually move in the same direction despite
the absence of centralized coordination, and that neighborhood of each agent will
change.
Recently, Jadbabaie, Lin, and Morse [3] offered a mathematical analysis of
this model and provided a theoretical explanation for the observed behavior.
They adopt a more conservative approach, which ignores how the neighbor-graphs
depend on the agent positions in the plane. Note that under this assumption,
the Vicsek model is a graphic example of a switched linear system. Their goal in
that paper was to determine for a certain large class of switching signals and for
any initial set of agent headings that the headings of all agents will converge into
the same steady heading.
Jadbabaie et al. [3] established sufficient conditions given in terms of neighbor-
graphs for coordination of agents. One main result of [3] shows that all agents
shall eventually move in the same heading if these graph are periodically jointly
connected, i.e., the union of any T sequential graphs is connected for some fixed
T . This is a nice result, but as Jadbabaie et al. put it [3, p990, below Theorem
2], what one would prefer instead is to show that coordination would be reached
eventually for every switching signal for which there is an infinite sequence of
bounded, non-overlapping (but not necessarily contiguous) intervals across which
the agents are linked together.
This paper will show above condition desired by Jadbabaie et al.
is indeed
a sufficient condition for asymptotic convergence. This, however, follows from a
more general observation: our main result in this paper shows that convergence
will be attained if these neighbor-graphs are finally jointly connected, i.e., the
union of all graphs started from any time is connected.
The structure of this paper is as follows. In Section II we give a formal descrip-
tion of the Vicsek model in terms of switching signals. Section III provides the
major results, for both leaderless coordination and leader-following coordination.
Conclusions and future work are given in the last section.
2 The Vicsek model and the nearest-neighbor
rule
In this section, we review some basic definitions concerning the Vicsek model.
2
The system studied by Vicsek et al.
[9] consists of n autonomous agents,
e.g., particles, robots, etc., labeled 1 through n. All agents move in the plane
with the same speed but with different headings. The system operates at discrete
time t = 0, 1, 2, · · · Let r > 0 and v > 0 be given numbers associated with the
system. The dynamics of agent i is described by the sequence {xi(t), yi(t), θi(t)},
where xi(t), yi(t) ∈ R are the coordinates of the agent in the plane, and θi(t) is
its heading taking value from [0, 2π). At any time t = 0, 1, 2, · · · , each agent's
heading is updated using a simple rule based on the average of its own heading
plus the headings of its neighbors. For any two agents i, j, we say j is a neighbor
of i at time t, written j ∈ Ni(t), if d(i, j) ≤ r,1 where
d(i, j) = q(xi(t) − xj(t))2 + (yi(t) − yj(t))2.
Then, agent i's next heading is defined as
θi(t + 1) =
θi(t) + Σj∈Ni(t)θj(t)
1 + ni(t)
(1)
(2)
where ni(t) is the number of agent i's neighbors at time t. Agent i's next coordi-
nates are defined as
xi(t + 1) = x(t) + vi(t) cos(θi(t))
yi(t + 1) = y(t) + vi(t) sin(θi(t)).
(3)
(4)
For any time t ≥ 0, we define the neighbor-graph of the system described
above as the simple undirected graph G(t) over V = {1, 2, · · · , n} where the
vertex i corresponds to agent i and two vertexes, i, j, are connected by an edge
in the graph if they are neighbors at time t, i.e., if j ∈ Ni(t). Since the neighbor
relation can change over time, so can the graph that describes them. In the sequel,
we write P for the collection of simple undirected graphs over V . A switching
signal is a function σ : N → P that assigns to each time t a neighbor-graph that
specifies the neighbor relation between agents. Clearly, for a Vicsek model, the
function that assigns to each time t the neighbor-graph G(t) is a switching signal.
Note that for the Vicsek model, the neighbor-graph is determined by the
initial positions and headings of all agents as well as the pre-specified r > 0 and
v > 0. A complete description of the model would have to explain explicitly how
σ changes over time. As it is difficult to take this into account in a convergence
analysis, Jadbabaie et al. adopt a more conservative approach, "which ignores
how σ depends on the agent positions in the plane and assumes instead that σ
might be any switching signal in some suitably defined set of interests."[3]
We in this paper follow this basic assumption and formalize the Vicsek model
as follows:
1In certain situation, choosing open neighborhood would give rise to more desirable results,
see [2].
3
Definition 2.1 (Vicsek model). Given n agents, labeled 1, 2, · · · , n, moving
in the plane at discrete time t ∈ N, let P be all simple undirected graphs over
V = {1, 2, · · · , n}. A Vicsek model is a pair (V, σ), where σ : N → P is a
switching signal.
For each agent i, define i's σ-neighborhood at time t, written Ni(t), to be the
set of agents that is connected to i by an edge in graph σ(t). That is, agent j is
a neighbor of agent i if and only if (i, j) is an edge in graph σ(t).
Given an initial heading θ(0) = hθi(0)in
i=1, agent i's heading θi(t) evolves in
discrete time according to Eq. 2. Namely, agent i's heading at time t + 1 is the
average of the headings of agent i and its neighbors at time t.
Remark 2.1. This definition of a multi-agent coordination model is very general
and more flexible. Several dimensions of extension/completion could be incor-
porated in this model: 1) We can choose either closed/open disk, a triangle-like
zone or any subset of V as the neighborhood; 2) The velocity could also change
in discrete time; 3) We could consider other state variables of agents besides
their headings; and 4) The neighbor-graph could also be directed. This flexibility
would be helpful in practical applications.
The goal of this paper is to show for a large class of Vicsek models (or switching
signals) and for any initial set of agent headings that the headings of all n agents
will converge into the same heading. Compared with the results obtained in [3],
ours are more general.
3 A sufficient condition for multi-agent coordi-
nation
3.1 Notations and preliminaries
Suppose (V, σ) is a Vicsek model. Following Savkin [6], we define a graph σ(∞)
over V = {1, 2, · · · , n} as follows: for any two nodes i, j, (i, j) is an edge in σ(∞)
if and only if for any K > 0, there exists some k ≥ K such that (i, j) is an edge in
graph σ(k). For convenience, given a collection of graphs {Gx : x ∈ X}, we write
Ux∈X Gx for the union of these graphs, i.e., any pair (i, j) is an edge in Ux∈X Gx
if and only if it is an edge in some Gx. Then it is easy to show that there exists
some K > 0 such that σ(∞) = Ut≥k σ(t) holds for all k ≥ K.
In this paper
we will show that all agents shall eventually move in the same heading provided
that σ(∞) is connected. This condition is more general than the one given in
[3], where the authors require that the σ(t)'s are periodically jointly connected.
In what follows, a switching signal σ is called finally jointly connected if σ(∞)
is connected. Clearly this is equivalent to saying that Ut≥k σ(t) is connected for
any k ∈ N.
For a sequence {f (k)} and a number u in R, we say u is an accumulation
point of {f (k)} if there is a subsequence of {f (k)} that converges to u. We write
4
Accu({f (k)}) for the set of accumulation points of {f (k)}.
Given a Vicsek model (V, σ) and an initial headings θ(0) = hθi(0)in
i=1, we now
fix some notations concerning the model.
For i = 1, 2, · · · , n, define
Θi = {θi(t) : t ∈ N}
mi = min AccuΘi
Mi = max AccuΘi
m =
M =
n
min
i=1
n
max
i=1
mi
Mi.
(5)
(6)
(7)
(8)
(9)
Note that Θi is a bounded set and, therefore, has minimum and maximum ele-
ments.
For any t ∈ N, define
θ(t) =
θ(t) =
n
min
i=1
n
max
i=1
θi(t)
θi(t).
(10)
(11)
The following lemma shows θ(t) ≤ m ≤ M ≤ θ(t).
Lemma 3.1. For any t ∈ N, we have θ(t) ≤ θ(t + 1) ≤ m ≤ M ≤ θ(t + 1) ≤ θ(t).
Consequently, we have limt→∞ θ(t) = m and limt→∞ θ(t) = M .
Proof. For any non-negative t, note that by Vicsek's nearest-neighbor rule (Eq. 2),
we have θ(t) ≤ θi(t+ 1) ≤ θ(t). In particular, we have θ(t) ≤ θ(t+ 1) ≤ θ(t+ 1) ≤
θ(t). Now since {θ(t)} ({θ(t)}, resp.) is a bounded ascending (descending, resp.)
sequence, it has a limit. We now show its limit is m (M, resp.). Take {θ(t)}
as an example. Since it is convergent, any subsequence of {θ(t)} also converges
to its limit. Suppose {f (k)} is a sequence such that limk→∞ θi(f (k)) = m for
some agent i. Note that θ(f (k)) ≤ θi(f (k)) for any k, we have limt→∞ θ(t) =
limk→∞ θ(f (k)) ≤ limk→∞ θi(f (k)) = m. On the other hand, since there exists
some agent i such that {t : θi(t) = θ(t)} is infinite, we have a sequence {g(k)}
such that θi(g(k)) = θ(g(k)). This shows that limt→∞ θ(t) = limk→∞ θ(g(k)) =
limk→∞ θi(g(k)) ≥ m since m is the minimum accumulation point. As a result,
we have limt→∞ θ(t) = m. Similarly, we can show limt→∞ θ(t) = M. So we have
θ(t) ≤ m ≤ M ≤ θ(t) for any t.
Note that as shown in the proof of the above lemma, we have a sequence, say
{f (k)}, such that θi(f (k)) = θ(f (k)) and limk→∞ θi(f (k)) = m for some agent i.
Similarly, we have a sequence, say g(k), such that limk→∞ θj(g(k)) = M for some
j.
5
3.2 Leaderless coordination
Theorem 3.1. Given a Vicsek model (V, σ), suppose σ(∞) is connected. Then
for any θ(0) = hθi(0)in
i=1, we have
lim
t→∞
θi(t) = θss (i = 1, 2, · · · , n)
(12)
where θss is a number depending only on θ(0) and σ.
To prove this theorem, we need several lemmas.
Recall V = {1, 2, · · · , n}. For a graph G over V and any two disjoint subsets
A, B of V , we say A and B are connected if there exist a ∈ A, b ∈ B such that
(a, b) is an edge in G. If A happens to be a singleton {a}, we also say node a is
connected to B. In this case, we say alternatively a has a neighbor in B.
The following lemma suggests that, if the agents are divided into two parts
such that the maximum heading of the first part is sufficiently smaller than the
minimum of the second part, then, after updating the headings using Eq. 2, the
agents will also form two parts such that one part is still sufficiently smaller than
the rest.
For a < b in R and any natural number t, we write Vt(a, b) = {i ∈ V : a <
θi(t) < b}.
Lemma 3.2. Given α < β < γ and set δ = β − α, ǫ = δ/nn, suppose Vt(α −
ǫ, α + ǫ) and Vt(β − ǫ, γ + ǫ) are two nonempty disjoint subsets of V such that
their union is V . We have
(1) If Vt(α − ǫ, α + ǫ) and Vt(β − ǫ, γ + ǫ) are disconnected at time t, then Vt+1(α −
ǫ, α + ǫ) = Vt(α − ǫ, α + ǫ) and Vt+1(β − ǫ, γ + ǫ) = Vt(β − ǫ, γ + ǫ).
(2) If Vt(α − ǫ, α + ǫ) and Vt(β − ǫ, γ + ǫ) are connected at time t, then Vt+1(α −
ǫ, α + ǫ) = Vt(α − ǫ, α + ǫ) − {i ∈ V : i has a neighbor in Vt(β − ǫ, γ + ǫ) at time t}
and V − Vt+1(α − ǫ, α + ǫ) = Vt+1(α + δ/n − ǫ, γ + ǫ).
Proof. If Vt(α − ǫ, α + ǫ) and Vt(β − ǫ, γ + ǫ) are disconnected at time t, then for
any i ∈ Vt(α − ǫ, α + ǫ), its neighbors are all in Vt(α − ǫ, α + ǫ). By Eq. 2, we
have α − ǫ < θi(t + 1) < α + ǫ. Similarly, for any j ∈ Vt(β − ǫ, γ + ǫ), we have
β − ǫ < θj(t + 1) < γ + ǫ.
On the other hand, suppose Vt(α − ǫ, α + ǫ) and Vt(β − ǫ, γ + ǫ) are connected
at time t. For i ∈ Vt(α − ǫ, α + ǫ), if all its neighbors are in Vt(α − ǫ, α + ǫ), then
6
i ∈ Vt+1(α − ǫ, α + ǫ); if i has a neighbor, say j0, in Vt(β − ǫ, γ + ǫ), then we have
θi(t + 1) =
θi(t) + Σj∈Ni(t)θj(t)
1 + ni(t)
θi(t) + θj0(t) + Σj∈Ni(t),j6=j0θj(t)
1 + ni(t)
(α − ǫ) + (β − ǫ) + (ni(t) − 1) × (α − ǫ)
1 + ni(t)
β − α + (1 + ni(t)) × (α − ǫ)
=
>
=
= (α − ǫ) +
1 + ni(t)
β − α
1 + ni(t)
≥ α − ǫ + δ/n.
Note that θi(t+1) < γ+ǫ holds for any i ∈ V . This shows that, if i ∈ Vt(α−ǫ, α+ǫ)
has a neighbor in Vt(β − ǫ, γ + ǫ), then i ∈ Vt+1(α + δ/n − ǫ, γ + ǫ). Similarly,
for any j ∈ Vt(β − ǫ, γ + ǫ), we can show θj(t + 1) > α + δ/n − ǫ. In summary,
we have i ∈ Vt+1(α − ǫ, α + ǫ) if and only if i ∈ Vt(α − ǫ, α + ǫ) and it has
a neighbor in Vt(β − ǫ, γ + ǫ) at time t. As for any other agent j, we have
j ∈ Vt+1(α + δ/n − ǫ, γ + ǫ).
Lemma 3.3. Suppose σ(∞) is connected and {f (k)} is a sequence. Then we
have a subsequence {g(k)} of {f (k)} such that all {θi(g(k))} are convergent for
i ∈ V .
Proof. This follows from the compactness of [0, 2π] and that θi(t) ∈ [0, 2π) for
any i, t.
Lemma 3.4. Suppose {g(k)} is a sequence such that {θi(g(k))} converges to li for
i = 1, 2, · · · , n. Then m = minn
i=1 li, where m = limt→∞ θ(t),
M = limt→∞ θ(t) and θ(t) = minn
i=1 li and M = maxn
i=1 θi(t), θ(t) = maxn
i=1 θi(t).
Proof. Take m = minn
i=1 li as an example. Note that there exists some i such
that {k : θi(g(k)) = θ(g(k))} is infinite. We have a subsequence {h(k)} of {g(k)}
such that θi(h(k)) = θ(h(k)) and li = limk→∞ θi(h(k)) = limk→∞ θ(h(k)) = m.
That M = maxn
i=1 li is similar.
Proof of Theorem 3.1. To show that these autonomous agents shall eventually
move into the same heading, we need only to show m = M. We prove this by
reduction to absurdity.
Suppose m < M and {g(k)} is a sequence such that {θi(g(k))} converges to li
for i = 1, 2, · · · , n. Recall by Lemma 3.4 that m = minn
i=1 li.
Set l = min{li : li > m} and take δ = l − m, ǫ = δ/nn. Then there exists K > 0
such that θi(g(k)) ∈ (li−ǫ, li+ǫ) for k ≥ K and i = 1, 2, · · · , n. Moreover, we have
Vg(k)(m − ǫ, m + ǫ) = {i ∈ V : li = m} and Vg(k)(l − ǫ, M + ǫ) = {i ∈ V : li ≥ l}.
Clearly Vg(k)(m−ǫ, m+ǫ) and Vg(k)(l−ǫ, M +ǫ) satisfy the condition of Lemma 3.2.
i=1 li and M = maxn
7
Now since Ut≥p σ(t) is connected for any p ∈ N, we have some k ≥ K such
that Vg(k)(m − ǫ, m + ǫ) and Vg(k)(l − ǫ, M + ǫ) are connected at time g(k) + w for
some 0 ≤ w < g(k + 1) − g(k). Fix one such k and suppose g(k + 1) − g(k) = W .
For each w = 0, 1, · · · , W , we define Aw = Vg(k)+w(m−ǫ, m+ǫ) and Bw = V −Aw.
Note that because AW = Vg(k+1)(m − ǫ, m + ǫ) = Vg(k)(m − ǫ, m + ǫ) = A0, it
is also true that m − ǫ < θ(g(k)) ≤ θ(g(k + 1)) < m + ǫ. Recall that because θ(t)
is an ascending chain (see Lemma 3.1), we also have θ(g(k) + w) ∈ (m − ǫ, m + ǫ)
for any w = 1, 2, · · · , W − 1.
Set C = {w ∈ [0, W ) : Aw and Bw are connected at time g(k) + w}. Clearly
C is not empty since there exists some w such that A0 is connected to B0 at time
g(k) + w. Suppose C = {w1, w2, · · · , wq} and 0 ≤ w1 < w2 < · · · < wq < W . We
claim
A0 ) Aw1+1 ) Aw2+1 ) · · · ) Awq−1+1 ) Awq+1
Bws+1 = Vg(k)+ws+1(m +
δ
ns − ǫ, M + ǫ) (s = 1, 2, · · · , q)
(13)
(14)
As the induction basis, note that A0 = Vg(k)(m − ǫ, m + ǫ) = {i : li = m} and
B0 = {i : li ≥ l} = Vg(k)(l − ǫ, M + ǫ) = Vg(k)(m + δ − ǫ, M + ǫ).
Note that w1 is the first index w such that Aw is connected to Bw; by
Lemma 3.2, we have A0 = Aw and Bw = Vg(k)+w(m + δ − ǫ, M + ǫ) for any
w ≤ w1. Moreover, since Aw1 = A0 is connected to Bw1 = B0 at time g(k) + w1,
by Lemma 3.2, we have
Aw1+1 = Aw1 − {i ∈ Aw1 : i has a neighbor in Bw1 at time g(k) + w1} ( A0 (15)
Bw1+1 = V − Aw1+1 = Vg(k)+w1+1(m + δ/n − ǫ, M + ǫ) (16)
Recall that Aw1+1 6= ∅ since θ(g(k) + w1 + 1) ∈ (m − ǫ, m + ǫ).
Suppose for s < q we have
A0 ) Aw1+1 ) Aw2+1 ) · · · ) Aws−1+1 ) Aws+1
Bwj+1 = Vg(k)+wj +1(m +
δ
nj − ǫ, M + ǫ) (j = 1, 2, · · · , s)
(17)
(18)
Note that s < n − 1 must hold since A0 contains at most n − 1 agents and
A0 ) Aw1+1 ) · · · ) Aws+1 6= ∅.
We now show Aws+1 ) Aws+1+1 and Bws+1+1 = Vg(k)+ws+1+1(m+ δ
ns+1 −ǫ, M +ǫ).
Note that ws+1 is the first index w > ws such that Aw is connected to Bw. By
Lemma 3.2, we have Aw = Aws+1 and Bw = Vg(k)+w(m + δ
ns − ǫ, M + ǫ) = Bws+1
for any w ∈ (ws, ws+1]. Moreover, since Aws+1 = Aws+1 is connected to Bws+1 =
Bws+1 at time g(k) + ws+1, by Lemma 3.2, we have
Aws+1+1 = Aws+1 −
{i ∈ Aws+1 : i has a neighbor in Bws+1 at time g(k) + ws+1} ( Aws+1 (19)
Bws+1+1 = V − Aws+1+1 = Vg(k)+ws+1+1(m +
δ
ns+1 − ǫ, M + ǫ) (20)
8
In summary, we have obtained that A0 = Aw1 ) Awq+1.
Note that if wq < W − 1, then Aw and Bw are disconnected for any w ∈
(wq, W ). By Lemma 3.2 again, we know Aw = AW for w ∈ (wq, W ]. In particular,
we have Awq+1 = AW . On the other hand, if wq = W − 1, we also have Awq+1 =
AW .
This suggests that if m < M, then A0 6= AW . This is a contradiction. So our
assumption that m < M cannot hold. This ends the proof of this theorem.
Remark 3.1. Note that if σ : N → P is a switching signal for which there exists
an infinite sequence of bounded, non-overlapping (but not necessarily contiguous)
intervals across which the n agents are linked together, then σ(∞) is connected.
By the above theorem, we know all agents would eventually move in the same
heading for this σ. Consequently, this theorem shows the desired condition given
in [3, p990, below Theorem 2] is a sufficient condition for asymptotic convergence.
The hypothesis of Theorem 3.1, however, is still not necessary. For example,
if some σ(t) is the complete graph over V , then a coordination could be achieved
at time t + 1. But if it is not connected, σ(∞) will have 1 < p ≤ n connected
components, say G1, G2, · · · , Gp. Similar to the argument given above for Theo-
rem 3.1, we can show for any h = 1, 2, · · · , p, there exists a heading θh such that
limt→∞ θi(t) = θh for any i ∈ Gh.
3.3 Leader-following coordination
In [3], Jadbabaie et al. also consider a modified version of Vicsek's discrete-time
system, which consists of the same group of n agents as before except that one
leader agent, labeled 0, is added. Agent 0 moves at the same constant speed as its
n followers but with a fixed heading θ0. Agent i then updates its heading using
the average of its own heading plus the headings of its neighbors. Note that this
time the leader may be in its neighborhood.
Our abstract Vicsek model with a leader now can be formulated as follows:
Definition 3.1. Suppose V + = {0, 1, · · · , n} and P + is the collection of simple
undirected graphs over V +. A leader-following Vicsek model is just a pair (V +, σ),
where σ : N → P + is a switching signal.
For each agent i > 0, define i's σ-neighborhood at time t, written Ni(t), to be
the set of agents that are connected to i by an edge in the graph σ(t). That is,
agent j is a neighbor of agent i if and only if (i, j) is an edge in the graph σ(t).
i=1 and a fixed heading θ0 in which
agent 0 moves at all times, for i > 0, agent i's heading evolves in discrete time
according to the following equation:
Given an initial heading θ(0) = hθi(0)in
θi(t + 1) =
θi(t) + Σj∈Ni(t)θj(t)
1 + ni(t)
(21)
where ni(t) is the number of agents in Ni(t).
9
For a leader-following Vicsek model, we have the following correspondence of
Theorem 3.1.
Theorem 3.2. Given a leader-following Vicsek model (V +, σ), suppose σ(∞) is
connected. Then for any θ(0) = hθi(0)in
i=1 and θ0, we have limt→∞ θi(t) = θ0 for
all i = 1, 2, · · · , n.
Proof. Note that Lemma 3.2 and Lemma 3.3 still hold for the leader-following
case; this theorem follows from a similar argument as given for Theorem 3.1.
4 Conclusions and further work
In [3], Jadbabaie et al. show that all agents shall move in the same heading if
the neighbor-graphs are periodically jointly connected, i.e., the union of any T
contiguous neighbor-graphs is connected for some fixed T . In the same paper,
they also ask whether this still holds when there exists a sequence of uniformly
bounded intervals over the discrete time such that the union of neighbor-graphs
across each interval is connected?
This paper has shown that all agents shall move in the same heading under a
very weak condition that requires the neighbor-graphs to be finally jointly con-
nected, i.e., the union of all graphs started from any time is connected. This
result gives an affirmative answer to the question raised in [3].
It should be emphasized that results obtained in this paper are valid for
many versions of the Vicsek model (or coordination multi-agent models that
use nearest-neighbor rule to update their state) (see Remark 2.1). As for some
specific versions of the Vicsek model, there have been some impressive results.
Recently, Jadbabaie [2] has shown that, if we choose the neighborhood region to
be open, then a necessary and sufficient condition for all headings to converge to
the same heading is that the neighbor-graph does not change after finite steps
and is connected.2 Jadbabaie also notes that [2, p.8, last paragraph] the problem
would be more complicated if a closed neighborhood region were chosen.
It
seems this method cannot be directly applied to other kinds of neighborhoods.
As a matter of fact, there are often situations when agents do not have disk-like
visibility but rather might have a cone-like field of view.3
Note that our results are based on the assumption that the switching signal is
pre-specified. In our future work, we shall plan to develop a model that can ex-
plain how the neighbor-graphs evolve over discrete time, and determine sufficient
conditions for coordination of multi-agents in terms of these agents' initial states.
Another thing that should be stressed is that coordination results obtained in
ours and in [3] are all asymptotic.
It will be interesting to devise other local
updating rules such that, using these rules, coordination will be reached quickly
and without centralized control. This work is currently being undertaken.
2A similar result was obtained by Savkin [6] for a simple version of Vicsek model where the
headings θi(t)'s take their values from a certain finite set.
3Such a situation has been investigated (though for continuous time) by Lin et al. [4].
10
References
[1] G. Flierl, D. Grunbaum, S. Levin, and D. Olson. From individuals to aggre-
gations: The interplay between behavior and physics. Journal of Theoretical
Biology, 196(4):397 -- 454, 1999.
[2] A. Jadbabaie. On distributed coordination of mobile agents with changing
nearest neighbors. Technical Report, University of Pennsylvania, Philade-
phia, PA, 2003.
[3] A. Jadbabaie, J. Lin, and A.S. Morse. Coordination of groups of mobile
IEEE Transactions on
autonomous agents using nearest neighbor rules.
Automatic Control, 48(6):988 -- 1001, 2003.
[4] Z. Lin, M. Broucke, and B. Francis. Local control strategies for groups
IEEE Transactions on Automatic Control,
of mobile autonomous agents.
49(4):622 -- 628, 2004.
[5] C.W. Reynolds. Flocks, birds, and schools: A distributed behavioral model.
Computer Graphics, 21:25 -- 34, 1987.
[6] A. Savkin. Coordinated collective motion of groups of autonomous mobile
robots: Analysis of vicsek's model. IEEE Transactions on Automatic Con-
trol, 49(6):981 -- 983, 2004.
[7] J. Toner and Y.H. Tu. Long range order in a two dimensional xy model:
How birds fly together? Physical Review Letters, 75:4326 -- 4329, 1995.
[8] J. Toner and Y.H. Tu. Flocks, herds, and schools: A quantitative theory of
flocking. Physical Review E, 58(4):4828 -- 4858, 1998.
[9] T. Vicsek, A. Czirok, E. Ben Jacob, I. Cohen, and O. Schochet. Novel type
of phaze transitions in a system of self-driven particles. Physical Review
Letters, 75:1226 -- 1229, 1995.
[10] K. Warburton and J. Lazarus. Tendency distance models of social cohesion
in animal groups. Journal of Theoretical Biology, 150(4):473 -- 488, 1991.
11
|
1911.06783 | 1 | 1911 | 2019-11-15T17:58:41 | A Turing Test for Crowds | [
"cs.MA"
] | The realism and believability of crowd simulations underpins computational studies of human collective behaviour, with implications for urban design, policing, security and many other areas. Realism concerns the closeness of the fit between a simulation and observed data, and believability concerns the human perception of plausibility. In this paper, we ask two questions, via a so-called "Turing Test" for crowds: (1) Can human observers distinguish between real and simulated crowds, and (2) Can human observers identify real crowds versus simulated crowds? In a study with student volunteers (n=384), we find convincing evidence that non-specialist individuals are able to reliably distinguish between real and simulated crowds. A rather more surprising result is that such individuals are overwhelmingly unable to identify real crowds. That is, they can tell real from simulated crowds, but are unable to say which is which. Our main conclusion is that (to the lay-person, at least) realistic crowds are not believable (and vice versa). | cs.MA | cs | A Turing Test for Crowds
Jamie Webster and Martyn Amos1
Abstract. The realism and believability of crowd simulations un-
derpins computational studies of human collective behaviour, with
implications for urban design, policing, security and many other ar-
eas. Realism concerns the closeness of the fit between a simulation
and observed data, and believability concerns the human perception
of plausibility. In this paper, we ask two questions, via a so-called
"Turing Test" for crowds: (1) Can human observers distinguish be-
tween real and simulated crowds, and (2) Can human observers iden-
tify real crowds versus simulated crowds? In a study with student
volunteers (n=384), we find convincing evidence that non-specialist
individuals are able to reliably distinguish between real and simu-
lated crowds. A rather more surprising result is that such individu-
als are overwhelmingly unable to identify real crowds. That is, they
can tell real from simulated crowds, but are unable to say which is
which. Our main conclusion is that (to the lay-person, at least) real-
istic crowds are not believable (and vice versa).
1
Introduction
The formal study of human crowds dates back to before the French
Revolution [8], but understanding collective behaviour is more ur-
gent than ever before, as populations migrate to high-density urban
centres, protests become more organized (and perhaps more com-
mon), and increasing numbers of individuals pass through large-scale
transportation hubs [9]. A number of computational techniques exist
to study the dynamics of crowd behaviour, but the most commonly-
used is simulation [41].
Crowd simulations (generally, but not exclusively, using an agent-
based approach) are now employed in many different domains, from
events planning and management
[5], to urban design [12], and
incident response and analysis [16, 37]. By studying flows of people
en masse, and their interactions with the environment and with one
another, researchers aim to better understand human collective social
behaviour, design more effective and enjoyable public spaces, and
improve levels of safety, security and well-being [15].
In this paper, we consider two related properties of crowd simula-
tions; (1) realism, and (2) believability. The first property concerns
how well a simulation's output matches the expected or observed
behaviour of a real crowd in the same scenario, under the same con-
ditions. The second property centres on how convincing a simulation
is to a human observer, and how closely it matches their expectations
of how a crowd will behave. The two properties are closely inter-
linked, and "believability" is often used as a synonym for "realistic".
However, as we will see, the two concepts require close examination
and careful handling.
1 Department of Computer and Information Sciences, Northumbria Uni-
versity, Newcastle upon Tyne, NE8 1ST, United Kingdom. Email:
{jamie.webster, martyn.amos}@northumbria.ac.uk
9
1
0
2
v
o
N
5
1
]
A
M
.
s
c
[
1
v
3
8
7
6
0
.
1
1
9
1
:
v
i
X
r
a
The rest of the paper is organized as follows: in Section 2 we
briefly review related work on crowd simulation and collective be-
haviour; in Section 3 we present our "Turing Test" for crowd be-
haviour, and in Section 4 we give the results of experimental trials.
We conclude with a brief discussion of the implications of our find-
ings.
2 Background
The study of human crowd dynamics [1] is motivated by the desire to
understand and predict the behaviour of individuals en masse, and en-
compasses a diverse range of crowd types, from large, mainly static
crowds at sporting events or concerts [44], to transitory and flowing
crowds, such as those found in train stations at rush hour [47], or at
religious events such as Hajj [7]. As urban centres grow in size (the
United Nations predicts that, by 2050, 68% of the global population
will live in cities [31]), we will need to understand and mitigate the
impact of crowds on infrastructure, safety, security, and quality of
life [2].
Early attempts to understand crowd behaviour were rooted in the
physical sciences, using metaphors and mathematical tools from
fluid dynamics [19], and modelled crowds at the macroscopic level
(i.e., without considering individuals) [21]. Subsequent work used
an entity-based approach, which treated crowds as individual "par-
ticles" [4, 42], along with the agent-based methodology, in which
individuals are treated as semi-autonomous actors [33].
As crowd simulations have become used more frequently, atten-
tion has become focused on issues of realism and believability.
Here, we define the realism of a simulation in terms of its validity
[22, 25, 36, 38]; how closely does the output of the model match data
obtained in the real world? It is straightforward to obtain statistical
properties of simulation outputs and compare these to the properties
of real-world crowds, and that is the approach we take in this paper.
The issue of believability is subtly different, and concerns the hu-
man perception of whether or not a crowd's behaviour is plausible.
As computer-generated imagery (CGI) becomes increasingly com-
mon in large-scale cinematic productions, it is being used to replace
human actors in large-scale crowd scenes, for reasons of cost and/or
feasibility (e.g., the 2001 film The Lord of the Rings: The Fellow-
ship of the Ring featured a prologue battle with 100,000 computer-
generated fighters). Many modern video games also feature large
numbers of simulated individuals. In these cases, the main concern
is to "fool" the observer into thinking that they are watching a real
crowd, without necessarily producing patterns that are behaviourily
valid [11, 34].
We seek, therefore, to study whether or not human observers may
be persuaded that a simulated crowd is actually a "real" crowd. This
may be thought of as a limited form of the famous "Turing Test",
named after Alan Turing, and described in his landmark paper on ar-
Figure 2. Distribution of walking speeds for pedestrians observed in Edin-
burgh Informatics Forum.
individuals whose trajectories were captured were actively partici-
pating in movement studies; the trajectories, therefore, are as close
to "natural" as possible (i.e., they have "behavioural ecological va-
lidity" [29]).
In what follows, we use the term "clip" to specifically refer to
a time-limited sequence of trajectory data (whether taken from the
Edinburgh dataset or from the output of a simulation), as opposed
to a movie visualisation. We wrote a utility to search the Edinburgh
dataset and extract clips of a specific duration containing a specific
number of individuals. This allowed us to ensure that the "real" and
"simulated" crowds contained the same number of individuals for
any single comparison.
3.2 Model calibration
In order to calibrate our simulation (and, later, to perform statisti-
cal analysis), we selected 20 clips at random from the Edinburgh
dataset (each of 60s duration), and calculated the average walking
speed of pedestrians observed traversing the Forum. The distribution
of speeds is shown in Figure 2, with a mean value of 1.17m/s. When
we simulated these scenarios (see next Section), the mean speed of
agents was higher (1.63m/s), due to the fact that simulated agents
were rarely impeded, did not encounter bottlenecks, and were free to
accelerate up to their maximum speed. However, as we will see from
our results, this did not affect the perception of the simulated crowds.
For the comparison experiments, we randomly selected six clips
taken from the Edinburgh dataset (the number of clips is the same
as in [20]; each was of 60s duration, and the number of individuals
in a clip ranged from 104 to 194 (with an average of 139). For each
clip, we extracted the route choice distribution and the entry time
distribution for all individuals. This allowed us to initialise our sim-
ulations with the same distributions, ensuring that the runs closely
matched the macroscopic properties of the real-world observations
(while leaving room for the microscopic differences in which we are
interested). We also calculated the average velocity of individuals in
each clip, and used this to scale the clip's length (by modifying the
video playback speed) to account for variability in camera capture
rate, thus normalizing the velocity of individuals relative to expected
walking speed [3].
Figure 1. Diagram of Edinburgh Informatics Forum (ingress and egress
points numbered).
tificial intelligence [43]. Turing proposed that if a human observer
was unable to distinguish between another person and a machine de-
signed to produce human-like responses in a conversational setting,
then the machine would be deemed to have "passed" the test. This
type of test has been proposed for biological modelling [17] and arti-
ficial life [6] as a way of capturing and interrogating life-like proper-
ties of artificial systems, and assessing the completeness and validity
of a model. Recently, a Turing test for collective motion in fish was
described [20], and we base our approach on this. Our overall aim is
to explore how a Turing-like test may be used to examine assump-
tions and preconceptions about the behaviour of human crowds, and
to establish the features of real crowds that must be emulated by a
simulation in order for it to be valid and/or "pass" the test. In the
next Section, we describe our methodology for doing so.
3 Experimental method
Our methodology is based on that of [20], but with in-person (as op-
posed to online) participants. We showed all participants six pairs of
videos, in which one randomly-selected video showed the movement
of a real crowd, and the other showed the movement of a simulated
crowd. Both real and simulated crowd movements were displayed us-
ing the same rendering method, and participants were asked to spec-
ify on a form which of each pair they thought was the real crowd.
3.1 Dataset
We use data on real pedestrians from the University of Edinburgh
School of Informatics [30]. This public dataset, captured in 2010,
contains over 299,000 individual trajectories corresponding to the
movement of individuals through the School Forum, and is one of
the largest open datasets of its type. A diagram of the Forum space is
shown in Figure 1. The Forum is rectangular in shape (measuring ap-
proximately 15.8 × 11.86 metres), has eleven ingress/egress points,
and is generally clear of obstructions. Images were captured (9 per
second) by a camera suspended 23m above the Forum floor, from
which individual trajectories ("tracks") were extracted and made
available (extraction was performed by the author of
[30]). This
dataset has been used in a large number of studies of pedestrian
movement/tracking, including [13, 29, 46], Importantly, none of the
2
3.3 Simulations
In order to produce the simulations to accompany each Edinburgh
clip, we simulated pedestrian movement using the Vadere package
[24]. This package is open-source, which means that (unlike com-
mercial software) its movement models are open to inspection, and
it also allows for easy exporting of simulating pedestrian trajectories
(which is important when we consider that we must use the same
renderer for both real and simulated videos).
A crucial component of the simulation is the crowd motion model.
This defines the rules of interaction between individuals (e.g., avoid-
ance), and between individuals and their environment (e.g., repulsion
from walls and physical obstacles), as well as route choice behaviour
and differential walking speed. Many different crowd motion mod-
els exist [10], but perhaps the most commonly-used type is based on
social forces. Inspired by the fluid flow paradigm of Henderson [19]
and others, Helbing and Molnar's social force model (SFM)
[18]
is a microscopic, continuous model which uses "attractive" and "re-
pulsive" force fields between individuals (and between individuals
and their environment) to guide movement. The SFM provides the
base movement model for a number of pedestrian simulation pack-
ages, including FDS+EVAC [26], PedSim [14], SimWalk [23] and
MassMotion [32], and it has been used extensively in movement re-
search. Additionally, the SFM has been validated using real-world
data [22, 38], and the comprehensive review of [10] recommends its
use in pedestrian movement studies. For all simulations, we used the
pre-supplied Vadere template for Helbing and Molnar's SFM, with
default attributes and parameters (listed in Table 1).
Table 1. Vadere simulation model parameters.
Parameter
Value
ODE Solver
Pedestrian body potential
Pedestrian recognition distance
Obstacle body potential
Obstacle repulsion strength
Pedestrian radius (m)
Pedestrian speed distribution mean (m/s)
Pedestrian minimum speed (m/s)
Pedestrian maximum speed (m/s)
Pedestrian acceleration (m/s)
Pedestrian search radius (m)
Dormand-Prince method
2.72
0.3
20.1
0.25
0.2
1.4
0.4
3.2
2.0
2.0
We added small amounts of noise to the simulated trajectories in
order to replicate noise in the real crowd data. As the Edinburgh
individuals were detected by an overhead camera running at 9fps,
occasional faulty detections caused very short-term errors in the ex-
tracted trajectories. Once rendered, this cause individuals to appear
to rapidly "flick" between two headings. As we had no reliable way
to quantify the (by inspection, small) amount of noise in the trajecto-
ries, we adjusted this by eye until the apparent noise in the simulated
data matched the noise level observed in the real data. At any time-
step, a simulated agent had a 15% chance of temporarily "flicking"
their heading by a randomly selected value up to 45 degrees (with-
out changing their trajectory). Importantly, as we will see from the
results, the addition of this noise had no effect on how the simulated
crowds were perceived.
A second artefact of inaccurate detections was that some trajec-
tories had missing sections for several time steps; once rendered,
these individuals would temporarily disappear from the frame and
then reappear. To fix this, we automatically detected such situations
and interpolated coordinates for the missing time-steps when parsing
the Edinburgh dataset. We also increased the number of frames per
second of both sets of trajectories (real and simulated), from 9 to 72,
by interpolating coordinates. This enabled smooth video playback
for the purpose of comparisons.
3.4 Comparisons
The trajectories of both the simulated and real individuals in each
pair of clips were then rendered in a uniform fashion, using a tool
coded in Java. This allowed us to produce "top down" visualisations
of both real and simulated clips that were uniform in appearance,
with individuals represented as filled circles, and headings depicted
by an arrow (see Figure 3).
Figure 3. Single frame render of an example crowd.
For each pair, the real and simulated videos were randomly as-
signed to position A (left) or B (right), and these were combined
side-by-side into a single video. Individual videos did not "loop",
and were made up of the first 30 seconds of the real and simulated
crowd clips in each pair. The total duration of the video, showing a
total of six comparisons, was 3m 18s (including a 3s pause between
each pair). The video is available at http://drives.media/google857,
and the real crowds are A, A, B, A, B and B.
3.5 Statistical properties
The first question we asked was whether or not participants could
distinguish between real and simulated crowds, even when their sta-
tistical properties were very similar. That is, are there features of real
(or simulated) crowds that somehow act as an discriminator, even
when there is no significant statistical difference between the two?
We used two metrics (as in [20]); polarization and nearest neigh-
bour distance (NND). Polarization measures the level of "order" in
a crowd, in terms of the heading alignment of members. Polarization
is zero when the crowd is completely disordered (everyone is point-
ing in a different direction), and has a maximum value of 1 when all
members of the crowd have the same heading:
(cid:42)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) i=1(cid:88)
N
(cid:43)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ϕ =
1
N
exp(ιθi)
,
(1)
where N is the number of individuals in the frame, ι is the imagi-
nary unit, and θi is the heading of each individual.
3
the order of 2%, and we do not believe that this is significant enough
to introduce any perceptible difference.
4 Experiments
We recruited 384 undergraduate students from Northumbria Univer-
sity, distributed over 9 groups taken from a mixture of computer sci-
ence and engineering courses. Of the participants who supplied their
details, the gender distribution was 78.83% male, 18.66% female,
2.5% non-binary/other, and the average age was 20.7 (we exclude
one outlier age value of 71, corresponding to a student's reader). All
trials took place at the beginning of a class, for which prior permis-
sion was obtained from the tutor. Students were informed about the
nature of the experiment, and told that they were under no obligation
to participate. Answer sheets were distributed, which consisted of a
simple numbered list of tickboxes (for each line, the choice was A or
B). Participants were asked to optionally specify their age and gen-
der. At the end of the trial, participants were also asked to provide
some optional narrative notes on any distinguishing features they
noticed that allowed them to tell the real crowd from the simulated
crowd. Each trial (from initial set-up to collection of answer sheets)
took around ten minutes. The experimental protocol was approved by
the Northumbria University Faculty of Engineering and Environment
Ethics Board, application number 16433.
4.1 Results
We define "score" in terms of correct identification of the real crowd;
so a score of zero means that a participant failed to identify any of
the real crowds, and a score of 6 means that the participant identified
the real crowd on every occasion. The average score for participants,
across all comparisons, was 1.6. That is, participants performed less
well than if they had guessed at random. The overall distribution of
scores is shown in Figure 5, overlaid with the expected binomial
distribution (as each comparison is a binary choice, we show this to
illustrate the expected distribution of scores if selections were made
at random)2.
If the real and simulated crowds were genuinely indistinguishable
(that is, the best strategy would be no better than random guessing),
then we would expect roughly 3% of participants (around 12 people)
to either guess none correctly, or to guess all six correctly. What we
actually found was that over 40% of participants (154 individuals)
obtained a score of either zero or six. That is, those individuals were
able to correctly partition all six pairs of videos into two sets. This
answers, in the affirmative, the first research question: can individu-
als distinguish between real and simulated crowds, even when they
have very similar statistical properties?
A highly striking result is that the most common score, by far,
was zero. That is, a significant proportion of participants (36.46%)
failed to identify a single real crowd. Only 3.65% of participants ob-
tained a perfect score of 6. The important implication of this is that
participants were reliably able to partition videos along the lines of
"real"/"artificial", but most of them were unable to say which was
which. This is a much stronger version of the result obtained in [20],
where participants were able to tell real fish from simulated fish, but
were not necessarily able to identify the real fish.
We conclude, therefore, that the second research question (can in-
dividuals identify real crowds versus simulated crowds, even when
they have very similar statistical properties?) must be answered in
2 Full results are available at https://doi.org/10.6084/m9.figshare.10308407
Figure 4. Crowd simulations/real crowd comparisons: Nearest Neighbour
Distance (NND) (top) and polarization (bottom) as a function of crowd size.
The outputs of both simulations have statistical properties that are close to
those of the real crowds.
Nearest-neighbour distance (NND) measures the level of "cluster-
ing in a crowd. The average NND for a single "frame" (derived from
either the real dataset or the simulation) is calculated from the sum
of nearest-neighbour distances of all N individuals:
i=1(cid:88)
N
ν =
1
N
di,
(2)
where di is the nearest neighbor distance between point i and the
closest individual in the frame, as calculated by the standard distance
formula,
di =(cid:112)(x2 − x1)2 + (y2 − y1)2
(3)
In order to confirm that we did not introduce implementation-
specific bias by choosing a specific software platform, we compared
the outputs of Vadere and JuPedSim [45], an alternative open-source
simulation package. We used each package to simulate the 20 real
crowd clips mentioned in the previous Section, and calculated aver-
age NND and polarization over 20 runs for each. The same statistics
were then calculated for the real clips (Figure 4).
These results confirmed that crowd simulations in Vadere and
JuPedSim display similar properties in terms of both NND and polar-
ization, so we selected Vadere as a representative example of crowd
simulation packages in general.
Importantly,
the statistical properties of the simulations also
matched the general properties of the real crowds, which confirmed
that they are essentially indistinguishable in those terms. In Figure 4
we notice a slight difference between the real and simulated crowds
in terms of polarization; the real crowds are generally slightly more
closely-aligned than the simulated crowds, but this difference is on
4
identifications, with an average score of 2.46. Group 9 (first year
computer science students) had the fewest correct identifications,
with an average score of 1.08.
The first
4.2 Narrative findings
In this Section, we analyse the free text supplied by participants. We
focus, in particular, on the large number of participants who scored
zero, as they consistently misidentified the real crowd. We highlight
themes and specific comments that may shed light on the assump-
tions and preconceptions held by these individuals, that led them to
consistently "flip" the real and simulated crowds in their perception.
theme that emerged concerned rapid or "random"
changes of movement in the real crowd, which many participants in-
correctly attributed to the simulated ("fake") crowd. Versions of this
included "Fake changed direction too quickly", "Fast change sug-
gests fake", "Generated crowd had too much random movement",
"Real seemed to change direction gently". Although the average
speed of the simulated agents was higher than that of the real people,
participants singled out rapid movement in the Edinburgh videos as
indicative of artificiality (when, in fact, the real people moved more
slowly). Overall, 72 participants mentioned a variant of this type of
observation. The underlying assumption here is that real people move
smoothly, at a uniform speed, and do not tend to deviate much from
their chosen path.
A second common theme concerned avoidance; many participants
incorrectly assumed that real people would avoid close contact with
one another, whereas the simulated individuals would "overlap" or
collide. Representative quotes included "Simulated people collided,
real crowds avoided each other", and "People overlapping". In re-
ality, the opposite is true, as the real dataset contains multiple in-
stances of individuals coming into close proximity. Moreover, the
social forces model explicitly tries to keep individuals apart unless
close proximity is unavoidable, so the behaviour (distance keeping)
that participants attributed to real people was actually an in-built fea-
ture of the simulation. This theme was mentioned by 22 participants.
Perhaps the most profound observations concerned perceived in-
tentionality and group-level behaviour; many participants believed
that "On the whole, people have relatively smooth and intentional
paths" (this was actually a feature of the simulation), "Real crowds
don't really stand around" (stationary groups were only present in the
real dataset), and "The real ones knew where they were going" (this
was actually a function of the simulation's path choice algorithm).
Variations on this theme were mentioned by 7 participants. The inter-
esting thing here is that participants (incorrectly) ascribed clear hu-
man intentionality and purpose to the simulated agents ("Real crowds
move more purposefully"), and failed to acknowledge it in the actual
humans that were observed.
Overall, we found that participants believed that individuals in
crowds are orderly, purposeful, respectful of personal space, and
consistent and uniform in their speed and direction. In fact, all of
these characteristics were features of the simulation. Participants also
failed to recognize features of real crowds such as rapid changes
in speed or direction, close proximity of individuals, and stationary
groups/individuals, all of which were discounted by participants as
being "glitchy" or "unrealistic". Our findings would, therefore, ap-
pear to contradict [28], in which the authors state that "However,
people do more than just walk. They talk to one another, they look
around, they scratch their heads or perform various other actions...
The absence of these mundane actions diminishes the credibility of
the simulated crowd." Such arguments about "diminished credibil-
Figure 5. Distribution of participant scores (the line represents the expected
binomial distribution).
the negative for this population. In the next Section, we analyse the
narrative text supplied by participants, in order to explore the under-
lying rationale for their decisions.
We now briefly explore secondary features of our findings. The
results for each comparison are shown in Figure 6, which we present
in terms of the proportion of participants who correctly selected the
real crowd. These results show that pair 6 presented the strongest
challenge to participants, and pair 2 was considered the least chal-
lenging. Overall, no clear trend emerged in terms of differential chal-
lenge across comparisons.
Figure 6. Success rate across individual pairwise comparisons.
Figure 7. Distribution of scores across participant groups.
In terms of variation across groups (Figure 7), Group 1 (second
year Engineering Mathematics students) obtained the most correct
5
ity" would seem to have little empirical basis in fact, and even when
attempts are made to integrate "realistic" behaviours into crowd sim-
ulations, no rigorous test is performed to assess whether or not they
have had any impact on perceived plausibility in human observers.
5 Conclusions
In this paper, we presented a Turing test for crowds that allowed us
to investigate issues of realism and believability in crowd simula-
tions by comparing them with visualisations representing data ob-
tained from real pedestrians. We performed trials with nearly 400
university students, and found that, while the students were gener-
ally able to discriminate between "real" and "artificial" crowds, they
were unable to say which was which. On the surface, this was a rather
surprising result, but it serves to emphasise that a simulation model
that is realistic (that is, accurately reflects reality) might be useful
for planning an evacuation scenario, where fidelity is paramount, but
it might be less useful where "believability" is more important (in a
video game, for example). Conversely, this serves as a warning that
software development platforms aimed primarily at entertainment ti-
tles should perhaps not be relied upon for safety-critical or poten-
tially expensive infrastructure studies without careful modification.
We acknowledge several potential limitations of our study; the use
of students as test subjects is the subject of ongoing debate [35], and
the computer science background of many of the students (and the
gender imbalance) may have biased our results. It may be the case
that our students have become conditioned to make certain assump-
tions about how crowds behave from playing games that use a rela-
tively unrealistic crowd model. However, this is merely speculation
on our part. Nonetheless, an important future development of this
work will be to rerun the trials using experts in crowd dynamics, to
find out whether they are better placed to identify the real crowd.
This is entirely consistent with Harel's expectation of how a bio-
logical modelling Turing test might work; "...our interrogators can't
simply be any humans of average intelligence. Both they and the ....
people responsible for 'running' the real organism and providing its
responses to probes, would have to be experts on the subject matter
of the model, appropriately knowledgeable about its coverage and
levels of detail." [17]
If (as we might expect) the experts are able to reliably identify the
real crowd, then this immediately suggests a mechanism for ascer-
taining the minimal set of crowd features that are necessary to "pass"
the test. If, for example we identified that "group-level movement"
was a "flag" for the experts, we might include such a behaviour in the
simulation and rerun the trial with a second group of experts. If the
success rate falls, and the experts are less able to tell the difference
between real and simulated crowds, then we might conclude that
group-level behaviour constitutes an important feature that should be
included in simulation packages. This would represent a formalised
methodology for implementing a number of recommendations that
have been recently made by a number of crowd scientists, who call
for the integration into software of a wider range of psychological
and inter-personal processes [27, 39, 40]. These recommendations
reflect a pressing need to revisit physics-based models of crowd be-
haviour which, though they may generate macroscopic behaviour
that is reasonably realistic, fail to capture the inherent "messiness"
and unpredictability of real human crowds.
ACKNOWLEDGEMENTS
We thank Jeremy Ellman for advice on statistical analysis.
6
REFERENCES
[1]
J. Adrian, M. Amos, M. Baratchi, et al., 'A glossary for research on
human crowd dynamics', Collective Dynamics, 4(A19), 1 -- 13, (2019).
[2] G.D.P.A. Aschwanden, S. Haegler, F. Bosch´e, L. Van Gool, and
G. Schmitt, 'Empiric design evaluation in urban planning', Automation
in Construction, 20(3), 299 -- 310, (2011).
[3] R.W. Bohannon, 'Comfortable and maximum walking speed of adults
aged 2079 years: reference values and determinants', Age and Ageing,
26(1), 15 -- 19, (1997).
[4] E. Bouvier, E. Cohen, and L. Najman, 'From crowd simulation to airbag
deployment: particle systems, a new paradigm of simulation', Journal
of Electronic Imaging, 6(1), 94 -- 108, (1997).
[5] L. Crociani, G. Lammel, and G. Vizzari, 'Multi-scale simulation for
crowd management: a case study in an urban scenario', in International
Conference on Autonomous Agents and Multiagent Systems (AAMAS),
pp. 147 -- 162. Springer, (2016).
[6] L. Cronin, N. Krasnogor, B.G. Davis, C. Alexander, N. Robertson,
J.H.G. Steinke, S.L.M. Schroeder, A.N. Khlobystov, G. Cooper, P.M.
Gardner, et al., 'The imitation gamea computational chemical approach
to recognizing life', Nature Biotechnology, 24(10), 1203, (2006).
[7] S. Curtis, S.J. Guy, B. Zafar, and D. Manocha, 'Virtual Tawaf: A case
study in simulating the behavior of dense, heterogeneous crowds', in
2011 IEEE International Conference on Computer Vision Workshops
(ICCV Workshops), pp. 128 -- 135. IEEE, (2011).
J. Drury and C. Stott, 'Contextualising the crowd in contemporary so-
cial science', Contemporary Social Science, 6(3), 275 -- 288, (2011).
J. Drury and C. Stott, Crowds in the 21st Century: Perspectives from
Contemporary Social Science, Routledge, 2015.
[8]
[9]
[10] D.C. Duives, W. Daamen, and S.P. Hoogendoorn, 'State-of-the-art
crowd motion simulation models', Transportation Research Part C:
Emerging Technologies, 37, 193 -- 209, (2013).
[11] F. Durupinar, N. Pelechano, J. Allbeck, U. Gudukbay, and N.I. Badler,
'How the ocean personality model affects the perception of crowds',
IEEE Computer Graphics and Applications, 31(3), 22 -- 31, (2009).
[12] T. Feng, L-F. Yu, S-K Yeung, K. Yin, and K. Zhou, 'Crowd-driven mid-
scale layout design.', ACM Transactions on Graphics, 35(4), 132 -- 1,
(2016).
[13] T. Fernando, S. Denman, S. Sridharan, and C. Fookes, 'Tracking by
prediction: A deep generative model for multi-person localisation and
tracking', in 2018 IEEE Winter Conference on Applications of Com-
puter Vision (WACV), pp. 1122 -- 1132. IEEE, (2018).
[14] C. Gloor.
PedSim: Pedestrian
crowd
simulation,
2016.
http://pedsim.silmaril.org.
J. Halatsch, A. Kunze, and G. Schmitt, 'Value Lab: A collabora-
tive environment for the planning of Future Cities', in Proceedings of
eCAADe, volume 27, (2009).
[15]
[16] P. Harding, S. Gwynne, and M. Amos, 'Mutual information for the de-
tection of crush', PloS ONE, 6(12), e28747, (2011).
[17] D. Harel, 'A Turing-like test for biological modeling', Nature Biotech-
nology, 23(4), 495, (2005).
[18] D. Helbing and P. Molnar, 'Social force model for pedestrian dynam-
ics', Physical Review E, 51(5), 4282, (1995).
[19] L.F. Henderson, 'On the fluid mechanics of human crowd motion',
Transportation research, 8(6), 509 -- 515, (1974).
J.E. Herbert-Read, M. Romenskyy, and D.J. Sumpter, 'A Turing test for
collective motion', Biology Letters, 11, 20150674, (2015).
[21] R.L. Hughes, 'The flow of human crowds', Annual Review of Fluid
[20]
Mechanics, 35(1), 169 -- 182, (2003).
[22] A. Johansson, D. Helbing, and P.K. Shukla, 'Specification of the social
force pedestrian model by evolutionary adjustment to video tracking
data', Advances in Complex Systems, 10(supp02), 271 -- 288, (2007).
[23] T. Kimura, H. Sekine, T. Sano, N. Takeichi, Y. Yoshida, and H. Watan-
abe, 'Pedestrian simulation system SimWalk', in Summaries of Techni-
cal Papers of Annual Meeting Architectural Institute of Japan, E-1, pp.
915 -- 916, (2003).
[24] B. Kleinmeier, B. Zonnchen, M. Godel, and G. Koster, 'Vadere: An
open-source simulation framework to promote interdisciplinary under-
standing', arXiv preprint arXiv:1907.09520, (2019).
[25] H. Klupfel, 'The simulation of crowd dynamics at very large events -
calibration, empirical data, and validation', in Pedestrian and Evacua-
tion Dynamics (PED) 2005, 285 -- 296, Springer, (2007).
[26] T. Korhonen, S. Hostikka, S. Heliovaara, and H. Ehtamo, 'FDS+Evac:
an agent based fire evacuation model', in Pedestrian and Evacuation
Dynamics (PED) 2008, 109 -- 120, Springer, (2010).
[27] S. Lemercier and J.M. Auberlet, 'Towards more behaviors in crowd
simulation', Computer Animation And Virtual Worlds, (2016).
[28] A. Lerner, E. Fitusi, Y. Chrysanthou, and D. Cohen-Or, 'Fitting behav-
iors to pedestrian simulations', in Proceedings of the 2009 ACM SIG-
GRAPH/Eurographics Symposium on Computer Animation, pp. 199 --
208. ACM, (2009).
[29] R. Lovreglio, C. Dias, X. Song, and L. Ballerini, 'Towards micro-
scopic calibration of pedestrian simulation models using open trajec-
tory datasets: the case study of the Edinburgh Informatics Forum',
in Conference on Traffic and Granular Flow, Washington DC, USA,
(2017).
[30] B. Majecka, Statistical models of pedestrian behaviour in the forum,
Master's thesis, School of Informatics, University of Edinburgh, 2009.
[31] United Nations. 2018 Revision of World Urbanization Prospects, 2018.
[32] Oasys. Mass motion product page, 2019.
https://www.oasys-
software.com/products/pedestrian-simulation/massmotion/.
[36]
[33] X. Pan, C.S. Han, K. Dauber, and K.H. Law, 'A multi-agent based
framework for the simulation of human and social behaviors during
emergency evacuations', AI & Society, 22(2), 113 -- 132, (2007).
[34] C. Peters and C. Ennis, 'Modeling groups of plausible virtual pedestri-
ans', IEEE Computer Graphics and Applications, 29(4), 54 -- 63, (2009).
[35] R.A. Peterson, 'On the use of college students in social science re-
search: Insights from a second-order meta-analysis', Journal of Con-
sumer Research, 28(3), 450 -- 461, (2001).
J. Pettr´e, J. Ondrej, A-H. Olivier, A. Cretual, and S. Donikian,
'Experiment-based modeling, simulation and validation of interac-
tions between virtual walkers', in Proceedings of the 2009 ACM SIG-
GRAPH/Eurographics Symposium on Computer Animation, pp. 189 --
198. ACM, (2009).
[37] M. Pretorius, S. Gwynne, and E.R. Galea, 'Large crowd modelling:
an analysis of the Duisburg Love Parade disaster', Fire and Materials,
39(4), 301 -- 322, (2015).
[38] S. Seer, C. Rudloff, T. Matyus, and N. Brandle, 'Validating social force
based models with comprehensive real world motion data', Transporta-
tion Research Procedia, 2, 724 -- 732, (2014).
[39] M.J. Seitz, A. Templeton, J. Drury, G. Koster, and A. Philippides, 'Par-
simony versus reductionism: How can crowd psychology be introduced
into computer simulation?', Review of General Psychology, 21(1), 95 --
102, (2017).
[40] A. Templeton, J. Drury, and A. Philippides, 'From mindless masses to
small groups : conceptualizing collective behavior in crowd modeling',
Review of General Psychology, 19(3), 215 -- 229, (2015).
[41] S. Thalmann and S.R. Musse, Crowd Simulation, Springer, 2013.
[42] A. Treuille, S. Cooper, and Z. Popovi´c, 'Continuum crowds', in ACM
Transactions on Graphics (TOG), volume 25, pp. 1160 -- 1168. ACM,
(2006).
[43] A.M. Turing, 'Computing machinery and intelligence', Mind, 59(236),
433, (1950).
[44] N. Wagner and V. Agrawal, 'An agent-based simulation system for con-
cert venue crowd evacuation modeling in the presence of a fire disaster',
Expert Systems with Applications, 41(6), 2807 -- 2815, (2014).
[45] A.K. Wagoum, M. Chraibi, J. Zhang, and G. Lammel, 'JuPedSim: an
open framework for simulating and analyzing the dynamics of pedes-
trians', in 3rd Conference of Transportation Research Group of India,
(2015).
[46] H. Wang and C. OSullivan, 'Globally continuous and non-Markovian
crowd activity analysis from videos', in European Conference on Com-
puter Vision, pp. 527 -- 544. Springer, (2016).
[47] Q. Zhang, B. Han, and D. Li, 'Modeling and simulation of passenger
alighting and boarding movement in Beijing metro stations', Trans-
portation Research Part C: Emerging Technologies, 16(5), 635 -- 649,
(2008).
7
|
1606.05837 | 1 | 1606 | 2016-06-19T07:16:28 | Acyclic Games and Iterative Voting | [
"cs.MA",
"cs.GT"
] | We consider iterative voting models and position them within the general framework of acyclic games and game forms. More specifically, we classify convergence results based on the underlying assumptions on the agent scheduler (the order of players) and the action scheduler (which better-reply is played).
Our main technical result is providing a complete picture of conditions for acyclicity in several variations of Plurality voting. In particular, we show that (a) under the traditional lexicographic tie-breaking, the game converges for any order of players under a weak restriction on voters' actions; and (b) Plurality with randomized tie-breaking is not guaranteed to converge under arbitrary agent schedulers, but from any initial state there is \emph{some} path of better-replies to a Nash equilibrium. We thus show a first separation between restricted-acyclicity and weak-acyclicity of game forms, thereby settling an open question from [Kukushkin, IJGT 2011]. In addition, we refute another conjecture regarding strongly-acyclic voting rules. | cs.MA | cs |
Acyclic Games and Iterative Voting ∗
Reshef Meir1
Maria Polukarov2
Jeffrey S. Rosenschein3
Nicholas R. Jennings4
1Technion-Israel Institute of Technology [email protected]
2University of Southampton, United Kingdom [email protected]
3The Hebrew University of Jerusalem, Israel [email protected]
4Imperial College, United Kingdom [email protected]
Abstract
We consider iterative voting models and position them within the general framework of acyclic games and
game forms. More specifically, we classify convergence results based on the underlying assumptions on the agent
scheduler (the order of players) and the action scheduler (which better-reply is played).
Our main technical result is providing a complete picture of conditions for acyclicity in several variations
of Plurality voting. In particular, we show that (a) under the traditional lexicographic tie-breaking, the game
converges for any order of players under a weak restriction on voters' actions; and (b) Plurality with randomized
tie-breaking is not guaranteed to converge under arbitrary agent schedulers, but from any initial state there is
some path of better-replies to a Nash equilibrium. We thus show a first separation between restricted-acyclicity
and weak-acyclicity of game forms, thereby settling an open question from [Kukushkin, 2011]. In addition, we
refute another conjecture regarding strongly-acyclic voting rules.
1
Introduction
Researchers in economics and game theory since Cournot [1838] had been developing a formal framework to study
questions about acyclicity and convergence of local improvement dynamics in games.
Intuitively put, strong-acyclicity means that the game will converge regardless of the order of players/voters
and how they select their action (as long as the moving agents are improving their utility in every step), i.e. that
there are no cycles of better-replies whatsoever; Weak-acyclicity means that while cycles may occur, from any
initial state (voting profile) there is at least one path of better-replies that leads to a Nash equilibrium; Restricted-
acyclicity is a middle ground, requiring convergence for any order of players (agent scheduler), but allowing the
action scheduler to restrict the way they choose among several available replies (e.g., only allowing best-replies).
Most relevant to us is the work of Kukushkin [1999; 2002; 2011], who studied general characterizations of game
forms that guarantee various notions of acyclicity.
A more recent field is iterative voting. In the iterative voting model, voters have fixed preferences and start
from some announcement (e.g., sincerely report their preferences). Votes are aggregated via some predefined rule
(e.g. Plurality), but can change their votes after observing the current announcements and outcome. The game
proceeds in turns, where a single voter changes his vote at each turn, until no voter has objections and the final
outcome is announced. This process is similar to online polls via Doodle or Facebook, where users can log-in at
any time and change their vote. Similarly, in offline committees the participants can sometimes ask to change their
vote, seeing the current outcome.
The formal study of iterative voting rules was initiated about 6 years ago in a AAAI paper that was a pre-
liminary version of this one [Meir et al., 2010]. Iterative voting papers typically focus on common voting rules
∗A preliminary version of this paper has been presented at AAAI-2010 [Meir et al., 2010].
1
such as Plurality and Borda, and study the conditions under which convergence of the iterative process to a Nash
equilibrium is guaranteed. Most results in the field consider best-reply dynamics [Lev and Rosenschein, 2012;
Reyhani and Wilson, 2012; Obraztsova et al., 2015].
While voting rules and game forms are essentially the same thing, the iterative voting literature has remained
largely detached from the more general literature on acyclicity in games. Bridging this gap is the main conceptual
contribution of this work, for two reasons. First, understanding the conditions that entail acyclicity of games and
game forms is crucial to the understanding of iterative voting scenarios, and to properly compare convergence
results (e.g. convergence of best-reply dynamics is a special case of restricted acyclicity). Likewise, convergence
results for specific voting rules under best/better-reply dynamics may shed light on more general questions re-
garding acyclicity. Building on the formalism of Kukushkin [2011] for strong/ restricted/ weak-acyclicity of game
forms, we re-interpret in this paper both known and new results on convergence of better- and best-reply in voting
games, and answer some open questions.
1.1 Related work
Kukushkin [2011] provided several partial characterizations for game forms with strong acyclicity. In particular, he
showed that if we further strengthen the acyclicity requirement to demand an ordinal potential, then this is attained
if and only if the game form is dictatorial, i.e., there is at most one voter that can affect the outcome. He further
characterized game forms that are strongly acyclic under coalitional improvements, and provided broad classes
of game forms that are "almost unrestricted acyclic," i.e.
restricted-acyclic under mild restrictions on voters'
actions. Other partial characterizations have been provided for acyclicity in complete information extensive-form
games [Boros et al., 2008; Andersson et al., 2010]. Some of this work is explained in more detail in the following
sections.
The study of classes of games (i.e. game forms with utilities) that are guaranteed to be acyclic or weakly acyclic
attracted much attention, in particular regarding the existence and properties of potential functions [Monderer and Shapley,
1996; Milchtaich, 1996; Fabrikant et al., 2010; Apt and Simon, 2012].
Strategic voting The notion of strategic voting has been highlighted in research on Social Choice as crucial to
understanding the relationship between preferences of a population and the final outcome of elections. In various
applications (ranging from political domains to artificial intelligence [AI]), the most widely used voting rule is
Plurality, in which each voter has one vote and the winner is the candidate who received the highest number of
votes. While it is known that no reasonable voting rule is completely immune to strategic behavior [Gibbard,
1973; Satterthwaite, 1975], Plurality has been shown to be particularly susceptible, both in theory [Saari, 1990;
Friedgut et al., 2011] and in practice [Forsythe et al., 1996]. This makes the analysis of any election campaign-
even one where the simple Plurality rule is used-a challenging task. As voters may speculate and counter-
speculate, it would be beneficial to have formal tools that would help us understand (and perhaps predict) the final
outcome.
In particular, natural tools for this task include the well-studied solution concepts developed for normal form
games, such as better/best responses, dominant strategies or different variants of equilibrium. Now, while vot-
ing settings are not commonly presented in this way, several natural formulations have been proposed in the
past [Dhillon and Lockwood, 2004; Chopra et al., 2004; Sertel and Sanver, 2004; Falik et al., 2012; Messner and Polborn,
2002]. These formulations are extremely simple for Plurality voting games, where voters have only a few available
ways to vote. Specifically, some of this previous work has been devoted to the analysis of solution concepts such
as elimination of dominated strategies [Dhillon and Lockwood, 2004] and strong equilibria [Sertel and Sanver,
2004]. There has been other multi-step voting procedures that have been proposed in the literature, such as iterated
majority vote [Airiau and Endriss, 2009] and extensive form games where voters vote one by one [Desmedt and Elkind,
2010].
In contrast to iterative voting, these models are inconsistent with the better-reply dynamics in normal
form games, and are analyzed via different techniques. A model more similar to ours was recently studied in
[Elkind et al., 2015], where voters can choose between voting truthfully, and manipulating under the assumption
that everyone else are truthful.
2
Convergence of better-reply dynamics in iterative voting for particular voting rules has been studied extensively
in the computational social choice literature. We summarize and compare these findings with ours in the concluding
section, and in particular in Table 1.
An important question in every model of strategic voting, including iterative voting, is whether the reached
equilibrium is good for the society according to various metrics. Branzei et al.[2013] showed bounds on the
dynamic price of anarchy, i.e. how far can the final outcome be from the initial truthful outcome. Other work used
simulations to show that iterative voting may improve the social welfare or Condorcet efficiency [Grandi et al.,
2013; Meir et al., 2014; Koolyk et al., 2016], but typically under the assumptions that voters use various heuristics.
Biased and sophisticated voting Some recent work on iterative voting deals with voters who are uncertain, truth-
biased, lazy-biased, bounded-rational, non-myopic, or apply some other restrictions and/or heuristics that diverge
from the standard notion of better-reply in games [Reijngoud and Endriss, 2012; Gohar, 2012; Grandi et al., 2013;
Obraztsova et al., 2013; Meir et al., 2014; Rabinovich et al., 2015; Obraztsova et al., 2015; Meir, 2015]. Although
the framework is suitable for studying such iterative dynamics as well, this paper deals exclusively with myopic
better-reply dynamics.1
1.2 Contribution and structure
The paper unfolds as follows. In Section 2, we define the iterative voting model within the more general framework
of game forms and acyclicity properties. In Section 3 we consider strong acyclicity, and settle an open question
regarding the existence of acyclic non-separable game forms. Section 4 focuses on order-free acyclicity of the
Plurality rule. Our main result in this section shows that to guarantee convergence, it is necessary and sufficient
that voters restrict their actions in a natural way that we term direct reply-meaning that a voter will only reassign
his vote to a candidate that will become a winner as a result. In Section 5, we use variations of Plurality to show
a strict separation between restricted acyclicity and weak acyclicity, thereby settling another open question. We
conclude in Section 6.
2 Preliminaries
We usually denote sets by uppercase letters (e.g., A, B, . . .), and vectors by bold letters (e.g., a = (a1, . . . , an)).
2.1 Voting rules and game forms
There is a set C of m alternatives (or candidates), and a set N of n strategic agents, or voters. A game form (also
called a voting rule) f allows each agent i ∈ N to select an action ai from a set of messages Ai. Thus the input to
f is a vector a = (a1, . . . , an) called an action profile. We also refer to ai as the vote of agent i in profile a. Then,
f chooses a winning alternative-i.e., it is a function f : A → C, where A = ×i∈N Ai. See Fig. 1 for examples.
A voting rule f is standard if Ai = A for all i, and A is either π(C) (the set of permutations over C)
or a coarsening of π(C). Thus most common voting rules except Approval are standard. Mixed strategies are
not allowed. The definitions in this section apply to all voting rules unless stated otherwise. For a permutation
P ∈ π(C), We denote by top(P ) the first element in P .
Plurality In the Plurality voting rule we have that A = C, and the winner is the candidate with the most votes.
We allow for a broader set of "Plurality game forms" by considering both weighted and fixed voters, and varying
the tie-breaking method. Each of the strategic voters i ∈ N has an integer weight wi ∈ N. In addition, there are n
"fixed voters" who do not play strategically or change their vote. The vector s ∈ Nm (called "initial score vector")
specifies the number of fixed votes for each candidate. Weights and initial scores are part of the game form.2
1We do consider however two standard ways to handle ties that slightly relax the better-reply definition. See Section 4.4.
2All of our results still hold if there are no fixed voters, but allowing fixed voters enables the introduction of simpler examples, and facilitates
some of the proofs, see Remark 4.1. For further discussion on fixed voters see [Elkind et al., 2015].
3
f1
a
b
c
a b
c
a a a
b
b
c
c
b
c
f2
a
b
c
a b
c
a a a
a b
b
c
a b
f3 x y
b
a
b
c
a
c
a
b
c
f4
a
b
c
z
y
x
w
ax ay az aw
bx by
bw
cw
cx cy
bz
cz
Figure 1: Four examples of game forms with two agents. f1 is a dictatorial game form with 3 candidates (the row
agent is the dictator). f2 is the Plurality voting rule with 3 candidates and lexicographic tie-breaking. f3 and f4
are non-standard game forms. In f3, A1 = C = {a, b, c}, A2 = {x, y}. Note that f4 is completely general (there
are 3 × 4 possible outcomes in C, one for each voting profile) and can represent any 3-by-4 game.
f P L
w,s
a
b
c
a
b
c
(14, 9, 3) {a}
(11, 12, 3) {b}
(11, 9, 6) {a}
(10, 13, 3) {b}
(7, 16, 3) {b}
(7, 13, 6) {b}
(10, 9, 7) {a}
(7, 12, 7) {b}
(7, 9, 10) {c}
Figure 2: A game form f P L
w,s, where N = {1, 2}, A1 = A2 = C = {a, b, c}, s = (7, 9, 3) and w = (3, 4) (i.e.,
voter 1 has weight 3 and voter 2 has weight 4). The table shows the final score vector s(a1,a2) for every joint action
of the two voters, and the respective winning candidate f P L
w,s(a1, a2) in curly brackets.
The final score of c for a given profile a ∈ An in the Plurality game form fw,s is the total weight of voters that
vote c. We denote the final score vector by ss,w,a (often just sa or s when the other parameters are clear from the
context), where s(c) = s(c) +Pi∈N :ai=c wi.
Thus the Plurality rule selects some candidate from W = argmaxc∈C ss,w,a(c), breaking ties according to
some specified method. The two primary variations we consider are f P L
s,w which breaks ties lexicographically, and
s,w which selects a winner from W uniformly at random. As with s, we omit the scripts w and s when they are
f P R
clear from the context.
For illustration, consider an example in Fig. 2, demonstrating a specific weighted Plurality game form with two
agents.
2.2 Incentives
Games are attained by adding either cardinal or ordinal utility to a game form. The linear order relation Qi ∈ π(C)
reflects the preferences of agent i. That is, i prefers c over c′ (denoted c ≻i c′) if (c, c′) ∈ Qi. The vector containing
the preferences of all n agents is called a preference profile, and is denoted by Q = (Q1, . . . , Qn). The game form
f , coupled with a preference profile Q, defines an ordinal utility normal form game G = hf, Qi with n agents,
where agent i prefers outcome f (a) over outcome f (a′) if f (a) ≻i f (a′). In standard game forms the action ai
may indicate the agent's preferences, hence their common identification with voting rules.
Improvement steps and equilibria Having defined a normal form game, we can now apply standard solution
concepts. Let G = hf, Qi be a game, and let a = (a−i, ai) be a joint action in G.
f (a−i, a′
context, only writing ai
We denote by a i→ a′ an individual improvement step, if (1) a, a′ differ only by the action of player i; and (2)
i) ≻i f (a−i, ai). We sometimes omit the actions of the other voters a−i when they are clear from the
i is an improvement
i is i's most preferred
i→ a′
i is called a best reply if a′
i. We denote by Ii(a) ⊆ Ai the set of actions a′
i). a i→ a′
i s.t. ai
i→ a′
step of agent i in a, and I(a) = Si∈NSa′
candidate in Ii(a).
i∈Ii(a)(a−i, a′
A joint action a is a (pure) Nash equilibrium (NE) in G if I(a) = ∅. That is, no agent can gain by changing
his vote, provided that others keep their strategies unchanged. A priori, a game with pure strategies does not have
to admit any NE.
4
(cid:10)f, Q1(cid:11)
* a
b
c
a
b
* c
{a} 3, 2
{b} 2, 1
{a} 3, 2
{b} 2, 1
{b} 2, 1
{b} 2, 1
* {a} 3, 2
{b} 2, 1
{c} 1, 3
Figure 3: A game G = (cid:10)f, Q1(cid:11), where f = f P L
w,s is as in Fig. 2, and Q1 is defined by a ≻1 b ≻1 c and
c ≻2 a ≻2 b. The table shows the ordinal utility of the outcome to each agent, where 3 means the best candidate.
Bold outcomes are the NE points. Here the truthful vote (marked with *) is also a NE.
(cid:10)f, Q2(cid:11)
* a
b
c
a
b
{a} 3, 1
{b} 1, 2
{a} 3, 1
{b} 1, 2
{b} 1, 2
{b} 1, 2
* c
* {a} 3, 1
{b} 1, 2
{c} 2, 3
Figure 4: This game has the same game form as in Fig. 2, and the preference profile Q2 is a ≻1 c ≻1 b and
c ≻2 b ≻2 a. In this case, the truthful vote a∗(Q2) is not a NE.
Now, observe that when f is a standard voting rule the preference profile Q induces a special joint action
i = top(Qi).
i equals (the coarsening of) Qi. E.g. in Plurality a∗
a∗ = a∗(Q), termed the truthful state, where a∗
We refer to f (a∗) as the truthful outcome of the game hf, Qi.
The truthful state may or may not be included in the NE points of the game, as can be seen from Tables 3 and 4
that demonstrate games that are induced by adding incentives to the game form shown in Fig. 2, and indicate the
truthful states and the NE points in these games.
2.3 Iterative Games
We consider natural dynamics in iterative games. Assume that agents start by announcing some initial profile a0,
and then proceed as follows: at each step t a single agent i may change his vote to a′
i ∈ Ii(at−1), resulting in a
new state (joint action) at = (at−1
i). The process ends when no agent has objections, and the outcome is set by
the last state.
−i , a′
Local improvement graphs and schedulers Any game G induces a directed graph whose vertices are all action
profiles (states) A, and edges are all local improvement steps [Young, 1993; Andersson et al., 2010]. The pure Nash
equilibria of G are all states with no outgoing edges. Since a state may have multiple outgoing edges (I(a) > 1),
we need to specify which one is selected in a given play.
A scheduler φ selects which edge is followed at state a at any step of the game [Apt and Simon, 2012]. The
scheduler can be decomposed into two parts, namely selecting an agent i to play (agent scheduler φN ), and se-
lecting an action in Ii(a) (action scheduler φA), where φ = (φN , φA). We note that a scheduler may or may not
depend on the history or other factors, but this does not affect any of our results.
Convergence and acyclicity Given a game G, an initial action profile a0 and a scheduler φ, we get a unique
(possibly infinite) path of steps.3 Also, it is immediate to see that the path is finite if and only if it reaches a Nash
finite.
Following [Monderer and Shapley, 1996; Milchtaich, 1996], a game G has the finite individual improvement
equilibrium (which is the last state in the path). We say that the triple(cid:10)G, a0, φ(cid:11) converges if the induced path is
property (we say that G is FIP), if(cid:10)G, a0, φ(cid:11) converges for any a0 and scheduler φ. Games that are FIP are also
known as acyclic games and as generalized ordinal potential games [Monderer and Shapley, 1996].
It is quite easy to see that not all Plurality games are FIP (see examples in Section 4). However, there are
alternative, weaker notions of acyclicity and convergence.
3By "step" we mean an individual improvement step, unless specified otherwise.
5
are known as weakly acyclic, or as φ-potential games [Apt and Simon, 2012].
• A game G is weakly-FIP if there is some scheduler φ such that(cid:10)G, a0, φ(cid:11) converges for any a0. Such games
• A game G is restricted-FIP if there is some action scheduler φA such that(cid:10)G, a0, (φN , φA)(cid:11) converges for
any a0 and φN [Kukushkin, 2011]. We term such games as order-free acyclic.
Intuitively, restricted FIP means that there is some restriction players can adopt s.t. convergence is guaranteed re-
gardless of the order in which they play. Kukushkin identifies a particular restriction of interest, namely restriction
to best-reply improvements, and defines the finite best-reply property (FBRP) and its weak and restricted analogs.
We emphasize that an action scheduler must select an action in Ii(a), if one exists. Thus restricted dynamics that
may disallow all available actions (as in [Gohar, 2012; Grandi et al., 2013]) do not fall under the definition of
restricted-FIP (but can be considered as separate dynamics).
We identify a different restriction, namely direct reply, that is well defined under the Plurality rule. For-
mally, a step a i→ a′ is a direct reply if f (a′) = a′
i, i.e., if i votes for the new winner (see labeled examples
in Section 4). Another rule where a natural direct reply exists is Veto, where a voter can veto the current win-
ner [Lev and Rosenschein, 2012].
φA is direct if it always selects a direct reply. We get the following definitions for a Plurality game G, where
FDRP stands for finite direct reply property:
• G is FDRP if(cid:10)G, a0, φ(cid:11) converges for any a0 and any direct φ.
• G is weakly-FDRP if there is a direct φ such that(cid:10)G, a0, φ(cid:11) converges for any a0.
• G is restricted-FDRP if there is a direct φA such that(cid:10)G, a0, (φN , φA)(cid:11) converges for any a0 and φN .
• FDBRP means that replies are both best and direct. Note that it is unique and thus cannot be further restricted.
Finally, a game form f has the X property (where X is any of the above versions of finite improvement) if hf, Qi
is X for all preference profiles Q ∈ (π(C))n. We have the following entailments, both for games and for game
forms. The third row is only relevant for Plurality/Veto.
FBRP
⇑
restricted-FBRP ⇒ weak-FBRP
⇓
⇓
ordinal potential ⇒ FIP ⇛ FDBRP ⇛ restricted-FIP ⇒ weak-FIP ⇒ pure Nash
exists
⇓
FDRP
⇑
⇑
exists
restricted-FDRP ⇒ weak-FDRP
Kukushkin notes that there are no known examples of game forms that are weak-FIP, but not restricted-FIP. We
settle this question later in Section 5.2.
Convergence from the truth We say that a game G is FIP from state a if hG, a, φi converges for any φ. Clearly
a game is FIP iff it is FIP from a for any a ∈ An. The definitions for other all other notions of finite improvement
properties are analogous.
We are particularly interested in convergence from the truthful state a∗. This is since: a. it is rather plausible to
assume that agents will start by voting truthfully, especially when not sure about others' preferences; and b. even
with complete information, they may be inclined to start truthfully, as they can always later change their vote.
Heuristic voting Much work on iterative voting deals with heuristics, rather than best- or better-replies. Strong,
Restricted, and Weak convergence properties can be defined the same way, where the only difference is the
way we define Ii(a) (i.e., all steps that are allowed for agent i at state a by the considered heuristics). For
example, truth-bias assumes that if a voter does not have any local improvement step, she reverts to her truth-
ful vote [Obraztsova et al., 2013]. Some heuristics are already restricted to a single action (for example, "k-
pragmatist" [Grandi et al., 2013]). In these cases the only meaningful distinction is between FIP and weak-FIP. In
this paper we do not consider heuristic voting.
6
3 Strong Acyclicity
An ordinal potential is a function that strictly increases if and only if some agent plays a better-reply [Monderer and Shapley,
1996]. A generalized ordinal potential is a function that strictly increases with every better-reply, but may also
increase with other steps. Clearly, a game is FIP if and only if it has a generalized potential (by a topological sort
of the better-reply graph).
Theorem 1 (Kukushkin [2011]). A game form f guarantees an ordinal potential (i.e. every derived game has an
ordinal potential function) if and only if f is a dictatorship.
We emphasize that this result does not preclude the existence of other game forms with FIP (generalized ordinal
potential). Indeed, Kukushkin provides a partial characterization of FIP game forms. For example, a rule where
there is a linear order L over C, and the winner is the first candidate according to L that is top-ranked by at least
one voter.
A game form f is called "separable" [Kukushkin, 2011] if there are mappings gi : Ai → C for i ∈ N s.t. for
all a ∈ A, f (a) ∈ {g1(a1), g2(a2), . . . , gn(an)}. That is, the vote of each voter is mapped to a single candidate
via some function gi, and the outcome is always one of the candidates in the range. Examples of separable rules
include Plurality and dictatorial rules, in both of which gi are the identity functions.
Conjecture 2 (Kukushkin [2011]). Any FIP game form is separable.
Some weaker variations of this conjecture have been proved. In particular, for game forms with finite coali-
tional improvement property [Kukushkin, 2011], and for FIP game forms with n = 2 voters [Boros et al., 2010]
(separable game forms are called "assignable" there). We next show that for sufficiently large n, there are non-
separable FIP game forms, thereby refuting the conjecture. Our proof uses the probabilistic method: we sample a
game form from some space, and prove that with positive probability it must be non-separable and FIP.
Theorem 3. For any n ≥ 20, there is a non-separable game form fn s.t. fn is FIP.
Proof. Let C = {a1, . . . , a2n}∪{z}. Let Ai = {x, y} for each voter. Thus fn is a function from the n dimensional
binary cube B = {x, y}n to C. We select 2n profiles a1, . . . , a2n uniformly at random, i.i.d. from B (allowing
repetitions), and define fn(aj ) = aj for all j ∈ {1, . . . , 2n}. For all other 2n − 2n profiles we define fn(a) = z.
For any two profiles a, a′, let d(a, a′) be the number of voters that disagree in a, a′ (the Manhattan distance
on the cube). Let B ⊆ B be all 2n profiles whose outcome is not z. For j, j′ ≤ 2n, denote by pj,j ′ the probability
) ≤ 2, and by Xj,j ′ the corresponding indicator random variable. Since both of aj, aj ′ were sampled
that d(aj , aj ′
uniformly i.i.d., and there are less than n2 profiles within distance 2 from aj, we get that pj,j ′ ≤ n2
2n .
Next, by the union bound,
P r(∃Xj,j ′ = 1) ≤ Xj≤2n,j ′≤2n
P r(Xj,j ′ ) = Xj≤2n,j ′≤2n
pj,j ′ ≤ (2n)2 n2
2n =
4n2
2n ,
which is strictly less than 1 for n ≥ 20. Thus w.p. > 0 we get Xj,j ′ = 0 for all j, j′. In particular there is at least
one such game form f ∗
n where Xj,j ′ = 0 for all j, j′. We argue that f ∗
n is both FIP and non-separable.4
Assume towards a contradiction that there is some cycle of better-replies in f ∗
n. Then there must be a path
containing at least 3 distinct outcomes, and thus at least 2 profiles from B. Denote these profiles by a, b. Since
Xj,j ′ = 0 for all j, j′, we have that any path between a and b is of length at least 3, and that the path must contain
at least two consequent states whose outcome is z. This path cannot be a better-reply path, since a better reply
must change the outcome. Hence we get a contradiction and f ∗
n is FIP.
Finally, note that since Xj,j ′ = 0 for all j, j′, in particular aj are all distinct profiles, and thus f ∗
n has 2n + 1 >
Pi≤n Ai possible outcomes. In contrast, for any separable rule f the size of the range of f is at mostPi≤n Ai,
since f (a) = gi(b) for some i ∈ N and b ∈ Ai. This means that f ∗
n is non-separable.
4Using the Hamming error-correcting code [Hamming, 1950], it is in fact possible to explicitly construct f ∗
n for as few as n = 7 voters.
The rest of the proof remains the same.
7
For most common voting rules, separable or not, it is easy to find examples where some cycles occur. Thus one
should focus on the weaker notions of convergence discussed in Section 1, which is what we do in the remainder
of the paper.
4 Order-Free Acyclicity: Plurality
Improvement steps in Plurality Recall that along a given path, at ∈ An = C n denotes the voting profile at
time t. We denote by st = sat the score vector at time t; by cwt = f P L(at) the candidate that wins at time t; and
by swt = st(cwt) the highest score at time t (including tie-breaking if it applies).
i→ at
steps into the following types (an example of such a step appears in parentheses):
Suppose that agent i has an improvement step (a.k.a. better reply) at−1
i at time t. We classify all possible
i
Type 1. from at−1
i
6= cwt−1 to at
i = cwt ; (step 1 in Ex.6a.)
Type 2. from at−1
i = cwt−1 to at
i = cwt ; (step 1 in Ex.6b.),
Type 3. from at−1
i = cwt−1 to at
i 6= cwt ; (step 2 in Ex.6a.)
Note that steps of type 1 and 2 are direct, whereas type 3 steps are indirect.
4.1 Lexicographic Tie-Breaking
In this section we assume that ties are broken lexicographically. Given some score vector s, we denote by s(c) ∈ R
the score of c ∈ C that includes the lexicographic tie-breaking component. One way to formally define it is by
setting s(c) = s(c) + 1
m+1 (m − L(c)), where L(c) is the lexicographic index of candidate c. However the only
important property of s is that s(c) > s(c′) if either s(c) > s(c′) or the score is equal and c has a higher priority
(lower index) than c′.
Thus for Plurality with lexicographic tie-breaking, a given weight vector w and a given initial score vector s,
we denote the outcome by
As with s, we omit the scripts w, s and P L when they are clear from the context.
f P L
s,w(a) = argmaxc∈C ss,w,a(c).
Lemma 4. Consider a gameDf P L
a direct best reply at state at−1.
w,s, QE. If there exists a better reply for a given agent i at state at−1, then i has
The proof is trivial under lexicographic tie-breaking, by letting i vote for her most preferred candidate among
all better replies. In this case the direct best reply is also unique.
One implication of the lemma is that it is justified and natural to restrict our discussion to direct replies and
focus on FDRP, as w.l.o.g. a voter always has a direct reply that is at least as good as any other reply.
Unweighted Voters Suppose all voters have unit weight. We start with our main result for this section.
Theorem 5. f P L
particular, Plurality is order-free acyclic.
s
is FDRP. Moreover, any path of direct replies will converge after at most m2n2 steps.
In
This extends a weaker version of the theorem that appeared in the preliminary version of this paper [Meir et al.,
2010], which only showed FDBRP. The bound on the number of direct-best-reply steps was recently improved to
O(mn) in [Reyhani and Wilson, 2012, Theorem 5.4].
Proof. By our restriction to direct replies, there can only be moves of types 1 and 2. We first consider moves of
type 1, and inductively prove two invariants that yield a bound on the total number of such moves. Next, we bound
the number of moves of type 2 by a given voter between any of his moves of type 1, which completes the proof.
8
i
a
b
cwt−1
st(b)
i
s
s
st(a)
b = cwt
a
Figure 5: An illustration of a type 1 move. Tie-breaking is in favor of the left most candidate.
Consider time t − 1 and denote the score of the current winner (including tie-breaking) by ¯s = swt−1. Suppose
that a move a i→ b of type 1 occurs at time t: that is, a 6= cwt−1 and b = cwt. We then have (see Fig. 5):
st(b) = swt ≥ swt−1 = ¯s ≥ st−1(a) = st(a) + 1.
(1)
We claim that at any later time t′ ≥ t the following two invariants hold:
I. Either there is a candidate c 6= a whose score is at least ¯s + 1, or there are at least two candidates c, c′ 6= a
whose score is at least ¯s. In particular it holds in either case that swt′
≥ ¯s.
II. The score of a does not increase: st′
(a) ≤ st(a).
Note that this, coupled with Eq. (1), implies that candidate a will never win again, as its score will stay strictly
below ¯s, and there will always be a candidate with a score of at least ¯s.
We now prove both invariants by induction on time t′. In the base case t′ = t, (I) holds since both cwt−1 and b
have a score of at least ¯s, and (II) holds trivially.
Assume by induction that both invariants hold until time t′ − 1, and consider step t′ by voter j. Due to (I), we
either have at least two candidates whose score is at least ¯s, or a candidate with a score of at least ¯s + 1. Due to
(II) and Eq. (1) we have that st′ (a) ≤ st(a) < ¯s − 1.
Let d j→ d′ be the step at time t′ by voter j (that is, d = at′−1
j ). We first argue that d′ 6= a: by adding
the vote of j to a its score will still be strictly less than ¯s, whereas by removing a vote from any other candidate d,
we still have at least one candidate c with score at least ¯s. Thus a cannot be a direct reply for any voter j, and (II)
still holds after step t′.
, d′ = at′
j
It remains to show that (I) holds. If d is not one of the candidates in (I) with the score of at least ¯s at time t′ − 1,
then their score does not decrease after step t′, and we are done. Otherwise, we divide into the following cases:
1. At t′ − 1, d is the (only) candidate with a score of at least ¯s + 1.
2. At t′ − 1, candidates c, c′ have scores of at least ¯s, and d is one of them (w.l.o.g. d = c).
In the first case, st′
(d) = st−1(d) − 1 ≥ ¯s + 1 − 1 = ¯s, whereas st′
(d′) > st′
(d) ≥ ¯s. Thus both d, d′ have
scores of at least ¯s at time t′, as required. In the second case, since only c = d can lose votes, then if d′ 6= c′,
st′
(d′) = swt′
≥ st′
(c′) = st′−1(c′) ≥ ¯s,
and thus both c′, d′ have scores of at least ¯s at time t, as required. If d′ = c′, then and thus both c′, d′ have scores
of at least ¯s at time t, as required. If d′ = c′, then
st′
(d′) = st′−1(d′) + 1 = st′−1(c′) + 1 ≥ ¯s + 1,
that is, d′ has a score of at least ¯s + 1, as required.
9
Next, we demonstrate that invariants (I) and (II) supply us with a polynomial bound on the rate of convergence.
Indeed, as we mentioned before, at every step of type 1, at least one candidate is ruled out permanently, and there
are at most n times that a vote can be withdrawn from a given candidate. Also note that, since a type 2 move by a
, each voter can make at most m − 1 type 2 moves before making a
given voter i implies that he prefers at
move of type 1. Hence, there are in total at most m2n2 steps until convergence.
i to at−1
i
Furthermore, it is easy to show that if all voters start from the truthful state then type 2 moves never occur.
Thus, the score of the winner never decreases, and the game converges in at most mn steps.
Next, we show that the restriction to direct replies is necessary to guarantee convergence, whereas a restriction
to best replies is insufficient.
Proposition 6. f P L is not FBRP, even from the truthful state. Moreover, there are: (a) a counterexample with
two strategic agents and an arbitrary initial state; (b) a counterexample with three strategic agents and a truthful
initial vote.
Remark 4.1. In this example and in others throughout the paper we use an initial score vector s. However, this is
w.l.o.g. since we could replace s with additional voters that do not participate in the cycle. Initial scores are only
useful to construct examples that are simpler and/or with fewer strategic agents. This holds for all negative results
in the paper.5 For positive results, we have to show convergence for every initial scores s.
Example 6a. C = {a, b, c}. We have a single fixed voter voting for a, thus s = (1, 0, 0). The preference profile is
defined as a ≻1 b ≻1 c, c ≻2 b ≻2 a. The following cycle consists of better replies (the vector denotes the votes
(a1, a2) at time t, the winner appears in curly brackets):
(b, c){a} 2→ (b, b){b} 1→ (c, b){a} 2→ (c, c){c} 1→ (b, c).
Note that all steps are best-replies, but the steps of agent 1 are indirect.
♦
Example 6b. C = {a, b, c, d}. Candidates a, b, and c have 2 fixed voters each, thus s = (2, 2, 2, 0). We use
3 agents with the following preferences: d ≻1 a ≻1 b ≻1 c, c ≻2 b ≻2 a ≻2 d and d ≻3 a ≻3 b ≻3 c.
Starting from the truthful state (d, c, d) the agents can make the following two improvement steps, which are direct
best-replies (showing only the outcome s and the winner): (2, 2, 3, 2){c} 1→ (2, 3, 3, 1){b} 3→ (3, 3, 3, 0){a},
after which agents 1 and 2 repeat the cycle shown in (6a).
♦
Thus for the non-weighted lexicographic case Theorem 5 and Proposition 6 provide a clear-cut rule: direct
replies guarantee convergence, whereas convergence is not guaranteed under other restrictions such as best reply
or initial truthful vote. However, as the following section demonstrates, in the presence of weighted agents even
direct replies may no longer converge.
Weighted Voters Next, we show that if the voters may have non-identical weights, then convergence to equilib-
rium is not guaranteed even if they start from the truthful state and use direct best replies.
Proposition 7. There is f P L
w that is not restricted-FDRP, even from the truthful state.
Example 7. The initial fixed score of candidates {a, b, c, d} is s = (0, 1, 2, 3). The weight of each voter i ∈ {1, 2, 3}
is i. The preference profile is as follows: c ≻1 d ≻1 b ≻1 a, b ≻2 c ≻2 a ≻2 d, and a ≻3 b ≻3 c ≻3 d. We
omit the rest of the proof. The initial truthful profile is thus a0 = (c, b, a), which results in the score vector
s0 = (3, 3, 3, 3) where a is the winner.
votes:
scores:
(c, b, a)
(3, 3, 3, 3){a}
↑ 3
(c, b, b)
1→
2←
(d, b, a)
(3, 3, 2, 4){d}
(c, c, b)
2→
1←
(d, c, a)
(3, 1, 4, 4){c}
↓ 3
(d, c, b)
(0, 6, 3, 3){b}
(0, 4, 5, 3){c}
(0, 4, 4, 4){b}
5Note that the remark does no longer hold if s is used to construct a counter example for weak-FIP. However we use no such examples in
this paper.
10
Our example shows a cycle of direct responses. Note that at every step there is only one direct reply available to
the agent, thus it is not possible to eliminate the cycle by further restricting the action scheduler.
♦
If there are only two weighted voters (and possibly other fixed voters), either restriction to direct reply or to a
truthful initial state is sufficient to guarantee convergence.
Theorem 8. f P L
s,w is FDRP for n = 2.
Proof. Clearly, in one of the two first states, the agents vote for distinct candidates. At any later state, they must
continue voting for distinct candidates, as every step must change the winner, and the other voter is always voting
for the current winner. This means that the score of the winner strictly increases with every step (possibly except
the first one).
Theorem 9. f P L
s,w is FIP from the truth for n = 2.
Proof. We show that the score of the winner can only increase. This clearly holds in the first step, which must be
of type 1. Once again, we have that the two agents always vote for different candidates, and thus only steps that
increase the score can change the identity of the winner.
Thus in either case convergence is guaranteed after at most 2m steps.
It remains an open question whether there is any restriction on better replies that guarantees order-free acyclic-
ity in weighted games, i.e.
w is restricted-FIP for n > 2. However Prop. 7 shows that if such restricted
dynamic exists, it must make use of indirect replies, which is rather unnatural. We thus conjecture that such
restricted dynamics does not exist.
if f P L
4.2 Arbitrary tie-breaking
Lev and Rosenschein [2012] showed that for any positional scoring rule (including Plurality), we can assign some
(deterministic) tie breaking rule, so that the resulting voting rule may contains cycles. For any positional scoring
rule fα with score vector α, denote by f LR
α the same rule with the Lev-Rosenschein tie-breaking.
Proposition 10 (Theorem 1 in [Lev and Rosenschein, 2012]). f LR
even from the truth. In particular, Plurality with the Lev-Rosenschein tie-breaking (f P LR) is not FBRP.
is not FBRP for any α, even for n = 2, and
α
In fact, a slight modification of their example (switching a and b in voter 2's preferences) yields the following:
Proposition 11. f P LR is not restricted-FIP, even for n = 2, and even from the truth.
4.3 Randomized tie-breaking
Compared to the previously considered deterministic rule, randomized tie-breaking has the advantage of being
neutral-no specific candidate or voter is preferred over another. Formally, the game form f P R
s,w maps any state
a ∈ An to the set argmaxc∈C ss,w,a(c). Since under randomized tie-breaking there are multiple winners, let
W t = f P R(at) ⊆ C denote the set of winners at time t.6 We define a direct reply at−1
i as one where
i ∈ W t.
at
i→ at
i
If ties are broken randomly, ≻i does not induce a complete order over outcomes. For instance, the order
a ≻i b ≻i c does not determine if i will prefer {b} over {a, c}. However, we can naturally extend Qi to a partial
preference order over subsets. There are several standard extensions, using the following axioms:7
K (Kelly [Kelly, 1977]): (1) (∀a ∈ X, b ∈ W, a ≻i b) ⇒ X ≻i W ; (2) (∀a ∈ X, b ∈ W, a (cid:23)i b) ⇒ X (cid:23)i W ;
G (Gardenfors [Gardenfors, 1976]): (∀b ∈ W, a ≻i b) ⇒ {a} ≻i ({a} ∪ W ) ≻i W ;
6This is a slight abuse of the notation we introduce in the beginning, where we defined the set of possible outcomes of f to be C. Here we
allow any W ∈ 2C \ {∅} as a possible outcome.
7We thank an anonymous reviewer for the references.
11
R (Responsiveness [Roth, 1985]): a ≻i b ⇐⇒ ∀W ⊆ C \ {a, b}, ({a} ∪ W ) ≻i ({b} ∪ W ) .
The axioms reflect various beliefs a rational voter may have on the tie-breaking procedure: the K axiom reflects
no assumptions whatsoever; The K+G axioms are consistent with tie-breaking according to a fixed and unknown
order [Geist and Endriss, 2011]; and K+G+R axioms are consistent with random tie-breaking with equal probabil-
ities (see Lemma 15 and Prop. 20). In this section we assume all axioms hold, however our results do not depend
on these interpretations, and we do not specify the voter's preferences in cases not covered by the above axioms.
Under strict preferences, it also holds that G entails K [Endriss, 2013]. We can also define "weak" variants G2 and
R2 for axioms G and R, by replacing all strict relations with weak ones, however as long as we restrict attention to
strict preferences over elements the weak variants are not required.
For the following lemma we only need Axiom K, i.e. it does not depend on the voter's tie-breaking assumptions.
Lemma 12. If there exists a better-reply in f P R
s,w for agent i at state at−1, then i has a direct best-reply.
i
i
Proof. Suppose there is a better reply at−1
arbitrary best reply. Let W = f RP
that at−1
reply (for the lexicographic case this follows immediately from W = {a′} and f P L(at−1
i→ b at time t − 1. As some best reply always exists, denote by b′ an
−i , b′), and let a′ be the most preferred candidate of i in W . Then we argue
i→ a′ is a direct best reply of i. Since a′ is a direct reply by definition, it is left to show that a′ is a best
s,w (at−1
−i , a′) = W = {a′}).
If b′ is a direct reply then b′ = a′ and we are done. Thus assume that b′ is not a direct reply from at−1
i→.
Then b′ /∈ W . By voting for a′ ∈ W , we get that f RP
−i , a′) = {a′}, i.e., a′ remains the unique winner. If
W = 1 then we are done as in the lexicographic case. Otherwise we apply Axiom K2 with X = {a′}, and get
that a′ (cid:23)i W . That is,
s,w (at−1
i
s,w (at−1
f RP
−i , a′) = {a′} (cid:23)i W = f RP
s,w (at−1
−i , b′),
which means that a′ is also a best-reply.
With weighted votes and and random tie-breaking, there may not be any pure Nash equilibrium at all [Meir et al.,
2010]. We therefore restrict attention in the rest of this section to unweighted votes.
Proposition 13. f P R is not FIP.
Example 13. C = {a, b, c} with initial score s = (0, 1, 0). The initial state is a0 = (a, a, b)-that is, s(a0) =
(2, 2, 0) and the outcome is the winner set {a, b}. The preferences are a ≻1 c ≻1 b, b ≻2 a ≻2 c and c ≻3 b ≻3 a.
We get the following cyclic sequence:
(2, 2, 0){a, b}
2→ (1, 2, 1){b}
1→ (0, 2, 2){b, c}
↑ 3
↓ 3
(1, 2, 1){b}
1← (2, 1, 1){a}
2← (1, 1, 2){c}
We emphasize that each step is justified as a better reply by either Axiom K or Axiom G. E.g, in the step of agent 2
in the top row, agent 2 prefers b ≻2 a, and thus b ≻2 {a, b} by Axiom G. This will be used later in Section 4.4. ♦
Theorem 14. f P R
s
is FBRP from the truth.
Proof. We denote the sets of winners and runnerups at time t as W t = f RP (at); Rt = {c : st(c) = swt − 1}. We
will show by induction that at any step at−1 i→ at:
1. W t ∪ Rt ⊆ W t−1 ∪ Rt−1 (i.e., candidates not in W t ∪ Rt will not be selected by any agent at a later time).
2. at
i is the most preferred candidate for i in W t ∪ Rt (in particular, a best reply is a direct reply).
3. at−1
i ≻i at
i (in the terminology of [Meir et al., 2014], this is a compromise step).
12
Since each voter can make at most m − 1 compromise steps, convergence is guaranteed within nm steps.
Assume that for some t ≥ 1, all of the above holds for any t′ < t (so we prove the base case together with the
other cases). Since a0 is truthful, the first step of any voter is always a compromise move. If i had already moved
at some previous time t′ < t, then at′
i
is most preferred in W t′
∪ Rt′.
i
By induction, a = at−1
is the most preferred candidate in some C′ that contains W t−1 ∪ Rt−1 (C′ = C
∪ Rt′ at any other step). Let x and y be i's most preferred candidates in W t−1
in i's first step, and C′ = W t′
i. Each of a or a′ may belong to W t−1, to Rt−1,
and in Rt−1, respectively, and denote the best reply by a′ = at
or to neither set. This means there are 3X3=9 cases to check. Fortunately, we can show that some of this cases
immediately lead to a contradiction, and in the other cases all invariants 1-3 will hold after step t.
Consider first the case a ∈ W t−1. Since a is most preferred in C′, it is strictly more preferred than any other
candidate in W t−1 or in Rt−1 (i.e., a = x). Thus if a′ ∈ W t−1 we get W t = {a′} ≺i W t−1 by Axiom G. If
a′ ∈ Rt−1 we get W t = (W t−1 \ {a}) ∪ {a′} ≺i W t−1 by Axiom R. In either case this is not an improvement
step for voter i.
Next, suppose a /∈ W t−1. We further split to subcases based on a′.
• If a′ ∈ W t−1 then f (a−i, a′) = {a′}. Then a′ = x, as otherwise f (a−i, x) = {x} ≻i {a′}, and i is
strictly better off by voting for x. This entails W t = {x}, Rt = W t−1 \ {x} so all invariants 1-3 hold: (1)
W t−1 = W t ∪ Rt; (2) follows from (1) since a′ = x is the most preferred in W t−1; and (3) follows from
(1) since a = at−1
is the most preferred in C′, and a′ ∈ C′.
i
• If a′ ∈ Rt−1 then f (a−i, a′) = {a′} ∪ W t−1. Then a′ = y, as otherwise f (a−i, y) = {y} ∪ W t−1 ≻i
{a′}∪W t−1 by Axiom R, which means i is strictly better off by voting for y. This entails W t = {y}∪W t−1,
Rt = Rt−1 \ {y}. We also get that a′ = y ≻i x or else x would have been a strictly better reply. Thus all
invariants 1-3 hold: (1) W t = W t−1 ∪ {y} ⊆ W t−1 ∪ Rt−1 and Rt = Rt \ {y}; (2) follows from (1) since
i = y is most preferred in Rt−1 and strictly preferred to x; (3) follows from (1) as in the previous
a′ = at
case.
• If a′ /∈ W t−1 ∪ Rt−1, then W t = f (a−i, a′) = W t−1. The outcome does not change so this cannot be an
improvement step for i.
Cardinal utilities A (cardinal) utility function is a mapping of candidates to real numbers u : C → R, where
ui(c) ∈ R is the utility of candidate c to agent i. We say that u is consistent with a preference relation Qi if
u(c) > u(c′) ⇔ c ≻i c′. The definition of cardinal utility naturally extends to multiple winners by setting
ui(W ) = 1
Lemma 15. Consider any cardinal utility function u and the partial preference order Q it induces on subsets by
random tie-breaking. Q holds Axioms K+G+R.
W Pc∈W ui(c) for any subset W ⊆ C.8
The proof is rather straight-forward, and is deferred to the appendix.
Proposition 16. f P R is not FIP from the truth.
Example 16. We use 5 candidates with initial score s = (1, 1, 2, 0, 0), and 2 agents with utilities u1 = (5, 3, 2, 8, 0)
and u2 = (4, 2, 5, 0, 8). In particular, {b, c} ≻1 c, {a, c} ≻1 {a, b, c}, and {a, b, c} ≻2 {b, c}, c ≻2 {a, c}, and the
following cycle occurs: (1, 1, 2, 1, 1){c} 1→ (1, 2, 2, 0, 1){b, c} 2→ (2, 2, 2, 0, 0){a, b, c} 1→ (2, 1, 2, 1, 0){a, c} 2→
(1, 1, 2, 1, 1){c}.
♦
Finally, in contrast to the lexicographic case, convergence is no longer guaranteed if agents start from an
arbitrary profile of votes, or are allowed to use direct-replies that are not best-replies. The following example
shows that in the randomized tie-breaking setting even direct best reply dynamics may have cycles, albeit for
specific utility scales.
8One interpretation is that we randomize the final winner from the set W , and hence the term randomized tie-breaking. For a thorough
discussion of cardinal and ordinal utilities in normal form games, see [Borgers, 1993].
13
Proposition 17. f P R is not restricted-FIP.
Example 17. There are 4 candidates {a, b, c, x} and 3 agents with utilities u1 = (7, 3, 0, 4), u2 = (0, 7, 3, 4)
and u3 = (3, 0, 7, 4).
In particular, the following preference relations hold: a ≻1 {a, b} ≻1 x ≻1 {a, c};
b ≻2 {b, c} ≻2 x ≻2 {a, b}; and c ≻3 {a, c} ≻3 x ≻3 {b, c}.
Consider the initial state a0 = (a, b, x) with s(a0) = (1, 1, 0, 1) and the outcome {a, b, x}. We have the
following cycle where every step is the unique reply of the playing agent.
(1, 1, 0, 1){a, b, x}
2→ (1, 0, 0, 2){x}
3→ (1, 0, 1, 1){a, x, c}
↑ 1
↓ 1
(0, 1, 0, 2){x}
3← (0, 1, 1, 1){x, b, c}
2← (0, 0, 1, 2){x}
♦
Proposition 18. f P R is not FDRP even from the truth.
Example 18. We take the game from Ex. 17, and add for each voter i ∈ {1, 2, 3} a candidate di, s.t. ui(di) =
8, ui(dj) = j for j 6= i. We also add an initial score of 3 to each of the candidates {a, b, c, x}. Voter 3 moves first
to a1
3 = x, which is a direct reply. Then voters 1 and 2 move to their best replies a, b, respectively. Now the cycle
continues as in Ex. 17.
♦
4.4 Stochastic Dominance and Local Dominance
While assigning cardinal utilities is one way to deal with ties, it is sometimes preferable not to assume a particular
cardinal utility scale. Denote by f P (a) ⊆ C the subset of candidates with maximal Plurality score, before any tie-
breaking takes place. We can still derive a well-defined dynamics from any partial order over subsets of candidates,
by assuming that a voter performs a better-response step if she strictly prefer the new outcome, and otherwise (if
the new outcome is same, worse, or incomparable) she does not move.
One example of such a partial order is stochastic dominance (SD), which was applied to tie-breaking by
[Reyhani and Wilson, 2012]. A different partial order is implied by local dominance (LD) which was defined for
voting with uncertainty about the outcome [Conitzer et al., 2011; Meir et al., 2014], when uncertainty is regarding
the tie breaking. We show how convergence results for LD/SD dynamics fit with other results.
Stochastic dominance Reyhani and Wilson assume that ties are broken uniformly at random, and that a voter
will only perform a step that stochastically dominates the current winner(s), if such exists.
Theorem 19 (Theorem 5.7 in [Reyhani and Wilson, 2012]). Plurality with stochastic dominance tie-breaking is
FDBRP.
We can show the following (see appendix):
Proposition 20. A step a i→ a′ is a better-response under random tie-breaking and stochastic dominance, if and
only if f P (a′) ≻i f P (a) is entailed by Qi, Axioms K+G+R, and transitivity.
In other words, while Theorem 14 allowed any moves consistent with the axioms, SD allows only moves that
follow from the axioms, and explicitly forbid any other step. Thus it is more restricted than expected-utility based
randomized tie-breaking.
Since any SD step is also a better-reply under any cardinal utility scale, any strong or restricted convergence
result for the latter applies to the former, but not vice-versa.
14
Local dominance Suppose that there are several candidates with maximal score. A voter may consider all of
them as "perhaps winners," without specifying how the actual winner is selected. If the voter is concerned about
making a move that will leave her worse off, she will only make moves that will improve her utility with certainty,
i.e.
that dominates her current action (where possible worlds are all strict tie-breaking orders) [Conitzer et al.,
2011; Meir et al., 2014; Meir, 2015].9
Theorem 21 (Theorem 11 in the full version of [Meir, 2015]). Plurality with Local-Dominance tie-breaking is
FDRP.
To see how this compares with other convergence results, we need the following proposition (see appendix).
Proposition 22. A step a i→ a′ is a better-response under unknown tie-breaking and local dominance, if and only
if f P (a′) ≻i f P (a) is entailed by Qi, Axioms K+G, and transitivity.
Note that since Axioms K+G+R include K+G, any LD step is also an SD step, so a restriction to LD can only
eliminate cycles. Thus FBDRP follows from Theorem 19. We note that with either SD or LD tie-breaking there
may be new stable states that are not Nash-equilibria. Even so, an analysis of Ex. 13 shows that all steps are
entailed by Axioms K+G (and thus by Axioms K+G+R). Thus neither game form is FIP.
What if we assume that voters are even more risk-averse and only follow steps that are better-replies by Ax-
iom K? Then it is easy to see that only moves to a more-preferred candidate can be better-replies (any move to or
from a tie cannot follow from Axiom K and is thus forbidden), which means there are trivially no cycles.
5 Weak Acyclicity
Except for Plurality and Veto, convergence is not guaranteed even under restrictions on the action scheduler and the
initial state. In contrast, simulations [Grandi et al., 2013; Meir et al., 2014; Koolyk et al., 2016] show that iterative
voting almost always converges even when this is not guaranteed by theory. We believe that weak acyclicity is an
important part of the explanation to this gap.
5.1 Plurality with Random tie-breaking
We have seen in Section 4 that while f P R is FDRP from the truthful initial state, this is no longer true from
arbitrary states, and in fact f P R is not restricted-FIP under any action scheduler. Our main theorem in this section
shows that under a certain scheduler (of agents+actions), convergence is guaranteed from any state. Further, this
still holds if actions are restricted to direct-replies.
Lemma 23. Consider any game G = (cid:10)f P R
s
x, y s.t. s0(x) ≥ s0(y) ≥ s0(a∗) + 2. Then for any sequence of direct replies, a∗ /∈ f (at).
, Q(cid:11). Consider some candidate a∗, and suppose that in a0, there are
Proof. We show that at any time t ≥ 0 there are xt, yt s.t. s0(x), s0(y) ≥ s0(a∗) + 2. For t = 0 this holds for
xt = x, yt = y. Assume by induction that the premise holds for at−1. Then there are two cases:
1. f (at−1) ≥ 2. Then since step t must be a direct reply, it must be to some candidate z with st−1(z) ≥
swt−1 − 1. Also, either xt−1 or yt−1 did not lose votes (w.l.o.g. xt−1). Thus st(x), st(z) ≥ swt−1 ≥
st−1(a∗) + 2 ≥ st(a∗) + 2.
2. f (at−1) = 1. Then suppose f (at−1) = {xt−1}, and we have that swt−1 ≥ st−1(a∗) + 3. The next step is
z where either st−1(z) = swt−1 − 1 (and then we conclude as in case 1), or st−1(z) = swt−1 − 2 and xt−1
loses 1 vote. In the latter case, st(xt−1) = st(z) = swt−1 − 1 ≥ st−1(a∗) + 2 ≥ st(a∗) + 2.
Theorem 24. f P R
s
is weak-FDRP.
9[Meir et al., 2014; Meir, 2015] consider more general uncertainty over candidates' score, and [Conitzer et al., 2011] considers arbitrary
information sets.
15
Proof. Consider a game G = (cid:10)f P R
reachable from a via paths of direct replies. Let B = B(a0), and assume towards a contradiction that B does not
contain a Nash equilibrium. For every b ∈ B, let C(b) = {c ∈ C : ∃a ∈ B(b) ∧ c ∈ f (a)}, i.e. all candidates
that are winners in some state reachable from b.
, Q(cid:11), and an initial state a0. For a state a, denote by B(a) ⊆ An all states
s
For any b ∈ B(a0), define a game Gb by taking G and eliminating all candidates not in C(b). Since we only
consider direct replies, for any a ∈ B(b), the set of outgoing edges I(a) is the same in G and in Gb (as any direct
reply must be to candidate in C(b)). Thus by our assumption, the set B(b) in game Gb does not contain an NE.
For any b ∈ B(a0), let b∗ be the truthful state of game Gb, and let T (b) ⊆ N be the set of agents who are
i .
truthful in b. That is, i ∈ T (b) if bi = b∗
Let b0 be some state b ∈ B(a0) s.t. T (b) is maximal, and let T 0 = T (b0). If T 0 = n then b0 is the truthful
state of Gb0, and thus by Theorem 14 all best-reply paths from b0 in Gb0 lead to an NE, in contradiction to B(b0)
not containing any NE. Thus T 0 < n. We will prove that there is a path from b0 to a state b′ s.t. T (b′) > T 0.
i at state b0. We
Let i /∈ T (b0) (must exist by the previous paragraph). Consider the score of candidate b∗
divide into 5 cases. All scores specified below are in the game Gb0.
Case 1. f (b0) > 1 and b∗
i is one of several winners). Then consider the step b0
i the unique winner, and thus it is a direct best-reply for i. In the new state b′ = (b0
i ∈ f (b0) (i.e b∗
i→ b∗
i . This
i ) we have
−i, b∗
make b∗
T (b′) = T (b0) ∪ {i}.
Case 2. s0(b∗
i ) = sw0 − 1 (i.e., b∗
i needs one more vote to become a winner). By Axioms G+R, i prefers f (b0
−i, b∗
i )
i , which results in a "more truthful" state b′.
over f (b0). Then similarly to case 1, i has a direct step b0 i→ b∗
Case 3. b∗
i is the unique winner). Then the next step b0
i = f (b0) (i.e b∗
previous cases. Moreover, it must hold that j /∈ T (b0) since otherwise b0
Ij(b0) = ∅. Thus T (b′) = T (b1) + 1 ≥ T (b0) + 1.
j→ b1 will bring us to one of the two
j = f (b0) which means
j = b∗
Case 4. f (b0) = x 6= b∗
i , and s0(x) = s0(b∗
i ) + 2. We further divide into:
Case 4.1. s0(b∗
i ) ≥ s0(y) for all y 6= x. Then the next step by j must be from x, which brings us to one of the
two first cases (as in Case 3).
Case 4.2. There is y 6= x s.t. s0(x) = s0(y) + 1 = s0(b∗
i ) + 2. Then we continue the sequence of steps until
the winner's score decreases. Since all steps that maintain swt select a more preferred candidate, this
most occur at some time t, and T (b0) ⊆ T (bt). Then at bt we are again in Case 1 or 2.
Case 4.3. There is y 6= x s.t. s0(x) = s0(y) = s0(b∗
i ) + 2. Then by Lemma 23 b∗
i can never be selected, in
contradiction to b∗
i ∈ C(b0).
Case 5. f (b0) = x 6= b∗
i , and s0(x) ≥ s0(b∗
i ) + 3. We further divide into:
Case 5.1. For all y 6= x, s0(y) ≤ s0(x) − 3. In this case no reply is possible.
Case 5.2. There is some y 6= x s.t. s0(y) ≥ s0(b∗
i ) + 2. Then by Lemma 23 b∗
i can never be selected, in
contradiction to b∗
i ∈ C(b0).
Case 5.3. There is some y 6= x s.t. s0(y) ≥ s0(b∗
means s1(x) = s1(y) = sw0 − 1 ≥ s0(b∗
contradiction.
i ) + 1 Then the next step must be from x to such y. Which
i ) + 2. Thus again by Lemma 23 we reach a
i ) + 2 = s1(b∗
Therefore we either construct a path of direct replies to b′ ∈ B(b0) with T (b′) > T (b0) in contradiction to
our maximality assumption, or we reach another contradiction. Thus B(b0) must contain some NE (both in Gb0
and in G), which means by construction that G is weakly-FDRP from b0. However since b0 ∈ B(a0), we get that
G is weakly-FDRP from a0 as well.
16
Remark 5.1. Theorem 24 and Ex. 17 provide a partial answer to an open question regarding whether there are
game forms that admit weak FIP but not restricted FIP [Kukushkin, 2011].
Indeed, the game form f P R for
m = 4, n = 3 is such an example, but one that uses randomization. However if we think of f P R as a deterministic
game form with 2m − 1 possible outcomes (all nonempty subsets of candidates), where players are restricted to m
actions each, then the allowed utility profiles are constrained (by Axioms G and R) and thus this result does not
settle Kukushkin's question completely.
5.2 Weighted Plurality
When voters are weighted, cycles of direct responses can emerge [Meir et al., 2010; Meir, 2016]. We conjecture
that such cycles must depend on the order of agents, and that certain orders will break such cycles and reach an
equilibrium, at least from the truthful state.
Conjecture 25. f P L
s,w is weak-FDRP (in particular weak-FIP).
Similar techniques to those used so far appear to be insufficient to prove the conjecture. For example, in
contrast to the unweighted case, a voter might return to a candidate she deserted in any scheduler, even if only two
weight levels are present. We thus leave the proof of the general conjecture for future work.
Yet, we want to demonstrate the power of weak acyclicity over restricted acyclicity, even when there are
no randomness or restrictions on the utility space. That is, to provide a definite (negative) answer to Kukushkin's
question of whether weak acyclicity entails restricted acyclicity. To do so, we will use a slight variation of Plurality
with weighted voters and lexicographic tie-breaking.
Theorem 26. There exist a game form f ∗ s.t. f ∗ is weak-FIP but not restricted-FIP.
Proof. Consider the following game G: The initial fixed score of candidates {a, b, c, d} is s = (0, 1, 2, 3). The
weight of each voter i ∈ {1, 2, 3} is i. The preference profile is as follows: c ≻1 d ≻1 b ≻1 a, b ≻2 c ≻2 a ≻2 d,
and a ≻3 b ≻3 c ≻3 d. This game was used in [Meir et al., 2010] to demonstrate that Plurality with weighted
voters is not FDRP, however it can be verified that G is restricted-FIP so it is not good enough for our use.
If we ignore agents' preferences, we get a particular game form f P L
s,w where N = {1, 2, 3}, M = {a, b, c, d},
s = (0, 1, 2, 3) and w = (1, 2, 3).
We define f ∗ by modifying f P L
s,w with the following restrictions on agents' actions: A1 = {c, d}, A2 =
{b, c}, A3 = {a, b, d}. Thus f ∗ is a 2 × 2 × 3 game form, presented in Figure 6(a).
We first show that f ∗ is not restricted-FIP. Indeed, consider the game G∗ accepted from f ∗ with the same
preferences from game G (Figure 6(b)). We can see that there is a cycle of length 6 (in bold). An agent scheduler
that always selects the agent with the bold reply guarantees that convergence does not occur, since in all 6 relevant
states the selected agent has no alternative replies.
Next, we show that f ∗ is weak-FIP. That is, for any preference profile there is some scheduler that guarantees
convergence. We thus divide into cases according to the preferences of agent 3. In each case, we specify a state
where the scheduler selects agent 3, the action of the agent, and the new state.
We note that since all thick edges must be oriented in the same direction, a ≻3 b if and only if b ≻3 c. Thus
the following three cases are exhaustive.
Q3
b ≻ d
d ≻ b & d ≻ a
a ≻ d ≻ b ≻ c
1
2
3
state
action
(d, b, a)
(c, b, b)
(d, c, b)
b
d
d
new state
(d, b, b)
(c, b, d)
(d, c, d)
In either case, agent 3 moves from a state on the cycle to a Nash equilibrium.
6 Conclusions and Future Work
The main conceptual contribution of this work was to provide a joint rigorous framework for the study of iterative
voting, as part of the broader literature on acyclicity of games and game forms.
17
(a) The game form f ∗
(c b d){d}
3
3
(c b b){b}
3
(c b a){a}
2
2
(c c a){c}
(c c b){c}
3
(c c d){d}
(c b d){d}
(d b a){d}
2
(d c a){c}
3
3
1
1
(b) The game G∗
(d b d){d}
3
(d b b){b}
3
(d c b){b}
3
(d c d){d}
(d b d){d}
(c b b){b}
(d b b){b}
(c b a){a}
(d b a){d}
(c c a){c}
(d c a){c}
(c c b){c}
(d c b){b}
(c c d){d}
(d c d){d}
Figure 6: In each state we specify the actions of all 3 agents, and the outcome in curly brackets. Agent 1 controls
the horizontal axis, agent 2 the vertical axis, and agent 3 the in/out axis. We omit edges between states with
identical outcomes, since such moves are impossible for any transitive preferences. A directed edge in (b) is a
better-reply in G∗.
On the technical level, this unified presentation enabled us to construct examples of voting rules that settle at
two open questions on acyclicity of game forms: first, showing that there may be non-separable game forms that
are FIP (Theorem 3); and second, that there are game forms that are weakly acyclic but not order-free acyclic
(Theorem 26).
In addition, we provide an extensive study of convergence properties of the common Plurality rule and its
variations. We summarize all known results on iterative voting that we are aware of in Table 1. Note that in
some cases we get positive results if we restrict the initial state or the number of voters (not shown in the table).
For Plurality we provide a more detailed picture in Figs. 7,8. Previous papers whose results are covered in the
Table 1 often use different terminology and thus theorems and examples need to be rephrased (and sometimes
slightly modified) to be directly comparable. These rephrasing and necessary modifications are explained in detail
in [Meir, 2016]. The only paper not covered in [Meir, 2016] is by Koolyk et al. [2016], which provided non-
convergence examples for a variety of common voting rules including Maximin, Copeland, Bucklin, STV, Second-
Order Copeland, and Ranked Pairs. All results demonstrate cycles under best-reply (and under several other
restrictions) from the truthful state, thereby proving that neither of these rules is FBRP (even from the truth).
Beyond the direct implication of various acyclicity properties on convergence in an interactive setting where
agents vote one-by-one, [strong/weak] acyclicity is tightly linked to the convergence properties of more sophis-
ticated learning strategies in repeated games [Bowling, 2005; Marden et al., 2007], which is another reason to
understand them.
Fabrikant et al. [2010] provide a sufficient condition for weak-acyclicity, namely that any subgame contains
a unique Nash equilibrium. Unfortunately, this criterion is not very useful for most voting rules, where typically
18
Voting rule
Dictator
Plurality (lex.)
Plurality (LD)
Plurality (SD)
Plurality (rand.)
Weighted Plurality (lex.)
Veto
k-approval (k ≥ 2)
Borda
PSRs (except k-approval)
Approval
Other common rules
FIP
V
X
X (Ex. 13)
X (Ex. 13)
X (Ex. 13)
X
X
X
X
X
X
X
FBRP
V
X (Ex. 6)
?
?
X
X
-
V (Thm. 5)
V [M15]
V [RW12]
X
X (Ex. 7)
X [M16]
V [RW12,LR12]
X [LR12,L15]
X [RW12,LR12]
X [LR12,L15]
X [M16]
X [KLR16]
-
-
-
-
-
FDBRP
restricted-FIP Weak-FIP
V
V
V
V
V
V
V
V
X (Ex. 17)
V (Thm. 24)
?
V
X
X
?
V [M16]
?
?
V
X [M16]
X [RW12]
?
V
?
Positive results carry to the right side, negative to the left side. We assume lexicographic tie
Table 1:
breaking in all rules except Plurality. FDBRP is only well-defined for Plurality and Veto. Reference codes:
RW12 [Reyhani and Wilson, 2012], LR12 [Lev and Rosenschein, 2012], M15 [Meir, 2015], L15 [Lev, 2015],
M16 [Meir, 2016], KLR16 [Koolyk et al., 2016].
FBRP (Ex. 6)
⇑
FIP
⇓
restricted-FBRP ⇒ weak-FBRP
⇓
⇓
restricted-FIP ⇒ weak-FIP
⇑
⇑
FDRP (Thm. 5) ⇛ FDBRP ⇛ restricted-FDRP ⇒ weak-FDRP
Figure 7: Convergence results for Plurality under lexicographic tie-breaking. Positive results (in light green) carry
with the direction of the arrows, whereas negative results (dark gray) carry in the opposite direction.
(at least) all unanimous votes form equilibria. Another sufficient condition due to Apt and Simon [2012] is by
eliminating never-best-reply strategies, and the prospects of applying it to common voting rules is not yet clear.
We can see that in the "standard" lexicographic domain, convergence is guaranteed from any initial state
provided that voters restrict themselves to direct replies. With randomized tie-breaking, we must also require
a truthful initial vote. On the other hand, we can also allow indirect best-replies, so the results are essentially
incomparable. However, we see the result in the lexicographic case as stronger, since it only requires a very mild
and natural behavioral restriction in the context of Plurality voting, whereas it is harder to justify assumptions on
the initial state.
Implications on social choice
Importantly, best-reply dynamics is a natural and straightforward process, and
requires little information. As such, and due to the convergence properties demonstrated in this work, it is an
attractive "baseline" candidate for predicting human voter behavior in elections and designing artificial agents
with strategic voting capabilities-two of the most important, and also the hardest, goals of social choice research.
However, the clear disadvantage of this approach is that in the vast majority of cases (especially when there
are more than a handful of voters), almost every voting profile (including the truthful one) is already a Nash
equilibrium. Given this, our analysis is particularly suitable when the number of voters is small, for two main
reasons. First, it is more practical to perform an iterative voting procedure with few participants. Second, the
question of convergence is only relevant when cases of tie or near-tie are common. In more complex situations
with many active voters who may change their vote, it is likely that a more elaborate game-theoretic model is
required, which takes into account voters' uncertainty and heuristic behavior (see Section 1.1).
19
FBRP from truth (Thm. 14) ⇐ FBRP
restricted-FBRP
⇒
weak-FBRP ?
⇑
⇑
⇓
FIP from truth
⇐ FIP ⇛ restricted-FIP (Ex. 17) ⇒
⇓
⇓
⇑
⇓
weak-FIP
⇑
FDRP from truth (Ex. 18) ⇐ FDRP
restricted-FDRP
⇒ weak-FDRP (Thm. 24)
Figure 8: Convergence results for Plurality under random tie-breaking.
Promising future directions Based on the progress made in this paper and the other results published since the
introduction of iterative voting in [Meir et al., 2010], we believe that research in this area should focus on three
primary directions:
1. Weak-acyclicity seems more indicative than order-free acyclicity to determine convergence in practice. Thus
theorists should study which voting rules are weak-FIP, perhaps under reasonable restrictions (as we demon-
strated, this property is distinct from restricted-FIP). We highlight that even in rules where there are counter
examples for weak acyclicity (k-approval, Borda), these examples use two voters and games with more
voters may well be weakly acyclic.
2. It is important to experimentally study how people really vote in iterative settings (both in and out of the
lab), so that this behavior can be formalized and behavioral models can be improved. The work of [Tal et al.,
2015] is a preliminary step in this direction, but there is much more to learn. Ideally, we would like to
identify a few types of voters, such that for each type we can relatively accurately predict the next action in a
particular state. It would be even better if these types are not specific to a particular voting rule or contextual
details.
3. We would like to know not only if a voting rule converges under a particular dynamics (always or often), but
also what are the properties of the attained outcome-in particular, whether the iterative process improves
welfare or fairness, avoids "voting paradoxes" [Xia et al., 2007] and so on. Towards this end, several re-
searchers (e.g., [Reijngoud and Endriss, 2012; Branzei et al., 2013; Meir et al., 2014; Bowman et al., 2014;
Koolyk et al., 2016]) have started to explore these questions via theory and simulations. However, a good un-
derstanding of how iterative voting shapes the outcome, whether the population of voters consists of humans
or artificial agents, is still under way.
References
St`ephane Airiau and Ulle Endriss. Iterated majority voting. In Proceedings of the 1st International Conference on
Algorithmic Decision Theory (ADT-2009), pages 38–49. Springer Verlag, 2009.
Daniel Andersson, Vladimir Gurvich, and Thomas Dueholm Hansen. On acyclicity of games with cycles. Discrete
Applied Mathematics, 158(10):1049–1063, 2010.
Krzysztof R. Apt and Sunil Simon. A classification of weakly acyclic games. In SAGT'12, pages 1–12. 2012.
Tilman Borgers. Pure strategy dominance. Econometrica, 61(2):423–430, 1993.
Endre Boros, Vladimir Gurvich, Kazuhisa Makino, and Wei Shao. Nash-solvable bidirected cyclic two-person
game forms. Technical report, Rutcor Research Report 26-2007 and DIMACS Technical Report 2008-13, Rut-
gers University, 2008.
Endre Boros, Vladimir Gurvich, Kazuhisa Makino, and D´avid Papp. Acyclic, or totally tight, two-person game
forms: Characterization and main properties. Discrete Mathematics, 310(6):1135–1151, 2010.
20
Michael Bowling. Convergence and no-regret in multiagent learning. Advances in neural information processing
systems, 17:209–216, 2005.
Clark Bowman, Jonathan K. Hodge, and Ada Yu. The potential of iterative voting to solve the separability problem
in referendum elections. Theory and decision, 77(1):111–124, 2014.
Simina Branzei, Ioannis Caragiannis, Jamie Morgenstern, and Ariel D. Procaccia. How bad is selfish voting? In
Twenty-Seventh AAAI Conference on Artificial Intelligence, 2013.
Samir Chopra, Eric Pacuit, and Rohit Parikh. Knowledge-theoretic properties of strategic voting. In Logics in
Artificial Intelligence, pages 18–30. Springer, 2004.
Vincent Conitzer, Toby Walsh, and Lirong Xia. Dominating manipulations in voting with partial information. In
AAAI, volume 11, pages 638–643, 2011.
Antoine-Augustin Cournot. Recherches sur les principes math´ematiques de la th´eorie des richesses par Augustin
Cournot. chez L. Hachette, 1838.
Yvo Desmedt and Edith Elkind. Equilibria of plurality voting with abstentions. In Proc. of 11th ACM-EC, pages
347–356, 2010.
Amrita Dhillon and Ben Lockwood. When are plurality rule voting games dominance-solvable? Games and
Economic Behavior, 46:55–75, 2004.
Edith Elkind, Umberto Grandi, Francesca Rossi, and Arkadii Slinko. Gibbard–satterthwaite games. In Proceedings
of the 24th International Joint Conference on Artificial Intelligence (IJCAI-15), 2015.
Ulle Endriss. Sincerity and manipulation under approval voting. Theory and Decision, 74(3):335–355, 2013.
Alex Fabrikant, Aaron D Jaggard, and Michael Schapira. On the structure of weakly acyclic games. In Algorithmic
Game Theory, pages 126–137. Springer, 2010.
Dvir Falik, Reshef Meir, and Moshe Tennenholtz. On coalitions and stable winners in plurality. In WINE'12, pages
256–269. 2012.
Robert Forsythe, Thomas Rietz, Roger Myerson, and Robert Weber. An experimental study of voting rules and
polls in three-candidate elections. International Journal of Game Theory, 25(3):355–83, 1996.
Ehud Friedgut, Gil Kalai, Nathan Keller, and Noam Nisan. A quantitative version of the gibbard-satterthwaite
theorem for three alternatives. SIAM Journal on Computing, 40(3):934–952, 2011.
Peter Gardenfors. Manipulation of social choice functions. Journal of Economic Theory, 13(2):217–228, 1976.
Christian Geist and Ulle Endriss. Automated search for impossibility theorems in social choice theory: Ranking
sets of objects. Journal of Artificial Intelligence Research, 40(1):143–174, 2011.
A. Gibbard. Manipulation of voting schemes. Econometrica, 41:587–602, 1973.
Neelam Gohar. Manipulative Voting Dynamics. PhD thesis, University of Liverpool, 2012.
Umberto Grandi, Andrea Loreggia, Francesca Rossi, Kristen Brent Venable, and Toby Walsh. Restricted manipula-
tion in iterative voting: Condorcet efficiency and Borda score. In Algorithmic Decision Theory, pages 181–192.
Springer, 2013.
Richard W Hamming. Error detecting and error correcting codes. Bell System technical journal, 29(2):147–160,
1950.
21
Jerry S Kelly. Strategy-proofness and social choice functions without singlevaluedness. Econometrica: Journal of
the Econometric Society, pages 439–446, 1977.
Aaron Koolyk, Omer Lev, and Jeffrey S Rosenschein. Convergence and quality of iterative voting under non-
scoring rules. In Proceedings of the 2016 International Conference on Autonomous Agents & Multiagent Sys-
tems, pages 1329–1330. International Foundation for Autonomous Agents and Multiagent Systems, 2016.
Nikolai S. Kukushkin. Congestion games: a purely ordinal approach. Economics Letters, 64:279–283, 1999.
Nikolai S. Kukushkin. Perfect information and congestion games. Games and Economic Behavior, 38:306–317,
2002.
Nikolai S. Kukushkin. Acyclicity of improvements in finite game forms. International Journal of Game Theory,
40(1):147–177, 2011.
Omer Lev and Jeffrey S. Rosenschein. Convergence of iterative voting. In Proc. of 11th AAMAS, pages 611–618,
2012.
Omer Lev. Agent Modeling of Human Interaction: Stability, Dynamics and Cooperation. PhD thesis, The Hebrew
University of Jerusalem, 2015.
Jason R Marden, Gurdal Arslan, and Jeff S Shamma. Regret based dynamics: convergence in weakly acyclic
games. In Proceedings of the 6th international joint conference on Autonomous agents and multiagent systems,
page 42. ACM, 2007.
Reshef Meir, Maria Polukarov, Jeffrey S. Rosenschein, and Nick Jennings. Convergence to equilibria of plurality
voting. In Proc. of 24th AAAI, pages 823–828, 2010.
Reshef Meir, Omer Lev, and Jeffrey S. Rosenschein. A local-dominance theory of voting equilibria. In EC'14,
2014.
Reshef Meir. Plurality voting under uncertainty. In AAAI'15, pages 2103–2109, 2015.
Reshef Meir. Strong and weak acyclicity in iterative voting. In COMSOC'16. 2016.
Matthias Messner and Mattias K. Polborn. Robust political equilibria under plurality and runoff rule, 2002. Mimeo,
Bocconi University.
Igal Milchtaich. Congestion games with player-specific payoff functions. Games and economic behavior,
13(1):111–124, 1996.
Dov Monderer and Lloyd S. Shapley. Potential games. Games and Economic Behavior, 14(1):124–143, 1996.
Svetlana Obraztsova, Evangelos Markakis, and David R. M. Thompson. Plurality voting with truth-biased agents.
In Algorithmic Game Theory, pages 26–37. Springer, 2013.
Svetlana Obraztsova, Evangelos Markakis, Maria Polukarov, Zinovi Rabinovich, and Nicholas R Jennings. On
In Twenty-Ninth AAAI
the convergence of iterative voting: How restrictive should restricted dynamics be?
Conference on Artificial Intelligence, 2015.
Zinovi Rabinovich, Svetlana Obraztsova, Omer Lev, Evangelos Markakis, and Jeffrey S. Rosenschein. Analysis
of equilibria in iterative voting schemes. In 29th National Conference on AI (AAAI), 2015.
Annemieke Reijngoud and Ulle Endriss. Voter response to iterated poll information. In Proc. of 11th AAMAS,
pages 635–644, 2012.
Reyhaneh Reyhani and Mark C Wilson. Best-reply dynamics for scoring rules. In 20th European Conference on
Artificial Intelligence. IOS Press, 2012.
22
Alvin E Roth. The college admissions problem is not equivalent to the marriage problem. Journal of economic
Theory, 36(2):277–288, 1985.
Donald G. Saari. Susceptibility to manipulation. Public Choice, 64:21–41, 1990.
M. Satterthwaite. Strategy-proofness and Arrow's conditions: Existence and correspondence theorems for voting
procedures and social welfare functions. Journal of Economic Theory, 10:187–217, 1975.
Murat R. Sertel and M. Remzi Sanver. Strong equilibrium outcomes of voting games are the generalized condorcet
winners. Social Choice and Welfare, 22:331–347, 2004.
Maor Tal, Reshef Meir, and Ya'acov Gal. A study of human behavior in voting systems. In AAMAS 15, pages
665–673, 2015.
Lirong Xia, J´erome Lang, and Mingsheng Ying. Sequential voting rules and multiple elections paradoxes. In
TARK'07, pages 279–288, 2007.
H Peyton Young. The evolution of conventions. Econometrica: Journal of the Econometric Society, pages 57–84,
1993.
23
A Proofs
Lemma 15. Consider any cardinal utility function u and the partial preference order Q it induces on subsets by
random tie-breaking. Q holds Axioms K+G+R.
Proof. Let u be any utility scale, we will show that all axioms hold. Let a, b ∈ C and W ⊆ C \ {a, b}.
u({a} ∪ W ) =
1
W + 1 u(a) + Xc∈W
u(c)! , u({b} ∪ W ) =
1
W + 1 u(b) + Xc∈W
u(c)! = u({b} ∪ W ),
thus {a} ∪ W ≻Q {b} ∪ W , and Axiom R holds.
Let a ∈ C, W ⊆ C s.t. ∀b ∈ W, a ≻b. Then
u(a) =
>
1
W + 1 u(a) + Xb∈W
W + 1 u(W ) + Xb∈W
1
u(a)! >
u(a)! =
1
W + 1 u(a) + Xb∈W
u(b)! = u({a} ∪ W )
1
W + 1
u(W ) +
W
W + 1
u(W ) = u(W ),
thus a ≻Q {a} ∪ W ≻Q W and Axiom G holds.
Axiom K1 follows immediately from G. K2 also follows if preferences are strict. Even if there are ties, and
a (cid:23) w for all a ∈ A, w ∈ W then:
i.e., A (cid:23)Q W .
u(A) ≥ min
a∈A
u(a) ≥ max
w∈W
u(w) ≥ u(W ),
Definition A.1. Suppose that X, Y ⊆ C, k = X ≤ Y = K. Sort X, Y in increasing order by Q. Let
rj =(cid:6) j
Y is partitioned into Y1 = {y1, y2, y3}, Y2 = {y4, y5}, Y3 = {y6, y7}).
k K(cid:7). Partition Y into sets Y1, . . . , Yk s.t. for j < K, Yj = {yrj−1+1, . . . , yrj } (e.g., if k = 3, K = 7, then
X match-dominates Y according to Q if:
• (I) ∀j ≤ k∀y ∈ Yj, xj (cid:23) y; and
• either (IIa) at least one relation is strict, or (IIb) K mod k 6= 0.
If X > Y , then X match-dominates Y if Y match-dominates X according to the reverse of Q.
Intuitively, match-domination means that for any q ∈ [0, 1], there is a fraction q of the set X that dominates a
fraction of 1 − q from the set Y : at least one x ∈ X dominates all of Y , at least 20% of X dominate at least 80%
of Y , and so on.
Lemma 27. Let a, a′ be two profiles that differ by a single vote, and define X = f (a), Y = f (a′).10
The following conditions are equivalent for any strict order Q over C:
1. X stochastically dominates Y under preferences Q and uniform lottery.
2. The relation X ≻ Y is entailed by Q and the Axioms K+G+R and transitivity.
3. u(X) > u(Y ) for every u that is consistent with Q.
4. X match-dominates Y according to Q.
10Without some restriction on X, Y , the lemma is incorrect. E.g. if x1 ≻ y1 ≻ y2 ≻ x2 ≻ y3 ≻ y4, then X stochastically dominates Y
but there is no way to derive X ≻ Y from the axioms K+G+R.
24
Proof. The equivalence of (1) and (3) is immediate, and used e.g. in [Reyhani and Wilson, 2012].
(2) ⇒ (3). If X ≻ Y follows from the axioms, then there is a sequence of sets X = X0 ≻ X1 ≻ · · · ≻ Xk = Y
such that each Xj ≻ Xj+1 follows from a single axiom K,G, or R. Thus it is sufficient to show for X ≻ Y that
follows from a single axiom.
If X ≻ Y follows from Axiom R, then X = {a} ∪ W, Y = {b} ∪ W for some W ⊆ C \ {a, b} and a ≻ b.
Thus
u(X) = u({a}∪W ) =
1
W + 1 u(a) + Xc∈W
u(c)! >
1
W + 1 u(b) + Xc∈W
u(c)! = u({b}∪W ) = u(Y ).
If X ≻ Y follows from Axiom G, then either X = Y ∪ {a} and a ≻ b for all b ∈ Y , or X = {x} and
Y = {x} ∪ W where x ≻ w for all w ∈ W . For the first case
u(X) =
1
Y + 1
u(a) +
1
Y + 1 Xy∈Y
u(y) =
1
Y + 1
1
Y Xy∈Y
u(a) +
1
Y + 1 Xy∈Y
u(y)
1
1
>
Y + 1
=(cid:18)1 +
u(y) +
Y Xy∈Y
Y (cid:19)
Y + 1 Xy∈Y
1
1
u(y)
1
1
u(y) =
Y + 1 Xy∈Y
Y Xy∈Y
Y u(x) + Xw∈W
1
For the second case,
u(X) = u(x) =
1
Y Xy∈Y
u(x) =
u(y) = u(Y ).
u(x)! >
1
Y u(x) + Xw∈W
u(w)! = u(Y ).
If X ≻ Y follows from Axiom K, then u(x) > u(y) for any x ∈ X, y ∈ Y which is a trivial case.
(3) ⇒ (4). Suppose that u(X) > u(Y ) for all u. Suppose first X ≤ Y . If X does not match-dominate
Y then either (I) there is an element xj ′ that is less preferred than some element y′ ∈ Yj ′ ; or (II) for all j and all
y ∈ Yj, xj =Q y and Yj = K
k = q for all j. We will derive a contradiction to (3) in either case. In the latter case,
we have u(xj ) = u(Yj) for all j and thus
u(Y ) =
In contradiction to (3).
1
K
Xj≤k
Yju(Yj)
= Pj≤k qu(xj)
K
= Pj≤k qu(xj)
kq
= u(X),
Thus we are left with case (I). That is, there are j′ ≤ k and y′ ∈ Yj ′ s.t. xj ′ ≺ y′. We define the (possibly
empty) set X ′ ⊆ X as all elements {x : x ≻ xj ′ }. We define Y ′ ⊆ Y as {y : y (cid:23) y′}. By construction, for any
j > j′, Yj ⊆ Y ′. Thus
k
k
Y ′ ≥ 1 +
Yj = 1 +
Xj=j ′+1
Xj=j ′+1
(rj − rj−1) = (K − rj ′ ) + 1 = (K −(cid:24) j′
k
K(cid:25)) + 1 > K −
j′
k
K = K(1 −
j′
k
),
whereas X ′ ≤ k − j′. We define u as follows: u(x) = 1, u(y) = 1 for all x ∈ X ′, y ∈ Y , and u(z) = 0 for
all other elements. Note that X ′, Y ′ contain the top elements of X, Y , respectively. In addition, y′ is the minimal
element in Y ′ and by transitivity y′ ≻ x for all x ∈ X \ X ′. Thus u is consistent with Q.
We argue that u(Y ) > u(X) in contradiction to (3). Indeed, u(X) = X ′
X ≤ k−j ′
k = 1 − j ′
k .
u(Y ) =
Y ′
Y
>
k )K
(1 − j ′
K
= 1 −
=
j′
k
k − j′
≥
X ′
X
= u(X),
k
so we get a contradiction to (3) again. Thus X matching-dominate Y .
(4) ⇒ (2). This is the only part of the proof where we use the profiles from which X, Y are obtained. When
a single voter moves, either the winner set changes by a single candidate (added, removed, or swapped), or X is a
single candidate, or Y is a single candidate. We prove case by case.
25
• The case where X = Y = 1 is immediate.
• Suppose X = 1 (i.e. X = {x}) and Y = K > 1. Then X match-dominates Y means that x (cid:23)
y for all y ∈ Y , with at least one relation being strict, w.l.o.g. yK (least preferred in Y ). Then X (cid:23)
{y1, . . . , yK−1} ≻ Y , where the first transition is by Axiom K2 and the second is by Axiom G.
• The case of Y = 1 is symmetric.
• Suppose X = Y = k. Then X match-dominates Y means that xi (cid:23) yi for all i. For all t ∈ {0, 1, . . . , k},
let X t = {x1, . . . , xt, yt+1, . . . , yk}. Then X t−1 = X t if xt = yt, and X t−1 ≻ X t otherwise from
Axiom R. In addition, X = X 0, Y = X k thus X ≻ Y from transitivity.
• Suppose X = k, Y = k + 1. Then X match-dominates Y means that Y1 = (cid:6) k+1
sets Yj are singletons Yj = yj. Consider the set Y ′ that includes the top k elements of Y . Since x1 is
(weakly) preferred to both candidates in Y1, Y ′ is match-dominated by X. By the previous bullet X (cid:23) Y ′
follows from Axiom R and transitivity. Finally, Y ′ ≻ Y = Y ′ ∪ {min Y } by Axiom G.
k (cid:7) = 2, and all other
The following is an immediate corollary:
Proposition 20. A step a i→ a′ is a better-response under random tie-breaking and stochastic dominance, if and
only if f (a′) ≻i f (a) is entailed by Qi, the Axioms K+G+R, and transitivity.
Proposition 22. A step a i→ a′ is a better-response under unknown tie-breaking and local dominance, if and only
if f (a′) ≻i f (a) is entailed by Qi, Axioms K+G, and transitivity.
Proof. Suppose that X = f (a′) locally-dominates Y = f (a). Let Z = X ∩ Y , and X ′ = X \ Z, Y ′ = Y \ Z.
We must have x ≻i y for any x ∈ X, y ∈ Y ′, otherwise, a tie-breaking order that selects y first and x second
would make i strictly lose when moving from Y to X. Similarly, x ≻i y for any x ∈ X ′, y ∈ Y . If Z = ∅ then
X = X ′ ≻i Y ′ = Y follows from Axiom K. Otherwise, by repeatedly applying Axiom G we get X (cid:23)i Z (cid:23)i Y
with at least one relation being strict.
In the other direction, since Axiom G can only be used to add elements lower (or higher) than all existing
elements, it may only induce relations of the form Z ≻ Z ∪ Y ′ where z ≻ y for all z ∈ Z, y ∈ Y ′; or relations of
the form Z ∪ X ′ ≻ Z where x ≻ z for all z ∈ Z, x ∈ X ′. Thus if X ≻ Y follows from Axiom G, they must have
the form X = Z ∪ X ′, Y = Z ∪ Y ′ where x ≻ z ≻ y for all x ∈ X ′, z ∈ Z, y ∈ Y ′. To see that this entails local
dominance, let xL = L(X) be the first element in X according to order L ∈ π(C), and likewise for Y . For any L,
xL (cid:23) yL (with equality iff L(X) = L(Y ) ∈ Z). Further, either X ′ or Y ′ are non-empty (w.l.o.g. X ′). Consider
an order L′ such that L′(X) ∈ X ′, then xL′ ≻ y for all y ∈ Y and in particular xL′ ≻ yL′.
26
|
1907.10782 | 2 | 1907 | 2019-07-26T20:57:47 | A Framework for Monitoring Human Physiological Response during Human Robot Collaborative Task | [
"cs.MA",
"cs.RO"
] | In this paper, a framework for monitoring human physiological response during Human-Robot Collaborative (HRC) task is presented. The framework highlights the importance of generation of event markers related to both human and robot, and also synchronization of data collected. This framework enables continuous data collection during an HRC task when changing robot movements as a form of stimuli to invoke a human physiological response. It also presents two case studies based on this framework and a data visualization tool for representation and easy analysis of the collected data during an HRC experiment. | cs.MA | cs | A Framework for Monitoring Human Physiological Response during
Human Robot Collaborative Task
Celal Savur1
Ferat Sahin3
Department of Electrical and Microelectronic Engineering
Shitij Kumar2
Rochester Institute of Technology
{cs13231, spk44222, feseee3}@rit.edu
Rochester, NY, 14623, USA
9
1
0
2
l
u
J
6
2
]
A
M
.
s
c
[
2
v
2
8
7
0
1
.
7
0
9
1
:
v
i
X
r
a
the use of 'psycophsiological' 1 methods to evaluate human
response and behavior during human robot
interaction. In
our opinion, continuous monitoring of physiological signals
during human-robot task is the first step in quantifying human
trust in automation. The inferences from these signals and
incorporating them in real-time to affect robot motion can
help in enhancing the human-robot interaction. Such a system
capable of 'physiological computing' 2 will result in a closed
human-in-the-loop system where both human and robot in
an HRC setup are monitored and information is shared. This
could result into better communication which would improve
trust in automation and increase productivity.
Abstract -- In this paper, a framework for monitoring hu-
man physiological response during Human-Robot Collaborative
(HRC) task is presented. The framework highlights the impor-
tance of generation of event markers related to both human
and robot, and also synchronization of data collected. This
framework enables continuous data collection during an HRC
task when changing robot movements as a form of stimuli to
invoke a human physiological response. It also presents two case
studies based on this framework and a data visualization tool for
representation and easy analysis of the collected data during an
HRC experiment.
Index Terms -- Physiological Signals, Psycophisiology, Human-
Robot Interaction, Collaborative Robots, Safety, Awareness,
Digital-Twin, Physiological Computing
I. INTRODUCTION
AcceptedinS M C 2019
The major challenges of any Human Robot Collaboration
(HRC) in industry are human safety, human trust in automa-
tion, and productivity [1]. Human safety has always been the
primary concern in robotics. One main aspect that concerns
safety is injuries due to human-robot collision. Different
strategies have been introduced to ensure human safety, one
is implementing physical and electronic safeguards according
to industrial standards [2]. However, new strategies and ap-
proaches are needed with human robot collaboration where
less standards are available to implement complex protection
schemes. Hence a new category of robots called collaborative
robots or cobots have been introduced in the market (e.g.
Universal Robots, Kuka lbr-iiwa, Rethink Robotics Sawyer;
to name a few). These robots are purposely designed to work
in direct cooperation with humans in a defined workspace by
lowering the severity and risks of injury due to collision.
Hence, in this work we propose a framework for a 'phys-
iological computing' system to monitor human physiological
responses during a human-robot collaboration task. This paper
highlights the aspects and challenges of collecting human-
physiological signals during a human-robot experiment. It
underscores the importance of a controlled HRC experiment
design, event marker generation related to both human and
robot, and the synchronization of data collected. In order
to verify this framework, a prototype implementation of the
system is shown as case studies of two HRC experiments.
The first case study is an experiment to monitor the effect
of change in robot acceleration and trajectory of motion on
human physiological signals and determine a human comfort
index. In this experiment the human is sitting and sharing
the workspace with a UR 5e robot. The second experiment is
monitoring the human-behavior for different safety algorithms
during human-robot collaborative task. This task is implemen-
tation of a speed and separation monitoring setup where a
human and a UR10 robot perform two separate tasks while
sharing a workspace [7]. Here, the human is not stationary
and moves in the workspace, which requires wireless data
acquisition of human physiological signals and representation
of human-robot shared workspace. The final objective of this
work is to generate a database that can be used to further the
1Psychophysiology is a branch of neuroscience that seeks to understand
how a persons mental state and physiological responses interact to affect one
another.
2Physiological computing represents a category of 'affective computing'
incorporates real-time software adaption to the psychophysiological
that
activity of the user.
Human trust in automation is about managing human ex-
pectations and how comfortable the human is sharing the
robot workspace. Even though cobots decrease the risk of
injury, any form of physical collision decreases the human
trust in automation. Thus, collision avoidance strategies such
as stopping or reducing speeds while human is in the operating
workspace of the robot have been implemented [3] [2]. How-
ever, the question arises how do we quantify human's trust in
automation?
In a human robot interaction setup, change in robot motion
can affect the human behavior. This was shown in experiments
done in [4] and [5]. The literature review in [6] highlights
The Intelligence represents the control of robot actions
during an HRC experiment. Programming experiment is part
of the Intelligence since it controls speed, acceleration, and
trajectory of the robot. The Intelligence module processes the
data from the Awareness module to generate event markers
as well as robot actions that can be used as stimuli to elicit
human response. In addition to Awareness it also receives input
from Compliance module, which is a form of interpretation
of human expectation. The Intelligence module interprets this
human command/feedback into actionable robot commands.
Using human physiological signal as feedback to the robot
or form of actionable control will help achieve a complete
human-in-the-loop closed loop system. Here, the Compliance
sub-module is responsible for inference from the physiological
signals or any form of commands from the human, that can
be used to modify the robot behavior. Thus achieving a higher
level of Compliance for the robot and managing the human
expectation by interpreting the human physiological state can
be a gateway for a more interactive human robot collaboration.
Intelligence, and Compliance are the main
parts of the framework [9], however to integrate these three
modules a communication layer for data transformation and
synchronization is required. This is critical as many sensor
devices and other systems do not have the same frequency
and timing clock. The communication layer is responsible
to transfer data in real time and also synchronize the data
from different sources such as physiological signal collection
devices, cameras, the representation of human-robot state in
the digital twin and robot state information.
Awareness,
AcceptedinS M C 2019
understanding of how human physiological responses can be
inferred to result in adaptive robot motion behavior.
The remainder of the paper is organized as follows: Section
II describes the proposed framework for creating a 'physi-
ological computing' system to monitor human physiological
responses during a human-robot collaborative task. Based on
this framework two case studies are implemented in Section
III and discussed in Section IV . Conclusions are drawn and
the future work mentioned in Section V.
II. PROPOSED APPROACH
In this section the key aspects and challenges of monitoring
human physiological response in Human Robot Collaboration
are presented. Asking questions to human subject during or af-
ter the experiment is common practice in human robot collab-
oration and interaction experiment [6], [8]. These response of
the subject allow researchers to quantify the subjective data of
the experiment. However, such methods that interrupt subject
during experiment may not be desirable for maintaining the
integrity of the desired physiological signals. In our opinion,
an alternative approach would be a system which is able to
generate event markers automatically during experiment and
enable the subject or the principle investigator to generate
markers as the experiment is being performed. Then these
event markers can be used during post processing by field
expert to identify response of the given input. In this way, it
could act as an alternative method to asking questions during
the experiment.
The block diagram of the proposed framework is shown in
Figure 1. The proposed framework is a solution for concur-
rently and continuously monitoring the state of human and
robot during an HRC task. The framework from a systems
perspective can be conceptually categorized further into three
sub modules: Awareness, Intelligence and Compliance [9]. The
communication layer between these sub-modules is equally
important as it
is responsible for data transformation and
synchronization.
The sub-module Awareness is the perception of system
which is generated from the physical world sensors and
digital represented in the virtual world. The physical world
is responsible to sense the environment through the sensor
information such as PPG sensor, GSR sensor, camera, motion
capture system etc. On the other hand virtual word is a digital-
twin representation of the physical world that mimics the
environment of the HRC task as well as the movements and
behavior of the robot and human agents [10]. The digital
twin can be used to calculate metrics such as human-robot
minimum distance, directed human-robot speeds, possible
collisions and changes in trajectory [7] [11]. The virtual world
updates its state constantly based on the sensory data received
from the physical world to update itself and generate new
data for the framework. Overall Awareness is responsible for
sensing physical and virtual world and provide this data to
rest of the system. Such a setup helps digitally represent a
combined human-robot state, which can then be associated
with the human physiological state.
Figure 1. An overview block diagram of the proposed framework for mon-
itoring Human Physiological Response during a Human Robot Collaborative
Task.
When designing human physiological signals related exper-
iment the following aspects are critical.
• Experiment design
• Event markers generation
• Synchronization
The importance of these is elaborated in the following Sec-
tions.
1) Experiment design: When designing an experiment, the
experiment and its parameters need to well defined. The task
need to be real or as realistic as possible to maintain the
integrity of the robot motion to act as stimuli to elicit the
human physiological response. For example, an industrial tasks
StorageSignal Transportation and SynchronizationRecorderFeedbackCompliance IntelligenceEvents Gen.ExperimentAwarenessSensorsPhysical WorldCameraRobot InfoRobot TwinVirtual WorldHuman TwinEnvironmentPrametersModel LearningCompliance system is good option for the experiment. Hence the industrial task
may improves the involvement of the subject sharing human
robot collaboration workspace. In addition the task need to be
simple and controlled to increase the repeatability of a human-
robot interaction scenario. A complex task may result into
more uncertainty.
2) Event Marker generation: The event markers generation
is part of experiment design. In the experiment, important
event need to be investigated and generated by the experiment.
Having markers during experiment gives more intuition about
experiment, such as Experiment Start/End, Task Start/End,
Robot Coming towards Human, etc. The event markers help
to synchronize signal across different channels. For example,
extracting Galvanic Skin Response and Heart Rate signal
between "Experiment Start" and "Experiment End" is trivial
when the event markers are present during signal recording.
Thus, the markers can be used during post-processing for
efficient data segmentation and epoching.
Figure 2. A picture of an subject that preparing for experiment, and device
placement
THE TABLE SHOWS THE PARAMETERS OF THE EACH TASK
Table I
3) Synchronization: Synchronization of signal from differ-
ent sensors is crucial for the human physiological response.
All the signal from human and robot need to be synchronized
with event markers. Thus a central synchronization system is
necessary. In proposed framework for the physiological com-
puting system, Lab Stream Layer (LSL) is used the interface
the subsystems, which integrates data from all different devices
being used. The Lab Stream Layer is a system for collection
of time series data over a local network with built-in time
synchronization [12]. The LSL stream is nearly real-time and
it is commonly used in biological signal collection system
such as OpenBCI, Pupil Lab, etc. Therefore, the LSL layer
is selected as the central core of the data acquisition system
in proposed framework. In the framework, each device has an
application node that is responsible to acquire signal from the
device in real-time and pushing it to the LSL stream. A node
is responsible to record all-time series data from LSL stream
into a local file for post-processing and analysis. Along with
LSL, Robot Operating System (ROS) and ZeroMQ is used to
monitor data in real time during the experiment [9].
AcceptedinS M C 2019
III. CASE STUDIES
Task 1
Task 2
Task 3
Task 4
A. Case Study I
Acceleration
Normal
High
Normal
High
Trajectory
Fixed
Fixed
Random
Random
will be minimized. Since the maximum speed is fixed, the
experiment is designed to have different accelerations and
trajectories. In the experiment acceleration has two modes:
fixed and random. The fixed mode indicates that robot has
fixed acceleration and random mode means the acceleration is
random.
The trajectory has two modes: simple and random. The
simple trajectory indicates there is no waypoint between pick
and place waypoints and the motion is fluent and predictable.
The random trajectory indicates multiple waypoints randomly
selected between pick and place waypoints. The Figure 3
shows an example of the trajectory in random mode in which
robot may take between Pick waypoint to Place waypoint or
vice-versa. The trajectory planner will generate a trajectory
from randomly selected waypoints from each plane.
Four type of tasks are performed by subjects. Each tasks
consists of two parts: loading inserts and unloading inserts.
The subject is responsible for loading inserts on plate shown
in Figure 2(top-left). There are two possible actions which
the human can take during tasks. The first one is load the
plate and wait for robot to unload all of the inserts from plate
then re-load the plate. The second action is, to increase the
productivity, while robot is unloading inserts, load inserts that
are taken by the robot. The subject has freedom to choose
whichever action is comfortable.
The robot is responsible for unloading i.e. picking all inserts
from the plate and placing them into the container. In order
to control the start the unloading, the robot checks master
The objective of the experiment is to monitor the effect
of acceleration and trajectory of the robot on human phys-
iological signals during collaborative task. The experiment
was performed using UR5e (Universal Robot) six degree of
freedom (DoF) arm robot, as shown in Figure 2. The UR5e is
a common collaborative robot with payload of 5 kg, which
is suitable for manufacturing environment and laboratories.
The experiment is a simplfied version of an industrial task for
loading inserts and unloading parts on at a plastic injection
molding plant. This experiment represents a scenario where
the human robot shared workspace is on a table and the human
is stationary.
The experiment consists of four sub tasks which are tab-
ulated in table I. In the experiment the max speed set 100
degree/seconds so in case of collusion any injury or pain
tracked using Pupil Labs and human physiological responses
such as pupil dilation, PPG, GSR, EEG & ECG recorded.
A system diagram showing the data collection and monitor-
ing is shown in Figure 5. The experiment setup is represented
as a digital-twin in order to represent human and robot state
during the experiment. This helps in generating the human-
robot interaction state data such as human-robot separation
distance (minimum distance), human head orientation, human
pose and velocity and action representation. This data is
monitored and collected along with the human-physiological
responses. It is used to represent a combined human-robot
state of the 'physiological computing' system and analyse the
stimulus and effect of human behavior during the experiment.
In this system, the event markers used for case studies I
and II, the physiological signals that can be used and the
communication and synchronization of data are discussed in
the following section.
Figure 4. A motion capture is used to monitor human motion, a camera is
used to record the experiment, the human-gaze is tracked using Pupil Labs
and human physiological responses such as pupil dilation, PPG, GSR, EEG
& ECG recorded.
IV. DISCUSSION
The auto generation of event markers during an HRC
experiment is critical. The choice of event markers depends
on the experiment setup and the objective of the experiment.
The biggest advantage of auto generation of event markers is
the experiment can be performed uninterrupted. These event
markers can be used to effectively post-process and analyse
the data as data segmentation and epoching of the collected
signals becomes easier. A list of events that are automatically
generated during the HRC task for Case Studies I and II are
listed in Table II.
B. Physiological Signals
In this Section, we list some of the human physiological
signals that have been used during human-robot experiments.
The devices for collecting these signals have been success-
fully interfaced in the implemented prototype system of the
proposed framework.
• Electroencephalogram (EEG) is the method to record the
brain's electrical activity via non-invasive electrode placed
Figure 3. Figure shows how the robot selects the waypoint between pick and
place positions. Solid line shows a possible random trajectory and striped line
shows fixed trajectory for the robot.
B. Case Study II
pin on plate every five seconds. This is helpful in generating
event-marker representing the start of the task. If the master
pin has inserts then the plate is full and the robot starts the
unloading process. It picks each item in order and place it
into a container. If there are no inserts on master pin, the
robot goes to its home position wait for five seconds. The
experiment setup and sensor placement can be seen in Figure
2.
AcceptedinS M C 2019
A. Event Marker
This experiment
is monitoring the human-behavior for
different safety algorithms during human-robot collaborative
task. This task is implementation of a speed and separation
monitoring setup where a human and a UR10 robot perform
two separate but related tasks while sharing a workspace [7].
Here, the human is not stationary and moves in the workspace,
which requires wireless acquisition of human physiological
signals and representation of human-robot shared workspace.
The experiment setup is a generic robot pick and place task
of placing 10 products in a box. The robot movement involves
moving the base joint 180◦ degrees between the pick and place
positions on the tables. The human has an assembly task for
threading a nut and a screw that are placed on the picking and
placing area. After threading the bolts and screws the human
puts the finished part on a table outside the robot workspace.
This human task was setup to control the human movement
and overlap of human-robot workspace. For more information
our previous work [3] and [7] can be referred. In order to
avoid collision, safety algorithms are implemented to detect
and anticipate the human motion, resulting into the robot
stopping, reducing speed or moving normally i.e. maximum
allowed speed for the task. The safety algorithms vary in terms
of parameters such as critical human-robot separation distance
and what sensors are used to calculate the separation distance.
This results into different robot motion behavior.
The objective of this experiment is to monitor human phys-
iological response and also see the overall task productivity
during this shared workspace task. During the experiment the
sensors used to monitor the human are shown in Figure 4.
Here the motion capture is used to monitor human motion, a
camera is used to record the experiment, the human-gaze is
Place waypointPick waypointSide ViewGalvanic SkinResponsePPG for HR andHRVMotion CaptureSystemCamera recordingSubjectPupil Dilation andGazeEEG&ECG G.tecNatilus Figure 5. A system diagram representing the data collection and monitoring during the experiment as described in Case Study II
THE TABLE SHOWS THE EVENT-MARKERS USED IN CASE STUDY I & II
Table II
AcceptedinS M C 2019
Experiment started
nth task initialized but subject has not
complete loading yet
nth task started robot unloading all the parts
nth task unloading is done
Each time robot comes toward human will
generate a event
Master pin is loaded
Master pin is not loaded
Experiment is complete
Experiment
When robot change state between Normal,
Reduced, and Stop
When robot going to complete stop
When robot is going to normal speed
Definition
activity. ECG can be used as a psychophysiological indicator
for physical stress, mental stress and fatigue. In an industrial
setup, robot behavior can be adjusted based on the state of
health of the operator. This can help in avoiding injuries that
may result from work exhaustion. [13].
• Electromyography (EMG) is method to record electrical
activity generated by muscles. EMG have been used as a
control input for basic robot interaction. A sense of control
is very important for building the trust of human. Another
example of EMG is using facial muscles to give information
about sudden emotional change or reaction. Placement of
these can be in safety glasses worn by the operator [8], [14],
[15] [16].
• Galvanic Skin Response (GSR) also known as Skin Con-
ductivity (SC) or Electro Dermal Activity (EDA), measures
skin conductivity which is triggered by the central nervous
system. This signal has been used in for emotion recognition,
lie detector and detecting physical and mental stress [8], [13],
[17], [18] [19].
• Heart Rate (HR) and Heart Rate Variability (HRV)
is a signal that can be extracted from the ECG and also
photoplethysmogram (PPG) signal. This information can give
the state of the person i.e. Resting or Active. HRV has been
used as a psychophysiological indicator.
• Pupil Dilation is a measurement of pupil diameter change.
The pupil dilation can be caused by ambient light change in
environment and emotional change. [20].
Event Marker
I
y
d
u
t
S
Experiment start
Task [n] init
Task [n] start
Task [n] end
Robot approaching
Pick up successful
Pick up failed
Experiment end
I Experiment start
e
s
a
C
I
y
d
u
t
S
e
s
a
C
Robot state change
Robot is stopping
Robot is speeding up
Robot is slowing down When robot is slowing down.
Experiment end
Experiment is complete
on the human head. EEG has been used for error related
potentials, emotional valence scale and evoked potentials. It
has also been used to detect alpha activity, which determines
attentiveness, stress, and other emotions. It can be questioned
that wearing an EEG cap while working can be uncomfort-
able. However, it must be noted that in industry, workers
can wear helmets or hats. With the advent of advance, IoT
systems wireless communication and small size factor of
EEG equipment make it plausible to get such data. e.g., g.Tec,
BioRadio, and openBCI.
• Electrocardiogram (ECG) measures the heart's electrical
Virtual World (Digital-Twin)Physical WorldRobot informationExperimentEvent generationGSR signal streamPPG signal streamECG signal streamPupil signal and Gaze streamMinimum distanceHead orientationHuman pose and velocityHuman Action representationLab Stream Layer (LSL)Camera Stream *notusing LSL layerData Visualization Tool for Real-Time Monitoring and AnalysisXDF File & ROS BagMonitoring CameraGaze Tracking CameraRight EyeLeft EyeDigital TwinLeft Eye Pupil DilationLeft Eye Pupil ConfidenceRight Eye Pupil DilationRight Eye Pupil ConfidenceGSR SignalPPG and Heart Beat SignalRobot Tool MomentumHuman-Robot Separation Distance C. Data Transfer and Signal Synchronization
REFERENCES
[1] S. Kumar and F. Sahin, "A framework for an adaptive human-robot
collaboration approach through perception-based real-time adjustments
of robot behavior in industry," in System of Systems Engineering
Conference (SoSE), 2017 12th.
IEEE, 2017, pp. 1 -- 6.
[2] ISO, "ISO/TS 15066:2016 - robots and robotic devices -- collaborative
robots." [Online]. Available: http://www.iso.org/
[3] S. Kumar, C. Savur, and F. Sahin, "Dynamic Awareness of an Industrial
Robotic Arm Using Time-of-Flight Laser-Ranging Sensors," in 2018
IEEE International Conference on Systems, Man, and Cybernetics
(SMC), Oct. 2018, pp. 2850 -- 2857.
[4] D. Kulic and E. Croft, "Anxiety detection during human-robot inter-
action," 2005 IEEE/RSJ International Conference on Intelligent Robots
and Systems, IROS, pp. 389 -- 394, 2005.
[5] D. Kul´ıc and E. Croft, "Affective state estimation for human-robot
interaction," in IEEE Transactions on Robotics, vol. 23, no. 5, 2007,
pp. 991 -- 1000. [Online]. Available: http://ieeexplore.ieee.org/document/
4339537/
[6] L. Tiberio, A. Cesta, and M. Belardinelli, "Psychophysiological
Methods to Evaluate User's Response in Human Robot Interaction: A
Review and Feasibility Study," Robotics, vol. 2, no. 2, pp. 92 -- 121,
2013. [Online]. Available: http://www.mdpi.com/2218-6581/2/2/92/
[7] S. Kumar, S. Arora, and F. Sahin, "Speed and separation mon-
laser -- ranging sensor arrays."
time -- of -- flight
[8] D. Kuli´c and E. Croft, "Physiological and subjective responses to
articulated robot motion," Robotica, vol. 25, no. 1, pp. 13 -- 27, 2007.
[9] C. Savur, S. Kumar, S. Arora, T. Hazbar, and F. Sahin, "HRC-SoS:
Human robot collaboration experimentation platform as system of sys-
tems," in 2019 14th Annual Conference System of Systems Engineering
(SoSE), May 2019, pp. 206 -- 211.
itoring using on-robot
arXiv:1904.07379v1, Apr. 2019.
AcceptedinS M C 2019
[16] R. Chalapathy and S. Chawla,
[10] T. Cichon and J. Rossmann, "Simulation-based user interfaces for digital
twins: Pre-, in-, or post-operational analysis and exploration of virtual
testbeds," 31st Annual European Simulation and Modelling Conference
2017, ESM 2017, pp. 365 -- 372, 2017.
[11] M. Safeea and P. Neto, "Minimum distance calculation using laser
interaction," Robotics and
scanner and IMUs for safe human-robot
Computer-Integrated Manufacturing, vol. 58, pp. 33 -- 42, Aug. 2019.
[12] SCCN, "Lab Stream Layer (LSL)," 2018.
[13] M. Ali, F. Al Machot, A. H. Mosa, M. Jdeed, E. Al Machot, and
K. Kyamakya, "A globally generalized emotion recognition system
involving different physiological signals," Sensors (Switzerland), vol. 18,
no. 6, pp. 1 -- 20, 2018.
[14] K. Gouizi, F. Bereksi Reguig, and C. Maaoui, "Emotion recognition from
physiological signals," Journal of Medical Engineering and Technology,
vol. 35, no. 6-7, pp. 300 -- 307, 2011.
[15] C. Savur and F. Sahin, "American Sign Language Recognition system
by using surface EMG signal," 2016 IEEE International Conference on
Systems, Man, and Cybernetics, SMC 2016 - Conference Proceedings,
pp. 2872 -- 2877, 2017.
Detection: A Survey," pp. 1 -- 50, 2019.
//arxiv.org/abs/1901.03407
"Deep Learning for Anomaly
[Online]. Available: http:
[17] S. Rohrmann, J. Hennig, and P. Netter, "Changing psychobiological
stress reactions by manipulating cognitive processes," International
Journal of Psychophysiology, vol. 33, no. 2, pp. 149 -- 161, 1999.
[18] K. H. Kim, S. W. Bang, and S. R. Kim, "Emotion recognition system
using short-term monitoring of physiological signals," Medical and
Biological Engineering and Computing, vol. 42, no. 3, pp. 419 -- 427,
2004.
[19] M. van Dooren, J. J. J. de Vries, and J. H. Janssen, "Emotional sweating
across the body: Comparing 16 different skin conductance measurement
locations," Physiology and Behavior, vol. 106, no. 2, pp. 298 -- 304, 2012.
[Online]. Available: http://dx.doi.org/10.1016/j.physbeh.2012.01.020
[20] P. Bonifacci, L. Desideri, and C. Ottaviani, "Familiarity of faces: Sense
or feeling? An exploratory investigation with eye movements and skin
conductance," Journal of Psychophysiology, vol. 29, no. 1, pp. 20 -- 25,
2015.
The proposed framework in Figure 1 for monitoring human
response during Human Robot Collaborative task uses LSL
layer as the core for transportation and synchronization. Using
LSL layer as the core brings many advantages. The first and
most important reason is that it has built in time synchro-
nization. In addition to synchronization, it allows developer
to use external timer as well. The second most important
feature is the LSL layer is operating system agnostic. This
bring flexibility to the proposed framework, since there are
sensor manufacturers have device drivers that supports only
certain operating systems.
Although LSL layer has the ability to record signal from the
stream as an XDF file, the proposed framework uses ROSbag
as an alternative for recording. Rosbag is a popular tool in
robotic application to record time-series data and replaying
data from collected bags. In addition, it has tools helps plotting
the stream from the bags. Hence it is selected as parallel
recording with LSL layer.
Figure 5 shows proposed framework. In the Figure each
device has an application node which push data to LSL layer.
Then LSL layer deliver data to two receivers, LabRecorder and
LSL2Bag application which are responsible to record data in
to a file.
V. CONCLUSION AND FUTURE WORK
In this research, a framework for monitoring and collecting
human physiological response during human robot collabora-
tive task is presented and a prototype implementation is shown.
The challenges of data communication, signal synchroniza-
tion and event markers are addressed and solution proposed.
The implementation shows the synchronized and continuous
collection of human-robot states and human physiological
responses. This system is expandable for additional sensors.
Although the framework designed for human robot collabora-
tion task, it is not limited to this setup. Similar approach can
be taken for other 'physiological computing' systems.
Future research will focus on developing a complete user
interface application of the 'physiological computing' system
for processing of recording signals, extracting information and
applying machine-learning algorithm to provide feedback to
the robot. The final objective of this work is to generate a
database that can be used to further the understanding of how
human physiological responses can be inferred to result in
adaptive robot motion behavior.
ACKNOWLEDGMENT
The authors would like to thank the Electrical Engineering
to the staff
Department at RIT. The authors are grateful
of Multi Agent Bio-Robotics Laboratory (MABL) and the
CM Collaborative Robotics Research (CMCR) Lab for their
valuable inputs.
|
1602.02032 | 1 | 1602 | 2016-02-05T14:19:34 | Track selection in Multifunction Radars for Multi-target tracking: an Anti-Coordination game | [
"cs.MA",
"cs.IT",
"cs.IT"
] | In this paper, a track selection problem for multi-target tracking in a multifunction radar network is studied using the concepts from game theory. The problem is formulated as a non-cooperative game, and specifically as an anti-coordination game, where each player aims to differ from what other players do. The players' utilities are modeled using a proper tracking accuracy criterion and, under different assumptions on the structure of these utilities, the corresponding Nash equilibria are characterized. To find an equilibrium, a distributed algorithm based on the best-response dynamics is proposed. Finally, computer simulations are carried out to verify the effectiveness of the proposed algorithm in a multi-target tracking scenario. | cs.MA | cs | TRACK SELECTION IN MULTIFUNCTION RADARS FOR MULTI-TARGET
TRACKING: AN ANTI-COORDINATION GAME
Nikola Bogdanovi´c, Hans Driessen, Alexander Yarovoy
Microwave Sensing, Signals and Systems
Delft University of Technology, The Nederlands
E-mails: {N.Bogdanovic, J.N.Driessen, A.Yarovoy}@tudelft.nl
6
1
0
2
b
e
F
5
]
A
M
.
s
c
[
1
v
2
3
0
2
0
.
2
0
6
1
:
v
i
X
r
a
ABSTRACT
In this paper, a track selection problem for multi-target tracking
in a multifunction radar network is studied using the concepts from
game theory. The problem is formulated as a non-cooperative game,
and specifically as an anti-coordination game, where each player
aims to differ from what other players do. The players' utilities are
modeled using a proper tracking accuracy criterion and, under dif-
ferent assumptions on the structure of these utilities, the correspond-
ing Nash equilibria are characterized. To find an equilibrium, a dis-
tributed algorithm based on the best-response dynamics is proposed.
Finally, computer simulations are carried out to verify the effective-
ness of the proposed algorithm in a multi-target tracking scenario.
Index Terms -- Multiple target tracking, track selection, non-
cooperative games, coordination, Nash equilibrium.
1. INTRODUCTION
Radar networks that employ multiple, distributed stations offer sig-
nificant advantages over standalone radars, in terms of providing di-
versities and enhancing tracking and detection performance. Fur-
thermore, recent advances in sensor technologies enabled a large
number of controllable degrees of freedom in modern radars. One
such system is the Multifunction Radar (MFR), which employs an
electronically scanned phased array composed of individually con-
trolled radiating elements [1]-[3]. Due to its beam and waveform
agility, the MFR is capable to track multiple targets and perform
new target search in the sector. Thus, the MFR is much more flexible
than conventional, dedicated radars by being capable of performing
different functions - volume surveillance, weapon control, and mul-
tiple target tracking to name a few. In this paper, we focus on the
latter function [4]-[5]; specifically, each MFR radar aims at tracking
several targets.
Even for a standalone MFR, the radar resource management
plays a crucial role so as to efficiently allocate resources to achieve
specified objectives while conforming to operational and techni-
cal constraints [6], [7]. Most of the existing approaches to MFR
radar resource management roughly fit into the following two cat-
egories [8], [9]. The first category consists of the rule-based tech-
niques [10]-[12], which control the resource allocation parameters
indirectly, under low computational burden. However, it is hard to
say what performance can be achieved since it highly depends on
The work reported in this paper has been conducted as part of the Sen-
sor Technology Applied in Reconfigurable systems for sustainable Security
(STARS) project, see the website www.starsproject.nl; the multi-site tracking
algorithm that is used in this paper has been developed under the SOS project
funded by the European Commission (FP7 GA no. 286105).
the application scenario and on the sensors being deployed. The
other category is related to the methods that formulate the problem
as an optimization one; and thus, they may achieve the optimal
performance, see [13]-[15], [1], [5] and the references therein.
Note that, in the network setting, the first category of approaches
is difficult to be extended, while the second one may involve ex-
cessive complexity due to the network dimension. To reduce such
complexity, in this work we propose a distributed approach based on
game theory so as to model track selection for multi-target tracking
in an MFR network.
Game theory is the mathematical study of conflict and cooper-
ation between intelligent rational decision-makers [16]. In addition
to its traditional research areas such as economics and political sci-
ences, over the last decade game theory (GT) is being applied to
signal processing and wireless communications. This is mainly due
to the issues dealing with (distributed) networking [17], [18], such as
power control [19], antennas' beamforming [20], multiple input mul-
tiple output (MIMO) communications [21], channel allocation [22],
adaptive estimation [23], [24], to name a few. More recently, GT
has been applied to solve certain radar problems, mostly related to
the MIMO radar networks. For instance, the problem of waveform
design has been tackled; in [25] by formulating a two player zero-
sum (TPZS) game between the radar design engineer and an oppo-
nent, and in [26] by a potential game in which the radars choose
among the pre-fixed transmit codes. Next, the interaction between
a smart target and a MIMO radar is modeled as a TPZS game [27],
where the mutual information criterion was used in the utility func-
tions. Also, the problem of transmission power management was
addressed in [28], [29]. Initially, the power control problem was for-
mulated in [28] assuming the presence of some interference due to
the other radars' transmissions. A non-cooperative game was used
for modeling and a distributed algorithm converging to a Nash equi-
librium was proposed. On the other hand, in reference [29], a coali-
tional game theoretic solution concept called the Shapley value was
employed to distribute a given power budget among all transmitting
radars. More related to the application scenario in this article, the
work in [30] utilize a market mechanism, called the continuous dou-
ble auction, in order to choose the global optimum parameters for
each individual task given the global (finite) resource constraint.
In contrast to the aforementioned literature, in this paper we
formulate a new problem of track selection for a multi-target track-
ing scenario in a resource-limited MFR network using the non-
cooperative games, the dominant branch of GT [16]. An example
of such a scenario is depicted in Fig. 1. To the best of our knowl-
edge, this is the first non-cooperative GT contribution dealing with
multi-target tracking. Due to its nature, the problem is modeled as a
coordination game which is known to have several Nash equilibria.
measurements {z(i)
j,k} related to the previously selected targets. The
number of transmissions each target j is tracked by at one time slot
is denoted as mt
j. For notational simplicity, in the rest of this section
we drop the index j for targets.
At each radar i and for each target j, the tracking process is per-
formed by an Extended-Kalman Filter (EKF). Firstly, the prediction
step occurs, i.e.,
xkk−1 = F · xk−1k−1
Pkk−1 = F Pk−1k−1F T + Q
(3)
(4)
where xkk−1 and Pkk−1 are the state estimate and the error co-
variance matrix for time step k given all measurements till time step
k − 1. Then, the updating step takes place where each available
measurement for target j of some radar n ∈ N is used in a cyclic
manner. In particular, for each p ∈ {1, . . . , mt
j},
k,n]T(cid:16)
(cid:16)
(cid:17)
k
[H (p)
k = P (p−1)
K (p)
kk
kk = x(p−1)
x(p)
kk + K (p)
I − K (p)
k H (p)
P (p)
kk =
(cid:16)
k,n
k,nP (p−1)
H (p)
k − h(n)(cid:16)
kk
z(n)
P (p−1)
kk
(cid:17)−1
k,n]T + Rn
(cid:17)(cid:17)
[H (p)
x(p−1)
kk
(5)
(6)
(7)
where P (p)
kk denotes the error covariance matrix after p incremental
updates at the same time step k, with P (0)
kk =
xkk−1. The linearized measurement matrix of radar n at time k
is H (p)
. Note that, due to the
fact that the position of each radar is known, the radars do not need
to exchange {Hk,n} matrices in order to implement the algorithm
above.
k,n = ∂h(n)/∂x evaluated at x(p−1)
kk
kk = Pkk−1 and x(0)
In the following, we study a natural game theoretic variant of
this problem. Specifically, we assume that the radars are autonomous
decision-makers interested in optimizing their own tracking perfor-
mance. The fact that each radar autonomously and rationally decides
to track the targets that increase its utility can be modeled as a one-
stage non-cooperative game in normal form.
3. GAME-THEORETIC MODEL
Here, we formulate the track selection problem in multi-target track-
ing as a non-cooperative game in normal form, which is the most
fundamental representation type in game theory [16]. Note that there
are many classes of normal-form games; however, in this work we
focus on coordination games, which do not rest solely upon conflict
among players. Instead, as their name suggests, more emphasis is
put on the coordination issue where players may have an incentive
to conform with or to differ from what others do. In the latter case,
this kind of games are usually called anti-coordination games [16],
[31]-[32].
We assume that the players are rational and their objective is to
maximize their payoff, i.e., the tracking accuracy of all targets. For-
mally, the track selection game (N ,S, u) has the subsequent com-
ponents:
• The players are the radars represented by the set N .
• The strategy of each radar i is represented by a T -tuple
si = (si,1, si,2, . . . , si,T ) where si,j = a if radar i de-
votes a transmission beams to a target j, with a ≤ m. Each
j=1 si,j ≤
i=1 si,j. Each vector s =
strategy-tuple has at most m transmissions, i.e.,(cid:80)T
j = (cid:80)N
m. Also, note that mt
Fig. 1. A track selection problem in multi-target tracking.
Then, the equilibria are characterized in terms of their existence con-
ditions and efficiency. Finally, to find an equilibrium, a distributed
algorithm based on the best-response dynamics is proposed and its
effectiveness for the track selection issue in multi-target tracking is
demonstrated.
2. PROBLEM FORMULATION
Suppose that there are multiple MFR radars and several targets to be
tracked whose number is known exactly and their current positions
approximately, see Fig. 1. We denote the set of radars by N , while
the set of targets is denoted by T . The targets are assumed to be
well-separated; thus there is no data association problem and differ-
ent transmission beams are required so as to illuminate the targets.
Consider that there is no fusion center and that each radar aims at
tracking all targets simultaneously. Next, each target j ∈ T , at each
discrete time k, follows the so-called white noise constant velocity
model [4], [13] given by
xj,k = F · xj,k−1 + wj,k−1
z(i)
j,k = h(i)
j (xj,k) + ν(i)
j,k
(1)
(2)
0 1
(cid:3) ⊗ I2, with ⊗ being the Kronecker
dynamics given as F = (cid:2) 1 tu
w ·(cid:104) t3
where the state vector x for each target j is comprised of the two di-
mensional coordinates and velocity, i.e., xj = [xj, yj, vj,x, vj,y]T ,
while F is a 4 × 4 matrix corresponding to the deterministic target
product, I2 stands for a 2 × 2 identity matrix and tu is the update
time that is fixed. The process noise w is Gaussian with zero mean
w models ma-
and covariance Q = σ2
neuverability. At each radar i ∈ N , the measurement vector z(i)
j,k
consists of range and azimuth, i.e., z(i)
, while the
nonlinear transformation h(i)
(cid:105) ⊗ I2, where σ2
(cid:105)T
(cid:21)
j (xj) is given by
u/3 t2
t2
u/2 tu
r(i)
j,k, a(i)
j,k =
(cid:104)
u/2
j,k
arctan((yj − yi)/(xj − xi))
(cid:20) (cid:112)(xj − xi)2 + (yj − yi)2
(cid:110)
j
.
The coordinates (xi, yi) of each radar i ∈ N are assumed known.
Finally, the measurement noise ν(i)
is zero-mean Gaussian with co-
rj ]2, [σ(i)
variance Rj,i = diag
aj ]2(cid:111)
[σ(i)
.
The radars have limited time budget in sense that they cannot
take measurements of all targets during the same time slot. The num-
ber of measurements per scan that each radar can make is given by
m < T . Since there is no central entity that may coordinate actions
of the radars, a distributed solution is needed. The interaction among
the radars is existing but limited to sharing, e.g., by broadcasting, the
h(i)
j (xj) =
{𝑧𝑧𝑗𝑗,𝑘𝑘(𝑖𝑖)} Target 1 Target 2 Target 3 Radar 1 Radar 2 ? ? (cid:26)
Tr
0,
(cid:27)
(a)
(b)
Fig. 2. An example of a track allocation where T = {a, . . . , e},
and N = m = 3. Each box represents a gain increment due to
a measurement, and the number of measurements per target, mt
j,
varies between 1 and 4 across targets in T . In case (a), the gains are
equal for the same number of measurements, while in case (b) they
differ.
(s1, . . . , sN ) ∈ S is called an action or strategy profile, and
s−i = (s1, . . . , si−1, si+1, . . . , sN ) is defined as a strategy
profile s without player i's strategy.
• The utility for each radar i is given by
(cid:0)gainj(mt
T (cid:88)
j=1
(cid:1) ,
(8)
j) − c · τj
ui(si, s−i) =
where the term gainj(mt
gain for target j ∈ T and it is defined by
j) represents the tracking accuracy
gainj(mt
j) =
Pj,kk−1 − P
(mt
j )
j,kk
,
if mt
j ≥ 1
otherwise
(9)
where all radars are assumed to have the same initial guesses
xj,00 and Pj,00. Finally, the normalization coefficient c in-
dicates delay importance, while τj is the delay function given
as
(cid:40)
τj =
if mt
j ≥ 1
0,
1, otherwise
.
(10)
In other words, the strategy of radar i defines the number of
transmissions per each target, at a given time slot. Due to the fact that
radars share their measurements, their tracking accuracy gains for a
specific target are dependent on all radars' measurements related to
that target. If set to be non-negative, the delay importance coefficient
can serve as a mechanism for punishing radars in case where not all
targets are being tracked. Otherwise, if set negative enough, it gives
incentives to radars to focus all their resources on less targets. In our
study, it is of interest to study the former case.
(cid:0)mt
(cid:1) =(cid:80)mt
p = Tr{P (p−1)
In practice, the gain function in (9) can be assumed to be increas-
ing in the number of measurements, at least in mean sense. Note
p=1 ∆g(j)
that the gain in (9) can be expressed as gainj
p ,
where ∆g(j)
1 = gainj(1). Also,
in a real system it reasonable to assume that an estimation accu-
racy gain increment ∆g(j)
p decreases as the order of measurements
p grows, i.e., ∆g(j)
p+1. For the analysis in the sequel, the
following two cases are distinguished:
j,kk} and ∆g(j)
j,kk − P (p)
p > ∆g(j)
j
j
p = ∆gp, for all j ∈ T and p ∈ {1, . . . , mt
j}
for j
(cid:54)= (cid:96), and minj∈T ∆g(j)
p
>
a) ∆g(j)
b) ∆g(j)
p
(cid:54)= ∆g((cid:96))
p ,
p+1.
maxj∈T ∆g(j)
Case a) represents an idealistic case where all nodes would have very
similar measurements among themselves and related to all targets,
see Fig 2(a). A more realistic scenario, corresponding to case b), is
illustrated in Fig 2(b). In the following section, the track selection
game for both cases will be analyzed.
4. NASH EQUILIBRIA
In general, one can reason about multiplayer games using solution
concepts, i.e., principles according to which interesting outcomes of
a game can be identified. Although there are many solution concepts
in the game-theory literature, a basic and the most widely accepted
one is the Nash equilibrium. Formally, in case in which players make
deterministic choices (pure strategies) the Nash equilibrium is de-
fined as follows [16].
Definition 1. A strategy profile s = (s1, . . . , sN ) is a pure-strategy
i (cid:54)=
Nash equilibrium (NE) if, for all players i and for all strategies s(cid:48)
si, it holds that ui(si, s−i) ≥ ui(s(cid:48)
i, s−i).
In other words, in an NE, no player can unilaterally improve its
utility by taking a different strategy. Also, it is important to define
Pareto domination and Pareto optimality.
Definition 2. Strategy profile s Pareto dominates strategy profile s(cid:48)
if ∀i ∈ N , ui(s) ≥ ui(s(cid:48)), and there exists some n ∈ N for which
un(s) > un(s(cid:48)). Also, strategy profile s is Pareto optimal (PO)
if there does not exist another strategy profile s(cid:48) ∈ S that Pareto
dominates s.
Generally, in a coordination game, there are multiple NE. If the
players have the same payoffs, and the equilibria are equal, the game
is a pure coordination one. In fact, in such a game, all NE are PO.
On the other hand, in a ranked one, the NE differ and usually there
is only one PO equilibrium [33].
Now, the main findings related to the NE for cases (a) and (b)
are provided.
Proposition 1. For case (a), any track assignment is a PO NE, if
c ≥ 0 and(cid:80)T
j=1 si,j = m, and if
j − mt
• mt
j ≤ 1, ∀j ∈ T , for a scenario where N · m ≤ T
• maxj,(cid:96)∈T {mt
N · m > T .
Firstly, let us assume that there is a radar i such that(cid:80)T
(cid:96)} ≤ 1, ∀j, (cid:96) ∈ T , for a scenario where
j=1 si,j <
m and that the corresponding s∗ is an NE. Then, radar i can change
its strategy by taking an additional measurement. Due to the fact
that the radar's gain function in (9) is increasing in the number of
measurements, its utility will be increased. But that contradicts our
initial assumption that s∗ is an NE; thus, as per our intuition, each
radar should make all possible transmissions toward the target(s) at
each time instant. Next, note that if the total number of measure-
ments is less than or equal to the number of targets, the condition
related to c ensures that the radars are punished if more than one
measurement in total is devoted to the same target. Also, due to the
structure of gain function, NE are precisely
(T −N·m)! outcomes
in which each measurement is devoted to a distinct target. On the
other hand, if N · m > T , condition c ≥ 0 promotes all targets
to be covered. Here, each NE corresponds to a balanced allocation.
For instance, the allocation in Fig 2(a) is not an NE since the payoffs
can be increased if some player moves its measurement from target
b to any other target. Finally, since the gain of any target is the same
for the same number of measurements, the game appears to be a
pure anti-coordination one. Thus, every NE is also Pareto optimal.
T !
targets gains a b c d e ∆𝑔𝑔4 𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔(1) } targets gains a b c d e Proposition 2. For case (b), any track assignment is an NE, if c ≥ 0
and(cid:80)T
j=1 si,j = m, and if,
• for N · m > T , the first
• for a scenario with N · m ≤ T , each radar chooses its
most accurate target that has not been selected,
(cid:108) N·mT
(cid:109) − 1, ∀j ∈ T , and for the
(cid:109) − 1 levels are filled in,
(cid:109)
(cid:108) N·mT
j ≥ (cid:108) N·mT
i.e., mt
-th
level each radar chooses its most accurate target that has not
been selected by others, where (cid:100)·(cid:101) is the ceiling function.
Similar arguments hold as for Prop. 1. Yet, the game above
seems to be a ranked anti-coordination game. Note that here there
are still multiple NE, but not all NE are necessarily equal, and hence,
not every NE is Pareto optimal (only one is). So, the conditions
above are not sufficient to have also a Pareto optimal NE.
To conclude this section, we provide a simple, distributed algo-
rithm, based on the best response dynamics [16], [17], to achieve an
NE. In the games above where N · m > T , in general, two types
of NE may arise, one where a radar illuminates only different targets
and the other where it chooses the same target more than once. In
practice, it is of interest to exploit the radars' diversity; thus, we fo-
cus on the former type. Let Ti denote the set of targets selected by
radar i. Then, for any initialization, at each time instant each radar
i ∈ N performs the following steps:
• Count mt
j, ∀j ∈ T , and reallocate the measurements for
∀j ∈ Ti satisfying si,j > 1 to a target argmin(cid:96)∈{T /Ti}mt
(cid:96).
• With probability α, reallocate the measurement from target j
to (cid:96)
(cid:108) N·mT
(cid:109)
-- if ∃j ∈ Ti such that mt
and the measure-
ment for (cid:96) is the most accurate one of those satisfying
argminq∈{T /Ti}mt
q, or
j >
-- if mt
j − mt
(cid:96) = 1, where mt
(cid:96) =
q, and if measurement for (cid:96) is more ac-
j = maxq∈Ti mt
q and mt
minq∈{T /Ti}mt
curate than the one for j.
• Transmit/receive measurements, and ∀j ∈ T , execute (3)-(4)
and employ all available measurements in (5)-(7).
To account for time-varying accuracy measures, e.g., range and/or
azimuth variances, or to deal with different model dynamics, algo-
rithm can be reinitialized every K time instants so as to search for
other NE during the tracking process.
5. SIMULATIONS
In this section, we will demonstrate the performance of the proposed
algorithm for track selection in multi-target tracking.
We consider an MFR network of N = 3 radars, each of
them making m = 2 measurements per scan and aiming at track-
ing T = 5 targets. The coordinates of radars are (x1, y1) =
[−10 km, 0 km], (x2, y2) = [3 km, 0 km] and (x3, y3) =
[10 km, 0 km]. The targets follow white noise constant veloc-
ity trajectories with initial x, y-coordinates and velocities x1,0 =
[1 km, 6 km, 0.5 km/s, 0.1 km/s]T , x2,0 = [0.5 km, 7 km,
0.35 km/s, −0.1 km/s]T , x3,0 = [1.5 km, 3.0 km, −0.3 km/s,
0 km/s]T , x4,0 = [2.0 km, 4.0 km, −0.2 km/s, 0.1 km/s]T
and x5,0 = [2.5 km, 5.0 km, 0.3 km/s, 0.2 km/s]T .
Initial
guesses xj,00 are noisy versions of the initial states xj,0 and ini-
tial covariances are equal to Pj,00 = P00 = diag(cid:8)(0.1 km)2,
(0.1 km)2, (0.1 km/s)2, (0.1 km/s)2(cid:9). The update time is tu =
Fig. 3. Sum of traces of error covariance matrices for all targets
during time.
0.25 s, and in order to model moderate maneuverability, σ2
w is set
to 2.5 · 10−5 km2/s3. Also, the standard deviation in azimuth
is σ(i)
aj = σa = 2 mrad, while the range accuracy varies among
rj = bi,j · σr, where σr = 15 m and
the radars and targets as σ(i)
bi,j is taken from the interval [1, 4.5]. Next, a comparison is made
among the following strategies: (a) the standalone radar that does
not send/receive measurements, and sequentially chooses m = 2
different targets at each time instant; (b) distributed strategy where
the radars exchange the measurements while each of them ran-
domly changes its selection at each K = 10 time instants; (c) same
as in (b), except that targets are being randomly chosen at each
time instant; (d) the proposed distributed algorithm seeking NE for
α = 0.4 while being reinitialized with K = 10; and (e) an approx-
imated centralized approach based on exhaustive search for optimal
measurements-to-target allocation every K = 10 time instants.
The results in Fig. 3 are averaged over 100 realizations. Not
surprisingly, due to the high process' dynamics, a standalone, non-
cooperative radar experiences weak performance since it utilizes
only its own measurements which are not sufficient to cover all tar-
gets. However, although approach in (b) uses 3x2=6 measurements,
due to the lack of coordination it performs poorly. Note that the dis-
tributed random strategy can be significantly improved if strategies
are constantly being changed, given that there are no track migration
costs involved. However, the proposed distributed algorithm that
learns underlying NE outperforms aforementioned strategies and
closely approaches to the performance of the centralized one while
being much more efficient in terms of complexity.
6. CONCLUSIONS
We have formulated a track selection problem for multi-target track-
ing in a network of MFR radars. The problem has been tackled using
the non-cooperative game theory. The Nash equilibria of the under-
lying anti-coordination games have been analyzed and a simple yet
effective distributed algorithm that looks for the equilibria points has
been proposed. It introduces a balancing effect in the track selection
which makes it be particularly convenient for the settings with high
dynamics. Also, it closely approximates the centralized performance
while mitigating its inherent complexity. Our future work may con-
sider extending the results for different communication topologies
and for cases where not all radars have the same interests.
0510152025303540455010−310−210−1100Time, kSum of Trace(P) Standalone − no shareDistributed random − K= 10Distributed random − K= 1Proposed distributed − K=10Centralized − K=10 7. REFERENCES
[1] A. O. Hero and D. Cochran,
"Sensor management: Past,
present, and future," Sensors Journal, IEEE, vol. 11, no. 12,
pp. 3064 -- 3075, 2011.
[2] S. Sabatini and M. Tarantino, Multifunction array radar,
Artech House, 1994.
[3] M. A. Richards, J. Scheer, and W. A. Holm, Principles of
modern radar, SciTech Publishing, 2010.
[4] S. S. Blackman and R. Popoli, Design and Analysis of Modern
Tracking Systems, Artech House radar library. Artech House,
1999.
[5] S. P. Sira, Y. Li, A. Papandreou-Suppappola, D. Morrell,
D. Cochran, and M. Rangaswamy, "Waveform-agile sensing
for tracking," IEEE Signal Processing Magazine, vol. 26, no.
1, pp. 53 -- 64, Jan 2009.
[6] Z. Ding, "A survey of radar resource management algorithms,"
in Proc. Canadian Conference on Electrical and Computer En-
gineering, May 2008, pp. 001559 -- 001564.
[7] F. Katsilieris, Sensor management for surveillance and track-
ing: An operational perspective, Ph.D. thesis, TU Delft, Delft
University of Technology, 2015.
[8] A. Charlish, Autonomous agents for multi-function radar re-
source management, Ph.D. thesis, UCL (University College
London), 2011.
[9] A. S. Narykov, O. A. Krasnov, and A. Yarovoy, "Algorithm
for resource management of multiple phased array radars for
target tracking," in Proc. 16th Int. Conference on Information
Fusion (FUSION 2013). IEEE, 2013, pp. 1258 -- 1264.
[10] G. van Keuk and S. S. Blackman,
tracking and parameter control,"
Aerospace Electronic Systems, vol. 29, pp. 186 -- 194, 1993.
"On phased-array radar
IEEE Transactions on
[11] W. Koch,
"On adaptive parameter control for phased-array
tracking," in Signal and Data Processing of Small Targets,
1999, pp. 444 -- 455.
[12] S. Coetzee, K. Woodbridge, and C. Baker,
"Multifunc-
tion radar resource management using tracking optimisation,"
Tech. Rep., DTIC Document, 2005.
[13] J. H. Zwaga and H. Driessen, "Tracking performance con-
strained mfr parameter control: applying constraints on pre-
diction accuracy," in Proc. 8th International Conference on
Information Fusion, 2005, July 2005, vol. 1, pp. 546 -- 551.
[14] J. Hansen, R. Rajkumar, J. Lehoczky, and S. Ghosh, "Resource
management for radar tracking," in IEEE Conference on Radar,
2006. IEEE, 2006, p. 8.
[15] V. Krishnamurthy and D. V. Djonin, "Optimal threshold poli-
cies for multivariate POMDPs in radar resource management,"
IEEE Transactions on Signal Processing, vol. 57, no. 10, pp.
3954 -- 3969, Oct 2009.
[16] Y. Shoham and K. Leyton-Brown, Multiagent Systems: Al-
gorithmic, Game-Theoretic, and Logical Foundations, Cam-
bridge University Press, 2008.
[17] Z. Han, D. Niyato, W. Saad, T. Baar, and A. Hjørungnes, Game
Theory in Wireless and Communication Networks: Theory,
Models, and Applications, Cambridge University Press, New
York, NY, USA, 1st edition, 2012.
[18] G. Bacci, S. Lasaulce, W. Saad, and L. Sanguinetti, "Game
arXiv preprint
theory for signal processing in networks,"
arXiv:1506.00982, 2015.
[19] C. U. Saraydar, N. B. Mandayam, and D. Goodman, "Efficient
power control via pricing in wireless data networks," IEEE
Transactions on Communications, vol. 50, no. 2, pp. 291 -- 303,
Feb 2002.
[20] E. G. Larsson and E. A. Jorswieck, "Competition versus coop-
eration on the MISO interference channel," IEEE Journal on
Selected Areas in Communications, vol. 26, no. 7, pp. 1059 --
1069, September 2008.
[21] G. Scutari, D. P. Palomar, and S. Barbarossa, "Competitive
design of multiuser mimo systems based on game theory: A
unified view," IEEE Journal on Selected Areas in Communica-
tions, vol. 26, no. 7, pp. 1089 -- 1103, 2008.
[22] M. Felegyhazi, M. Cagalj, and J.-P. Hubaux, "Efficient mac
in cognitive radio systems: A game-theoretic approach," IEEE
Transactions on Wireless Communications, vol. 8, no. 4, pp.
1984 -- 1995, April 2009.
[23] C. Yu, M. van der Schaar, and A.H. Sayed, "Reputation de-
sign for adaptive networks with selfish agents," Proc. 2013
IEEE 14th Workshop on Signal Processing Advances in Wire-
less Communications (SPAWC), pp. 160 -- 164, June 2013.
[24] N. Bogdanovi´c, D. Ampeliotis, and K. Berberidis, "Coalitional
game theoretic approach to distributed adaptive parameter es-
timation," in Proc. 2015 IEEE Int. Conference on Acoustics,
Speech and Signal Processing (ICASSP), April 2015.
[25] S. Gogineni and A. Nehorai, "Game theoretic design for po-
larimetric MIMO radar target detection," Signal Processing,
vol. 92, no. 5, pp. 1281 -- 1289, 2012.
[26] M. Piezzo, A. Aubry, S. Buzzi, A. De Maio, and A. Farina,
"Non-cooperative code design in radar networks: a game-
theoretic approach," EURASIP Journal on Advances in Signal
Processing, vol. 2013, no. 1, pp. 1 -- 17, 2013.
[27] X. Song, P. Willett, S. Zhou, and P. B. Luh, "The MIMO radar
and jammer games," IEEE Transactions on Signal Processing,
vol. 60, no. 2, pp. 687 -- 699, Feb 2012.
[28] G. Bacci, L. Sanguinetti, M.S. Greco, and M. Luise, "A game-
theoretic approach for energy-efficient detection in radar sen-
sor networks," in Proc. 7th Sensor Array and Multichannel
Signal Processing Workshop (SAM), June 2012, pp. 157 -- 160.
[29] H. Chen, S. Ta, and B. Sun, "Cooperative game approach to
power allocation for target tracking in distributed MIMO radar
sensor networks," IEEE Sensors Journal, vol. PP, no. 99, pp.
1 -- 1, 2015.
[30] A. Charlish, K. Woodbridge, and H. Griffiths, "Multi-target
tracking control using continuous double auction parameter se-
lection," in Proc. 15th International Conference on Informa-
tion Fusion (FUSION 2012). IEEE, 2012, pp. 1269 -- 1276.
[31] N. Nisan, T. Roughgarden, E. Tardos, and V. V. Vazirani, Al-
gorithmic Game Theory, Cambridge University Press, 2007.
[32] Y. Bramoull´e,
"Anti-coordination and social interactions,"
Games and Economic Behavior, vol. 58, no. 1, pp. 30 -- 49,
2007.
[33] E. Rasmusen, Games and information, Blackwell Publishers,
2001.
|
1912.06880 | 1 | 1912 | 2019-12-14T16:57:56 | Spatial Influence-aware Reinforcement Learning for Intelligent Transportation System | [
"cs.MA",
"cs.AI",
"cs.LG"
] | Intelligent transportation systems (ITSs) are envisioned to be crucial for smart cities, which aims at improving traffic flow to improve the life quality of urban residents and reducing congestion to improve the efficiency of commuting. However, several challenges need to be resolved before such systems can be deployed, for example, conventional solutions for Markov decision process (MDP) and single-agent Reinforcement Learning (RL) algorithms suffer from poor scalability, and multi-agent systems suffer from poor communication and coordination. In this paper, we explore the potential of mutual information sharing, or in other words, spatial influence based communication, to optimize traffic light control policy. First, we mathematically analyze the transportation system. We conclude that the transportation system does not have stationary Nash Equilibrium, thereby reinforcement learning algorithms offer suitable solutions. Secondly, we describe how to build a multi-agent Deep Deterministic Policy Gradient (DDPG) system with spatial influence and social group utility incorporated. Then we utilize the grid topology road network to empirically demonstrate the scalability of the new system. We demonstrate three types of directed communications to show the effect of directions of social influence on the entire network utility and individual utility. Lastly, we define "selfish index" and analyze the effect of it on total group utility. | cs.MA | cs |
Spatial Influence-aware Reinforcement Learning for
Intelligent Transportation System
Wenhang Bao+, Xiao-Yang Liu∗
∗Electrical Engineering, Columbia University,
+Department of Statistics, Columbia University,
Emails: {WB2304, XL2427}@columbia.edu
Abstract
Intelligent transportation systems (ITSs) are envisioned to be crucial for smart
cities, which aims at improving traffic flow to improve the life quality of urban res-
idents and reducing congestion to improve the efficiency of commuting. However,
several challenges need to be resolved before such systems can be deployed, for
example, conventional solutions for Markov decision process (MDP) and single-
agent Reinforcement Learning (RL) algorithms suffer from poor scalability, and
multi-agent systems suffer from poor communication and coordination. In this
paper, we explore the potential of mutual information sharing, or in other words,
spatial influence based communication, to optimize traffic light control policy.
First, we mathematically analyze the transportation system. We conclude that
the transportation system does not have stationary Nash Equilibrium, thereby re-
inforcement learning algorithms offer suitable solutions. Secondly, we describe
how to build a multi-agent Deep Deterministic Policy Gradient (DDPG) system
with spatial influence and social group utility incorporated. Then we utilize the
grid topology road network to empirically demonstrate the scalability of the new
system. We demonstrate three types of directed communications to show the effect
of directions of social influence on the entire network utility and individual utility.
Lastly, we define "selfish index" and analyze the effect of it on total group utility.
1
Introduction
Emerging intelligent transportation systems (ITSs) [1 -- 4] are expected to play an instrumental role in
improving traffic flow, thus optimizing fuel efficiency, reducing delays and enhancing the general
driving experience. ITSs are designed to resolve traffic congestion, which is an exceedingly complex
and important issue faced by metropolitan areas around the world, as a result of global urbanization.
The urbanization process makes traffic a serious problem in the urban area, as there are so many
commuting objects and vehicles. Street interactions in dense urban areas can be critical bottlenecks
in urban road networks, which affect commuters' efficiency. The competing nature of different
intelligent routing applications is not aiming at improving the efficiency of the city as a whole. They
optimize the routing paradigm for their customers locally, which might not be the global optimal
solution.
ITSs are expected to resolve these issues, as advanced communication and computing technologies are
developed to allow efficient information sharing [5, 6] among commuting objects, making intelligent
routing and traffic light control possible [7, 8]. Current traffic light systems are typically hard-coded
based on investigation results or pure experience. The light control policy is optimized based on
historical data and adapted according to daily patterns or drivers' feedback. The opportunity of using
Artificial Intelligence (AI) for adaptation to real-time conditions, e.g. through detection wires in the
Machine Learning for Autonomous Driving Workshop at the 33rd Conference on Neural Information Processing
Systems (NeurIPS 2019), Vancouver, Canada.
pavement, tends to be fairly rudimentary. Evolving technologies offer the option of using information
collected by cameras, including fine-grained knowledge of the positions and speeds of the vehicles.
Such comprehensive real-time information can be leveraged to improve traffic flow through more
agile traffic light control systems. While the potential benefits are immense, so are the technical
challenges that arise in solving such real-time actuation problems on an unprecedented scale in terms
of intrinsic complexity, geographic range, and number of objects involved.
Under suitable assumptions, the problem of optimal dynamic traffic control may be formulated
as a Markov decision process (MDP) [9 -- 11]. The MDP framework provides a rigorous notion of
optimally along the basis for computational techniques such as value iteration, policy iteration or
dynamic programming. However, an MDP formulation may not always be satisfied in reality. The
knowledge of relevant environmental parameters may not be available. Also, the environment is not
stationary. Moreover, in terms of computational cost, an MDP approach suffers from the curse of
dimensionality, resulting in excessively large state spaces in the realistic traffic system. RL algorithms
overcome some of these limitations and have been previously considered in the context of optimal
dynamic traffic light control. Researchers tried to use single-agent algorithms, for example, DDPG
[12], to resolve this issue, but these methods are still prone to prohibitively large state spaces and
action spaces, implying poor scalability beyond a single-interaction scenario.
In this paper, we explore the potential of using multi-agent reinforcement learning algorithms,
particularly Multi-agent Deep Deterministic Policy Gradients (MADDPG) [13], to optimize real-time
traffic light control policies in large-scale systems [14]. First, we formulate the ITSs environment
and the optimization goal. Then we analyze the properties of the Transportation system. We
provide mathematical proof that stationary Nash Equilibrium can not be achieved in this case.
Therefore, Reinforcement Learning algorithms are needed to resolve it. Thirdly, we consider a grid
network topology with multiple rows and columns to examine the scalability properties of MADDPG
algorithms and their performance in the presence of highly complex interactions created by the flow
of vehicles along the main artery. Under this scenario, the influencing network would have complexity
similar to a real city. We develop inward, outward and fully connected spacial influencing flows, to
verify the effect of influencing directions in optimizing traffic signal control. We conclude that the
directions of the spatial influence would not affect total network utility but affect individual utility.
Finally, we define and investigate the "selfish index" of agents to check its influence on group utility.
The remainder of the paper is organized as follows. We first present a detailed traffic environment
description and problem statement. Then we provide a specification of our customized deep rein-
forcement learning algorithm MADDPG for grid topology road network with border interactions.
The properties of the Intelligent Transportation System are also discussed. Also, we empirically
evaluate the performance of the proposed spatial influence based MADDPG algorithms and illustrate
the scalability of this algorithm. Social group utility optimization through reward function adjustment
is also investigated. Finally, we conclude this paper and point out some future directions.
2 Problem Description
We model the road intersections and formulate our problem similar to [15]. We consider a multi-
agent reinforcement learning environment where each agent is responsible for one intersection.
Our central aim here is to design a system and explore the performance and scalability of multi-
agent reinforcement learning algorithms in optimizing real-time traffic light control policies, rather
than design a productive level policy for a specific location. We adopt a discrete time formulation
to simplify the description and allow the direct application of conventional MDP techniques for
comparison. The methods can be easily converted to continuous time operation as well.
2.1 Traffic Model
We consider the simplest meaningful intersection setup with bidirectional traffic flows. Regardless of
the number of intersections, all intersections follow the same settings. Vehicles are coming from 4
directions as 4 queues. We use Xni(t), i = 1, 2, 3, 4, to denote the number of vehicles of traffic flow
i waiting to pass the intersection n at time t and Yn(t) ∈ {0, 1, 2, 3} to indicate the configuration of
the traffic lights for intersection n at time t. The traffic light L has four configurations:
• "0": green light for flow 1 and hence red light for flow 2;
2
(a) Traffic Light Configuration
(b) Traffic Flow
Figure 1: The state transition diagram (left) and the traffic flows of a grid topology road network
(right).
• "1": yellow light for flow 1 and hence red light for flow 2;
• "2": green light for flow 2 and hence red light for flow 1;
• "3": yellow light for flow 2 and hence red light for flow 1.
Each configuration Yn(t) can either simply be continued in the next time slot or must otherwise be
switched to the natural subsequent configuration (Yn(t) + 1) mod 4. This is determined by the
action A(t) ∈ {0, 1} selected at the end of time slot t, which is represented by a binary variable as
follows:"0" for continue, and "1" for switch; then we have
Y (t + 1) = (Y (t) + A(t)) mod 4.
These rules give rise to a strictly cyclic control sequence. And the evolution of the queue state over
time is governed by the recursion
Xni(t + 1) = Xni(t) + Cni(t) − Dni(t),
with Ci(t) denoting the number of vehicles of traffic flow i appearing at the intersection during time
slot t and Di(t) denoting the number of departing vehicles of traffic flow i crossing the intersection
during time slot t and D1(t) = 0 if Y (t) (cid:54)= 0 and D2(t) = 0 if Y (t) (cid:54)= 2.
Traffic flows can either come from outside of the network or from neighbouring intersections
belonging to the system. Vehicles coming from outside of the system follow a stochastic process and
those from neighbouring intersections are controlled by the states and actions of those intersections.
2.2 MDP Formulation
The target Intelligent Transportation System is a Markov Decision Process (MDP). Under the MDP
frame work, the state space S, action space A, reward r, policy π, and state transition probability P
of our problem are defined as follows:
• State S = (Xni(t), Yn(t)) where n = 1, 2, . . . , N and N is the number of intersections,
i = 1, 2, 3, 4 are queue length at west, north, east and south directions, respectively. The
environment return the full state to the network. However, each agent would only receive
local observation, which is sn(t) = (Xni(t), Yn(t) for agent n.
• Action A ∈ {0, 1} where 0 denotes continue and 1 denotes switch, as described in traffic
• Reward r(st, at) = −F (Xni(t)), i = 1, 2, 3, 4, where F (Xni(t)) = (cid:80)4
model.
ni(t). The
negative sign means that the queue should be penalized. In this paper we use quadratic norm
because we want to encourage short queue length. Long queue length would be penalized
by the quadratic term of the reward function.
• Policy π is the signal controlling command from the ITS, which maps a state to a probability
distribution over the actions π : S → P(A), where P(·) denotes probability distribution.
i=1 X 2
3
• State transition probability P is the probability of the network entering the new state st+1,
after taking the action at at the current state st. At the current state, by taking action at,
vehicles are transmitted to neighbouring intersections or to outside of the system, while
stochastic traffic flow from outside of the system during time t to t + 1 would come into the
system.
3 Deep Reinforcement Learning Algorithms
In this section, we first present an overview of the DDPG algorithm, then analyze the transportation
system. We propose the way to apply spatial influence and reward adjustment [16] in multi-agent
DDPG algorithm in solving signal control problem, and finally describe how they are implemented.
3.1 Deep Deterministic Policy Gradient Algorithm
Deep deterministic policy gradient (DDPG) [12] uses deep neural networks to approximate both
action policies and value functions (state value or action-value). This method has two advantages. 1)
This reduces the dimension of the state space and action space, since it only uses a limited number of
parameters to approximate them. 2) Gradient descent can be used to get the optimum, which greatly
speeds up the convergence and reduces the computational time.
While traditional DDPG algorithms have a continuous action space, the traffic control signals in our
model are discrete, i.e., A(t) ∈ {0, 1}. We will apply the discretization process to transform the
continuous outputs of the actor network in DDPG to discrete ones. The output layer of actor network
has a modified sigmoid function as activation:
y = sigmoid(tx),
where x, y are inputs and outputs of the final layer, and t is the ratio for steepening the sigmoid
function, which is 1000 in our experiment settings. Combined the modified sigmoid activation
function with a node-wise binarization process, our discrete DDPG algorithm can reduce the errors
caused by the continuous-to-discrete transformation to a great degree.
3.2
Intelligent Transportation System Analysis
Theorem 1. Consider a road network that has N intersections thus controlled by N agents, the total
group network utility can not be improved by single agent changing its action unilaterally. i.e.,
a∗
n = argmax
an∈An
Sn(an, a−n),∀n ∈ N,
where a−n = (x1, . . . , xn−1, xn+1, . . . , xN ) is the set of actions chosen by other agents except agent
n, Sn is the network group utility, or total traffic congestion cost in our case, and an is the action
taken by agent n.
It is worth noting that a traffic system always follows a cooperative game formulation. Also, it is
always a total network maximization problem. Optimizing one intersection is of course of great value
to the local community. However, to resolve the urban traffic issue in general, that is far from enough.
The ITSs need to optimize the road network as a whole to reduce traffic congestion and improve
commuting efficiency.
If we assume that one agent can improve total group utility by changing its action unilaterally in
a dynamically stationary road network, that means the vehicles' coming rate and leaving rate are
equal for each intersection in a full state transition cycle. If one agent simply changes its strategy, the
balance would be broken, queue length at certain directions would increase, thus increasing the total
road network utility.
Theorem 2. There is no stationary Nash Equilibrium for Transportation System.
Proof. Here we prove by contradicting the following two steps.
• If there is Nash Equilibrium, it must be a = (1, 1, . . . , 1), if traffic flow from outside of the
system is not 0;
4
• a = (1, 1, . . . , 1) is not a Nash Equilibrium.
To prove Step 1, we only need to assume that there is a 0 in a. Similar to the proof of Theorem 1, if
there is a 0 in a as the solution of Nash Equilibrium, that means for such intersections, the traffic
signal never changes. However, if the traffic signal never changes, the length of 2 queues at that
intersection would be strictly increasing, thus the expected reward would increase for that intersection
as well as the traffic network. Therefore, if there is Nash Equilibrium, it must be a = (1, 1, . . . , 1).
To prove Step 2, we first introduce the conclusion of Step 1. Consider a linear road topology with 2
intersections. So (X1i(t), Y1(t)) and (X2i(t), Y2(t)), i = 1, 2, 3, 4 would be the queue length and
light states for agent 1 and 2, respectively. And the evolution of the queue state over time is governed
by the recursion
Xni(t + 1) = Xni(t) + Cni(t) − Dni(t),
where n = 1, 2, with Ci(t) denoting the number of vehicles of traffic flow i appearing at the
intersection during time slot t and Di(t) denoting the number of departing vehicles of traffic flow i
crossing the intersection during time slot t, as described in traffic model. Here we make a simple
assumption that the number of vehicles appearing at and departing from one intersection at a constant
rate c and d. Note that as Ci(t) from outside of the system is a constant, so we have a stationary
environment. Also note that Di(t) = d only when there is green light. Also, as a1 = a2 = 1 for all t,
we conclude that 4 actions complete one cycle. If there is Nash Equilibrium, following equation must
hold:
Xni(t + 4) ≤ Xni(t).
Therefore,
Xni(t + 4) ≤ Xni(t + 3) + Cni(t + 3) − Dni(t + 3)
. . .
= Xni(t) + 4c − d ⇒ c ≤ 1
4
d.
If c = 1
transportation network.
2 d > 1
4 d, the above inequality doesn't hold. So a = (1, 1, . . . , 1) is not Nash Equilibrium for
2 d, we only need to have 5 actions in a cycle, to make sure
However, we can easily infer that if c = 1
that there are 2 green light slot in a cycle, thus Xni(t + 5) ≤ Xni(t).
Note that we used two assumptions in the proof of no Nash Equilibrium: 1. Constant appearing rate
of vehicles from outside of the transportation system; 2. Constant departing rate of vehicles at the
intersection. These two assumptions are too naive and too strong to be satisfied in real traffic flow.
In real life, it is more reasonable to assume that the appearing rate of vehicles is stochastic and the
departing rate of vehicles obeys certain probability distribution.
It is counter-intuitive but reasonable to find out that there is no stationary Nash Equilibrium solution
for the transportation system. However, it supports the fact that reinforcement Learning algorithm is
necessary to learn optimal signal control policy. Otherwise, the hard-coded policy would be enough.
3.3 Spatial Influence Based Solution
We propose a unified method for achieving both coordination and communication in Multi-agent
reinforcement learning (MARL). For road networks with multiple intersections, we apply multi-agent
DDPG in our system. Opposed to the single-agent DDPG method, multi-agent methods do not have
computational issues when there are huge observation space and action space. As one agent is only
responsible for optimizing traffic signal control for one intersection, it significantly improves the
scalability of the system. Also, MADDPG methods would make no assumption about the shape of
the network, make modeling mixed direction and shape road networks possible.
The MADDPG method needs a way to communicate with other agents to compliment the lack of
global view for more advanced control like "greenwave". Prior work often resorts to centralized
training to ensure that agents learn to coordinate. However, centralized training is not feasible in
this case, as intersections are not identical. Some intersections are at the border of the network and
others are in the middle of the network. Border intersections observe stochastic traffic flows, which
5
means that they need to learn to adapt. Central intersections observe traffic flows coming from other
intersections within the network, so they need to cooperate. Inspired by [17], we propose to use social
influence as the way to accomplish it. In our work, we make no assumption that agents could view
one anothers rewards, as that is often not practical in real life. We relying only on agents viewing
each other's actions. We model neighboring agents' actions into the target agent's observation space.
The observation space for agent n now changes to (Xni, Yn, a), where a is the actions taken by
neighbouring agents at the last time step. The intuition behind this is that, by observing neighboring
agents' actions and the observation space of itself, the agent would be able to learn the policy of
traffic signal control and balance between optimizing the reward of itself, its neighbourhood and
indirectly the whole system.
3.4 Rewards Adjustment
As we have shown in the previous option, total group network utility can not be improved by a single
agent changing its action unilaterally. Therefore, we propose to adjust the reward function of the
agent to make sure that each agent would not only consider its own reward, but also the rewards of its
neighbouring agents. We change the reward function from:
rn(st, at) = −F (Xn(t))
rn(st, at) = −F (Xn(t)) − (cid:88)
m∈Mn
to
(cid:80)
(1)
where Mn is the neighbouring agents of agent n, wnm is the social tie weight. We define
m∈M wnm as "Selfish Index" as it determines the importance of neighbouring agents to the
target agent. Based on this definition, the agent would not only take its own reward into account,
but its neighbourhood as well. By directly linking agent with its neighbourhood and thus indirectly
with agents far away, the agent is expected to learn to cooperate and communicate with each other to
optimize the social group total utility.
wnmF (Xm(t)),
3.5 DDPG for Grid Road Networks
For road networks with multiple intersections, we apply the multi-agent DDPG algorithm in the
network. We have one agent for each intersection. The number of cars and lights states at all
intersections are inputs to the MADDPG system. To ensure that the MADDPG can possess a non-
local view regarding the total network utility, spatial influence and social group utility methods are
considered, as we have specified in the previous sections. The algorithm is defined in Algorithm 1:
4 Performance Evaluation
In this section, we discuss the numerical experiments implemented to evaluate the performance of
the social influence based MADDPG algorithms as detailed above. In the experiments, we apply
a 4-layer fully-connected neural network for both actor and critic in both the linear topology and
grid scenario, with 400 neurons for the first 2 layers, following with two layers with 600 and 200
neurons, respectively. All agents share the same architecture, while they have difference observation
space, as spatial influence is applied to them. Noting that the actor needs to output near-binary values
as action values as mentioned above, the actor neural network has one modified sigmoid activation
function for the output layer. As usual, a copy of the actor and critic neural network are taken as the
target networks with "soft" updates. The outputs of the actor network are clipped to binary values
0 and 1 indicating light state changing and remaining the same. We take an episode length of 150
steps of simulation for collecting the learning samples, and train both actor and critic networks with a
batch size of 64 and discount factor γ = 0.99. The OrnsteinUhlenbeck noise [18] is applied for the
explorations, with a variance of 0.3.
As for the traffic environment settings in the experiments, the vehicle coming and passing rates are set
differently for main roads and branch roads. the vehicle coming rate indicates the number of vehicles
coming from outside of the road networks into the networks on each road per time step, which is set
to be a random value with upper bound Cm for main roads or Cb for branch roads; and the vehicle
passing rates indicate the number of vehicles passing one intersection within one time step, set to be
the fixed numbers of 16 and 4 for main road and branch road, respectively.
6
Algorithm 1 Multi-agent DDPG for ITSs
Input: number of episodes M, time frame T , minibatch size N, learning rate λ, and number of
1: for j = 1, J do
2:
agents or intersections J
Randomly initialize critic network Qj(Oj, aθQ
weight θQ
Initialize target network Q(cid:48)
Initialize replay buffer Bj for each agent j;
j for agent j;
j and µ(cid:48)
j and θµ
j ) and actor network µj(Ojθµ
j ) with random
j with weights θQ(cid:48)
j ← θQ
j , θµ(cid:48)
j ← θµ
j for each agent j;
3:
4:
5: end for
6: for episode = 1, M do
7:
8:
9:
10:
11:
for j = 1, J do
Initialize a random process N for action exploration;
Receive initial observation state s0;
for t = 1, T do
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
j ) + Nt according to the current policy and exploration
Select action aj,t = µj(Oj,tθµ
noise;
end for
Each agent executes action aj,t, market state changes to st+1;
Each agent observes reward rj,t and observation Oj,t+1, where observation Oj,t+1 is
adjusted according to spatial influence;
Each agent adjust reward according to Equation 1
for j = 1, J do
Store transition (Oj,t, aj,t, rj,t, Oj,t+1) in Bj;
Sample a random minibatch of N transitions (Oj,i , aj,i , rj,i , Oj,i+1) from Bj;
j θQ(cid:48)
Set yj,i = rj,i + γQ(cid:48)
Update the critic by minimizing the loss: L = 1
N
Update the actor policy by using the sampled policy gradient:
(cid:80)
i(yj,i − Qj(Oj,i, aj,iθQ
j )) for i = 1, . . . , N;
j(Oj,i+1θµ(cid:48)
j(st+1, µ(cid:48)
j ))2;
∇θµπ
≈
1
N
∇aQj(O, aθQ
j )O=Oj,i,a=µj (Oj,i) × ∇θµµj(Ojθµ)si;
(cid:88)
i
Update the target networks: θQ(cid:48)
j ← τ θQ
j + (1 − τ )θQ(cid:48)
j ,
22:
23:
24:
25: end for
end for
end for
θµ(cid:48)
j ← τ θµ
j + (1 − τ )θµ(cid:48)
j .
4.1 Grid Road Topology
We assume the numbers of arriving vehicles in eastern and western direction to the arterial road
in each time step to be independent and Bernoulli distributed with parameter p1. The numbers of
arriving vehicles in southern and northern directions in each time step are also independent and
Bernoulli distributed with parameter p2. We extend the road network to be a grid topology. The
intersection setting is described in Section 2 and the road network has more than one intersections
in south-north and east-west directions. The quadratic congestion cost function is of the form
ni. In our experiments, we design a road network with three columns and
three rows, therefore 9 intersections in total. Fig. 2 shows that the algorithm with spatial influence
and reward adjustment integrated performs the best, no matter the directions of the influencing flow.
F (X) =(cid:80)N
(cid:80)4
i=1 X 2
n=1
4.1.1 Spatial Influence
As we have described in Section 3, we design spatial influence flow to affect the behavior of
neighbouring agents.
In inward and outward scenarios, for any two agents that are physically
connected, there will be only one influencer and one influencee. For a fully connected influence
scenario, all agents can observe the last actions from all its neighbours.
• Inward Influence: agents at the four corners are independent.
7
(a) Fully Connected
(b) Outward
(c) Inward
Figure 2: Different influencing flow. We can see for all of the influencing direction flow, by adding
spatial influence and reward adjustment functions, the traffic system would have higher return.
(a) Directions
(b) Selfish index
Figure 3: The influencing directions do not affect the performance level at the convergence (left).
The "Selfish Index" affects the system performance (right). If agents are cooperative, the total group
utility would be better.
• Outward Influence: agent at the center is independent.
• Fully Connected Influence: all agents are not independent.
As we can see from Fig. 3, the total utility converges to the same level. So all the methods can
optimize total network utility. However, we can see from the detailed analysis of individual utility
that in inward influence flow, agents at the corners have less congestion compared to other scenarios,
while in outward flow scenario, the agent at the center has relatively less congestion. Therefore,
in a city where the importance of interactions varies, the direction of influencing flow should be
considered. Thus the city traffic flow can be optimized both globally and locally.
4.1.2 Social Group Utility
We adjust the reward function of each agent according to Equation (1). As we can see from Fig. 3,
the total network utility increases as the social tie weights of neighbouring agents increases. This
confirms our anticipation that as neighbouring agents' rewards become more important to the target
agent, it will learn to cooperate with its neighbours to increase the road network utility.
5 Conclusion
We have explored the scope of using social influence based Multi-agent Deep Deterministic Policy
Gradient (MADDPG) to optimize real-time traffic signal control policies in emerging large-scale
Intelligent Transportation Systems. We compared the performance between social influence based
MADDPG and naive MADDPG, and demonstrated that social influence based algorithm has the
maximum potential in scaling without the issue of large observation space and action space and would
help communication between agents. We verified it the scalability properties of DDPG algorithms in
both a linear topology and a grid topology, and demonstrated the emergence of intelligent behavior
such as "green wave" patterns, confirming that the agents learned to cooperate through social influence
communication.
8
References
[1] Mohammed M Vazifeh, P Santi, G Resta, SH Strogatz, and C Ratti. Addressing the minimum
fleet problem in on-demand urban mobility. Nature, 557(7706):534, 2018.
[2] Ming Zhu, Xiao-Yang Liu, and Xiaodong Wang. An online ride-sharing path-planning strategy
for public vehicle systems. IEEE Transactions on Intelligent Transportation Systems, 2018.
[3] Ming Zhu, Xiao-Yang Liu, Feilong Tang, Meikang Qiu, Ruimin Shen, Wei Wennie Shu, and
Min-You Wu. Public vehicles for future urban transportation. IEEE Transactions on Intelligent
Transportation Systems, 17(12):3344 -- 3353, 2016.
[4] Ming Zhu, Xiao-Yang Liu, and Xiaodong Wang. Joint transportation and charging scheduling
IEEE Transactions on Intelligent
in public vehicle systems -- a game theoretic approach.
Transportation Systems, 19(8):2407 -- 2419, 2018.
[5] Ardi Tampuu, Tambet Matiisen, Dorian Kodelja, Ilya Kuzovkin, Kristjan Korjus, Juhan Aru,
Jaan Aru, and Raul Vicente. Multiagent cooperation and competition with deep reinforcement
learning. PloS One, 12(4):e0172395, 2017.
[6] Jakob N Foerster, Christian A Schroeder de Witt, Gregory Farquhar, Philip HS Torr, Wendelin
Boehmer, and Shimon Whiteson. Multi-agent common knowledge reinforcement learning.
arXiv preprint arXiv:1810.11702, 2018.
[7] Muhammad Alam, Joaquim Ferreira, and José Fonseca. Introduction to intelligent transportation
systems. In Intelligent Transportation Systems, pages 1 -- 17. Springer, 2016.
[8] Yisheng Lv, Yanjie Duan, Wenwen Kang, Zhengxi Li, Fei-Yue Wang, et al. Traffic flow predic-
tion with big data: A deep learning approach. IEEE Transactions on Intelligent Transportation
Systems, 16(2):865 -- 873, 2015.
[9] Simona Onori, Lorenzo Serrao, and Giorgio Rizzoni. Dynamic programming. In Hybrid
Electric Vehicles, pages 41 -- 49. Springer, 2016.
[10] Martin L Puterman. Markov decision processes: discrete stochastic dynamic programming.
John Wiley & Sons, 2014.
[11] Sheldon M Ross. Introduction to stochastic dynamic programming. Academic press, 2014.
[12] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,
David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. ICLR,
2016.
[13] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-
agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural
Information Processing Systems, pages 6379 -- 6390, 2017.
[14] Tianshu Chu, Jie Wang, Lara Codecà, and Zhaojian Li. Multi-agent deep reinforcement learning
for large-scale traffic signal control. IEEE Transactions on Intelligent Transportation Systems,
2019.
[15] Xiao-Yang Liu, Zihan Ding, Sem Borst, and Anwar Walid. Deep reinforcement learning for
intelligent transportation systems. NeurIPS Workshop on Machine Learning for Intelligent
Transportation Systems, 2018.
[16] Xu Chen, Xiaowen Gong, Lei Yang, and Junshan Zhang. Exploiting social tie structure for
cooperative wireless networking: A social group utility maximization framework. IEEE/ACM
Transactions on Networking (ToN), 24(6):3593 -- 3606, 2016.
[17] Natasha Jaques, Angeliki Lazaridou, Edward Hughes, Caglar Gulcehre, Pedro Ortega,
Dj Strouse, Joel Z Leibo, and Nando De Freitas. Social influence as intrinsic motivation
for multi-agent deep reinforcement learning. In International Conference on Machine Learning,
pages 3040 -- 3049, 2019.
[18] Werner Horsthemke and René Lefever. A perturbation expansion for external wide band
markovian noise: Application to transitions induced by ornstein-uhlenbeck noise. Zeitschrift
für Physik B Condensed Matter, 40(3):241 -- 247, 1980.
9
|
1803.08867 | 1 | 1803 | 2018-03-23T16:25:05 | Testing demand responsive shared transport services via agent-based simulations | [
"cs.MA",
"physics.soc-ph"
] | Demand Responsive Shared Transport DRST services take advantage of Information and Communication Technologies ICT, to provide on demand transport services booking in real time a ride on a shared vehicle. In this paper, an agent-based model ABM is presented to test different the feasibility of different service configurations in a real context. First results show the impact of route choice strategy on the system performance. | cs.MA | cs | TESTING DEMAND RESPONSIVE SHARED TRANSPORT
SERVICES VIA AGENT-BASED SIMULATIONS
Giuseppe Inturri, Nadia Giuffrida, Matteo Ignaccolo, Michela Le
Pira, Alessandro Pluchino, Andrea Rapisarda
University of Catania
Abstract
Demand Responsive Shared Transport (DRST) services take advantage of
Information and Communication Technologies (ICT), to provide "on demand"
transport services booking in real time a ride on a shared vehicle. In this paper, an
agent-based model (ABM) is presented to test different the feasibility of different
service configurations in a real context. First results show the impact of route choice
strategy on the system performance.
Keywords: shared mobility, flexible transit, dynamic ride sharing, demand
responsive transport, agent-based model
Introduction
This paper focuses on the potential contribution of innovative Demand Re-
sponsive Shared Transport (DRST) services provided by a fleet of vehicles, booked
by users via mobile device applications and scheduled in real-time to pick up and
drop off passengers in accordance with their needs [1]. The system stands between
an expensive conventional exclusive-ride door-to-door service (like a conventional
taxi) and a flexible system where a dynamic sharing of trips makes users experiment
longer travel distances and times while the vehicles drop off and pick up other pas-
sengers (like a conventional transit).
From the operator's point of view, it is important to select the optimal strat-
egy to assign vehicles to passengers' requests, so to perform high load factor and
low driven distance (to reduce operation costs), while minimizing the additional
time and distances travelers have to experience (to assure the expected level of ser-
vice).
2
The city of Ragusa (Italy) is chosen as case study, where an innovative
DRST has already been tested. The model is used as a realistic environment where
to simulate different scenarios, with simple rules assigned to agents' behavior, in
order to explore the transport demand and supply variables that make the DRST
service feasible and convenient. The aim is to understand, starting from the micro-
interaction between demand and supply agents (i.e. passengers and vehicles), the
macroscopic behavior of the system so to monitor, via appropriate indicators, its
performance and give suggestions on its planning, management and optimization.
In the last years an increasing attention has been paid on shared transport
services. Optimization models have been proposed to solve a dial-a-ride (Stein,
1978) or multiple depot vehicle scheduling problems (Bodin and Golden, 1981).
More recently, simulation models have been developed to study the usability and
profitability of dial-a-ride with fixed-route problems (Shinoda et al, 2004), the effi-
cient scheduling of dynamic DRT services (Diana, 2006), the dynamics of a taxi-
sharing system (D'Orey and Fernandes, 2012), the effects of using a zoning vs. a
no-zoning strategy and time-window settings (Quadrifoglio et al, 2008).
Agent-based simulations has proved to be a good technique in this context,
also to overcome some limitations linked to a top-down approach. They are suitable
to reproduce the interaction among stakeholders involved in transport decision-
making (Le Pira et al, 2018; Marcucci et al, 2017). They have been proposed to
study taxi fleet operations (Cheng and Nguyen, 2011), car sharing (Lopes et al,
2014) and to investigate DRT services, and developing an open-source simulation
testbed (Čertický et al, 2016).
Main benefits of ABM are (1) possibility to capture emergent phenomena,
(2) providing of a natural description of a system, (3) flexibility, for the easiness to
add more entities to the model, to modify behavior, degree of rationality, ability to
learn and evolve, and rules of interactions of agents (Bonabeau, 2002).
In this paper a new ABM is presented to test the operation of a DRST
service under different dispatching configurations. The interaction among vehicles
traveling along a semi-flexible route and users walking from their origins to stop
nodes to get a transport service to their destinations is simulated. The main novelty
of the model relies on the implementation of a GIS-based demand model implying
an easy transferability to other contexts. The ABM allows exploring the emergence
of optimal operation configuration by identifying ad-hoc indicators to monitor the
system's performance.
Description of the model
An ABM has been built within NetLogo (Wilensky, 1999) to test the im-
pact of different vehicle dispatching strategies on the service efficiency and effec-
tiveness.
The model can be described according to its main features, i.e. (i) transport
network, (ii) demand model, (iii) agent (user and vehicle) dynamics, (iv) route
choice strategies, (v) performance indicators.
3
Transport network. The network consists of a fixed route and three op-
tional routes, composed of network nodes and links, stop nodes and diversion nodes.
The network follows the actual road network and a GIS map, reproducing census
tracts, is used to implement in the model georeferenced socio-economic data about
population, through the GIS extension of NetLogo.
Demand model. A user group's (with a maximum prefixed size) trip re-
quest is randomly generated with a negative exponential distribution with an aver-
age trip rate from an origin i with trip rate proportional to density population, to a
destination j with a gravitationally distributed probability.
Agent (user and vehicle) dynamics.
If the origin or destination of the trip request is more distant than a prefixed
threshold, it assumes the status "rejected". Otherwise, the user group moves to the
nearest stop and assumes the status "waiting", until a vehicle (with an appropriate
number of available seats) reaches the stop; in this case each user boards and alights
at the nearest stop to its required destination, assumes the status of "satisfied" pas-
senger; if the maximum waiting time is overcome, each user gives up and assumes
the status of "unsatisfied".
A given number of vehicles, with prefixed seat capacity, generated at ran-
dom stops, starts traveling along the fixed route at constant speed. At each stop,
waiting users are loaded following the First-Come-First-Served queue rule, if the
group size is not greater than the available seats. At each diversion node a vehicle
can shift to an optional route if waiting users or on-board passengers' destinations
are present. The available vehicle seats are updated at each event of passenger load-
ing/unloading.
Route choice strategies. All vehicles drive on the fixed route. At diversion
nodes a vehicle may drive on a flexible route according with the Route Choice Strat-
egy (RCS) it is assigned to.
In this first version of the model, there are three RCSs:
"Fully at Random" (FR);
"All Vehicles drive on All flexible Routes" (AVAR); a prefixed percentage of
vehicles can be assigned at random;
"Each Vehicle is Assigned to a flexible Route" (EVAR); a prefixed percentage
of vehicles can be assigned at random.
The randomness component has been considered to add some "noise" to
the system, since it has been shown that random strategies can have a beneficial role
in increasing the efficiency of social and economic complex systems (Pluchino et
al, 2010).
Performance indicators. The local strategies determining interaction be-
tween passengers and vehicles give raise to global patterns that can be monitored
via appropriate performance indicators: total number of passengers transported NP,
total driven distance TDD (km), average passenger travel distance APTD (km), av-
erage vehicle load factor ALF, passenger travel time in terms of average waiting
time AWT (min), average on-board time AoBT (min), average total travel time
APTT (min), average vehicle speed AVS (km/h), transport intensity CI (km/pax) as
ratio of total driven distance and number of passengers, total user travel time TPTT
4
(h) (including a penalty of 60 min for each unsatisfied user), vehicle operation cost
OC (€) and total unit cost TUC (€/pax), evaluated according to equation (1):
(1)
Case study
The case study is located at Ragusa, a small-medium city (70,000 inh.) in
the south-eastern part of Sicily (Italy), where an innovative DRST service, called
MVMANT (http://www.mvmant.com/portfolio-view/ragusa/) has already been ex-
perimented in 2016. The city of Ragusa is characterized by two distinct areas, the
upper town and the lower and older town of Ragusa Ibla, with a high touristic vo-
cation. MVMANT has connected several park-and-ride facilities with the main
destinations in Ragusa Ibla, which is scarcely connected to the center, offering a
continuous service with midsize passenger vans.
Fig. 1. shows the fixed (blue) and flexible (orange) routes and census
zones colored according to population (from light to dark green).
Fig. 1. Virtual map (left) and satellite map (right).
The main input variables of the system are:
service variables, i.e. total simulation time (h), number of vehicles, vehicle max-
imum capacity (seats), vehicle average speed (km/h);
demand variables, i.e. demand rate (request/hour), maximum number of passen-
gers per demand, maximum waiting time (min)
route choice strategy, i.e. FR, EVAR, AVAR, with a variable percentage of ran-
domness.
Preliminary Results and Conclusions
For a first test of the model, simulations were performed by considering
different route choice strategies (i.e. FR, EVAR, AVAR) with increasing levels of
randomness, so to test the overall system performance.
Fig. 2. shows on the left the total hours spent by all passengers (in yellow),
while waiting (in light blue), on board (orange) plus a penalty of 60 minutes for
€()(€)€()TPTThVOTOChTUCpaxNPpax 5
each users that waited more than a prefixed threshold at the stop (grey). In dark blue
the total number of passengers transported. They decrease with 15 and 30 vehicles
because group requests with 3 passengers cannot be satisfied by vehicles with low
capacity. On the right, Fig. 2. shows the total unit cost TUC (€/pax) by the number
of vehicles. It is calculated by attributing a monetary value to each hour spent by
the passengers in the system (10 €/h), adding the operation cost of vehicles (in the
range of 0.5-1.0 €/km according with the vehicle size) and drivers' cost (20 €/h) and
dividing the sum by the number of passengers. It can be considered a measure of
the total cost sustained by the society (demand and supply) for the mobility of one
person. In this case an optimal range of shared services can be identified within a
range between 5 and 10 vehicles.
Fig. 2. Passenger travel time and number of passengers (on the left) and total unit cost
TUC (on the right) (route choice strategy EVAR with 30% randomness; maximum group size=3;
total seat capacity=30).
Comparing TUC for EVAR and AVAR strategies with variable random-
ness (Fig. 3. ), best results are found with EVAR and no randomness, while AVAR
is the worst. This is because the assignment of vehicles to specific routes (EVAR)
reduces the empty driven distances. By increasing randomness, it can be seen that
the two strategies get closer in terms of TUC. With randomness more than 40% the
two strategies are the same with intermediate and almost constant performance. It
can be concluded that a certain rate of randomness is beneficial for the AVAR strat-
egy, since it implies that not all vehicles will simultaneously explore all the routes
if demand is present (thus reducing the empty driven distance). Vice versa, EVAR
strategy works better without randomness.
6
Fig. 3. Total unit cost by randomness (number of vehicles= 5; seat capacity=8; maxi-
mum group size= 1)
This paper presented an agent-based model able to simulate flexible de-
mand responsive shared transport (DRST) services. Different strategies can be
tested by changing the service variables and assigning simple rules to the agents so
to explore, starting from the microinteraction between passengers and vehicles, the
macroscopic behavior of the system. For a first test, the city of Ragusa (Italy) was
chosen as case study, where an innovative DRST has already been experimented
and different scenarios have been reproduced by changing fleet composition and
vehicle dispatching strategy.
Simulation results show that the service quality and performance consid-
erably vary in relation to the number and capacity of vehicles. In particular, given a
fixed supply (in terms of total seat capacity), many vehicles with low capacity de-
crease the passenger travel time (and cost) and increase the operator costs, while
few high-capacity vehicles perform better from an operator's point of view. An op-
timal range can be found by considering a total unit cost accounting both for pas-
senger travel time and vehicle operation cost.
Besides, different vehicle dispatching strategies have been simulated to
test different system configurations (from more flexible to more fixed route strate-
gies). Results show that assigning vehicles to specific routes (i.e. avoiding that all
vehicles drive on all routes) reduces travelled distances by empty vehicles and im-
proves the overall system performance. If all vehicles are allowed to drive on all the
routes, then a certain level of randomness in agent's choice is found to be beneficial
for the system performance.
The first results show that the model is able to reproduce different system
configurations and to monitor, via appropriate indicators, its performance.
Further research will focus on: (a) comparing DRST with pure taxi and
pure bus services; (b) testing other strategies to optimize the service (i.e. increase
7
load factor, reduce vehicle-km), e.g. re-balancing/idle strategies; (c) testing reac-
tive/adaptive agent behaviors for route choice strategies based on system states; (d)
testing pricing strategies and public subsidies to increase the service effectiveness
(in terms of satisfied demand); (e) testing the performance of the system with au-
tonomous vehicles; (f) including elasticity of demand to price; (g) improving the
demand model (e.g. including socio-demographics characteristics, data from sur-
veys).
The final aim is to have a reliable decision-support tool for planning, man-
agement and optimization of DRST services, which can help to reduce the burden
of transport in our cities and contribute to sustainable mobility.
REFERENCES
1. Ambrosino, G., Nelson, J.D., and Romanazzo, M. Demand Responsive Transport Services:
Towards the Flexible Mobility Agency. ENEA. Italian National Agency for New Tech-
nologies, Energy and the Environment, ISBN 88-8286-043-4, 2003
Bodin, L. and Golden, B. Classication in vehicle routing and scheduling. Networks, 11(2):97-
108, 1981
Bonabeau, E. Agent-based modeling: Methods and techniques for simulating human systems.
PNAS May Vol 99 Suppl 3 7280-7287, 2002.
Čertický, M., Jakob, M., Píbil, R. Simulation Testbed for Autonomic Demand-Responsive
Mobility Systems. Autonomic Road Transport Support Systems, 147-164, 2016
Cheng, S. F. and Nguyen, T. D. Taxisim: A multiagent simulation platform for evaluating
taxi fleet operations. In Proceedings of the 2011 IEEE/WIC/ACM International Confer-
ences on Web Intelligence and Intelligent Agent Technology- Volume 02, pages 14-21,
2011.
Diana, M. The importance of information flows temporal attributes for the efficient schedul-
ing of dynamic demand responsive transport services. Journal of advanced Transporta-
tion, 40(1):23-46, 2006.
D'Orey, P. M., Fernandes, R. and Ferreira, M. Empirical evaluation of a dynamic and distrib-
uted taxi-sharing system. In Proceedings of ITSC 2012, pages 140-146. IEEE, 2012.
Le Pira, M., Inturri, G., Ignaccolo, M., Pluchino, A. Dealing with the Complexity of Stake-
holder Interaction in Participatory Transport Planning. In Zak, J., Hadas, Y., Rossi, R.
(eds.) "Advanced Concepts, Methodologies and Technologies for Transportation and Lo-
gistics". Advances in Intelligent Systems and Computing 572. Springer International Pub-
lishing. DOI 10.1007/978-3-319-57105-8_3, 2018.
Lopes, M. M., Martineza, L. M., de Almeida Correia, G. H. Simulating carsharing operations
through agent-based modelling: an application to the city of Lisbon, Portugal. Transpor-
tation Research Procedia, 3, 2014
Marcucci, E., Le Pira, M., Gatta, V., Ignaccolo, M., Inturri, G. & Pluchino, A. Simulating
participatory urban freight transport policy-making: Accounting for heterogeneous stake-
holders' preferences and interaction effects. Transportation Research Part E 103, 69-86,
2017
Pluchino, A., Rapisarda, A., Garofalo, C. The Peter Principle revisited: a computational
study. Physica A, 389 (3): 467-472, 2010.
Quadrifoglio, L., Dessouky, M., Ordonez, F. A simulation study of demand responsive transit
system design. Transportation Research Part A, 42, 718–737, 2008.
8
Shinoda, K., Noda, I., Ohta, M., Kumada, Y. and Nakashima, H. Is dial-a-ride bus reasonable
in large scale towns? evaluation of usability of dial-a-ride systems by simulation. In Multi-
agent for mass user support, pages 105-119. Springer, 2004.
Stein, D. M. Scheduling dial-a-ride transportation systems. Transportation Science,
12(3):232-249, 1978
Wilensky, U. NetLogo. Center for Connected Learning and Computer Based Modeling.
Northwestern University, Evanston, IL, 1999. In: http://ccl.northwestern.edu/netlogo/.
|
1805.05631 | 2 | 1805 | 2018-05-17T17:00:18 | Complexity Reduction in the Negotiation of New Lexical Conventions | [
"cs.MA",
"cs.CL",
"cs.SI"
] | In the process of collectively inventing new words for new concepts in a population, conflicts can quickly become numerous, in the form of synonymy and homonymy. Remembering all of them could cost too much memory, and remembering too few may slow down the overall process. Is there an efficient behavior that could help balance the two? The Naming Game is a multi-agent computational model for the emergence of language, focusing on the negotiation of new lexical conventions, where a common lexicon self-organizes but going through a phase of high complexity. Previous work has been done on the control of complexity growth in this particular model, by allowing agents to actively choose what they talk about. However, those strategies were relying on ad hoc heuristics highly dependent on fine-tuning of parameters. We define here a new principled measure and a new strategy, based on the beliefs of each agent on the global state of the population. The measure does not rely on heavy computation, and is cognitively plausible. The new strategy yields an efficient control of complexity growth, along with a faster agreement process. Also, we show that short-term memory is enough to build relevant beliefs about the global lexicon. | cs.MA | cs | Complexity Reduction in the Negotiation of New Lexical Conventions
William Schueller ([email protected])1, 2, Vittorio Loreto3, 4, 5, and Pierre-Yves Oudeyer2
2INRIA Bordeaux Sud-Ouest/ Ensta ParisTech : Flowers Project-team, Bordeaux, France
1University of Bordeaux, Bordeaux, France
3SONY Computer Science Lab, Paris, France
4Physics Dpt., Sapienza University of Rome, Rome, Italy
5Complexity Science Hub Vienna (CSHV), Vienna, Austria
8
1
0
2
y
a
M
7
1
]
A
M
.
s
c
[
2
v
1
3
6
5
0
.
5
0
8
1
:
v
i
X
r
a
Abstract
In the process of collectively inventing new words for new con-
cepts in a population, conflicts can quickly become numerous,
in the form of synonymy and homonymy. Remembering all of
them could cost too much memory, and remembering too few
may slow down the overall process. Is there an efficient be-
havior that could help balance the two? The Naming Game is
a multi-agent computational model for the emergence of lan-
guage, focusing on the negotiation of new lexical conventions,
where a common lexicon self-organizes but going through a
phase of high complexity. Previous work has been done on
the control of complexity growth in this particular model, by
allowing agents to actively choose what they talk about. How-
ever, those strategies were relying on ad hoc heuristics highly
dependent on fine-tuning of parameters. We define here a new
principled measure and a new strategy, based on the beliefs
of each agent on the global state of the population. The mea-
sure does not rely on heavy computation, and is cognitively
plausible. The new strategy yields an efficient control of com-
plexity growth, along with a faster agreement process. Also,
we show that short-term memory is enough to build relevant
beliefs about the global lexicon.
Keywords: language emergence, active learning, multi-agent
model, control of complexity growth
Motivations
Lexical conventions constitute an important element of social
interactions. They can emerge, evolve, or be learnt within a
population, without necessarily having a centralized control.
In other words, they can be negotiated through local inter-
actions between individuals. In practice, this happens con-
tinuously in human societies, being the spread of new words
and conventions, the acquisition of those conventions by in-
fants or other learners, or even the emergence of new forms of
communication. Despite the high complexity of the processes
involved, humans deal with these issues quite efficiently.
Learning of high complexity tasks in individuals can in
general be facilitated by an active control of the complexity of
learning situations , often driven by intrinsic motivation, like
for example maximization of the learning progress (Gottlieb,
Oudeyer, Lopes, & Baranes, 2013; Baldassarre & Mirolli,
2013; Barto, 2013; Oudeyer, Kaplan, & Hafner, 2007). This
type of mechanism is also argued to be an evolutionary ad-
vantage for cognitive abilities (Oudeyer & Smith, 2014), and
can also be found in lexicon acquisition at the individual level
(Partridge, McGovern, Yung, & Kidd, 2015). But does it have
a significant impact on population-wide learning and conven-
tions negotiation dynamics?
The Naming Game (Steels & Kaplan, 1998; Wellens, 2012;
Loreto, Baronchelli, Mukherjee, Puglisi, & Tria, 2011; Ke,
Minett, Au, & Wang, 2002) is an adapted framework to test
this hypothesis.
It is a class of multi-agent models of lan-
guage emergence and evolution, where pairs of randomly se-
lected individuals try to communicate by referring to some
pre-defined meanings using words. At the beginning, they do
not share any convention about word-meaning associations.
Through repeated decentralized interactions, a common lex-
icon self-organizes. However, the process can be slow and
pass through a high-complexity phase where agents memo-
rize a lot of conflictual information, in the form of synonyms
and homonyms.
It has already been shown that active learning mecha-
nisms can increase convergence speed towards a shared lex-
icon in different language emergence models (Cornudella,
Van Eecke, & Van Trijp, 2015; Schueller & Oudeyer, 2016).
The main idea behind those mechanisms is to allow agents
to actively choose the topic of their communication, based
on information collected during their past interactions and
driven by control of complexity growth. However, the algo-
rithms used so far are based on ad hoc heuristics, constrained
interaction scenarios and can depend heavily on fine-tuning
of parameters themselves depending on population size and
number of words and meanings.
In previous work, an approximation of the global state is
built by each agent using the information of past interac-
tions, in the form of an average vocabulary of the population
(Oliphant & Batali, 1997; De Vylder, 2007). Is it possible to
design a new principled algorithm for an active topic choice
based on such a representation? Could decisions be driven by
both the compatibility of an agent's own lexicon with this av-
erage vocabulary, and a reduction of both their complexities?
Such an algorithm should rely on a time scale for the mem-
ory of past interactions: Indeed, in the case of uncentralized
negotiation of a lexicon, conflictual conventions will neces-
sarily appear and have to be forgotten in order to converge
to a functional global vocabulary. Remembering them could
slow down the self-organization process.
In this work, we define a principled measure of correlation
between an agent's lexicon and a local approximation of the
average lexicon of the population. We build a strategy driven
by the maximization of this value without being computation-
ally hard, to be cognitively plausible. We study and discuss
the impact of this strategy on convergence time and complex-
ity growth, depending on a time scale used for memory.
Methods
The Naming Game
We define here precisely the Naming Game model that we
used (see fig.1 for an overview). We need to explicit:
- The interaction scenario itself
- How agents represent their lexicon
- How they update their lexicon at the end of each interaction
It is a simple modification of the standard Naming Game sce-
nario (Loreto et al., 2011; Wellens, 2012).
Figure 1: Illustration of the Naming Game model. Out of a popu-
lation of simulated agents, two are picked and try to communicate,
using/inventing words to refer to meanings. After repeated such in-
teractions, a common lexicon self-organizes. In this example, there
are M=2 possible meanings, W =4 possible words and N=12 agents.
Interaction process We re-use a previously defined in-
teraction process called Speaker's Choice (Schueller &
Oudeyer, 2016). It allows one of the interacting agents, called
the speaker, choose actively the topic of the interaction.
Each interaction involves two agents, that are picked ran-
domly from the population. One of them is assigned the role
speaker, and the other the role hearer. The speaker chooses
a topic and picks up a word for this topic. If it does not have
a word associated so far to the meaning used as topic, it just
invents a new meaning-word association. It utters this word,
which is interpreted by the hearer as a meaning, if it knows
this word. If the interpreted meaning is the same as the topic,
i.e. the meaning intended by the speaker, the communication
is considered successful. Otherwise, it is considered a failure.
See fig.2 for a detailed illustration of the interaction process.
Vocabulary Representation Vocabularies, or lexicons, are
a set of associations between meanings and words. In this
work, we consider only a finite set of meanings M and a finite
set of words W . In this context, vocabularies can be repre-
sented as associations matrices, where each row corresponds
to a meaning, and each column to a word. This representa-
tion has been extensively used in related work (Oliphant &
Batali, 1997; Steels & Kaplan, 1998; Ke et al., 2002). Two
parts of the lexicon are distinguished, the coding or produc-
tion part, which maps a meaning to a set of words weighted by
Figure 2: Interaction process: Beforehand, 2 individuals have been
randomly selected among a population, an designated as speaker (S)
and hearer (H). 1. S chooses a topic, 2. S checks its vocabulary to
find or invent an associated word, 3. S utters the word, 4. H guesses
the intended meaning, 5. S indicates the intended meaning.
probabilities of usage, and a decoding or interpretation part,
mapping a word to a set of meanings that can be interpretated
from this word, also weighted by probabilities.
We represent the vocabulary of an agent A as a matrix V (A)
of size M×W , with values of 1 for each word-meaning asso-
ciation used by the agent. Each agent starts with an empty vo-
cabulary, a matrix filled with zeros. The coding matrix V c(A)
and decoding matrix V d(A) are derived from V (A) by nor-
malizing respectively over rows and columns:
V c(A)mw =
V (A)mw
∑
V (A)mw(cid:48)
w(cid:48)
V d(A)mw =
V (A)mw
∑
V (A)m(cid:48)w
m(cid:48)
(1)
(cid:1)
(cid:0)V d(A)mw
Normalization factors are used only if V (A)mw (cid:54)= 0. In prac-
tice, when coding a meaning m, a word wi is sampled using
the distribution (V c(A)mw)w∈W . When decoding a word w,
a meaning m j is interpreted, sampled from the distribution
m∈M. In our case, these distributions are uniform
either on the set of words associated to m for coding, or on
the set of meanings associated to w for decoding. Those 2
sets change over time, during the vocabulary update.
Vocabulary Update Policy At the end of each interaction,
each agent takes into account the result of the interaction by
modifying its lexicon. There exists various policies that have
been described and studied in previous work (Wellens, 2012).
We are using the one called Minimal Naming Game.
In this policy, updates work this way: when the communi-
cation fails, both agents add the used word-meaning associa-
tion (meaning used as a topic by the speaker, and word uttered
by the speaker) to their lexicon, and do nothing if they already
had it. If the communication is successful, not only do they
add this association to their respective lexicons, they also re-
move any conflicting synonyms and homonyms. See fig.3 for
an illustration of the update policy in both cases.
Typically, among existing policies, Minimal NG and an-
other one called Basic Lateral Inhibition are used: they are
more realistic as they allow synonymy/homonymy and yield
faster agreement. Moreover, Minimal NG has been shown to
yield similar dynamics as Basic Lateral Inhibition, yet being
simple and not depending on any parameter, while the latter
depends on 3. This is the reason why we are using the Mini-
mal NG as vocabulary update policy.
Figure 3: Vocabulary update (Minimal NG). Failure (when the
hearer interpreted the word as another meaning than the topic):
the word-meaning association used by the speaker is added to the
hearer's vocabulary. The speaker adds it as well if it was just in-
vented. Success (when the hearer interpreted correctly the word as
the topic): both hearer and speaker remove synonyms/homonyms in
conflict with the word-meaning association used during the interac-
tion. In both examples the topic is the apple, and the word rimi.
Measures
The self-organization process happening while simulating the
Naming Game has complex dynamics, and goes through
various states before reaching global consensus. We talk
about those dynamics as a convergence process, towards the
state where all agents share the exact same lexicon, with
exactly one word for each meaning without synonymy and
homonymy. This state is stable, lexicons will not change any-
more whatever are the modalities of the interaction – which
agent is the speaker, which is the hearer, and which mean-
ings and words are used. Convergence and stability for dif-
ferent types of Naming Games has been proved analytically
(De Vylder, 2007). In this paper, we do not focus on whether
the model converges or not, but on the speed and complex-
ity properties of the dynamics before convergence. Measures
for each of those aspects, used to describe the system while in
this intermediate state, were defined in previous work (Loreto
et al., 2011). We distinguish local measures –accessible to
each agent– from global measures, computed on the whole
population.
TCS: Theoretical Communicative Success The Theoret-
ical Communicative Success is a measure of distance to the
fully converged state. First, for each meaning, we can con-
sider the probability of having a successful communication
when using this meaning as a topic, given a state of the pop-
ulation. The TCS is the average of those probabilities, over
all possible meanings. In the case of Random Topic Choice,
this measure coincides with the general probability of having
a successful interaction. By definition, it is a global measure,
not accessible to individual agents. To retrieve its value, we
can either estimate it using a snapshot of the population and
a Monte Carlo method with random topic choice, or compute
it. To detail the exact computation formula, we need to first
define the probability of success between two given vocabu-
laries of agents A and B. As detailed in the previous section,
a vocabulary has 2 components: a coding part, used to find
words associated to a meaning, and a decoding part, used to
find meanings associated to a word. For vocabulary V (A),
we would then have the 2 matrices V c(A) and V d(A). If A
is the speaker and B the hearer, A is coding and B decoding,
hence the formula of the probability of success in this case,
averaged over all possible meanings:
TCSs(A,B) =
1
M ∑
m
∑
w
V c(A)m,w ·V d(B)m,w
(2)
Because before an interaction we do not necessarily know
which agent will be the speaker and which will be the hearer,
the 2 situations (A speaker and B hearer / B speaker and A
hearer) are to be considered, as equiprobable. The final value
TCS(A,B) is the mean of TCSs(A,B) and TCSs(B,A).
To scale up to population level, one can compute an aver-
age vocabulary for the whole population V (P), and then the
probability of success for an interaction between this lexicon
and itself. For a large enough population, this value is indeed
a good approximation of the probability of success. V (P) is
an element-wise average of the lexicon matrices of all agents.
When using random topic choice, this value abruptly goes
from 0 to 1 after a certain number of interactions. These dy-
namics can be seen on fig.4, where the random topic choice
is represented – among active strategies that are explained in
a following section. In practice, we use Monte Carlo estima-
tion for the values at population level over time, and the exact
computation for the active topic choice strategy (see follow-
ing section), as it requires more precision and the population
vocabulary is already built.
Local Complexity The starting state of an agent's vocabu-
lary is empty (all-zero matrices), and the end state is identical
coding and decoding matrices, with exactly one distinct word
per meaning. But between those 2 situations, through which
states goes the vocabulary? How much conflictual informa-
tion (synonymy and homonymy) has to be considered?
For each agent, we can define a local complexity measure,
by counting the number of distinct associations present in the
vocabulary.
In our case, this is exactly the sum of all ele-
ments of the matrix V (A). At the beginning of a simulation,
while the vocabulary is empty, this measure equals 0. At the
end, its value is the number of meanings M. When using ran-
dom topic choice, there is a fast growth to a maximum, before
a slow decrease to the final value M (can be seen in fig.4).
This measure is nearly proportional to the minimal memory
needed to represent the lexicon (as a sparse matrix or a list
of word-meaning associations), and therefore should remain
low in a cognitively plausible situation.
Active Topic Choice Strategy
The main contribution of this work is the definition of an ac-
tive strategy for the choice of the topic in each interaction,
by comparison to the usual choice of picking meanings ran-
domly (with a uniform distribution over the space of mean-
ings). The strategy has to be local, i.e. use only information
available to the agent, namely its own vocabulary and results
of past interactions it was involved in.
To both converge quickly and control complexity, behav-
ior should be driven by maximization at each interaction of
the Theoretical Communicative Success. However, this value
is a global measure, therefore not accessible at agent level.
Agents only sample information about the global state of the
population, or the average vocabulary V (P), through their in-
teractions as hearer or speaker.
The strategies for active topic choice found in previous
work are separated in two levels of decision (Schueller &
Oudeyer, 2016). First, a decision between exploring a new
meaning (that is associated to no words in the vocabulary
so far) and choosing (exploiting) a meaning among those al-
ready used before. Then, if exploiting, deciding which known
meaning to use depending on past interaction results.
The strategy introduced in this work keeps those two levels,
while basing both decisions on a new measure called Local
Approximated Probability of Success (LAPS), using a local
representation of V (P).
LAPS: Local Approximated Probability of Success
Here, we define an approximation of V (P), (cid:101)V (P), using in-
construct independently the coding and decoding parts(cid:101)V c(P)
and(cid:101)V d(P). For every meaning m (and every word w), we use
formation sampled by agents during their interactions. We
a sliding window over the recent past interactions – of maxi-
mal length τ, the time scale parameter– and count the number
of times it is associated to each word w(cid:48) (or meaning m(cid:48)). This
value divided by τ is the local estimation of the probability of
an other agent coding m using w(cid:48) (or decoding w as m(cid:48)). With
this, we retrieve the values of both matrices(cid:101)V c(P) and(cid:101)V d(P).
using the meaning m. We can now build(cid:101)V c(P):
Let Mc(m) be the memory of the past interactions where
m was the topic, if there has been Tm such interactions. wt
denotes the word used during the tth interaction of the agent
Mc(m) = (wt )1≤t≤Tm
(cid:101)V c(P)mw =
Tm
∑
t=Tm−τ+1
τ
δw,wt
(3)
Similarly, by defining Md(w) be the memory of the past
interactions where w was the topic, with Tw such interactions,
we can build(cid:101)V d(P):
Md(w) = (mt )1≤t≤Tw
Until τ interactions have been done with a given meaning
(cid:101)V d(P)mw =
(cid:101)V d(P)mw do not sum to 1. The re-
(cid:101)V c(P)mw and ∑
t=Tw−τ+1
τ
δm,mt
Tw
∑
(4)
or word, ∑
w
m
maining probability weight is assumed to be associated with
failure. If we would normalize to 1, with a single interaction
an agent would already estimate as 100% sure that the same
word-meaning association would be used again with the same
topic for example. Without the normalization, this happens
only after τ interactions. In other words, this reflects lack of
information due to small sample size. We define a Local Ap-
proximated Probability of Success, or a local equivalent of
the Theoretical Communicative Success for an agent A with
vocabulary V (A):
LAPSA = TCS(V (A),(cid:101)V (P)A)
(5)
For some vocabulary update policies called lateral inhibition,
similar matrices are computed, but used directly as an agent's
own representation of the lexicon. This usage does not pre-
vent the complexity burst (Wellens, 2012).
Exploration vs. Exploitation The first choice of our new
strategy is between exploring new meanings or exploiting al-
ready known ones. Exploration should happen when agents
are confident enough about their agreement with the rest
of the population over their known meanings (Schueller &
Oudeyer, 2016). The LAPS measure in itself is a measure of
confidence, and the simplest way to take this into account is
to only explore when reaching the maximum value K
M where
K is the number of known meanings and M the total number
of meanings in the world. This value can actually be reached,
thanks to the sliding window of parameter τ.
Multi-Armed Bandit The second decision process con-
cerns the exploitation part, when picking the topic among the
known meanings. We designed a behavior driven by the in-
crease of the LAPS measure.
In other words, agents seek
the meaning that would yield the greatest increase of LAPS.
However, computing the expectancy of this value ∆LAPS is
hard computationally speaking, and therefore not suitable for
a model of a cognitive process. We can only consider the pro-
cess a black box, where following a decision between a finite
set of options, a reward value is obtained. This is exactly the
definition of the Multi-Armed Bandit problem, associated to
a class of reinforcement learning algorithms that have been
extensively studied (Bubeck, Cesa-Bianchi, et al., 2012). The
name comes from an analogy with a person trying to maxi-
mize their gain while facing a set of slot-machines (also called
one armed bandit), and being able to use only one at a time.
The probability distribution of the reward of each machine
is unknown, and the player has to both collect information
by playing and exploit the highest rewarding machine – with
limited knowledge of its reward distribution – hence keep bal-
ance between exploration and exploitation. In our problem,
we can see known meanings as the possible arms, and the re-
ward ∆LAPS. Our case is quite specific, as: 1) distributions
are non stationary, 2) they depend on past choices, 3) and the
number of arms grows over time (and starts at 0). This spe-
cific situation led us to choose an algorithm, where weights
associated to each arm undergo a decay over time, which let
them stay at the same order of magnitude of the initial weights
of new arms (Clement, Roy, Oudeyer, & Lopes, 2015). Our
algorithm depends on 2 parameters: integrated balance be-
tween reward-driven exploitation and random exploration be-
tween arms through the parameter γ, and time scale n for the
decay of weights. As a reward, we consider the increase of
LAPS yielded by the interaction, ∆LAPS, or 0 if the latter is
negative in order to avoid negative weights. See algorithm 1.
Algorithm 1 LAPSmax Bandit Multi-Armed Bandit algo-
rithm used as a topic choice strategy maximizing the LAPS
measure. New arms are created with weights wa equal to the
reward ri obtained at the end of an interaction with a new
meaning.
Require: γ rate of exploration for bandit
Require: n time scale for weights decay
Require: vocabularies V and(cid:101)V , #meanings M
K ← V.known meanings()
if LAPSA = K
M then
m ← random (V.unknown meanings())
for a ∈ Arms do
else
wa = wa
∑ j w j
pa = (1− γ)· wa + γ
K
end for
Sample m ∈ Arms using distribution (pa)a∈Arms
end if
return m
{Interact using topic m and compute reward r}
if m ∈ Arms then
wm ← n
n+1 · wm + r
Add m to Arms with wm = r
else
end if
Results
For all simulations, we set N=M=W =40, compute up to
80,000 interactions and take the mean over 8 trials. The situ-
ation M=W is the most constrained and complex to solve, as
synonymy and homonymy are more probable. We ran simu-
lations for 1 ≤ τ ≤ 50, and set n=τ. For the exploration rate,
if the condition 0 < γ (cid:28) 1 is respected, the actual value of
γ does not matter much, as its only function is to avoid rare
cases where some weights reach a value of 0 and cannot be
selected anymore. We set γ=0.01. However, we also ran sim-
ulations with pure random choices at the bandit level, to be
able to study the influence of each level of our algorithm. This
case identifies as γ=1.
The evolution of the TCS and complexity over time is rep-
resented on fig.4, for several values of the time scale τ. They
are compared on the same plots with Random Topic Choice.
We can see that convergence is faster for low values of τ, the
fastest being for τ=2, which is 4 times faster than Random
Topic Choice. As for complexity, for all configurations ex-
Figure 4: Theoretical success rate over time and Local Com-
plexity for Random Topic Choice and Active Topic Choice
with several values of the time scale τ used in the LAPS mea-
sure. N=M=W =40, γ=0.01, mean over 8 trials.
cepted τ=1 values stay below the final level 40. After reach-
ing a first threshold, they increase linearly with time, the slope
being smaller for higher values of τ. For τ=1, the maximum
value is only half of the maximum reached by Random Topic
Choice.
in this
case LAPS is an autocorrelation with the current interaction,
by definition older interactions are not taken into account.
It is understandable that τ=1 is an outlier:
On fig.5, we can see the dependency of convergence time
on the parameter τ, plotted for configurations γ=0.01, γ=1 and
the value for Random Topic Choice as a reference. Both have
dynamics consistently faster than Random Topic Choice for
low values of τ, however γ=0.01 performs better. Excepted
for τ=1, convergence time increases linearly with τ for both,
with a minimum at τ=2, and a smaller slope for γ=0.01.
Discussion
Results show that the new strategy presented in this paper
1) allows fast convergence, 2) controls efficiently complexity
growth, 3) its dynamics are consistent and highly correlated
with parameter change, 4) the 2 levels of the algorithm each
contribute to the increased performance.
Barto, A. G. (2013). Intrinsic motivation and reinforcement
learning. In Intrinsically motivated learning in natural and
artificial systems (pp. 17–47). Springer.
Bubeck, S., Cesa-Bianchi, N., et al. (2012). Regret analysis of
stochastic and nonstochastic multi-armed bandit problems.
Foundations and Trends R(cid:13) in Machine Learning, 5(1), 1–
122.
Clement, B., Roy, D., Oudeyer, P.-Y., & Lopes, M. (2015).
Multi-Armed Bandits for Intelligent Tutoring Systems.
Journal of Educational Data Mining (JEDM), 7(2), 20–48.
Cornudella, M., Van Eecke, P., & Van Trijp, R. (2015). How
intrinsic motivation can speed up language emergence. In
Proceedings of the european conference on artificial life
(pp. 571–578).
De Vylder, B. (2007). The evolution of conventions in multi-
agent systems. Unpublished doctoral dissertation, Vrije
Universiteit Brussel, Brussels.
Gottlieb, J., Oudeyer, P.-Y., Lopes, M., & Baranes, A. (2013).
Information-seeking, curiosity, and attention: computa-
tional and neural mechanisms. Trends in cognitive sciences,
17(11), 585–593.
Ke, J., Minett, J. W., Au, C.-P., & Wang, W. S.-Y. (2002).
Self-organization and selection in the emergence of vocab-
ulary. Complexity, 7(3), 41–54.
Loreto, V., Baronchelli, A., Mukherjee, A., Puglisi, A., &
Tria, F. (2011). Statistical physics of language dynamics
(Vol. 2011) (No. 04).
Oliphant, M., & Batali, J. (1997). Learning and the emer-
gence of coordinated communication. Center for research
on language newsletter, 11(1), 1–46.
Oudeyer, P.-Y., Kaplan, F., & Hafner, V. V. (2007). Intrinsic
motivation systems for autonomous mental development.
Evolutionary Computation, IEEE Transactions on, 11(2),
265–286.
Oudeyer, P.-Y., & Smith, L.
(2014). How evolution may
work through curiosity-driven developmental process. Top-
ics Cogn. Sci.
Partridge, E., McGovern, M., Yung, A., & Kidd, C. (2015).
Young Childrens Self-Directed Information Gathering on
Touchscreens. In Proceedings of the 37th annual meeting
of the cognitive science society.
Schueller, W., & Oudeyer, P.-Y. (2016). Active control of
complexity growth in naming games: Hearer's choice. In
The evolution of language: Proceedings of the 11th inter-
national conference (evolangx11).
Steels, L., & Kaplan, F. (1998). Stochasticity as a source of
innovation in language games. In Proceedings of artificial
life vi (pp. 368–376).
Wellens, P. (2012). Adaptive Strategies in the Emergence of
Lexical Systems. Unpublished doctoral dissertation, Vrije
Universiteit Brussel, Brussels.
Figure 5: Convergence time depending on time scale τ used
in LAPS measure, for both γ=0.01 and γ=1, compared with
Random Topic Choice. N=M=W =40, mean over 8 trials.
With τ=2, each agent on average only speaks 15 times
about each meaning before convergence (i.e.
less than half
the population), and information has already been both con-
veyed between all agents and disambiguated. The linearity of
the evolution of TCS and complexity lets think that this algo-
rithm may as well scale efficiently to other values of N, M and
W . Compared to previous work, this topic choice algorithm
is more robust, and optimal parameters are easier to find. It
generalized well to Minimal Naming Game and can be used
for all other Naming Game models.
LAPS is coherent from a cognitive point of view, and cor-
responds to an actual internal confidence about quality of
communication with the rest of the population. As stated in
the results section, the case τ=1 is an outlier, being a sim-
ple autocorrelation with the current interaction. The optimal
value τ=2 is then the lowest possible value taking into account
past interactions, i.e.(cid:101)V (P) takes the lowest possible memory,
which is therefore credible for humans. Further work will be
needed to determine for which values of N, M and W τ=2
stays the optimal value.
Acknowledgments
The IdEx program (Univ. de Bordeaux) allowed W. Schueller
to visit V. Loreto. We thank Miguel Ibanez Berganza and
Benjamin Cl´ement for the fruitful discussions.
Source code The Python code used for
the simula-
tions of this paper is available as open source software:
https://github.com/wschuell/notebooks cogsci2018
References
Baldassarre, G., & Mirolli, M.
Intrinsically moti-
vated learning systems: an overview. In Intrinsically moti-
vated learning in natural and artificial systems (pp. 1–14).
Springer.
(2013).
|
1208.2503 | 1 | 1208 | 2012-08-13T06:05:01 | Distributed Pareto Optimization via Diffusion Strategies | [
"cs.MA",
"math.OC"
] | We consider solving multi-objective optimization problems in a distributed manner by a network of cooperating and learning agents. The problem is equivalent to optimizing a global cost that is the sum of individual components. The optimizers of the individual components do not necessarily coincide and the network therefore needs to seek Pareto optimal solutions. We develop a distributed solution that relies on a general class of adaptive diffusion strategies. We show how the diffusion process can be represented as the cascade composition of three operators: two combination operators and a gradient descent operator. Using the Banach fixed-point theorem, we establish the existence of a unique fixed point for the composite cascade. We then study how close each agent converges towards this fixed point, and also examine how close the Pareto solution is to the fixed point. We perform a detailed mean-square error analysis and establish that all agents are able to converge to the same Pareto optimal solution within a sufficiently small mean-square-error (MSE) bound even for constant step-sizes. We illustrate one application of the theory to collaborative decision making in finance by a network of agents. | cs.MA | cs |
Distributed Pareto Optimization via Diffusion
1
Strategies
Jianshu Chen, Student Member, IEEE, and Ali H. Sayed, Fellow, IEEE
Abstract
We consider solving multi-objective optimization problems in a distributed manner by a network
of cooperating and learning agents. The problem is equivalent to optimizing a global cost that is the
sum of individual components. The optimizers of the individual components do not necessarily coincide
and the network therefore needs to seek Pareto optimal solutions. We develop a distributed solution that
relies on a general class of adaptive diffusion strategies. We show how the diffusion process can be
represented as the cascade composition of three operators: two combination operators and a gradient
descent operator. Using the Banach fixed-point theorem, we establish the existence of a unique fixed
point for the composite cascade. We then study how close each agent converges towards this fixed point,
and also examine how close the Pareto solution is to the fixed point. We perform a detailed mean-square
error analysis and establish that all agents are able to converge to the same Pareto optimal solution
within a sufficiently small mean-square-error (MSE) bound even for constant step-sizes. We illustrate
one application of the theory to collaborative decision making in finance by a network of agents.
Index Terms
Distributed optimization, network optimization, diffusion adaptation, Pareto optimality, mean-square
performance, convergence, stability, fixed point, collaborative decision making.
I. INTRODUCTION
We consider solving a multi-objective optimization problem in a distributed manner over a network of
k (w);
N cooperative learners (see Fig. 1). Each agent k is associated with an individual cost function J o
The authors are with Department of Electrical Engineering, University of California, Los Angeles, CA 90095. Email: {jshchen,
sayed}@ee.ucla.edu.
This work was supported in part by NSF grants CCF-1011918 and CCF-0942936. A preliminary short version of this work
is reported in the conference publication [1].
August 21, 2018
DRAFT
and each of these costs may not be minimized at the same vector wo. As such, we need to seek a solution
that is "optimal" in some sense for the entire network. In these cases, a general concept of optimality
known as Pareto optimality is useful to characterize how good a solution is. A solution wo is said to be
2
Pareto optimal if there does not exist another vector w that is able to improve (i.e., reduce) any particular
l (w)}l(cid:54)=k. To illustrate the
k (w), without degrading (increasing) some of the other costs {J o
cost, say, J o
idea of Pareto optimality, let
O (cid:44) {(J o
1 (w), . . . , J o
(1)
denote the set of achievable cost values, where W denotes the feasible set. Each point P ∈ O represents
attained values for the cost functions {J o
l (w)} at a certain w ∈ W. Let us consider the two-node case
(N = 2) shown in Fig. 2, where the shaded areas represent the set O for two situations of interest. In Fig.
N (w)) : w ∈ W} ⊆ RN
Fig. 1. A network of N cooperating agents; a cost function J o
node k is denoted by Nk (including node k itself); this set consists of all nodes with which node k can share information.
k (w) is associated with each node k. The set of neighbors of
2(a), both J o
1 (w) and J o
1 (wo), J o
1 (w) attains its minimum at point P1, while J o
2 (w) achieve their minima at the same point P = (J o
2 (wo)), where wo is
the common minimizer. In comparison, in Fig. 2(b), J o
2 (w)
attains its minimum at point P2, so that they do not have a common minimizer. Instead, all the points
on the heavy red curve between points P1 and P2 are Pareto optimal solutions. For example, starting at
point A on the curve, if we want to reduce the value of J o
2 (w),
then we will need to move out of the achievable set O. The alternative choice that would keep us on the
curve is to move to another Pareto optimal point B, which would however increase the value of J o
2 (w).
1 (w). For this reason, the curve from P1 to
In other words, we need to trade the value of J o
P2 is called the optimal tradeoff curve (or optimal tradeoff surface if N > 2) [2, p.183].
1 (w) without increasing the value of J o
2 (w) for J o
August 21, 2018
DRAFT
12487k5936NkNkak;1ck;1Jok(w)Jo1(w)Jo2(w)Jo3(w)Jo5(w)Jo6(w)Jo4(w)Jo7(w)Jo9(w)Jo8(w)a1;kc1;k 3
(a)
(b)
Fig. 2. Optimal and Pareto optimal points for the case N = 2: (Left) P denotes the optimal point where both cost functions
are minimized simultaneously and (Right) Pareto optimal points lie on the red boundary curve.
To solve for Pareto optimal solutions, a useful scalarization technique is often used to form an aggregate
cost function that is the weighted sum of the component costs as follows:
N(cid:88)
J glob(w) =
πlJ o
l (w)
(2)
l=1
where πl is a positive weight attached with the lth cost. It was shown in [2, pp.178 -- 180] that the minimizer
of (2) is Pareto optimal for the multi-objective optimization problem. Moreover, by varying the values
of {πl}, we are able to get different Pareto optimal points on the tradeoff curve. Observing that we can
always define a new cost Jl(w) by incorporating the weighting scalar πl,
Jl(w) (cid:44) πlJ o
l (w)
it is sufficient for our future discussions to focus on aggregate costs of the following form:
N(cid:88)
J glob(w) =
Jl(w)
(3)
(4)
l=1
constraint of the form pT
If desired, we can also add constraints to problem (4). For example, suppose there is additionally some
k w < bk at node k, where pk is M × 1 and bk is a scalar. Then, we can consider
using barrier functions to convert the constrained optimization problem to an unconstrained problem [2],
[3]. For example, we can redefine each cost Jk(w) to be Jk(w) ← Jk(w) + φ(pT
k w − bk), where φ(x)
is a barrier function that penalizes values of w that violate the constraint. Therefore, without loss of
generality, we shall assume W = RM and only consider unconstrained optimization problems. Moreover,
l (w)} are differentiable and, for each given set of positive weights {πl}, the cost
we shall assume the {J o
August 21, 2018
DRAFT
Jo2(w)Jo2(w)Jo1(w)Jo1(w)OOPPJo2(w)Jo2(w)Jo1(w)Jo1(w)OOP1P1P2P2AABBCC 4
J glob(w) in (2) or (4) is strongly convex so that the minimizer wo is unique [4]. Note that the new cost
Jl(w) in (3) depends on πl so that the wo that minimizes J glob(w) in (4) also depends on {πl}.
One of the most studied approaches to the distributed solution of such optimization problems is the
incremental approach -- see, e.g., [5] -- [12]. In this approach, a cyclic path is defined over the nodes
and data are processed in a cyclic manner through the network until optimization is achieved. However,
determining a cyclic path that covers all nodes is generally an NP-hard problem [13] and, in addition,
cyclic trajectories are vulnerable to link and node failures. Another useful distributed optimization
approach relies on the use of consensus strategies [5], [14] -- [20]. In this approach, vanishing step-size
sequences are used to ensure that agents reach consensus and agree about the optimizer in steady-state.
However, in time-varying environments, diminishing step-sizes prevent the network from continuous
learning; when the step-sizes die out, the network stops learning.
In [21], we generalized our earlier work on adaptation and learning over networks [22], [23] and
developed diffusion strategies that enable the decentralized optimization of global cost functions of the
form (4). In the diffusion approach, information is processed locally at the nodes and then diffused
through a real-time sharing mechanism. In this manner, the approach is scalable, robust to node and
link failures, and avoids the need for cyclic trajectories. In addition, compared to the aforementioned
consensus solutions (such as those in [16], [19], [24]), the diffusion strategies we consider here employ
constant (rather than vanishing) step-sizes in order to endow the resulting networks with continuous
learning and tracking abilities. By keeping the step-sizes constant, the agents are able to track drifts in
the underlying costs and in the location of the Pareto optimal solutions. One of the main challenges in
the ensuing analysis becomes that of showing that the agents are still able to approach the Pareto optimal
solution even with constant step-sizes; in this way, the resulting diffusion strategies are able to combine
the two useful properties of optimality and adaptation.
In [21], we focused on the important case where all costs {Jl(w)} share the same optimal solution
wo (as was the case with Fig. 2(a)); this situation arises when the agents in the network have a common
objective and they cooperate to solve the problem of mutual interest in a distributed manner. Examples
abound in biological networks where agents work together, for example, to locate food sources or
evade predators [25], and in collaborative spectrum sensing [26], system identification [27], and learning
applications [28]. In this paper, we develop the necessary theory to show that the same diffusion approach
(described by (9) -- (10) below) can be used to solve the more challenging multi-objective optimization
problem, where the agents need to converge instead to a Pareto optimal solution. Such situations are
common in the context of multi-agent decision making (see, e.g., [3] and also Sec. IV where we discuss
August 21, 2018
DRAFT
5
one application in the context of collaborative decision in finance). To study this more demanding scenario,
we first show that the proposed diffusion process can be represented as the cascade composition of three
operators: two combination (aggregation) operators and one gradient-descent operator. Using the Banach
fixed-point theorem [29, pp.299 -- 303], we establish the existence of a unique fixed point for the composite
cascade. We then study how close each agent in the network converges towards this fixed point, and also
examine how close the Pareto solution is to the fixed point. We perform a detailed mean-square error
analysis and establish that all agents are able to converge to the same Pareto optimal solution within a
sufficiently small mean-square-error (MSE) bound. We illustrate the results by considering an example
involving collaborative decision in financial applications.
Notation. Throughout the paper, all vectors are column vectors. We use boldface letters to denote random
quantities (such as uk,i) and regular font to denote their realizations or deterministic variables (such as
uk,i). We use diag{x1, . . . , xN} to denote a (block) diagonal matrix consisting of diagonal entries (blocks)
x1, . . . , xN , and use col{x1, . . . , xN} to denote a column vector formed by stacking x1, . . . , xN on top of
each other. The notation x (cid:22) y means each entry of the vector x is less than or equal to the corresponding
entry of the vector y.
II. DIFFUSION ADAPTATION STRATEGIES
In [21], we motivated and derived diffusion strategies for distributed optimization, which are captured
by the following general description:
φk,i−1 =
N(cid:88)
N(cid:88)
l=1
l=1
a1,lkwl,i−1
N(cid:88)
ψk,i = φk,i−1 − µk
wk,i =
l=1
a2,lkψl,i
clk∇wJl(φk,i−1)
(5)
(6)
(7)
where wk,i is the local estimate for wo at node k and time i, µk is the step-size parameter used by node k,
and {φk,i−1, ψk,i} are intermediate estimates for wo. Moreover, ∇wJl(·) is the (column) gradient vector
of Jl(·) relative to w. The non-negative coefficients {a1,lk}, {clk}, and {a2,lk} are the (l, k)-th entries of
matrices A1, C, and A2, respectively, and they are required to satisfy:
AT
1
1 = 1, AT
2
1 = 1, C1 = 1
a1,lk = 0, a2,lk = 0, clk = 0 if l /∈ Nk
(8)
DRAFT
August 21, 2018
6
where 1 denotes a vector with all entries equal to one. Note from (8) that the combination coefficients
{a1,lk, a2,lk, clk} are nonzero only for those l ∈ Nk. Therefore, the sums in (5) -- (7) are confined within the
neighborhood of node k. Condition (8) requires the combination matrices {A1, A2} to be left-stochastic,
while C is right-stochastic. We therefore note that each node k first aggregates the existing estimates
from its neighbors through (5) and generates the intermediate estimate φk,i−1. Then, node k aggregates
gradient information from its neighborhood and updates φk,i−1 to φk,i through (6). All other nodes in the
network are performing these same steps simultaneously. Finally, node k aggregates the estimates {φl,i}
through step (7) to update its weight estimate to wk,i.
Algorithm (5) -- (7) can be simplified to several special cases for different choices of the matrices
{A1, A2, C}. For example, the choice A1 = I, A2 = A and C = I reduces to the adapt-then-combine
(ATC) strategy that has no exchange of gradient information [21] -- [23], [30]:
ψk,i = wk,i−1 − µk∇wJk(wk,i−1)
wk,i =
alkψl,i
(ATC, C = I)
(9)
while the choice A1 = A, A2 = I and C = I reduces to the combine-then-adapt (CTA) strategy, where
the order of the combination and adaptation steps are reversed relative to (9) [22], [23], [30]:
(cid:88)
l∈Nk
(cid:88)
alkwl,i−1
ψk,i−1 =
wk,i = ψk,i−1 − µk∇wJk(ψk,i−1)
l∈Nk
(CTA, C = I)
(10)
Furthermore, if in the CTA implementation (10) we enforce A to be doubly stochastic, replace ∇wJk(·) by
a subgradient, and use a time-decaying step-size parameter (µk(i) → 0), then we obtain the unconstrained
version used by [24]. In the sequel, we continue with the general recursions (5) -- (7), which allow us
to examine the convergence properties of several algorithms in a unified manner. The challenge we
encounter now is to show that this same class of algorithms can still optimize the cost (4) in a distributed
manner when the individual costs {Jl(w)} do not necessarily have the same minimizer. This is actually a
demanding task, as the analysis in the coming sections reveals, and we need to introduce novel analysis
techniques to be able to handle this general case.
August 21, 2018
DRAFT
7
A. Modeling Assumptions
III. PERFORMANCE ANALYSIS
In most situations in practice, the true gradient vectors needed in (6) are not available. Instead, perturbed
versions are available, which we model as
(cid:91)∇wJl(w) = ∇wJl(w) + vl,i(w)
(11)
where the random noise term, vl,i(w), may depend on w and will be required to satisfy certain conditions
given by (16) -- (17). We refer to the perturbation in (11) as gradient noise. Using (11), the diffusion
algorithm (5) -- (7) becomes the following, where we are using boldface letters for various quantities to
highlight the fact that they are now stochastic in nature due to the randomness in the noise component:
N(cid:88)
N(cid:88)
l=1
N(cid:88)
(cid:104)∇wJl(φk,i−1) + vl,i(φk,i−1)
(cid:105)
clk
(12)
(13)
(14)
φk,i−1 =
a1,lkwl,i−1
ψk,i = φk,i−1 − µk
wk,i =
l=1
a2,lkψl,i
l=1
Using (12) -- (14), we now proceed to examine the mean-square performance of the diffusion strategies.
Specifically, in the sequel, we study: (i) how fast and (ii) how close the estimator wk,i at each node k
approaches the Pareto-optimal solution wo in the mean-square-error sense. We establish the convergence
of all nodes towards the same Pareto-optimal solution within a small MSE bound. Since we are dealing
with individual costs that may not have a common minimizer, the approach we employ to examine the
convergence properties of the diffusion strategy is fundamentally different from [21]; we follow a system-
theoretic approach and call upon the fixed-point theorem for contractive mappings [29, pp.299 -- 303].
To proceed with the analysis, we introduce the following assumptions on the cost functions and gradient
noise. As explained in [21], these conditions are weaker than similar conditions in the literature of
distributed optimization; in this way, our convergence and performance results hold under more relaxed
conditions than usually considered in the literature.
Assumption 1 (Bounded Hessian). Each component cost function Jl(w) has a bounded Hessian matrix,
i.e., there exist nonnegative real numbers λl,min and λl,max such that, for each k = 1, . . . , N:
λl,minIM ≤ ∇2
wJl(w) ≤ λl,maxIM
with(cid:80)N
l=1 clkλl,min > 0.
August 21, 2018
(15)
(cid:4)
DRAFT
Assumption 2 (Gradient noise). There exist α ≥ 0 and σ2
v ≥ 0 such that, for all w ∈ Fi−1:
E{vl,i(w) Fi−1} = 0
E(cid:8)(cid:107)vl,i(w)(cid:107)2(cid:9) ≤ α · E(cid:107)∇wJl(w)(cid:107)2 + σ2
v
for all i, l, where Fi−1 denotes the past history of estimators {wk,j} for j ≤ i − 1 and all k.
8
(16)
(17)
(cid:4)
If we choose C = I, then Assumption 1 implies that the cost functions {Jl(w)} are strongly convex1.
This condition can be guaranteed by adding small regularization terms. For example, we can convert a non-
l (w) to a strongly convex one by redefining Jl(w) as Jl(w) ← Jl(w) + (cid:107)w(cid:107)2,
strongly convex function J(cid:48)
where > 0 is a small regularization factor. We further note that, assumption (17) is a mix of the
"relative random noise" and "absolute random noise" model usually assumed in stochastic approximation
[4]. Condition (17) implies that the gradient noise grows when the estimate is away from the optimum
(large gradient). Condition (17) also states that even when the gradient vector is zero, there is still some
residual noise variance σ2
v.
B. Diffusion Adaptation Operators
To analyze the performance of the diffusion adaptation strategies, we first represent the mappings
performed by (12) -- (14) in terms of useful operators.
Definition 1 (Combination Operator). Suppose x = col{x1, . . . , xN} is an arbitrary N × 1 block column
vector that is formed by stacking M × 1 vectors x1, . . . , xN on top of each other. The combination
operator TA : RM N → RM N is defined as the linear mapping:
TA(x) (cid:44) (AT ⊗ IM ) x
(18)
where A is an N × N left stochastic matrix, and ⊗ denotes the Kronecker product operation.
(cid:4)
Definition 2 (Gradient-Descent Operator). Consider the same N × 1 block column vector x. Then, the
1A differentiable function f (x) on Rn is said to be strongly convex if there exists a λmin > 0 such that f (x + y) ≥
f (x)+yT∇f (x)+λmin(cid:107)y(cid:107)2/2 for any x, y ∈ Rn. And if f (x) is twice-differentiable, this is also equivalent to ∇2f (x) ≥ λminI
[4, pp.9-10]. Strong convexity implies that the function f (x) can be lower bounded by some quadratic function.
August 21, 2018
DRAFT
gradient-descent operator TG : RM N → RM N is the nonlinear mapping defined by:
x1 − µ1
xN − µN
(cid:80)N
l=1 cl1∇wJl(x1)
...
(cid:80)N
l=1 clN∇wJl(xN )
TG(x) (cid:44)
9
(19)
(cid:4)
Definition 3 (Power Operator). Consider the same N × 1 block vector x. The power operator P :
RM N → RN is defined as the mapping:
where (cid:107) · (cid:107) denotes the Euclidean norm of a vector.
P [x] (cid:44) col{(cid:107)x1(cid:107)2, . . . ,(cid:107)xN(cid:107)2}
(20)
(cid:4)
We will use the power operator to study how error variances propagate after a specific operator TA(·)
or TG(·) is applied to a random vector. We remark that we are using the notation "P [·]" rather than "P (·)"
to highlight the fact that P is a mapping from RM N to a lower dimensional space RN . In addition to
the above three operators, we define the following aggregate vector of gradient noise that depends on the
(cid:40)
N(cid:88)
l=1
cl1vl(x1), . . . , µN
N(cid:88)
l=1
v(x) (cid:44)−col
µ1
(cid:41)
clN vl(xN )
(21)
state x:
With these definitions, we can now represent the two combination steps (12) and (14) as two combination
operators TA1(·) and TA2(·). We can also represent the adaptation step (13) by a gradient-descent operator
perturbed by the noise operator (21):
We can view (cid:98)TG(x) as a random operator that maps each input x ∈ RM N into an RM N random vector,
(22)
and we use boldface letter to highlight this random nature. Let
wi (cid:44) col{w1,i, w2,i, . . . , wN,i}
(23)
denote the vector that collects the estimators across all nodes. Then, the overall diffusion adaptation steps
(12) -- (14) that update wi−1 to wi can be represented as a cascade composition of three operators:
(cid:98)TG(x) (cid:44) TG(x) + v(x)
(cid:98)Td(·) (cid:44) TA2 ◦ (cid:98)TG ◦ TA1(·)
August 21, 2018
(24)
DRAFT
10
(a) TA1 (·), TA2 (·) and TG(·).
(b) Cascade representation of diffusion adaptation.
Fig. 3. Representation of the diffusion adaptation strategy (12) -- (14) in terms of operators. Each diffusion adaptation step can
then (cid:98)Td(·) becomes Td(·).
be viewed as a cascade composition of three operators: TA1 (·), TG(·), and TA2 (·) with gradient perturbation v(·). If v(·) = 0,
where we use ◦ to denote the composition of any two operators, i.e., T1 ◦ T2(x) (cid:44) T1(T2(x)). If there
is no gradient noise, then the diffusion adaptation operator (24) reduces to
Td(·) (cid:44) TA2 ◦ TG ◦ TA1(·)
(25)
In other words, the diffusion adaptation over the entire network with and without gradient noise can be
described in the following compact forms:
wi = (cid:98)Td(wi−1)
(26)
wi = Td(wi−1)
(27)
Fig. 3(a) illustrates the role of the combination operator TA(·) (combination steps) and the gradient-
descent operator TG(·) (adaptation step). The combination operator TA(·) aggregates the estimates from
the neighborhood (social learning), while the gradient-descent operator TG(·) incorporates information
from the local gradient vector (self-learning). In Fig. 3(b), we show that each diffusion adaptation step
can be represented as the cascade composition of three operators, with perturbation from the gradient
noise operator. Next, in Lemma 1, we examine some of the properties of the operators {TA1, TA2, TG},
which are proved in Appendix A.
Lemma 1 (Useful Properties). Consider N×1 block vectors x = col{x1, . . . , xN} and y = col{y1, . . . , yN}
with M × 1 entries {xk, yk}. Then, the operators TA(·), TG(·) and P [·] satisfy the following properties:
1) (Linearity): TA(·) is a linear operator.
2) (Nonnegativity): P [x] (cid:23) 0.
August 21, 2018
DRAFT
26664377752666437775+ + + (¢)26664377752666437775(¢)26664377752666437775+ + + (¢)wi¡1wi¡1Ái¡1Ái¡1Ái¡1Ái¡1Ãi¡1Ãi¡1Ãi¡1Ãi¡1wiwi(¢)(¢)(¢)+ v(¢)bT(¢)bT(¢)w¡wbT(¢)ww…… bT(¢)ww¡bT(¢)w 11
3) (Scaling): For any scalar a ∈ R, we have
(28)
4) (Convexity): suppose x(1), . . . , x(K) are N × 1 block vectors formed in the same manner as x, and
P [ax] = a2P [x]
let a1, . . . , aK be non-negative real scalars that add up to one. Then,
(29)
5) (Additivity): Suppose x = col{x1, . . . , xN} and y = col{y1, . . . , yN} are N × 1 block random
P [a1x(1) + ··· + aKx(K)] (cid:22) a1P [x(1)] + ··· + aKP [x(K)]
vectors that satisfy ExT
k yk = 0 for k = 1, . . . , N. Then,
EP [x + y] = EP [x] + EP [y]
P [TA(x)] (cid:22) AT P [x]
P [TG(x) − TG(y)] (cid:22) Γ2P [x − y]
6) (Variance relations):
where
Γ (cid:44) diag{γ1, . . . , γN}
γk (cid:44) max{1 − µkσk,max, 1 − µkσk,min}
σk,min (cid:44)
clkλl,min, σk,max (cid:44)
N(cid:88)
l=1
7) (Block Maximum Norm): The ∞−norm of P [x] is the squared block maximum norm of x:
clkλl,max
N(cid:88)
(cid:107)xk(cid:107)(cid:17)2
l=1
(cid:107)P [x](cid:107)∞ = (cid:107)x(cid:107)2
b,∞ (cid:44)(cid:16)
max
1≤k≤N
(30)
(31)
(32)
(33)
(34)
(35)
(36)
8) (Preservation of Inequality): Suppose vectors x, y and matrix F have nonnegative entries, then
(cid:4)
x (cid:22) y implies F x (cid:22) F y.
C. Transient Analysis
Using the operator representation developed above, we now analyze the transient behavior of the
diffusion algorithm (12) -- (14). From Fig. 3(b) and the previous discussion, we know that the stochastic
recursion wi = (cid:98)Td(wi−1) is a perturbed version of the noise-free recursion wi = Td(wi−1). Therefore, we
first study the convergence of the noise free recursion, and then analyze the effect of gradient perturbation
on the stochastic recursion.
August 21, 2018
DRAFT
12
Intuitively, if recursion wi = Td(wi−1) converges, then it should converge to a vector w∞ that satisfies
w∞ = Td(w∞)
(37)
In other words, the vector w∞ should be a fixed point of the operator Td(·) [29, p.299]. We need to
answer four questions pertaining to the fixed point. First, does the fixed point exist? Second, is it unique?
Third, under which condition does the recursion wi = Td(wi−1) converge to the fixed point? Fourth, how
far is the fixed point w∞ away from the minimizer wo of (4)? We answer the first two questions using
the Banach Fixed Point Theorem (Contraction Theorem) [29, pp.2 -- 9, pp.299 -- 300]. Afterwards, we study
convergence under gradient perturbation. The last question will be considered in the next subsection.
Definition 4 (Metric Space). A set X, whose elements we shall call points, is said to be a metric space
if we can associate a real number d(p, q) with any two points p and q of X, such that
(a) d(p, q) > 0 if p (cid:54)= q; d(p, p) = 0;
(b) d(p, q) = d(q, p);
(c) d(p, q) ≤ d(p, r) + d(r, q), for any r ∈ X.
Any function d(p, q) with these three properties is called a distance function, or a metric, and we denote
a metric space X with distance d(·,·) as (X, d).
(cid:4)
Definition 5 (Contraction). Let (X, d) be a metric space. A mapping T : X −→ X is called a contraction
on X if there is a positive real number δ < 1 such that d(T (x), T (y)) ≤ δ · d(x, y) for all x, y ∈ X
Lemma 2 (Banach Fixed Point Theorem [29]). Consider a metric space (X, d), where X (cid:54)= ∅. Suppose
that X is complete2 and let T : X → X be a contraction. Then, T has precisely one fixed point.
(cid:4)
As long as we can prove that the diffusion operator Td(·) is a contraction, i.e., for any two points
x, y ∈ RM N , after we apply the operator Td(·), the distance between Td(x) and Td(y) scales down by a
scalar that is uniformly bounded away from one, then the fixed point w∞ defined in (37) exists and is
unique. We now proceed to show that Td(·) is a contraction operator in X = RM N when the step-size
parameters {µk} satisfy certain conditions.
Theorem 1 (Fixed Point). Suppose the step-size parameters {µk} satisfy the following conditions
0 < µk <
2
σk,max
,
k = 1, 2, . . . , N
(38)
2A metric space (X, d) is complete if any of its Cauchy sequences converges to a point in the space; a sequence {xn} is
Cauchy in (X, d) if ∀ > 0, there exists N such that d(xn, xm) < for all n, m > N.
August 21, 2018
DRAFT
Then, there exists a unique fixed point w∞ for the unperturbed diffusion operator Td(·) in (25).
13
Proof: Let x = col{x1, . . . , xN} ∈ RM N×1 be formed by stacking M × 1 vectors x1, . . . , xN on
top of each other. Similarly, let y = col{y1, . . . , yN}. The distance function d(x, y) that we will use is
induced from the block maximum norm (36): d(x, y) = (cid:107)x − y(cid:107)b,∞ = max1≤k≤N (cid:107)xk − yk(cid:107). From the
definition of the diffusion operator Td(·) in (25), we have
P [Td(x) − Td(y)]
(cid:104)
(cid:16)
(cid:17)(cid:105)
(cid:105)
TG ◦ TA1(x) − TG ◦ TA1(y)
TG ◦ TA1(x) − TG ◦ TA1(y)
TA2
(a)
= P
(cid:104)
(b)(cid:22) AT
2 P
2 Γ2P [TA1(x) − TA1(y)]
(c)(cid:22) AT
2 Γ2P [TA1(x − y)]
(d)
= AT
(e)(cid:22) AT
(39)
where steps (a) and (d) are because of the linearity of TA1(·) and TA2(·), steps (b) and (e) are because
of the variance relation property (31), and step (c) is due to the variance relation property (32). Taking
the ∞−norm of both sides of (39), we have
2 Γ2AT
1 P [x − y]
(cid:107)P [Td(x) − Td(y)](cid:107)∞ ≤ (cid:107)AT
2 Γ2AT
1 (cid:107)∞ · (cid:107)P [x − y](cid:107)∞
≤ (cid:107)Γ(cid:107)2∞ · (cid:107)P [x − y](cid:107)∞
1 (cid:107)∞ = (cid:107)AT
(40)
where, in the second inequality, we used the fact that (cid:107)AT
2 are
right-stochastic matrices. Using property (36), we can conclude from (40) that: (cid:107)Td(x) − Td(y)(cid:107)b,∞ ≤
(cid:107)Γ(cid:107)∞ · (cid:107)x − y(cid:107)b,∞. Therefore, the operator Td(·) is a contraction if (cid:107)Γ(cid:107)∞ < 1, which, by substituting
(33) -- (34), becomes
2 (cid:107)∞ = 1 since AT
1 and AT
1 − µkσk,max < 1,
1 − µkσk,min < 1,
k = 1, . . . , N
and we arrive at the condition (38) on the step-sizes In other words, if condition (38) holds for each
k = 1, . . . , N, then Td(·) is a contraction operator. By Lemma 2, the operator Td(·) will have a unique
fixed point w∞ that satisfies equation (37).
Given the existence and uniqueness of the fixed point, the third question to answer is if recursion
wi = Td(wi−1) converges to this fixed point. The answer is affimative under (38). However, we are not
going to study this question separately. Instead, we will analyze the convergence of the more demanding
noisy recursion (26). Therefore, we now study how fast and how close the successive estimators {wi}
August 21, 2018
DRAFT
TA2
(cid:104)
(cid:105)
TA2 ◦ (cid:98)TG ◦ TA1(wi−1) − TA2 ◦ TG ◦ TA1(w∞)
(cid:17)(cid:105)
(cid:104)
(cid:16)(cid:98)TG ◦ TA1(wi−1) − TG ◦ TA1(w∞)
(cid:104)(cid:98)TG ◦ TA1(wi−1) − TG ◦ TA1(w∞)
(cid:105)
(cid:16)
(cid:17)
(cid:104)
(cid:104)
(cid:110)EP
(cid:17) − TG
(cid:16)
(cid:17) − TG
TA1(wi−1)
TA1(wi−1)
TA1(w∞)
TA1(w∞)
(cid:17)(cid:105)
(cid:16)
(cid:16)
(cid:16)
EP
EP
TG
TG
+ EP
(cid:17)(cid:105)
+ v
TA1(wi−1)
(cid:104)
(cid:16)
(cid:17)(cid:105)(cid:111)
v
TA1(wi−1)
2 Γ2EP [TA1(wi−1) − TA1(w∞)] + AT
2
EP [v(TA1(wi−1))]
= EP
(a)
= EP
(b)(cid:22) AT
2
(c)
= AT
2
(d)
= AT
2
(e)(cid:22) AT
(f )(cid:22) AT
14
generated by recursion (26) approach w∞. Once this issue is addressed, we will then examine how close
w∞ is to the desired wo. Introduce the following mean-square-perturbation (MSP) vector at time i:
MSPi (cid:44) EP [wi − w∞]
(41)
The k-th entry of MSPi characterizes how far away the estimate wk,i at node k and time i is from wk,∞
in the mean-square sense. To study the closeness of wi to w∞, we shall study how the quantity MSPi
evolves over time. By (26), (37) and the definitions of (cid:98)Td(·) and Td(·) in (24) and (25), we obtain
MSPi = EP [wi − w∞]
2 Γ2AT
1 · EP [wi−1 − w∞] + AT
1 · MSPi−1 + AT
2
2
EP [v(TA1(wi−1))]
= AT
2 Γ2AT
(42)
where step (a) is by the linearity of TA1(·), steps (b) and (f) are by property (31), step (c) is by the
substitution of (22), step (d) is by Property 5 in Lemma 1 and assumption (16), and step (e) is by (32).
EP [v(TA1(wi−1))]
To proceed with the analysis, we establish the following lemma to bound the second term in (42).
Lemma 3 (Bound on Gradient Perturbation). It holds that
1·Ω2AT
EP [v(TA1(wi−1))] (cid:22) 4αλ2
max(cid:107)C(cid:107)2
1 ·EP [wi−1−w∞]+(cid:107)C(cid:107)2
1Ω2bv
where
August 21, 2018
λmax (cid:44) max
1≤k≤N
bv (cid:44) 4αλ2
maxAT
λk,max
1 P [w∞ − 1N ⊗ wo]
{2α(cid:107)∇wJk(wo)(cid:107)2 + σ2
v}
+ max
1≤k≤N
Ω (cid:44) diag{µ1, . . . , µN}
(43)
(44)
(45)
(46)
DRAFT
Proof: By the definition of v(x) in (21) with x = TA1(wi−1) being a random vector, we get
For each block in (47), using Jensen's inequality, we have
µ2
1
l=1 cl1vl(x1),
EP [v(x)] =
µ2
N
l=1 clN vl(xN )
E(cid:13)(cid:13)(cid:13)(cid:80)N
E(cid:13)(cid:13)(cid:13)(cid:80)N
...
(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)2
E(cid:13)(cid:13)(cid:13) N(cid:88)
l=1
clkvl(xk)
=
(cid:16) N(cid:88)
≤(cid:16) N(cid:88)
l=1
l=1
≤ (cid:107)C(cid:107)1
l=1
clk
clk
(cid:17)2 · E(cid:13)(cid:13)(cid:13) N(cid:88)
clk(cid:80)N
(cid:17)2 · N(cid:88)
clk(cid:80)N
(cid:104)
N(cid:88)
(cid:104)(cid:90) 1
l=1 clk
(cid:16)
clk
l=1
l=1
vl(xk)
l=1 clk
E(cid:107)vl(xk)(cid:107)2
αE(cid:107)∇wJl(xk)(cid:107)2 + σ2
(cid:13)(cid:13)(cid:13)2
(cid:105)
(cid:17)
v
(cid:105)
15
(47)
(48)
(49)
(50)
(51)
where (cid:107)·(cid:107)1 denotes the maximum absolute column sum, and in the last step, we used (17). Using (123),
∇wJl(xk) = ∇wJl(wo) +
dt
From (124) and the norm inequality (cid:107)x + y(cid:107)2 ≤ 2(cid:107)x(cid:107)2 + 2(cid:107)y(cid:107)2, we obtain
0
∇2
wJl
wo + t(xk − wo)
(xk − wo)
(cid:107)∇wJl(xk)(cid:107)2 ≤ 2(cid:107)∇wJl(wo)(cid:107)2 +2λ2
≤ 2(cid:107)∇wJl(wo)(cid:107)2 +2λ2
l,max·(cid:107)xk − wo(cid:107)2
max·(cid:107)xk − wo(cid:107)2
Substituting (50) into (48), we obtain
E(cid:13)(cid:13)(cid:13) N(cid:88)
l=1
(cid:13)(cid:13)(cid:13)2 ≤ (cid:107)C(cid:107)1
(cid:104)
N(cid:88)
clk
l=1
2αλ2
max
E(cid:107)xk−wo(cid:107)2 +2α(cid:107)∇wJl(wo)(cid:107)2 +σ2
v
(cid:105)
clkvl(xk)
≤ 2αλ2
max(cid:107)C(cid:107)2
1 · E(cid:107)xk − wo(cid:107)2 + (cid:107)C(cid:107)2
1 · σ2
v
where σ2
v
(cid:44) max
1≤l≤N
EP [v(TA1(wi−1))] (cid:22) Ω2(cid:110)
= Ω2(cid:110)
= Ω2(cid:110)
(c)(cid:22) Ω2(cid:110)
(a)
(b)
{2α(cid:107)∇wJl(wo)(cid:107)2 + σ2
v}. Substituting (51) and x = TA1(wi−1) into (47) leads to
(cid:111)
2α(cid:107)C(cid:107)2
1λ2
2α(cid:107)C(cid:107)2
2α(cid:107)C(cid:107)2
(cid:105)
max · EP [TA1(wi−1) − 1N ⊗ wo] + (cid:107)C(cid:107)2
(cid:16)
TA1(wi−1) − TA1(1N ⊗ wo)
max · EP
1λ2
max · EP
(cid:104)
(cid:104)
1 · EP [wi−1 − 1N ⊗ wo] + (cid:107)C(cid:107)2
wi−1 − 1N ⊗ wo(cid:17)(cid:105)
maxAT
TA1
1λ2
1λ2
1σ2
v
2α(cid:107)C(cid:107)2
+ (cid:107)C(cid:107)2
1σ2
v
1N
(cid:111)
1N
1σ2
1N
v
+ (cid:107)C(cid:107)2
1σ2
v
1N
(cid:111)
(cid:111)
August 21, 2018
DRAFT
(cid:40)
(e)(cid:22) Ω2(cid:110)
(d)
= Ω2
1 ·4EP
1 ·(cid:16)
2α(cid:107)C(cid:107)2
1λ2
maxAT
2α(cid:107)C(cid:107)2
1λ2
maxAT
max·Ω2AT
(cid:34)
wi−1−w∞
2
+
(cid:35)
w∞−1N ⊗ wo
2
(cid:41)
1N
16
(cid:111)
+ (cid:107)C(cid:107)2
1σ2
v
(cid:17)
2EP [wi−1−w∞]+2P [w∞−1N ⊗ wo]
+(cid:107)C(cid:107)2
1σ2
v
1N
= 4α(cid:107)C(cid:107)2
1 ·EP [wi−1−w∞]+(cid:107)C(cid:107)2
1Ω2·bv
1λ2
(52)
1 is right-stochastic so that TA1(1N ⊗ wo) = 1N ⊗ wo, step (b) is
where step (a) is due to the fact that AT
because of the linearity of TA1(·), step (c) is due to property (31), step (d) is a consequence of Property
3 of Lemma 1, and step (e) is due to the convexity property (29).
Substituting (43) into (42), we obtain
MSPi (cid:22) AT
2 ΓdAT
1 · MSPi−1 + (cid:107)C(cid:107)2
1 · AT
2 Ω2bv
where
Γd (cid:44) Γ2 + 4αλ2
max(cid:107)C(cid:107)2
1 · Ω2
(53)
(54)
The following theorem gives the stability conditions on the inequality recursion (53) and derives both
asymptotic and non-asymptotic bounds for MSP.
Theorem 2 (Mean-Square Stability and Bounds). Suppose AT
2 ΓdAT
1. Then, the following non-asymptotic bound holds for all i ≥ 0:
1 is a stable matrix, i.e., ρ(AT
2 ΓdAT
1 ) <
MSPi (cid:22) (AT
2 ΓdAT
1 )i[MSP0 − MSPub∞ ] + MSPub∞
where MSPub∞ is the asymptotic upper bound on MSP defined as
1(IN − AT
And, as i → ∞, we have the following asymptotic bound
MSPub∞ (cid:44) (cid:107)C(cid:107)2
2 ΓdAT
1 )−1AT
2 Ω2bv
MSPi (cid:22) MSPub∞
lim sup
i→∞
(cid:40)
(cid:41)
Furthermore, a sufficient condition that guarantees the stability of the matrix AT
2 ΓdAT
1 is that
0 < µk < min
σk,max
σ2
k,max+4αλ2
max(cid:107)C(cid:107)2
1
,
σk,min
σ2
k,min+4αλ2
max(cid:107)C(cid:107)2
1
for all k = 1, . . . , N, where σk,max and σk,min were defined earlier in (35).
Proof:
Iterating inequality (53), we obtain
MSPi (cid:22) (AT
2 ΓdAT
1 )iMSP0 + (cid:107)C(cid:107)2
August 21, 2018
j=0
1 ·(cid:104) i−1(cid:88)
1 )j(cid:105)
(AT
2 ΓdAT
AT
2 Ω2bv
(55)
(56)
(57)
(58)
(59)
DRAFT
17
matrix so that (I − X) is invertible, then it leads to(cid:80)i−1
For the second term in (59), we note that (I + X + ··· + X i−1)(I − X) = I − X i. If X is a stable
j=0 X j = (I − X i)(I − X)−1. Using this relation
and given that the matrix AT
1 is stable, we can express (59) as
1 ·(cid:104)
2 ΓdAT
1 )iMSP0 + (cid:107)C(cid:107)2
IN −(AT
1 )i[MSP0 − MSPub∞ ] + MSPub∞
1 )i(cid:105)
MSPi (cid:22) (AT
2 ΓdAT
= (AT
2 ΓdAT
2 ΓdAT
(IN −AT
2 ΓdAT
1 )−1AT
2 Ω2bv
Letting i → ∞ on both sides of the above inequality, we get lim sup
i→∞
we need to show the conditions on the step-sizes {µk} that guarantee stability of the matrix AT
Note that the spectral radius of a matrix is upper bounded by its matrix norms. Therefore,
(60)
MSPi (cid:22) MSPub∞ . In the last step,
1 .
2 ΓdAT
ρ(AT
2 ΓdAT
1 (cid:107)∞
2 ΓdAT
2 (cid:107)∞ · (cid:107)Γd(cid:107)∞ · (cid:107)AT
1 (cid:107)∞
1 ) ≤ (cid:107)AT
≤ (cid:107)AT
= (cid:107)Γd(cid:107)∞
(cid:13)(cid:13)(cid:13)Γ2 + 4αλ2
=
1 · Ω2(cid:13)(cid:13)(cid:13)∞
max(cid:107)C(cid:107)2
If the right-hand side of the above inequality is strictly less than one, then the matrix AT
Using (33) -- (34), this condition is satisfied by the following quadratic inequalities on µk :
2 ΓdAT
2 is stable.
(1 − µkσk,max)2 + µ2
(1 − µkσk,min)2 + µ2
k · 4αλ2
k · 4αλ2
max(cid:107)C(cid:107)2
max(cid:107)C(cid:107)2
1 < 1
1 < 1
(61)
(62)
for all k = 1, . . . , N. Solving the above inequalities, we obtain condition (58).
The non-asymptotic bound (55) characterizes how the MSP at each node evolves over time. It shows that
2 ΓdAT
the MSP converges to steady state at a geometric rate determined by the spectral radius of the matrix
1 . The transient term is determined by the difference between the initial MSP and the steay-state
AT
MSP. At steady state, the MSP is upper bounded by MSPub∞ . We now examine closely how small the
steady-state MSP can be for small step-size parameters {µk}. Taking the ∞−norm of both sides of (57)
and using the relation (IN − AT
1 )j, we obtain
(cid:107)MSPub∞(cid:107)∞ =
2 ΓdAT
1 · (IN − AT
1 )−1 =(cid:80)∞
(cid:32) ∞(cid:88)
(cid:32) ∞(cid:88)
(cid:33)
(cid:107)AT
2 ΓdAT
j=0
(cid:107)Γd(cid:107)j∞
·(cid:16)
j=0
(cid:13)(cid:13)(cid:13)(cid:107)C(cid:107)2
≤ (cid:107)C(cid:107)2
1 ·
(a)≤ (cid:107)C(cid:107)2
1 ·
2 ΓdAT
j=0(AT
1 )−1 · AT
2 Ω2bv
max
1≤k≤N
µk
(cid:13)(cid:13)(cid:13)∞
(cid:33)
(cid:17)2 · (cid:107)bv(cid:107)∞
· (cid:107)AT
2 (cid:107)j∞ · (cid:107)Γd(cid:107)j∞ · (cid:107)AT
1 (cid:107)j∞
2 (cid:107)∞ · (cid:107)Ω(cid:107)2∞ · (cid:107)bv(cid:107)∞
August 21, 2018
DRAFT
·(cid:16)
(cid:107)C(cid:107)2
1 · (cid:107)bv(cid:107)∞
1 − (cid:107)Γd(cid:107)∞
1 and AT
(cid:17)2
=
µk
max
1≤k≤N
(63)
2 are right-stochastic matrices so that their ∞−norms (maximum
where step (a) is because AT
absolute row sum) are one. Let µmax and µmin denote the maximum and minimum values of {µk},
respectively, and let β (cid:44) µmin/µmax. For sufficiently small step-sizes, by the definitions of Γd and Γ in
(54) and (33), we have
18
(cid:107)Γd(cid:107)∞ ≤ (cid:107)Γ(cid:107)2∞ + 4αλmax(cid:107)C(cid:107)2
1 · (cid:107)Ω(cid:107)2∞
{1 − µkσk,min2} + 4αλmaxµ2
max(cid:107)C(cid:107)2
1
(a)
= max
1≤k≤N
≤ 1−2µminσmin +µ2
= 1−2βµmaxσmin +µ2
max(σ2
max +4αλmax(cid:107)C(cid:107)2
1)
max +4αλmax(cid:107)C(cid:107)2
1)
(64)
where σmax and σmin are the maximum and minimum values of {σk,max} and {σk,min}, respectively,
and step (a) holds for sufficiently small step-sizes. Note that (63) is a monotonically increasing function
of (cid:107)Γd(cid:107)∞. Substituting (64) into (63), we get
max(σ2
(cid:107)MSPi(cid:107)∞ ≤ (cid:107)MSPub∞(cid:107)∞ ≤
lim sup
i→∞
(cid:107)C(cid:107)2
1 · (cid:107)bv(cid:107)∞ · µmax
2βσmin−µmax(σ2
max +4αλmax(cid:107)C(cid:107)2
1)
∼ O(µmax)
(65)
Note that, for sufficiently small step-sizes, the right-hand side of (65) is approximately (cid:107)C(cid:107)2
µmax,
which is on the order of O(µmax). In other words, the steady-state MSP can be made be arbitrarily small
for small step-sizes, and the estimators wi = col{w1,i, . . . , wN,i} will be close to the fixed point w∞ (in
the mean-square sense) even under gradient perturbations. To understand how close the estimate wk,i at
each node k is to the Pareto-optimal solution wo, a natural question to consider is how close the fixed
point w∞ is to 1N ⊗ wo, which we study next.
1·(cid:107)bv(cid:107)∞
2βσmin
D. Bias Analysis
Our objective is to examine how large (cid:107)1N ⊗ wo − w∞(cid:107)2 is when the step-sizes are small. We carry
out the analysis in two steps: first, we derive an expression for w∞ (cid:44) 1N ⊗ wo − w∞, and then we
derive the conditions that guarantee small bias.
To begin with, recall that w∞ is the fixed point of Td(·), to which the recursion wi = Td(wi−1)
converges. Also note that Td(·) is an operator representation of the recursions (5) -- (7). We let i → ∞ on
both sides of (5) -- (7) and obtain
August 21, 2018
l=1
a1,lk wl,∞
(66)
DRAFT
N(cid:88)
φk,∞ =
N(cid:88)
clk∇wJl(φk,∞)
ψk,∞ = φk,∞ − µk
wk,∞ =
l=1
a2,lk ψl,∞
N(cid:88)
19
(67)
(68)
where wk,∞, φk,∞ and ψk,∞ denote the limits of wk,i, φk,i and ψk,i as i → ∞, respectively. Introduce
the following bias vectors at node k
l=1
wk,∞ (cid:44) wo−wk,∞, φk,∞ (cid:44) wo−φk,∞, ψk,∞ (cid:44) wo−ψk,∞
(69)
Subtracting each equation of (66) -- (68) from wo and using relation ∇wJl(φk,∞) = ∇wJl(wo)−Hlk,∞ φk,∞
that can be derived from Lemma 4 in Appendix A, we obtain
l=1
N(cid:88)
(cid:34)
N(cid:88)
l=1
φk,∞ =
ψk,∞ =
wk,∞ =
a1,lk wl,∞
N(cid:88)
IM −µk
l=1
a2,lk
ψl,∞
(cid:35)
clkHlk,∞
φk,∞ +µk
N(cid:88)
l=1
clk∇wJl(wo)
(70)
(71)
(72)
(73)
(74)
(75)
(76)
(77)
(78)
(79)
where Hlk,∞ is a positive semi-definite symmetric matrix defined as
(cid:16)
(cid:90) 1
0
N(cid:88)
l=1
Hlk,∞ (cid:44)
∇2
wJl
wo−t
a1,lk wl,∞)
dt
(cid:17)
Introduce the following global vectors and matrices
w∞ (cid:44) 1N ⊗ wo − w∞ = col{ w1,∞, . . . , wN,∞}
A1 (cid:44) A1 ⊗ IM , A2 (cid:44) A2 ⊗ IM ,
M (cid:44) diag{µ1, . . . , µN} ⊗ IM
C (cid:44) C ⊗ IM ,
N(cid:88)
(cid:110)
R∞ (cid:44)
diag
cl1Hl1,∞,··· , clN HlN,∞
l=1
go (cid:44) col{∇wJ1(wo), . . . ,∇wJN (wo)}
(cid:111)
,
Then, expressions (70), (72) and (71) lead to
IM N −AT
w∞ =
(cid:104)
2 (IM N −MR∞)AT
1
(cid:105)−1AT
2 MCT go
Theorem 3 (Bias at Small Step-sizes). Suppose that AT
1 is a regular right-stochastic matrix, so that
its eigenvalue of largest magnitude is one with multiplicity one, and all other eigenvalues are strictly
2 AT
August 21, 2018
DRAFT
20
smaller than one. Let θT denote the left eigenvector of AT
2 AT
1 of eigenvalue one. Furthermore, assume
the following condition holds:
where Ω (cid:44) diag{µ1, . . . , µN} was defined earlier in Lemma 3, and c0 is some constant. Then,
θT AT
2 ΩCT = c01T
(cid:107) w∞(cid:107)2 = (cid:107)1N ⊗ wo − w∞(cid:107)2 ∼ O(µ2
max)
(80)
(81)
Proof: See Appendix B.
Therefore, as long as the network is connected (not necessarily fully connected) and condition (80) holds,
AT
the bias would become arbitrarily small. For condition (80) to hold, one choice is to require the matrices
2 to be doubly stochastic, and all nodes to use the same step-size µ, namely, Ω = µIN . In that
2 is doubly-stochastic so that the left eigenvector of eigenvalue one is θT = 1T
case, the matrix AT
1 and AT
1 AT
and (80) holds.
Finally, we combine the results from Theorems 2 and 3 to bound the mean-square-error (MSE) of the
estimators {wk,i} from the desired Pareto-optimal solution wo. Introduce the N × 1 MSE vector
MSEi (cid:44) EP [ wi]
Using Properties 3 -- 4 in Lemma 1, we obtain
= col
= EP [1N ⊗ wo − wi]
(cid:110)E(cid:107) w1,i(cid:107)2, . . . , E(cid:107) wN,i(cid:107)2(cid:111)
(cid:16) 1N ⊗ wo−w∞
w∞−wi
(cid:104)
(cid:17)(cid:105)
MSEi = EP
2
+
(cid:22) 2P [ w∞]+2 EP [w∞−wi]
2
2
Taking the ∞−norm of both sides of above inequality and using property (36), we obtain
= 2P [ w∞]+2 MSPi
lim sup
i→∞
(cid:107)MSEi(cid:107)∞ ≤ 2(cid:107)P [ w∞](cid:107)∞ + 2 lim sup
i→∞
(cid:107)MSPi(cid:107)∞
(cid:107)MSPi(cid:107)∞
= 2(cid:107) w∞(cid:107)2
∼ O(µ2
b,∞ + 2 lim sup
i→∞
max) + O(µmax)
(82)
(83)
(84)
where in the last step, we used (65) and (81), and the fact that all vector norms are equivalent. Therefore,
as the step-sizes become small, the MSEs become small and the estimates {wk,i} get arbitrarily close
August 21, 2018
DRAFT
R∞ ≈ N(cid:88)
(cid:110)
to the Pareto-optimal solution wo. We also observe that, for small step-sizes, the dominating steady-state
error is MSP, which is caused by the gradient noise and is on the order of O(µmax). On the other hand,
the bias term is a high order component, i.e., O(µ2
max), and can be ignored.
The fact that the bias term w∞ is small also gives us a useful approximation for R∞ in (77). Since
w∞ = col{ w1,∞, . . . , wN,∞} is small for small step-sizes, the matrix Hlk,∞ defined in (73) can be
approximated as Hlk,∞ ≈ ∇2
wJl(wo). Then, by definition (77), we have
diag
cl1∇2
wJl(wo), . . . , clN∇2
wJl(wo)
(85)
Expressing (85) is useful for evaluating closed-form expressions of the steady-state MSE in sequel.
l=1
E. Steady-State Performance
So far, we derived inequalities (84) to bound the steady-state performance, and showed that, for small
step-sizes, the solution at each node k approaches the same Pareto-optimal point wo. In this section,
we derive closed-form expressions (rather than bounds) for the steady-state MSE at small step-sizes.
Introduce the error vectors3
21
(86)
(87)
(88)
(89)
(90)
(91)
(cid:111)
(cid:111)
(cid:111)
φk,i (cid:44) wo−φk,i, ψk,i (cid:44) wo−ψk,i, wk,i (cid:44) wo−wk,i
and the following global random quantities
wi (cid:44)col{ w1,i, . . . , wN,i}
N(cid:88)
(cid:90) 1
N(cid:88)
l=1
0
Ri−1 (cid:44)
Hlk,i−1 (cid:44)
gi (cid:44)
(cid:110)
(cid:32)
(cid:110)
N(cid:88)
l=1
diag
cl1Hl1,i−1,··· , clN HlN,i−1
∇2
wJl
wo−t
a1,lk wl,i−1
dt
(cid:33)
col
cl1vl(φ1,i−1),··· , clN vl(φN,i−1)
l=1
Then, extending the derivation from [21, Sec. IV A], we can establish that
2 MCT go +AT
2 [IM N −MRi−1]AT
1 wi−1 +AT
wi = AT
2 Mgi
According to (84), the error wk,i at each node k would be small for small step-sizes and after long enough
time. In other words, wk,i is close to wo. And recalling from (12) that φk,i−1 is a convex combination
3 In this paper, we always use the notation w = wo − w to denote the error relative to wo. For the error between w and the
fixed point w∞, we do not define a separate notation, but instead write w∞ − w explicitly to avoid confusion.
August 21, 2018
DRAFT
22
of {wl,i}, we conclude that the quantities {φl,i−1} are also close to wo. Therefore, we can approximate
(cid:90) 1
Hlk,i−1, Ri−1 and gi in (88) -- (90) by
wJl(wo)dt =∇2
∇2
(cid:110)
Ri−1≈ N(cid:88)
(cid:111)≈R∞
wJl(wo), . . . , clN∇2
Hlk,i−1≈
wJl(wo)
wJl(wo)
cl1∇2
(92)
(93)
diag
0
Then, the error recursion (91) can be approximated by
l=1
wi = AT
2 [IM N −MR∞]AT
1 wi−1 +AT
2 MCT go +AT
2 Mgi
(94)
First, let us examine the behavior of E wi. Taking expectation of both sides of recursion (94), we obtain
E wi = AT
2 [IM N − MR∞]AT
1
E wi−1 + AT
2 MCT go
(95)
This recursion converges when the matrix AT
(see Appendix C of [21]). Let i → ∞ on both sides of (95) so that
2 [IM N − MR∞]AT
1 is stable, which is guaranteed by (38)
E w∞ (cid:44) lim
E wi
i→∞
IM N −AT
=
(cid:104)
2 (IM N −MR∞)AT
1
(cid:105)−1AT
2 MCT go
(96)
Note that E w∞ coincides with (79). By Theorem 3, we know that the squared norm of this expression
max) at small step-sizes -- see (81). Next, we derive closed-form expressions for
is on the order of O(µ2
the MSEs, i.e., E(cid:107) wk,i(cid:107)2. Let Rv denote the covariance matrix of gi evaluated at wo:
Rv = E
col
cl1vl,i(wo),··· , clN vl,i(wo)
col
cl1vl,i(wo),··· , clN vl,i(wo)
(97)
(cid:111)(cid:35)T(cid:41)
(cid:40)(cid:34) N(cid:88)
l=1
(cid:110)
(cid:111)(cid:35)(cid:34) N(cid:88)
l=1
(cid:110)
In practice, we can evaluate Rv from the expressions of {vl,i(wo)}. Equating the squared weighted
Euclidean "norm" of both sides of (94), applying the expectation operator with assumption (16), and
following the same line of reasoning from [21], we can establish the following approximate variance
relation at small step-sizes:
E(cid:107) wi(cid:107)2
Σ ≈ E(cid:107) wi−1(cid:107)2
Σ(cid:48) + Tr(ΣAT
2 MCT go)T ΣAT
2 MRvMA2) + Tr{ΣAT
2 (IM N −MR∞)AT
1
E wi−1
+ 2(AT
2 MCT go(AT
2 MCT go)T}
Σ(cid:48) ≈ A1 (IM N −MR∞)A2ΣAT
2 (IM N −MR∞)AT
1
(98)
(99)
where Σ is a positive semi-definite weighting matrix that we are free to choose. Let σ = vec(Σ) denote
the vectorization operation that stacks the columns of a matrix Σ on top of each other. We shall use the
August 21, 2018
DRAFT
notation (cid:107)x(cid:107)2
σ and (cid:107)x(cid:107)2
Σ interchangeably. Following the argument from [21], we can rewrite (98) as
23
E(cid:107) wi(cid:107)2
σ ≈ E(cid:107) wi−1(cid:107)2
F σ + rT σ + σT Q E wi−1
where
r (cid:44) vec(cid:0)AT
F (cid:44) A1[IM N −MR∞]A2 ⊗ A1[IM N −MR∞]A2
2 MCTgo
(cid:1)+AT
2 MRvMA2
2 MCT go⊗AT
2 MCT go
1 ⊗ AT
Q (cid:44) 2AT
2 (IM N −MR∞)AT
(100)
(101)
(102)
(103)
We already established that E wi−1 on the right-hand side of (100) converges to its limit E w∞ under
condition (38). And, it was shown in [31, pp.344-346] that such recursion converges to a steady-state
value if the matrix F is stable, i.e., ρ(F ) < 1. This condition is guaranteed when the step-sizes are
sufficiently small (or chosen according to (38)) -- see the proof in Appendix C of [21]. Letting i → ∞
on both sides of expression (100), we obtain:
E(cid:107) wi(cid:107)2
(I−F )σ ≈ (r + Q E w∞)T σ
(104)
lim
i→∞
We can now resort to (104) and use it to evaluate various performance metrics by choosing proper
weighting matrices Σ (or σ). For example, the MSE of any node k can be obtained by computing
limi→∞ E(cid:107) wi(cid:107)2
T with a block weighting matrix T that has an identity matrix at block (k, k) and
T . Denote the vectorized version of this matrix by tk (cid:44)
zeros elsewhere:
vec(diag(ek)⊗ IM ), where ek is a vector whose kth entry is one and zeros elsewhere. Then, if we select
σ in (104) as σ = (I − F )−1tk, the term on the left-hand side becomes the desired limi→∞ E(cid:107) wk,i(cid:107)2
and the MSE for node k is therefore given by:
E(cid:107) wk,i(cid:107)2 = lim
i→∞
E(cid:107) wi(cid:107)2
lim
i→∞
MSEk (cid:44) lim
i→∞
E(cid:107) wk,i(cid:107)2 ≈ (r + Q E w∞)T (I−F )−1tk
If we are interested in the average network MSE, then it is given by
N(cid:88)
k=1
MSE (cid:44) 1
N
MSEk
(105)
(106)
IV. APPLICATION TO COLLABORATIVE DECISION MAKING
We illustrate one application of the framework developed in the previous sections to the problem of
collaborative decision making over a network of N agents. We consider an application in finance where
each entry of the decision vector w denotes the amount of investment in a specific type of asset. Let
the M × 1 vector p represent the return in investment. Each entry of p represents the return for a
August 21, 2018
DRAFT
24
unit investment in the corresponding asset. Let p and Rp denote the mean and covariance matrix of p,
respectively. Then, the overall return by the agents for a decision vector w is pT w. Note that, with decision
w, the return pT w is a (scalar) random variable with mean pT w and variance var(pT w) = wT Rpw, which
are called the expected return and variance of the return in classical Markowitz portfolio optimization [2,
p.155], [32] -- [35]. These two metrics are often used to characterize the quality of the decision w: we want
to maximize the expected return while minimizing the variance. However, solving the problem directly
requires all agents to know the global statistics p and Rp. What is available in practice are observations
that are collected at the various nodes. Suppose a subset U of the agents observes a sequence of return
vectors {uk,i} with Euk,i = p. The subscripts k and i denote that the return is observed by node k at
time i. Then, we can formulate the cost functions for the nodes in set U as follows:
Ju,k(w) = −E[uT
k,iw] = −pT w
k ∈ U ⊂ {1, . . . , N}
(107)
We place a negative sign in (107) so that minimizing Ju,k(w) is equivalent to maximizing the expected
return. Similarly, suppose there is another subset of nodes, exclusive from U and denoted by S, which
observes a sequence of centered return vectors {sk,i}, namely, vectors that have the same distribution as
p − Ep so that E[sk,isT
k,i] = Rp. Then, we can associate with these nodes the cost functions:
= wT Rpw
(108)
Additionally, apart from selecting the decision vector w to maximize the return subject to minimizing
k ∈ S ⊂ {1, . . . , N}
its variance, the investment strategy w needs to satisfy other constraints such as: i) the total amount of
investment should be less than a maximum value that is known only to an agent k0 ∈ K (e.g., agent k0
is from the funding department who knows how much funding is available), ii) the investment on each
asset be nonnegative (known to all agents), and iii) tax requirements and tax deductions4 known to agents
in a set H. We can then formulate the following constrained multi-objective optimization problem:
Js,k(w) = E(cid:104)sT
k,iw2(cid:105)
(cid:40)(cid:88)
k∈U
min
w
(cid:41)
(cid:88)
k∈S
Ju,k(w),
Js,k(w)
(109)
4For example, suppose the first and second entries of the decision vector w denote the investments on charity assets. When the
charity investments exceed a certain amount, say b, there would be a tax deduction. We can represent this situation by writing
hT w ≥ b, where h (cid:44) [1 1 0 ··· 0]T .
August 21, 2018
DRAFT
J glob(w) =
Ju,k(w) +
(cid:35)
(cid:35)
M(cid:88)
M(cid:88)
m=1
φ(−eT
mw)
φ(−eT
mw)
k∈U
(cid:34)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
k∈S
k∈H
k∈K
+
+
+
(cid:34)
(cid:34)
(cid:34)
Js,k(w) +
m=1
φ(bk − hT
k w) +
φ(−eT
mw)
(cid:35)
(cid:35)
M(cid:88)
M(cid:88)
m=1
m=1
φ(1T w − b0) +
φ(−eT
mw)
s.t. 1T w ≤ b0
k w ≥ bk,
hT
w (cid:23) 0
k ∈ H
25
(110)
(111)
(112)
Using the scalarization technique and barrier function method from Sec. I, we convert (109) -- (112) into
the following unconstrained optimization problem (for simplicity, we only consider π1 = ··· = πN = 1):
The term(cid:80)M
where φ(·) is a barrier function to penalize the violation of the constraints -- see [3] for an example, and
the vector em ∈ RM is a basis vector whose entries are all zero except for a value of one at the mth entry.
mw) is added to each cost function to enforce the nonnegativity constraint (112),
which is assumed to be known to all agents. Note that there is a "division of labor" over the network:
the entire set of nodes is divided into four mutually exclusive subsets {1, . . . , N} = U ∪ S ∪ H ∪ K,
and each subset collects one type of information related to the decision. Diffusion adaptation strategies
m=1 φ(−eT
allow the nodes to arrive at a Pareto-optimal decision in a distributed manner over the network, and each
subset of nodes influences the overall investment strategy.
In our simulation, we consider a randomly generated connected network topology. There are a total
of N = 10 nodes in the network, and nodes are assumed connected when they are close enough
geographically. The cardinalities of the subsets U, S, H and K are set to be 3, 4, 2 and 1, respectively. The
nodes are partitioned into these four subsets randomly. The dimension of the decision vector is M = 5.
The random vectors uk,i and sk,i are generated according to the Gaussian distributions N (1, IM ) and
N (0, IM ), respectively. We set b0 = 5 and the parameters {hk, bk} for k ∈ H to
hk1 = [1 2 ··· 5] ,
hk2 = [5 4 ··· 1] ,
bk1 = 2
bk2 = 3
(113)
(114)
DRAFT
August 21, 2018
26
(a) Topology of the network.
(b) Learning curve (µ = 10−2).
(c) MSE for different values of step-sizes.
(d) Error of fixed point for different values of step-
sizes.
Fig. 4. Simulation results for collaborative decision making.
where k1 and k2 are the indices of the two nodes in the subset H. Furthermore, we use the barrier
function given by (15) in [3] in our simulation with t = 10, ρ = 0.1 and τ = 0.1. We set the combination
coefficients {alk} to the Metropolis rule (See Table III in [23]) for both ATC and CTA strategies. The
weights {clk} are set to clk = 1 for l = k and zero otherwise, i.e., there is no exchange of gradient
information among neighbors. According to Theorem 3, such a choice will always guarantee condition
(80) so that the bias can be made arbitrarily small for small step-sizes. In our simulation, we do not
assume the statistics of {uk,i} and {sk,i} are known to the nodes. The only information available is their
realizations and the algorithms have to learn the best decision vector w from them. Therefore, we use
August 21, 2018
DRAFT
02004006008001000−25−20−15−10−50Number of IterationsNetwork MSE (dB) ATCCTAConsensusCentralized10−310−210−1−35−30−25−20−15−10−505Step−Size µNetwork MSE (dB) ATCATC theoryCTACTA theoryConsensusCentralized10−310−210−1−50−40−30−20−100Step−Size µError of Fixed Point (dB) ATCATC theoryCTACTA theoryConsensus the following stochastic gradient vector5 at each node k:
(cid:92)∇wJk(w) =
∇wφ(bk − hT
m=1 ∇wφ(−eT
m=1 ∇wφ(−eT
−uk,i +(cid:80)M
2sk,i +(cid:80)M
k w) +(cid:80)M
∇wφ(1T w − b0) +(cid:80)M
(cid:88)
mw)
k ∈ U
mw)
m=1 ∇wφ(−eT
m=1 ∇wφ(−eT
k ∈ S
mw) k ∈ H
mw) k ∈ K
alkwl,i−1 − µ(cid:92)∇wJk(wk,i−1)
wk,i =
l∈Nk
27
(115)
(116)
To compare the performance with other algorithms, we also simulate the consensus-based approach from
[16] with the same stochastic gradient6 as (115). The algorithm is listed below:
N(cid:88)
k=1
Furthermore, we also simulate the conventional centralized approach to such optimization problem, which
collects data from all nodes and implements stochstic gradient descent at the central node:
wi = wi−1 − µ
1
N
(cid:92)∇wJk(wi−1)
(117)
where the factor of 1/N is used to make the convergence rate the same as the distributed algorithms.
The simulatin results are shown in Fig. 4(a) -- 4(d). Fig. 4(a) shows the network topology, and Fig.
4(b) shows the learning curves of different algorithms. We see that ATC outperforms CTA and CTA
outperforms consensus. To further compare the steady-state performance, we plot the steady-state MSE
for different values of step-sizes in Fig. 4(c). We also plot the theoretical curves from (105) -- (106) for
ATC and CTA algorithms. We observe that all algorithms approach the performance of the centralized
solution when the step-sizes are small. However, diffusion algorithms always outperform the consensus-
based strategy; the gap between ATC and consensus algorithm is about 8 dB when µ = 0.1. We also see
that the theoretical curves match the simulated ones well. Finally, we recall that Theorem 3 shows that
the error between the fixed point w∞ and 1 ⊗ wo can be made arbitrarily small for small step-sizes, and
the error (cid:107)w∞ − 1 ⊗ wo(cid:107)2 is on the order of O(µ2). To illustrate the result, we simulate the algorithms
using true gradients {∇wJk(w)} so that they converge to their fixed point w∞, and we get different
values of w∞ for different step-sizes. The theoretical values for ATC and CTA can be computed from
(79). The results are shown in Fig. 4(d). We see that the theory matches simulation, and the power of
5For nodes in H an K, the cost functions are known precisely, so their true gradients are used.
6The original algorithm in [16] does not use stochastic gradients but the true gradients {∇wJk(w)}.
August 21, 2018
DRAFT
28
the fixed point error per node7 decays at 20dB per decade, which is O(µ2) and is consistent with (81).
Note that diffusion algorithms outperform the consensus. Also note from (79) and (96) that the bias and
the fixed point error have the same expression. Therefore, diffusion algorithms have smaller bias than
consensus (the gap in Fig. 4(d) is as large as 5dB between ATC and consensus).
V. CONCLUSION
This paper generalized diffusion adaptation strategies to perform multi-objective optimization in a
distributed manner over a network of nodes. We use constant step-sizes to endow the network with
continuous learning and adaptation abilities via local interactions. We analyzed the mean-square-error
performance of the diffusion strategy, and showed that the solution at each node gets arbitrarily close to
the same Pareto-optimal solution for small step-sizes.
APPENDIX A
PROPERTIES OF THE OPERATORS
Properties 1-3 are straightforward from the definitions of TA(·) and P [·]. We therefore omit the proof
for brevity, and start with property 4.
(Property 4: Convexity)
We can express each N × 1 block vector x(k) in the form x(k) = col{x(k)
Then, the convex combination of x(1), . . . , x(N ) can be expressed as
1 , . . . , x(k)
N } for k = 1, . . . , N.
al x(k) = col
alx(k)
1 , . . . ,
alx(k)
N
According to the definition of the operator P [·], and in view of the convexity of (cid:107) · (cid:107)2, we have
K(cid:88)
(cid:34) K(cid:88)
k=1
k=1
(cid:35)
k=1
(cid:40) K(cid:88)
(cid:40)(cid:13)(cid:13)(cid:13) K(cid:88)
(cid:40) K(cid:88)
K(cid:88)
k=1
k=1
=
al P [x(k)]
(cid:41)
k=1
K(cid:88)
(cid:13)(cid:13)(cid:13) K(cid:88)
K(cid:88)
k=1
k=1
(cid:41)
(cid:13)(cid:13)(cid:13)2
(cid:41)
(cid:13)(cid:13)(cid:13)2
P
al x(k)
= col
alx(k)
1
, . . . ,
alx(k)
N
(cid:22) col
al(cid:107)x(k)
1 (cid:107)2, . . . ,
al(cid:107)x(k)
N (cid:107)2
7The power of the fixed point error per node is defined as 1
N (cid:107)w∞ − 1 ⊗ wo(cid:107)2 = 1
N
k=1
(cid:80)N
k=1 (cid:107)wk,∞ − wo(cid:107)2.
August 21, 2018
(118)
(119)
DRAFT
(Property 5: Additivity)
By the definition of P [·] and the assumption that ExT
k yk = 0 for each k = 1, . . . , N, we obtain
EP [x + y] = col{E(cid:107)x1 + y1(cid:107)2, . . . , E(cid:107)xN + yN(cid:107)2}
= col{E(cid:107)x1(cid:107)2 + E(cid:107)y1(cid:107)2, . . . , E(cid:107)xN(cid:107)2 + E(cid:107)yN(cid:107)2}
= EP [x] + EP [y]
29
(120)
(Property 6: Variance Relations)
We first prove (31). From the definition of TA(·) in (18) and the definition of P [·] in (20), we express
P [TA(x)] = col
al1xl
, . . . ,
alN xl
(121)
Since (cid:107)·(cid:107)2 is a convex function and each sum inside the squared norm operator is a convex combination
of x1, . . . , xN (AT is right stochastic), by Jensen's inequality [2, p.77], we have
(cid:13)(cid:13)(cid:13)2
(cid:40)(cid:13)(cid:13)(cid:13) N(cid:88)
(cid:40) N(cid:88)
l=1
(cid:13)(cid:13)(cid:13) N(cid:88)
l=1
N(cid:88)
(cid:41)
(cid:13)(cid:13)(cid:13)2
(cid:41)
P [TA(x)] (cid:22) col
al1(cid:107)xl(cid:107)2, . . . ,
alN(cid:107)xl(cid:107)2
l=1
l=1
= AT col{(cid:107)x1(cid:107)2, . . . , (cid:107)xN(cid:107)2}
= AT P [x]
(122)
Next, we proceed to prove (32). We need to call upon the following useful lemmas from [4, p.24], and
Lemmas 1 -- 2 in [21], respectively.
Lemma 4 (Mean-Value Theorem). For any twice-differentiable function f (·), it holds that
(cid:20)(cid:90) 1
(cid:16)
(cid:21)
(cid:17)
dt
where ∇2f (·) denotes the Hessian of f (·), and is a symmetric matrix.
0
∇f (y) = ∇f (x) +
∇2f
x + t(y − x)
(y − x)
(123)
(cid:4)
Lemma 5 (Bounds on the Integral of Hessian). Under Assumption 1, the following bounds hold for any
vectors x and y:
(cid:90) 1
N(cid:88)
0
clk
λl,minIM ≤
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)I − µk
l=1
wJl(x + ty)dt ≤ λl,maxIM
∇2
(cid:20)(cid:90) 1
∇2
wJl(x + ty)dt
(cid:21)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ γk
0
where (cid:107) · (cid:107) denotes the 2−induced norm, and γk, σk,min and σk,max were defined in (34) -- (35).
August 21, 2018
(124)
(125)
(cid:4)
DRAFT
By the definition of the operator TG(·) in (19) and the expression (123), we express TG(x)− TG(y) as
TG(x) − TG(y) =
(cid:34)
l=1
(cid:34)
IM −µ1
N(cid:88)
N(cid:88)
P [TG(x) − TG(y)] (cid:22) col(cid:8)γ2
IM −µN
l=1
cl1
(cid:90) 1
(cid:90) 1
(cid:35)
wJl(y1 +t(x1−y1))dt
∇2
(cid:35)
wJl(yN +t(xN −yN ))dt
∇2
...
0
clN
0
(x1−y1)
(xN −yN )
30
(126)
(127)
(128)
Therefore, using (125) and the definition of P [·] in (20), we obtain
1 · (cid:107)x1 − y1(cid:107)2, . . . , γ2
N · (cid:107)xN − yN(cid:107)2(cid:9)
= Γ2P [x − y]
(Property 7: Block Maximum Norm)
According to the definition of P [·] and the definition of block maximum norm [21], we have
(cid:107)P [x](cid:107)∞ =
(cid:13)(cid:13)(cid:13)col{(cid:107)x1(cid:107)2, . . . , (cid:107)xN(cid:107)2}(cid:13)(cid:13)(cid:13)∞
(cid:16)
(cid:107)xk(cid:107)2
(cid:107)xk(cid:107)(cid:17)2
max
1≤k≤N
= max
1≤k≤N
=
= (cid:107)x(cid:107)2
b,∞
(Property 8: Preservation of Inequality)
To prove F x (cid:22) F y, it suffices to prove 0 (cid:22) F (y − x). Since x (cid:22) y, we have 0 (cid:22) y − x, i.e., all entries
of the vector y − x are nonnegative. Furthermore, since all entries of the matrix F are nonnegative, the
entries of the vector F (y − x) are all nonnegative, which means 0 (cid:22) F (y − x).
It suffices to show that
APPENDIX B
BIAS AT SMALL STEP-SIZES
(cid:107)1 ⊗ wo − w∞(cid:107)
µmax
= ξ
lim
µmax→0
(129)
where ξ is a constant independent of µmax. It is known that any matrix is similar to a Jordan canonical
1 = Y JY −1, where J is the Jordan
form [36]. Hence, there exists an invertible matrix Y such that AT
1 , and the columns of the matrix Y are the corresponding right
canonical form of the matrix AT
principal vectors of various degrees [36, pp.82 -- 88]; the right principal vector of degree one is the right
2 AT
2 AT
August 21, 2018
DRAFT
eigenvector. Obviously, the matrices J and Y are independent of µmax. Using the Kronecker product
property [36, p.140]: (A ⊗ B)(C ⊗ D) = AC ⊗ BD, we obtain
31
AT
2 AT
1 ⊗ IM
2 AT
1 = AT
= (Y ⊗ IM )(J ⊗ IM )(Y −1 ⊗ IM )
(130)
Denote µk = βkµmax, where βk is some positive scalar such that 0 < βk ≤ 1. Substituting (130) into
(79), we obtain
1 ⊗ wo − w∞ =(cid:2)IM N − AT
2 AT
1 + AT
2 MR∞AT
1
(cid:3)−1 AT
2 MCT go
−1 (Y −1 ⊗ IM )AT
2 MCT go
= (Y ⊗ IM ) [IM N − J ⊗ IM + µmaxE]
where
E = (Y −1 ⊗ IM )AT
M0 (cid:44) M/µmax = diag{β1, . . . , βN}
2 M0R∞AT
(cid:123)(cid:122)
1 (Y ⊗ IM )
(cid:125)
⊗IM
(cid:124)
(cid:44)Ω0
1 is right-stochastic, and since AT
2 AT
By (8), the matrix AT
1 is regular, it will have an eigenvalue of
one that has multiplicity one and is strictly greater than all other eigenvalues [37]. Furthermore, the
corresponding left and right eigenvectors are θT and 1, with θT (cid:31) 0 (all entries of the row vector θT are
real positive numbers). For this reason, we can partition J, Y −1 and Y in the following block forms:
2 AT
J = diag{1, J0},
Y −1 = col
(134)
where J0 is an (N −1)×(N −1) matrix that contains the Jordan blocks of eigenvalues strictly within unit
circle, i.e., ρ(J0) < 1. The first row of the matrix Y −1 in (134) is normalized by θT 1 so that Y −1Y = I.
(Note that Y −1Y = I requires the product of the first row of Y −1 and the first column of Y to be one:
Y = [1 YL]
, YR
θT 1
,
(cid:26) θT
(cid:27)
θT
θT 1
1 = 1.) Substituting these partitionings into (132), we can express E as
(131)
(132)
(133)
(135)
(136)
(137)
(138)
DRAFT
E21 E22
E11 E12
(cid:17)AT
(cid:17)AT
2 M0R∞AT
2 M0R∞AT
1 (1 ⊗ IM )
1 (YL ⊗ IM )
2 M0R∞AT
1 (1 ⊗ IM )
E =
E11 (cid:44)(cid:16) θT
E12 (cid:44)(cid:16) θT
⊗ IM
⊗ IM
E21 (cid:44) (YR ⊗ IM )AT
θT 1
θT 1
where
August 21, 2018
E22 (cid:44) (YR ⊗ IM )AT
2 M0R∞AT
1 (YL ⊗ IM )
32
(139)
Observe that the matrices E11, E12, E21 and E22 are independent of µmax. Substituting (134) and (135)
into (131), we obtain
1 ⊗ wo−w∞ = (Y ⊗ IM )
Let us denote
−1 1
µmaxE12
I−J0 ⊗ IM +µmaxE22
θT 1 (θT ⊗ IM )AT
(YR ⊗ IM )AT
2 MCT go
2 MCT go
µmaxE11
µmaxE12
µmaxE21 I−J0 ⊗ IM +µmaxE22
−1
µmaxE21
µmaxE11
(cid:44)
G11 G12
N(cid:88)
G21 G22
Furthermore, recalling that wo is the minimizer of the global cost function (4), we have
∇wJl(wo) = 0 ⇔ (1T ⊗ IM ) go = 0
which, together with condition (80), implies that
l=1
(θT ⊗ IM )AT
2 MCT go = (θT AT
2 ΩCT ⊗ IM )go
= c0(1T ⊗ IM )go
where we also used the facts that AT
product property: (A ⊗ B)(C ⊗ D). Substituting (141) and (143) into (140) and using (133) lead to
(143)
2 ⊗ IM , CT = CT ⊗ IM , M = Ω ⊗ IM and the Kronecker
2 = AT
= 0
(140)
(141)
(142)
(144)
(145)
To proceed with analysis, we need to evaluate G12 and G22. We call upon the relation from [36, pp.48]:
(YRAT
G12
G22
1 ⊗ wo − w∞ = µmax · (Y ⊗ IM )
2 Ω0CT ⊗ IM )go
P Q
U V
=
−1
P −1 + P −1QSU P −1 −P −1QS
(cid:16) θT
−SU P −1
(cid:17)R∞(AT
2 Ω0 ⊗ IM
AT
1 ⊗ IM )
S
1
E11 =
θT 1
where S = (V − U P −1Q)−1. To apply the above relation to (141), we first need to verify that E11 is
invertible. By definition (136),
= (zT ⊗ IM )R∞(1 ⊗ IM )
August 21, 2018
DRAFT
N(cid:88)
zk
N(cid:88)
k=1
l=1
=
clkHlk,∞
33
(146)
where zk denotes the kth entry of the vector z (cid:44) Ω0A2θ/θT 1 (note that all entries of z are non-negative,
i.e., zk ≥ 0). Recall from (73) that Hlk,∞ is a symmetric positive semi-definite matrix. Moreover, since
zk and clk are nonnegative, we can conclude from (146) that E11 is a symmetric positive semi-definite
matrix. Next, we show that E11 is actually strictly positive definite. Applying (124) to the expression of
Hlk,∞ in (73), we obtain Hlk,∞ ≥ λl,minIM . Substituting into (146) gives:
N(cid:88)
(cid:33)
l=1
(cid:34) N(cid:88)
(cid:32) N(cid:88)
k=1
k=1
E11 ≥
≥
=
(cid:35)
(cid:40) N(cid:88)
(cid:40) N(cid:88)
l=1
l=1
zk
clkλl,min
· IM
zk
min
1≤k≤N
clkλl,min
· IM
(cid:41)
(cid:41)
1T Ω0A2θ
θT 1
· min
1≤k≤N
clkλl,min
· IM
(147)
Noting that the matrices Ω0 and A0 have nonnegative entries with some entries being positive, and that
all entries of θ are positive, we have (1T Ω0A2θ)/(θT 1) > 0. Furthermore, by Assumption 1, we know
l=1 clkλl,min > 0 for each k = 1, . . . , N. Therefore, we conclude that E11 > 0 and is invertible.
(cid:80)N
Applying (145) to (141), we get
11 ET
11 E12G22
G12 = −E−1
12)(cid:3)−1
G22 =(cid:2)I−J0 ⊗ IM +µmax(E22−E21E−1
G22(YRAT
G22(YRAT
−E−1
−E−1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(Y ⊗ IM )
= lim
µmax→0
11 E12
11 E12
I
I
(148)
(149)
(150)
(151)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Substituting (149) into (144) leads to
1 ⊗ wo − w∞ = µmax · (Y ⊗ IM )
2 Ω0CT ⊗ IM )go
Substituting expression (150) into the left-hand side of (129), we get
(cid:107)1 ⊗ wo − w∞(cid:107)
lim
µmax→0
µmax
2 Ω0CT ⊗ IM )go
Observe that the only term on the right-hand side of (151) that depends on µmax is G22. From its
expression (149), we observe that, as µmax → 0, the matrix G22 tends to (I − J0 ⊗ IM )−1, which is
independent of µmax. Therefore, the limit on the right-hand side of (151) is independent of µmax.
August 21, 2018
DRAFT
REFERENCES
34
[1] J. Chen and A. H. Sayed, "Distributed Pareto-optimal solutions via diffusion adaptation," in Proc. IEEE Workshop on
Statistical Signal Process. (SSP), Ann Arbor, MI, Aug. 2012, pp. 1 -- 4.
[2] S. P. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
[3] Z. J. Towfic, J. Chen, and A. H. Sayed, "Distributed throughput optimization over P2P mesh networks using diffusion
adaptation," in Proc. IEEE International Conf. Commun. (ICC), Ottawa, Canada, June 2012, pp. 1 -- 5.
[4] B. Polyak, Introduction to Optimization, Optimization Software, NY, 1987.
[5] D.P. Bertsekas and J.N. Tsitsiklis, Parallel and Distributed Computation: Numerical Methods, Athena Scientific, Belmont,
1997.
[6] D. P. Bertsekas, "A new class of incremental gradient methods for least squares problems," SIAM J. Optim., vol. 7, no.
4, pp. 913 -- 926, 1997.
[7] A. Nedic and D. P. Bertsekas, "Incremental subgradient methods for nondifferentiable optimization," SIAM J. Optim., vol.
12, no. 1, pp. 109 -- 138, 2001.
[8] A. Nedic and D. P. Bertsekas,
"Convergence rate of incremental subgradient algorithms," Stochastic Optimization:
Algorithms and Applications, S. Uryasev and P. M. Pardalos, Eds., pp. 263 -- 304, 2000.
[9] B. T. Polyak and Y. Z. Tsypkin, "Pseudogradient adaptation and training algorithms," Automation and Remote Control,
vol. 12, pp. 83 -- 94, 1973.
[10] M. G. Rabbat and R. D. Nowak, "Quantized incremental algorithms for distributed optimization," IEEE J. Sel. Areas
Commun., vol. 23, no. 4, pp. 798 -- 808, 2005.
[11] C. G. Lopes and A. H. Sayed, "Incremental adaptive strategies over distributed networks," IEEE Trans. Signal Process.,
vol. 55, no. 8, pp. 4064 -- 4077, Aug. 2007.
[12] L. Li and J. A. Chambers, "A new incremental affine projection-based adaptive algorithm for distributed networks," Signal
Processing, vol. 88, no. 10, pp. 2599 -- 2603, Oct. 2008.
[13] R. M. Karp, "Reducibility among combinational problems," Complexity of Computer Computations, R. E. Miller and J.
W. Thatcher, Eds., pp. 85 -- 104, 1972.
[14] J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans, "Distributed asynchronous deterministic and stochastic gradient optimization
algorithms," IEEE Trans. Autom. Control, vol. 31, no. 9, pp. 803 -- 812, 1986.
[15] S. Barbarossa and G. Scutari, "Bio-inspired sensor network design," IEEE Signal Process. Mag., vol. 24, no. 3, pp. 26 -- 35,
2007.
[16] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multi-agent optimization," IEEE Trans. Autom. Control,
vol. 54, no. 1, pp. 48 -- 61, 2009.
[17] S. Kar and J. M. F. Moura, "Convergence rate analysis of distributed gossip (linear parameter) estimation: Fundamental
limits and tradeoffs," IEEE J. Sel. Topics. Signal Process., vol. 5, no. 4, pp. 674 -- 690, Aug. 2011.
[18] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione, "Gossip algorithms for distributed signal
processing," Proc. IEEE, vol. 98, no. 11, pp. 1847 -- 1864, 2010.
[19] K. Srivastava and A. Nedic, "Distributed asynchronous constrained stochastic optimization," IEEE J. Sel. Topics Signal
Process., vol. 5, no. 4, pp. 772 -- 790, Aug. 2011.
[20] U. A. Khan and A. Jadbabaie, "Networked estimation under information constraints," Arxiv preprint arXiv:1111.4580,
Nov. 2011.
August 21, 2018
DRAFT
35
[21] J. Chen and A. H. Sayed, "Diffusion adaptation strategies for distributed optimization and learning over networks," IEEE
Trans. Signal Process., vol. 60, no. 8, pp. 4289 -- 4305, Aug. 2012.
[22] C. G. Lopes and A. H. Sayed, "Diffusion least-mean squares over adaptive networks: Formulation and performance
analysis," IEEE Trans. Signal Process., vol. 56, no. 7, pp. 3122 -- 3136, July 2008.
[23] F. S. Cattivelli and A. H. Sayed, "Diffusion LMS strategies for distributed estimation," IEEE Trans. Signal Process., vol.
58, no. 3, pp. 1035 -- 1048, March 2010.
[24] S. S. Ram, A. Nedic, and V. V. Veeravalli,
"Distributed stochastic subgradient projection algorithms for convex
optimization," J. Optim. Theory Appl., vol. 147, no. 3, pp. 516 -- 545, 2010.
[25] S.-Y. Tu and A. H. Sayed, "Mobile adaptive networks," IEEE J. Sel. Topics. Signal Process., vol. 5, no. 4, pp. 649 -- 664,
Aug. 2011.
[26] P. Di Lorenzo and S. Barbarossa, "A bio-inspired swarming algorithm for decentralized access in cognitive radio," IEEE
Trans. Signal Process., vol. 59, no. 12, pp. 6160 -- 6174, Dec. 2011.
[27] S. Chouvardas, K. Slavakis, and S. Theodoridis, "Adaptive robust distributed learning in diffusion sensor networks," IEEE
Trans. Signal Process., vol. 59, no. 10, pp. 4692 -- 4707, Oct. 2011.
[28] S. Theodoridis, K. Slavakis, and I. Yamada, "Adaptive learning in a world of projections," IEEE Signal Process. Mag.,
vol. 28, no. 1, pp. 97 -- 123, Jan. 2011.
[29] E. Kreyszig, Introductory Functional Analysis with Applications, Wiley, NY, 1989.
[30] N. Takahashi and I. Yamada,
"Link probability control for probabilistic diffusion least-mean squares over resource-
constrained networks," in IEEE ICASSP, Dallas, TX, Mar. 2010, pp. 3518 -- 3521.
[31] A. H. Sayed, Adaptive Filters, Wiley, NJ, 2008.
[32] H. Markowitz, "Portfolio selection," The Journal of Finance, vol. 7, no. 1, pp. 77 -- 91, 1952.
[33] M. Rubinstein, "Markowitz's "Portfolio Selection": A Fifty-Year Retrospective," The Journal of Finance, vol. 57, no. 3,
pp. 1041 -- 1045, Jun. 2002.
[34] A. D. Fitt, "Markowitz portfolio theory for soccer spread betting," IMA Journal of Management Mathematics, vol. 20, no.
2, pp. 167 -- 184, Apr. 2009.
[35] E. J. Elton, M. J. Gruber, and C. R. Blake, "Applications of Markowitz portfolio theory to pension fund design," in
Handbook of Portfolio Construction, J. B. Guerard, Ed., pp. 419 -- 438. Springer, US, 2010.
[36] A. J. Laub, Matrix Analysis for Scientists and Engineers, SIAM, PA, 2005.
[37] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press, 1990.
August 21, 2018
DRAFT
|
1902.06996 | 1 | 1902 | 2019-02-19T11:13:38 | Agent Madoff: A Heuristic-Based Negotiation Agent For The Diplomacy Strategy Game | [
"cs.MA"
] | In this paper, we present the strategy of Agent Madoff, which is a heuristic-based negotiation agent that won 2nd place at the Automated Negotiating Agents Competition (ANAC 2017). Agent Madoff is implemented to play the game Diplomacy, which is a strategic board game that mimics the situation during World War I. Each player represents a major European power which has to negotiate with other forces and win possession of a majority supply centers on the map. We propose a design architecture which consists of 3 components: heuristic module, acceptance strategy and bidding strategy. The heuristic module, responsible for evaluating which regions on the graph are more worthy, considers the type of region and the number of supply centers adjacent to the region and return a utility value for each region on the map. The acceptance strategy is done on a case-by-case basis according to the type of the order by calculating the acceptance probability using a composite function. The bidding strategy adopts a defensive approach aimed to neutralize attacks and resolve conflict moves with other players to minimize our loss on supply centers. | cs.MA | cs | Agent Madoff: A Heuristic-Based Negotiation Agent For The
Diplomacy Strategy Game
School of Computer Science and Engineering, Nanyang Technological University, Singapore
[email protected]
Tan Hao Hao
Abstract -- In this paper, we present the strategy of Agent
Madoff, which is a heuristic-based negotiation agent
competing in the upcoming Automated Negotiating
Agents Competition (ANAC 2017). Agent Madoff is
implemented to play the game Diplomacy, which is a
strategic board game that mimics the situation during
World War I. Each player represents a major European
power which has to negotiate with other forces and win
possession of a majority supply centers on the map. We
propose a design architecture which consists of 3
components: heuristic module, acceptance strategy and
bidding strategy. The heuristic module, responsible for
evaluating which regions on the graph are more worthy,
considers the type of region and the number of supply
centers adjacent to the region and return a utility value
for each region on the map. The acceptance strategy is
done on a case-by-case basis according to the type of the
order by calculating the acceptance probability using a
composite function. The bidding strategy adopts a
defensive approach aimed to neutralize attacks and
resolve conflict moves with other players to minimize
our loss on supply centers.
Keywords -- Automated Negotiation, Multi-Issue
Negotiation, Multi-Agent System, Diplomacy
1 INTRODUCTION
Negotiation is described as a process of reaching an
agreement between two or more individuals. Technically,
we could also treat negotiation as a distributed search
through a space of possible agreements. This serves as
the fundamental perspective adopted by automated
negotiation, which agents devise algorithms to search for
Pareto optimal deals within a given negotiation
agreement space. Automated negotiation has been a
growing area of research in the recent years with an
increasing number of applications in domains such as e-
commerce, board games and even human-agent
negotiation.
The Automated Negotiation Agents Competition
(ANAC) is one of the competitions which fuels research
interest in developing practical, state-of-the-art agents
that can negotiate under various circumstances. As it
evolves towards a more practical approach, the latest
ANAC 2017 introduces a new negotiation league named
as "Negotiation Strategies for the Diplomacy Strategy
Game", which requires the participating agents to be able
to negotiate over a large agreement space. Under this
league, the Diplomacy game almost simulates how
human individuals interact during a negotiation process.
Hence, this is a much more surreal condition which is
closer to the human negotiation environment, and with
no doubt the complexity should increase significantly.
This paper first gives an introduction of the above stated
league by specifying how the Diplomacy game is played,
and how negotiation is done in Diplomacy under the
BANDANA framework. Then, it would focus on
describing the negotiation strategy used by Agent
Madoff, an agent implemented in accordance with the
regulations proposed in the above stated league. In
addition,
the agent's
evaluation are provided, and further improvements that
could be done on the agent are highlighted.
the experimental results of
2 MAIN CONTENT
2.1 RULES OF ANAC 2017
The
upcoming Automated Negotiating Agents
Competition (ANAC 2017) has implemented some
drastic changes on its rules and regulations compared to
the past few years [1]. The league that we are focusing
on in this paper is called "Negotiation Strategies for the
Diplomacy Strategy Game", which requires participants
to implement a negotiation algorithm on top of a ready-
made strategic module, combining both modules to form
an agent that can play the classical Diplomacy board
game.
Unlike the GENIUS framework used by previous years
of ANAC, the main differences between Diplomacy
Game League are as follow:
There is no explicit formula to determine an
agent's utility function. No graph could be
plotted out explicitly like scenarios in the
GENIUS framework as before.The assumptions
which were made in the study
Only heuristic approach can be used to
estimate the value of a deal proposed and the
value of the agent's current utility.
BANDANA framework does not allow the
agents to learn opponent's strategy by now,
unlike last year's GENIUS framework
Hence, a totally different paradigm is needed to deal with
the BANDANA framework, however common ground
between these two frameworks should also be found so
that research progress in the GENIUS framework can be
further adopted into agents developed in the BANDANA
framework.
2.2 THE DIPLOMACY GAME
Diplomacy is a strategy game published by Avalon Hill
[2] designed for 7 players. Each player represents one of
the 7 "Great Powers in Europe" in the years prior to
World War I, namely England, France, Germany, Italy,
Austria, Turkey and Russia. Each player has a number of
armies and fleets positioned on the map of Europe, and
the goal is to conquer half of the "Supply Centers" across
Europe, (in normal cases, more than 18 supply centers).
Other than conquering more than 18 supply centers to
win the game (in this case called a "solo victory"),
players can also choose to propose draw with other
players existing on board to "share" the victory.
Figure 1 below shows a classical map initial setting of
the Diplomacy game. The black dots on some of the
provinces represents that the province is a supply center.
A "hold" order (implemented as HLDOrder
class) -- where the unit tries to stay where it is
by now
A "support move" order (implemented as
SUPMTOOrder) -- where the unit does not
move, but gives extra strength to another unit
that is moving towards the province adjacent
to the current unit
A "support hold" order (implemented as
SUPOrder) -- where the unit does not move,
but gives extra strength to another unit to hold
its position on the province adjacent to the
current unit
If 2 or more units are to move into the same province, the
unit which gains more support from other units can
successfully move into the intended province. If both
units have the same strength, then both units will bounce
back to their original position and no one gets in.
However, if another unit is moving into the location of
the supporting unit, then the support order is "cut", and
hence unsuccessful.
2.2.2 SUMMER AND AUTUMN PHASES
For Summer and Autumn Phases, there may be some
units that are "kicked out" from their original provinces
as they are conquered by other units. These units are said
to be "dislodged". For such units, they should retreat to
another province which is adjacent to its current location
(this is also done by the "move to" order). If no such
province exists, then the unit must be disbanded, and the
player lost one unit.
Figure 1: Initial game setting for Diplomacy game
2.2.2 WINTER PHASES
At the start of the game, each powers are allocated with
3 units (except Russia is allocated with 4 units). The
game then iterates through 5 types of phases in the
following order, starting from year 1901:
Spring phase
Summer phase
Fall phase
Autumn phase
Winter phase
Hence the game will develop in the manner of "Spring
1901, Summer 1901, Fall 1901, Winter 1901, Spring
1902 …" and so on. In each round, all players should
submit different types of "orders" for all of their units,
depending on which phase they are in.
2.2.1 SPRING AND FALL PHASES
For Spring and Fall phases, each player must choose
among the following 3 orders to be submitted for each of
his unit:
A "move to" order (in BANDANA framework,
it is implemented as an MTOOrder class) --
where the unit tries to move from one province
to another adjacent province.
If a player successfully conquers a new supply center
throughout the year, then the player is the owner of the
supply center. If the player has more supply centers
than units, he can build new units (either army unit or
fleet unit) at this phase until the number of units equal
the number of conquered supply centers. However if the
player has more units than supply centers, then some of
the units would be disbanded. Players who lost all his
units will be eliminated from the game.
Note that army units can only move on inland
provinces, and fleet units can only move on sea
provinces. There are also coastal provinces which allow
to have 1 army unit and 1-2 fleet unit in the province at
the same time.
2.3 NEGOTIATION IN DIPLOMACY
It is extremely hard for one single player to win the
Diplomacy Game without gaining support from the other
players, hence negotiation plays a very important role in
this game to form alliances among players and agree on
certain commitments and promises in order to reach each
player's own objective.
Negotiation is done using the BAsic eNvironment for
Diplomacy playing Automated Negotiating Agents
(BANDANA) framework, which is a Java based
focus on implementing the tactical part of the agent
instead of researching on the negotiation part.
The most relevant literature would be a paper introducing
an automated negotiation agent, the Diplominator, where
in the paper some existing automated agents are
introduced, though a large number of them do not have
negotiating capabilities.
2.4.1 DUMBBOT
Although DumbBot is not a negotiating bot, it provides
reference on deriving heuristics for strategies and tactics.
It first assign values to each province which consider
factors such as the owner of the province, their strength,
the competitiveness of the province and the chances for
own units to move there. Then, it implements an
algorithm that "averages the board" so that the value of
a certain province will also be affected by its adjacent
province. The tactics developed by the bot will be based
on the values of the provinces.
Although DumbBot's heuristic is only used for tactical
purpose, it can still be referenced by negotiation
algorithms because negotiation deals can only be made
after determining a route plan on how to conquer the
supply centers on the board. To determine such route
plan, we may need a similar heuristic like DumbBot's.
2.4.2 DIPLOMAT
Diplomat uses an economic based view of Diplomacy,
which negotiation is used to exchange resources and the
results of the trades are focused. A simulated "meeting"
is also set up and powers that are involved in the deal
undergoes an "auction" to determine the winner that
makes the deal eventually. It also implements a part of
strategy on deceit, but it would not be applicable to our
bot.
Diplomat is said to perform worse than DumbBot,
despite its sophisticated design on the negotiation
algorithm. However, The Market Based approach and
the "auction" approach are still interesting and worth
looking into further in future.
2.4.3 THE ISRAELI DIPLOMAT
The Israeli Diplomat is a sophisticated Diplomacy player
which uses a multi-agent architecture designed based on
real life war-time government structures. The 'Prime
Minister' acts as the representative and opponents only
talk to him. If needed, the Prime Minister will pass the
suggested negotiation to another role to gain opinion
before replying to the opponent. With such architecture,
the task of negotiation could be split up to different
departments, each focusing on different criteria.
framework to allow negotiation capabilities in the
Diplomacy game [3]. Under this framework, the main
types of deals that would be proposed by players among
each other in Diplomacy are:
An order commitment: a proposal to a certain
power to impose a certain order on a certain
phase. For example: you may need France to
support your move from Vienna to Galacia on
Spring 1903.
In BANDANA, it is implemented by a tuple,
oc = (y, ᶲ, o)
where y is "year", ᶲ is either "Spring" or "Fall",
and o is the order object as specified before.
A demilitarized zone (DMZ): an agreement
among a number of powers to not occupy (or
move to) a province on a certain phase. For
example: France, Germany and Italy agreed on
not to occupy the supply centre in Belgium on
Spring 1903.
In BANDANA, it is implemented by a tuple,
dmz = (y, ᶲ, A, B)
where y is "year", ᶲ is either "Spring" or "Fall",
A is a list of powers, B is a list of provinces
intended to be demilitarized.
Hence, a BasicDeal object is implemented to contain a
list of order commitments and a list of demilitarized
zones, as below:
d = {List<oc>, List<dmz>}
which is to be sent by the agents in the end of each phase
to the corresponding powers involved in the deal to
accept or reject it.
In ANAC, we are required to only implement the
negotiation algorithm of the agent and not its tactical
module, as our algorithm will be built on top of another
tactical module (the DBrane tactical module) to play the
game. Moreover, the same tactical module will be used
by other negotiation agents as well. Also in ANAC,
promises cannot be broken. This means that if France
previously agreed to move from Belgium to Holland, it
cannot make another deal to move from Belgium to
Picardy. The DBrane tactical module will always obey
the deals made, which means the unit in Belgium will
indeed move to Holland finally.
As reneging on the deals isn't allowed to be made in this
framework, it greatly affects the strategy of the agent
since now all agents must obey the deals made. This
means that there is no chance for an agent to "cheat" on
another agent to gain benefit from it.
2.4 LITERATURE REVIEW
As the BANDANA framework is still in its early stage,
there are limited existing literatures that fully described
a negotiation model which suits the scenario of the
Diplomacy game stated above, as a large number of them
maximizing the number of supply centers on every phase
may not be an optimal strategy.
2.5 AGENT STRATEGY
We have designed Agent Madoff, which is a negotiating
agent to participate in the upcoming Diplomacy League
in ANAC 2017.
The design architecture of Agent Madoff mainly consists
of 3 components: heuristic module, acceptance strategy
and bidding strategy. The pseudo code of the negotiate()
method
below:
of Agent Madoff
is
as
Figure 2: Israeli Diplomat's multi-agent architecture
while(has message){
while (deadline is not reached){
This multi-agent architecture is claimed to be very
successful in the past, and it is truly interesting to be
looked into further.
2.4.4 DIPLOMINATOR
Diplominator uses a banking protocol as its negotiation
strategy, modelling the same idea of how banks decide
whether or not to lend money to individuals. The key
components of the strategy includes:
Utility of Cooperation - As a currency is needed in
banking protocols, it simply uses DumbBot's heuristic
to calculate the destination values of all provinces, and
uses it as the currency to measure the values of moves
requested by other bots, in order to deduce our utility of
cooperating with them.
should estimate whether the opponent is trustworthy for
negotiation or not. A simple way to implement this is to
count the percentage of times the opponent has assisted
us. Another way will be to calculate the enemy factor,
which is defined by
Creditworthiness of Opposition Players -- A bot
𝑒𝑛𝑒𝑚𝑦 𝑓𝑎𝑐𝑡𝑜𝑟 =
#𝑡𝑟𝑢𝑒 𝑎𝑡𝑡𝑎𝑐𝑘𝑠
#𝑐ℎ𝑎𝑛𝑐𝑒𝑠 𝑡𝑜 𝑎𝑡𝑡𝑎𝑐𝑘
When chances = 0, enemy factor equals the number of
true attacks by the player.
Credit Limit - We should stop assisting opponents
at some point until they pay us back, in case we are being
too kind to be exploited by opponents in the end.
2.4.5 ANACEXAMPLENEGOTIATOR
ANACExampleNegotiator is a random negotiator. It first
checks whether there are incoming messages in the
message queue, and check what type of message it is. If
there are proposed deals coming in, it accepts the deal on
a 50% probability. After handling all messages, it picks
among 10 randomly generated orders that yields the
higher number of supply centers if obeyed by others.
This bot shows a simple two-layered architecture which
focus on the acceptance strategy and the bidding
strategy, where two strategies may be loosely coupled.
However, the framework manual did remind that
handle incoming message;
if (near to deadline) break;
}
bidding strategy;
}
Update hostility and strength;
The algorithm enters a while loop which terminates when
the deadline is reached. Within the loop, we first handle
incoming messages if there are any, and apply the
acceptance strategy here. We break the loop if it is near
to deadline to prevent opponents from flooding us with
incoming messages. Next, we execute our bidding
strategy to propose deals to other powers. Before the turn
ends, we update the hostility and strength of our
opponents to keep track of the game's situation.
In the following discussion, we will further describe our
strategy according to the 3 components specified in the
design architecture.
2.5.1 HEURISTIC MODULE
The heuristic module aids the agent's decision making
by evaluating which provinces/regions on the graph are
more worthy, and hence have a higher utility for the
agent to either move into it or to protect it from
opponent's invasion.
Figure 3: Map representation around Austria (the red
region)
Take an example of Austria, certainly supply centers
(with a black dot) such as Vienna, Budapest and Trieste
have a higher utility for the power itself. However,
although provinces like Bohemia, Galacia and Tyrolia
are not SCs, they are crucial for Austria because they are
adjacent to other SCs which Austria may want to
conquer. Hence, the number of adjacent SCs should also
be a factor of determining the actual utility of a certain
province.
With the idea of adjacent provinces giving influence to
its neighbors to result in the final utility value, we
propose a heuristic calculation as below:
1.
Initialize the utility value for each region: 1 for
non-supply centers and 10 for supply centers.
2. Then, we add a value of the sum of utility values
of adjacent regions, multiplied by a discount
factor of 0.3 to the utility value of each region.
3. We divide the utility value by the maximum
utility value among all regions to normalize the
value between [0, 1]. This is because in the
future calculation of acceptance probability it
needs to yield a result between [0, 1], hence all
parameters need to be normalized.
An example will be like Figure 3 as below:
Figure 4: An example of showing the utility calculation
In Figure 4, assume that the regions in red are supply
centers. Since B itself is a supply center, it has an initial
utility of 10. The sum of the utility value of its adjacent
regions are 1+10+1 = 12. Hence the final utility value of
B is 10 + 0.3 x 12 = 13.6. Assume that utility value of A
= 3, C = 4, D = 15, then the normalized utility value of B
is
13.6
15
= 0.9067.
between [0, 1], and we called it as the acceptance
probability for that particular order.
There are 2 types of parameters that will be frequently
used in the evaluations below:
(1) Hostility
It shows the measure of friendliness of a particular power
towards you (e.g. have they attacked you, have they
support you before, etc.).
Throughout the game, a hostility list is maintained. For
each power, the hostility value is initially set to 0. If a
power supports us, its hostility value increase by 5. If a
power attacks us (steal our provinces), its hostility value
is decreased by 10. The hostility value, h is normalized
linearly:
𝐼𝑓 ℎ < 0, ℎ𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑠𝑒𝑑 = 0.5 (
)
ℎ − hmin
−ℎmin
ℎ
ℎmax
𝐼𝑓 ℎ ≥ 0, ℎ𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑠𝑒𝑑 = 0.5 (
) + 0.5
Where ℎmin and ℎmax denotes the minimum and
maximum hostility among the list.
(2) Strength
It means how strong the particular power is, as we argue
that we should team up with the weaker teams to fight
the stronger teams so that they will be a balance of power
among all players.
The strength of a team is determined by the number of
SCs that they have currently conquered. In the beginning
each team is given 3 SCs, and with 18 SCs one can
declare a solo victory. Hence we need a mapping
function to map values from [3, 18] to [0, 1] to represent
its strength. In this case, the mapping function used is
𝑠𝑡𝑟𝑒𝑛𝑔𝑡ℎ𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑠𝑒𝑑 = 0.5 sin (
𝜋(#𝑆𝐶𝑠 − 9)
18
) + 0.5
Support move order
For our agent, all regions would have a fixed utility value
throughout the game. Hence, it could be pre-computed
once the agent is initialized.
2.5.2 ACCEPTANCE STRATEGY
The acceptance strategy is done on a case-by-case basis
according to the type of the order. For each type of order,
we determine several parameters which are essential to
the acceptance of the order. A composite function that
takes in the parameters is derived to calculate a value
Figure 5: Support move order
We establish 5 parameters: Supportee's hostility,
supportee's strength, unit's neediness, target's hostility,
target's strength. All parameters have a value between [0,
1].
Unit's neediness, in this case, means how much is the
army needed for my team. The higher the unit's
neediness, the more reluctant the unit is to support other
unit's move. With this, we need to know the original plan
of the army which is available in the strategic module of
the BANDANA framework:
If the army is planned to move to some other
places, the unit's neediness depends on the
utility value of the intended region
If the army is planned to support my own team,
unit's neediness = 1
If the army is planned to support other teams, its
neediness is higher if the supporting team is
more friendly towards us
Figure 7: Move-to order
If the army is planned to just hold, unit's
neediness = 0.5
We establish 3 parameters: target's hostility, and whether
the new order is better (called newIsBetter below).
With these 5 parameters, the acceptance probability is
then equal to:
𝑎𝑐𝑐𝑒𝑝𝑡𝑎𝑛𝑐𝑒 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦
newIsBetter is to determine whether moving to the
proposed target is better than its original plan. With this,
we check the original plan of the army.
= 0.2 × ℎ𝑜𝑠𝑡𝑖𝑙𝑖𝑡𝑦𝑠𝑢𝑝𝑝𝑜𝑟𝑡𝑒𝑒 + 0.1
× (1 − 𝑠𝑡𝑟𝑒𝑛𝑔𝑡ℎ𝑠𝑢𝑝𝑝𝑜𝑟𝑡𝑒𝑒) + 0.5
× (1 − 𝑢𝑛𝑖𝑡𝑁𝑒𝑒𝑑𝑖𝑛𝑒𝑠𝑠) + 0.1
× 𝑠𝑡𝑟𝑒𝑛𝑔𝑡ℎ𝑡𝑎𝑟𝑔𝑒𝑡 + 0.1
× ℎ𝑜𝑠𝑡𝑖𝑙𝑖𝑡𝑦𝑡𝑎𝑟𝑔𝑒𝑡
If no unit resides in the target region, then target's
hostility and strength equals to 0, in this case:
𝑎𝑐𝑐𝑒𝑝𝑡𝑎𝑛𝑐𝑒 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦
= 0.3 × ℎ𝑜𝑠𝑡𝑖𝑙𝑖𝑡𝑦𝑠𝑢𝑝𝑝𝑜𝑟𝑡𝑒𝑒 + 0.2
× (1 − 𝑠𝑡𝑟𝑒𝑛𝑔𝑡ℎ𝑠𝑢𝑝𝑝𝑜𝑟𝑡𝑒𝑒) + 0.5
× (1 − 𝑢𝑛𝑖𝑡𝑁𝑒𝑒𝑑𝑖𝑛𝑒𝑠𝑠)
Support hold order
We initialized newIsBetter = 0.2.
If the army is to hold, and the target to move in
has a higher utility value, then newIsBetter =
0.8; or else newIsBetter remains as 0.2.
If the army is planning to move to the other
destination, we compare the utility of the
original destination to the target region. If the
target has a higher utility, newIsBetter = 0.8; or
else newIsBetter remains as 0.2.
Hence,
the acceptance probability
is as below:
𝑎𝑐𝑐𝑒𝑝𝑡𝑎𝑛𝑐𝑒 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦
= 0.3 × ℎ𝑜𝑠𝑡𝑖𝑙𝑖𝑡𝑦𝑡𝑎𝑟𝑔𝑒𝑡 + 0.7
× 𝑛𝑒𝑤𝐼𝑠𝐵𝑒𝑡𝑡𝑒𝑟
If no unit resides in the target region, then target's
hostility equals to 0, in this case the acceptance
probability simply equals to newIsBetter.
Hold order
Figure 6: Support hold order
Support hold order works in a similar way with support
move-to order without target node, only that the
supportee is to hold. Hence,
𝑎𝑐𝑐𝑒𝑝𝑡𝑎𝑛𝑐𝑒 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦
= 0.3 × ℎ𝑜𝑠𝑡𝑖𝑙𝑖𝑡𝑦𝑠𝑢𝑝𝑝𝑜𝑟𝑡𝑒𝑒 + 0.2
× (1 − 𝑠𝑡𝑟𝑒𝑛𝑔𝑡ℎ𝑠𝑢𝑝𝑝𝑜𝑟𝑡𝑒𝑒) + 0.5
× (1 − 𝑢𝑛𝑖𝑡𝑁𝑒𝑒𝑑𝑖𝑛𝑒𝑠𝑠)
Move-to order
Figure 8: Hold order
We first check the unit's original plan.
If its plan is to hold, simply accept the proposal.
If its plan is to move to another region, check
the utility of the intended region. If the province
has a utility greater than 0.7, reject the proposal.
Or else, return a default probability of 0.4. This
shows a slight reluctance to accept a hold order
for our agent.
If its plan is to support other units, it is highly
improbable to accept a hold order. Hence we
return a probability of 0.1.
DMZ order (demilitarized zone order)
DMZ proposal is made between several powers such that
all the powers agree not to invade the specified
provinces.
We establish 2 parameters needed: utility of the region,
and the competitiveness of the DMZ proposal.
The higher the utility of the region, the more reluctant we
would demilitarize it since it is important to us.
Competitiveness is measured by counting the number of
powers involved in the DMZ. It is smoothen by a
logarithm function to map the value between [0, 1]. We
argue that the more number of powers involved, the
harder it is for you to fight for the region if you do not
accept the DMZ, hence the more competitive it is the
more probable we should demilitarize the region.
For DMZs, we check the original plan of each units and
see which provinces we intend to move to.
If the DMZ zone is not within our intended
provinces, simply accept it as it cause no harm
to us
Else, we compute the acceptance probability for
the DMZ proposal as below:
𝑎𝑐𝑐𝑒𝑝𝑡𝑎𝑛𝑐𝑒 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦
= 0.6 × 𝑐𝑜𝑚𝑝𝑒𝑡𝑖𝑡𝑖𝑣𝑒𝑛𝑒𝑠𝑠
+ 0.4 × (1
− 𝑢𝑡𝑖𝑙𝑖𝑡𝑦𝑂𝑓𝑅𝑒𝑔𝑖𝑜𝑛)
COMPOSITE ORDER PROPOSAL
Proposals usually contain more than 1 type of order. For
example, A could send a proposal to B such that A
request B to support his move, and A will DMZ one of
the regions in conflict with B as a return of favor. In this
case, this proposal contains a support order and a DMZ
order.
For our agent, after calculating
the acceptance
probability of each order, we calculate the mean
probability of all orders and treat it as the acceptance
probability of the composite order proposal. Denote
mean probability as x:
If x > 0.8, accept the deal;
If x < 0.4; reject the deal;
If 0.4 ≤ x ≤ 0.8, we flip a coin so that the deal
is accepted with a probability of x.
In addition, our agent only takes in account the proposals
regarding the current phase only. For proposals regarding
the future phases, we would return an acceptance
probability of 0, as we argue that there is no reason for
us to bind to a "virtual" agreement made for the
unpredictable future.
2.5.4 BIDDING STRATEGY
Agent Madoff does not speculate moves further than the
current phase. Hence, each proposal of the agent is done
only according to the current phase and game setting.
This is because binding to deals which happen in the
future may be invalid and hence unnecessary.
The bidding strategy that our agent adopted is a
defensive, de-conflict strategy. It consists of 2 stages:
neutralize the attack regions, and resolve the conflict
regions.
Order Calculator
We first discuss about the "order calculator" which is
used within the bidding strategy. By assuming that for
each type of order, the opponent considers same
parameters as our agent, we could use our acceptance
strategy module to mimic an order calculator which
calculates the acceptance probability of a power for a
given order in its own perspective. With this, we could
propose orders with higher acceptance probability to
achieve a higher chance of agreement.
Neutralizing attack regions
The strategic module within BANDANA is able to return
the opponent's move given its power and the current
game state, before negotiation is made. Hence for each
power, we could speculate the pre-determined move of
each of its unit. If we find that the unit's move is to
invade our supply center, we should neutralize the attack
through negotiation.
The neutralization steps are as below:
We generate all possible orders for the attacking
unit, and calculate all acceptance probability of
the orders.
If there exists orders which has a higher
probability than the attacking order, we pick the
highest one and propose it to the opponent.
If no alternative orders are found, we request for
support from other powers, and propose a
"favor returning" deal (which will be discussed
later) to increase the chance of agreement.
If we have no units to move into / hold our
supply centers, we could only ask for a DMZ
with the attacking unit.
Resolving conflict regions
Since the strategic module is able to anticipate the
opponents move given the opponent's power and the
current game state, we could then see if our agent's
moves are in conflict with other agent's move before
negotiation takes place.
A "conflict" means that there are more than 1 opponent
that want to move to a region that you intend to move
into. If there is, then negotiation proposals are needed to
resolve the conflict.
By supporting it to hold: some units may want
to hold their position to prevent invasion into
their regions by other powers. If we have a unit
which is adjacent to the supporter's unit, we
could support them to hold.
Figure 9: Conflict move
The resolving steps are as below:
Figure 11: Return the favor by support hold order
For each power, we check if we have any in-
conflict moves with them.
We generate all possible orders for
the
opponent's conflicting unit, and calculate all
acceptance probability of the orders.
If there exists orders which has a higher
probability than the conflict order, we pick the
highest one and propose it to the opponent.
If no alternatives are found, we request for
support from other powers, and propose a
"favor returning" deal (which will be discussed
later) to increase the chance of agreement.
We keep a DMZ deal with the opponent's unit
for that region as a "reservation deal" which
will be proposed should the previous deals
failed to be accepted.
Return the favor
Normally within a proposal, we would add a "favor
returning" deal apart from a pure request of support so
that the proposal is more likely to be accepted. In this
case, we specify a "return credit", which is the number
of our units which could return the favor to the powers
which have supported us. This is to prevent most of our
units being used for returning the favor, which further
weakens our strength. We allow 1/3 of the total number
of units that we currently have to be our "return credit".
We propose 2 ways to return the favor to our supporters:
By supporting its move: we search for units
situated adjacent to the power's army's adjacent
neighbors, as shown in the top part of the
diagram above. If there is, and if the neighbor
(region A as in Figure 9) has a very high utility
for the power, we could propose to support the
power to move into A in return.
3 CONCLUSION
In this paper, we first introduce the new Diplomacy
league in ANAC and observe how negotiation is
implemented in the league using the BANDANA
framework. We give a review on existing Diplomacy
agents, which provide useful ideas for us to implement
an efficient agent. We present the design of Agent
Madoff, whose architecture comprises of 3 main
components: the heuristic module, acceptance strategy
and bidding strategy. Our agent is submitted to
participate in ANAC 2017 Diplomacy league, and the
results are forthcoming.
Future work will focus on improving each of the
modules: for the heuristic module, we could adopt a
dynamic model which changes according to the game
situation instead of the current static model; for the
acceptance strategy, the parameters could be adjusted to
further mimic the thought process of a human agent
making negotiation decisions.; for the bidding strategy,
it could be improved to search for tempting bids that are
more likely to be accepted by the opponents. Instead of
adopting a defensive strategy, we could also propose
deals to form attacks and arrange battle plans against our
opponents, acting as a real diplomat during war.
ACKNOWLEDGEMENT
I would like to thank my supervisor, Prof Bo An for
providing detailed guidance and support for this project.
His experienced thoughts in the field of automated
negotiation has spurred exciting discussions between us.
Moreover throughout the year, he has showed me by
himself all the essential qualities of being a dedicated and
successful researcher.
REFERENCES
[1] R. Aydogan, "Eighth International Automated
Negotiation Agents Competition", May 2016;
http://web.tuat.ac.jp/~katfuji/ANAC2017/
[2] Avalon Hills, "The Rules of Diplomacy", 2008;
https://www.wizards.com/avalonhill/rules/diploma
cy_rulebook.pdf
[3] D. D. Jonge, "The BANDANA Framework v1.3"
2016;
December
Figure 10: Return the favor by support move-to order
http://www.iiia.csic.es/~davedejonge/bandana/file
s/Bandana%201.3%20Manual.pdf.
[4] A. Webb, J. Chin and T. Wilkins, J. Payce,
"Automated Negotiation
In The Game Of
Diplomacy", Technical report, Imperial College
London, 2008.
[5] D. D. Jonge, "Negotiations over Large Agreement
Spaces", PhD thesis, Universitat Autonoma de
Barcelona, 2015.
[6] T. Baarslag, K. Hindriks, C. Jonker, S. Kraus, R.
Lin, "The First Automated Negotiating Agents
Competition (ANAC 2010)", New Trends in Agent-
Based Complex Automated Negotiations, SCI, vol.
383, pp. 113-135, Springer, Heidelberg (2010)
|
1909.04615 | 1 | 1909 | 2019-09-10T16:45:16 | On Re-Balancing Self-Interested Agents in Ride-Sourcing Transportation Networks | [
"cs.MA",
"math.OC"
] | This paper focuses on the problem of controlling self-interested drivers in ride-sourcing applications. Each driver has the objective of maximizing its profit, while the ride-sourcing company focuses on customer experience by seeking to minimizing the expected wait time for pick-up. These objectives are not usually aligned, and the company has no direct control on the waiting locations of the drivers. In this paper, we provide two indirect control methods to optimize the set of waiting locations of the drivers, thereby minimizing the expected wait time of the customers: 1) sharing the location of all drivers with a subset of drivers, and 2) paying the drivers to relocate. We show that finding the optimal control for each method is NP-hard and we provide algorithms to find near-optimal control in each case. We evaluate the performance of the proposed control methods on real-world data and show that we can achieve between 20% to 80% improvement in the expected response. | cs.MA | cs | On Re-Balancing Self-Interested Agents in Ride-Sourcing Transportation Networks
Armin Sadeghi
Stephen L. Smith
9
1
0
2
p
e
S
0
1
]
A
M
.
s
c
[
1
v
5
1
6
4
0
.
9
0
9
1
:
v
i
X
r
a
Abstract -- This paper focuses on the problem of controlling
self-interested drivers in ride-sourcing applications. Each driver
has the objective of maximizing its profit, while the ride-
sourcing company focuses on customer experience by seeking to
minimizing the expected wait time for pick-up. These objectives
are not usually aligned, and the company has no direct control
on the waiting locations of the drivers. In this paper, we provide
two indirect control methods to optimize the set of waiting
locations of the drivers, thereby minimizing the expected wait
time of the customers: 1) sharing the location of all drivers with
a subset of drivers, and 2) paying the drivers to relocate. We
show that finding the optimal control for each method is NP-
hard and we provide algorithms to find near-optimal control in
each case. We evaluate the performance of the proposed control
methods on real-world data and show that we can achieve
between 20% to 80% improvement in the expected response.
I. INTRODUCTION
In recent years, ride-sourcing services such as UberX and
Lyft have emerged as an alternative mode of urban trans-
portation. The reduced wait times for pickup of these services
is the compelling feature compared to the conventional taxi
services [1]. A key factor that affects the response time of the
service is where the drivers wait to respond to the next ride
request. Ride-sourcing companies do not have control over
the position of the drivers as they are self-interested units
maximizing their objectives. Therefore, a challenge is to
ensure the drivers are distributed throughout the city in order
to minimize the expected wait time of the customers. This
must be done either by providing information to drivers, or
through incentives (payment) that make relocation attractive.
As a method for both re-balancing and increasing the
supply of drivers, Uber introduced surge pricing in high
demand areas. This reduces the expected response time of
servicing the requests by drawing more drivers to those areas.
However, the surge pricing can draw drivers away from lower
demand areas, resulting in higher wait times in those areas
and more imbalance [2], [3].
The problem of servicing requests in ride-sourcing net-
works can be divided into two major problems: 1) assignment
of the ride request to the drivers; and 2) re-balancing of the
drivers for future ride requests. In this paper, we focus on
the re-balancing problem for a subset of drivers to service
ride requests that arrive sequentially in an environment. The
drivers' motion in the environment is captured as a road-
map (i.e., graph), and each ride request arrives at a node of
the graph according to a known arrival rate (see Figure 1).
The drivers are self-interested units maximizing the expected
profit of their workday. Hence, the ride-sourcing company
This research is partially supported by the Natural Sciences and Engi-
neering Research Council of Canada (NSERC).
The authors are with the Department of Electrical and Computer
Engineering, University of Waterloo, Waterloo ON, N2L 3G1 Canada
([email protected]; [email protected])
Fig. 1: A set of drivers in the ride-sourcing system and a set of
locations with high probability of ride request arrival.
has no direct control over the waiting locations of the drivers.
The objective of the ride-sourcing company is to incentivize
the drivers to relocate to a set of waiting locations that
minimizes the expected wait-time of the customers.
Related Work: The problem of dispatching taxis to service
ride requests arriving sequentially over time has been the
subject of extensive research [4], [5], [6], [7]. These studies
focus on policies to optimally assign the ride requests to
the taxis. In contrast, we focus on the waiting locations
of the drivers that minimize the expected wait time of the
customers. We assume the ride requests are assigned in a
first-come-first-serve fashion to the closest available driver.
Assigning the closest available driver to each request
is
the common method employed by the ride-sourcing com-
panies [8].
The problem of re-balancing service units in the environ-
ment has been studied for various applications. In mobility-
on-demand problem (MOD) [9], [10], [11], a group of
vehicles are located at a set of stations. The customers arrive
at the stations, hire the vehicles for ride, and then drop
the vehicles off at their destination stations. The objective
is to balance the vehicles at the stations to minimize the
expected wait time of the customer. In comparison to MOD,
we consider the customer wait-time as the time between
the request arrival and the pick-up time, which incorporates
the distance of the closest available vehicle to the pick-up
location.
The facility location problem [12], [13] and its extension
to the mobile facility location problem (MFL) [14] is the
problem of distributing facilities in a set of locations to
respond to the demands arriving at different locations. The
objective is to minimize the time to respond to the demands
and the total cost of opening facilities. A special case of
the facility location problem is the k-median problem [13]
where the number of open facilities is limited and the
cost of opening a facility is zero. In [15], we addressed
a multi-stage MFL problem where we relocate a set of
autonomous vehicles to minimize expected response time for
future requests in a receding horizon manner.
In the aforementioned studies, the actions of service units
are controlled by a central unit and the objective of the
service units is aligned with the global objective. However,
we consider self-interested units maximizing their own profit
such that the ride-sourcing companies have no direct control
on their decisions. A closely related problem is that of
Voronoi games on graphs [16] where requests arrive on the
vertices of a graph and the objective of each self-interested
service unit is to maximize the number vertices assigned
to them. They cast the problem as a game between the
service units prove that the problem of finding the pure
Nash equilibrium on general graphs is NP-hard. In [17], the
authors provide the best response strategy for each driver and
they approximate Nash equilibria, which can be utilized by
the ride-sourcing companies to incentivize the drivers in a
way to maximizes a global objective. These studies focus on
the strategies of the self-interested service units, in contrast,
we focus on finding the optimal policy for the ride-sourcing
company to optimally respond to the ride requests.
Contributions: The contributions of this paper are three-
fold. First, we formulate the re-balancing problem of self-
interested service units. Second, we propose two indirect
method to relocate the service units to minimize the expected
response time and provide algorithms with near-optimal
solutions. Third and finally, we evaluate the performance
of the two proposed control methods on real-world ride-
sourcing data.
The paper is organized as follows. In Section II, we for-
mulate the problem of minimizing customer wait-time with
self-interested service units. In Section III, we provide the
first indirect control method based on sharing the information
on the location of the drivers. Section IV.I consists of the
second control method based on incentive pay for the drivers
to relocate to desired waiting locations. Finally in Section V,
we provide an extensive set of experiments, on real-world
ride-sourcing data, characterizing the performance of the two
control methods.
II. PROBLEM FORMULATION
Consider a set of m drivers and a set of pick-up and drop-
off locations V . Let G = (V, E, c) be a metric graph on
vertices V , let E be the set of edges between the locations
and c : E → R+ be the function assigning a travel time to
each edge of the graph. The drivers wait on a subset of the
vertices for the next request, which we call the configuration
of the drivers Q. The set of all configurations of the drivers
is denoted by Q. Each driver i is aware of the position of a
We assume that
subset of the drivers Ii ⊆ Q, for instance, each driver may
be aware of the location of the other drivers in its vicinity.
the requests arrive at each vertex u
according to an independent Poisson process with arrival
rate λu. Upon a request arrival, the closest driver to the
vertex of the request
is assigned to service the request.
Let pa(u) denote the arrival probability, which is the ratio
of number of requests arriving at u to the total number
of requests arriving in a period of time. Let the drop-off
probability pd(dropoff = wpickup = v) be the probability
that a request with pickup location at vertex v has a drop-off
location at vertex w.
Driver i's perception of her expected profit is a function
of her information Ii on the location of other drivers,
environment parameters such as arrival times, her waiting
location qi and the period of working time Bi, denoted by
Vi(u, Bi). For the development of our main control methods
we do not assume any specific form of this function. We do
assume, however, that the ride-sourcing company has access
to this function, obtained through data of driver behavior.
In Section V we present one potential model of Vi(u, Bi),
which is then used for simulating the two control methods.
Drivers' objective: Each driver is a self-interested unit,
therefore, they will wait at a location that maximizes their
expected profit, i.e.,
−σc(qi, u) + Vi(u, Bi − c(qi, u)),
(1)
arg max
u∈V
where σ is the cost per minute of driving.
Global objective: In addition to the objective of each
driver, there is a global objective for the service providers
such as Uber and Lyft to maximize the service quality by
minimizing the expected wait time of the customers until
pick-up, i.e.,
(cid:88)
u∈V
min
Q∈Q D(Q) =
min
qi∈Q
pa(u)c(qi, u),
(2)
The main challenge in optimizing the global objective
is that the drivers are self-interested units and the service
provider does not have any direct control on the configuration
of the vehicles. Therefore, the service provider is not able
to minimize the expected response time to the requests
directly. The two indirect control methods proposed in this
paper incentivize the drivers to relocate to desired waiting
locations. The first control method exploits the dependency
of the expected profit of the drivers on their information
Ii. The service provider can share more information on the
location of drivers with a subset of them to manipulate their
decision towards relocating to a desired waiting location. We
refer to this as the sharing information control method. The
second proposed control method, incentivizes the drivers to
relocate to desired waiting locations with payments, which
we refer to as the pay-to-control method. These control
methods are applicable to various models of driver behaviour
V.
In the following sections, we provide a detailed description
of the two control methods and propose algorithms to find
near optimal controls.
(a) Equal information sharing
(b) Partial information sharing
Fig. 2: Instance of ride-sourcing problem with two vehicles and two
request arrival locations.
III. CONTROL BY SHARING INFORMATION
that
In this section, we provide an indirect control on the
configuration of the drivers exploiting the fact
the
optimal waiting location for each driver in Equation (1) is a
function of the information provided to the driver regarding
the position of the other drivers, i.e., Ii.
Figure 2 demonstrates the importance of information on an
instance of the ride-sourcing problem with two vehicles and
two request locations. The locations are within unit distance
apart and the arrival rate at locations v1 and v2 are 0.1
and 0.2, respectively. The vehicles are initially located at
v1 and will relocate to the best waiting location, namely
optimizing Equation (1). Figure 2a shows the two scenarios
where both vehicles are provided the same information, i.e.,
I1 = I2 = ∅ and I1 = I2 = Q. Note that the configuration
of the vehicles when they are provided the same information
is the worst possible configuration for the global objective.
However, illustrated in Figure 2b, providing the information
to a subset of the vehicles results in the optimal configuration
for the global objective.
Let q(cid:48)
The information sharing problem consists of deciding
the subset of drivers we share information with and the
information shared with each driver. However, for this work,
we consider the binary decision where either full information
or no information is provided.
i,Q (resp. q(cid:48)
i,∅) be the new waiting location selected
by driver i from Equation (1) with information Ii = Q (resp.
i,∅} be the set of candidate waiting
Ii = ∅). Let Fi = {q(cid:48)
i,Q, q(cid:48)
locations for driver i. If there exist i, j ∈ [m] such that q(cid:48)
=
q(cid:48)
, then we create a duplicate vertex u for q(cid:48)
such that
, v) = c(u, v) for all v ∈ V . The formal definition of
c(q(cid:48)
the problem of sharing information with drivers is given as
follows:
Problem III.1. Consider a metric graph G = (∪m
i=1Fi ∪
V, E, c). Find a new configuration Q(cid:48) by picking only one
Q(cid:48) ∩ Fi = 1 for
vertex from each Fi,
each i, while minimizing the global objective D(Q(cid:48)) =
i.e., such that
j,Ij
i,Ii
j,Ij
i,Ij
u∈V minqi∈Q(cid:48) pa(u)c(q, u).
Figure 3 shows an instance of the information sharing
problem. The green vertices are the waiting locations of the
drivers if they have no information on the position of the
other drivers, and the red vertices are the waiting locations
of the drivers if the information is provided to each driver by
(cid:80)
Fig. 3: An instance of Problem III.1. The green vertices represent
the desired waiting location of each driver if Ii = ∅, and the red
vertices represent the waiting location of the drivers if Ii = Q.
the service provider. Let Q(cid:48) be the solution to Problem III.1
in which if qi,Q ∈ Q(cid:48) then the driver i is provided complete
information of the position of the other drivers, and no
information is available for driver i if q(cid:48)
i,∅ ∈ Q(cid:48).
First, we analyze the complexity of Problem III.1, and then
we provide an LP-rounding algorithm to find the controls.
A. Proof of Hardness
We now prove the NP-hardness of Problem III.1 with a
reduction from CNF-SAT [18] as follows:
Consider an instance of CNF-SAT with n Boolean vari-
ables and m clauses. We will reduce this problem to Prob-
lem III.1.
(i) Let ∪m
i , vF
i=1Fi contain 2n vertices, partitioned into n
sets of size two. The set Fi contains two vertices,
i }, where vT
{vT
i will correspond to setting the ith
SAT variable to true (i.e., the positive literal) and vF
i
will correspond to setting it to false (i.e., the negative
literal).
V .
clause in the SAT formula.
(ii) We let V contain m vertices, one representing each
(iii) Let E contain an edge for each v ∈ ∪m
i=1Fi and w ∈
(iv) For each e = (v, w) ∈ E, we set its cost to 1 if the
literal v appears in the clause w, and 2 if the literal
does not. Note that the costs are metric.
(v) Let Pu = 1/m for each u ∈ V .
Now, we solve the instance of Problem III.1. If it returns
a subset of ∪m
i=1Fi with cost exactly 1, then for each clause
c ∈ V , there is literal in ∪m
i=1Fi with edge cost of 1 to c.
This implies that the literal chosen from each subset in the
partition of ∪m
i=1Fi gives a satisfying truth assignment for
the SAT instance. If the subset returned has a cost greater
than 1, then there exists a clause w ∈ V for which every
chosen literal has edge cost of 2. Thus, this clause is not
satisfied and no satisfying instance exists.
B. Linear-Program Rounding Algorithm
Given that Problem III.1 is NP-hard, we turn our focus
to suboptimal algorithms. In particular, we provide a simple
Linear Program (LP)-rounding algorithm for Problem III.1.
Although we do not provide bound on the performance of
the LP-rounding algorithm, we evaluate the performance of
the algorithm on an extensive set of real-world ride sourcing
data in Section V and we show that the proposed algorithm
is on average within 0.014% of the optimal.
First we cast Problem III.1 as an integer linear program
(ILP), then we propose a rounding algorithm based on the
solution to the relaxation of the ILP. Let integer parameter
xu,v ∈ {0, 1} denote the assignment of a request at v to a
driver at vertex u if xu,v = 1, and xu,v = 0 otherwise. Let
the integer parameter yu ∈ {0, 1} for all u ∈ ∪m
i Fi represent
if there is a driver assigned to wait for next request arrival
at u. Then we write the ILP for Problem III.1 as follows:
(cid:88)
minimize (cid:88)
subject to: (cid:88)
v∈V
i Fi
u∈∪m
xu,v ≥ 1,
i Fi
u∈∪m
yu ≥ xu,v,
yu + yv = 1,
yu, xu,v ∈ {0, 1},
(3)
(4)
(5)
(6)
pa(v)c(u, v)xu,v
∀v ∈ V
∀v ∈ V, u ∈ ∪m
i Fi
Fi = {u, v}, i ∈ [m]
∀v ∈ V, u ∪m
i Fi
By constraint (4), a feasible solution assigns each request
location to a driver. Equation (5) ensures that a request is
assigned to u only if there is a driver located at u, and finally
Equation (6) shows that in a feasible solution only one of
the candidate waiting locations is chosen from each subset
Fi, which represent that either the information is provided
to a driver or otherwise.
Now we propose our LP-rounding algorithm for Prob-
lem III.1. Let (x(cid:48), y(cid:48)) be the solution to the LP relaxation of
ILP (3). Without loss of generality for all Fi = {u, v}, i ∈
[m], let yu ≥ 1/2 and yv ≤ 1/2. Given solution (x(cid:48), y(cid:48)) we
construct an integer solution to ILP (3) by setting yu = 1
for each vertex u with y(cid:48)
u > 1/2 and yv = 0. In a case,
Fi = {u, v} and y(cid:48)
u = y(cid:48)
v = 1/2, we set yu = 1 where u is
the optimal waiting location of driver i with Ii = ∅. Then we
assign each vertex v ∈ V to the closest vertex u in ∪m
i=1Fi
with yu = 1 by setting xu,v = 1. Note that the constructed
solution (x, y) satisfies the constraint of ILP (3), therefore,
it is a feasible solution to Problem III.1. Also, observe that
the optimal objective value to the LP relaxation is a lower-
bound on the optimal value of ILP (3) and provides a bound
on the performance of the LP-rounding algorithm.
In the solution to the information sharing problem, if driver
i is selected to receive information on the location of drivers,
a snapshot of the location of drivers is presented to driver
i and the driver can calculate their expected profit based
on complete information. This method employed at each
time step and presents information to a driver if there is
an opportunity to improve the expected response time.
The problem of information sharing indirectly controls
the configuration of the drivers by providing information to
a subset of them, however, the possible configurations are
limited to the candidate waiting locations of the drivers. In
the following section, we provide the details on the pay-
to-control method for the service provider to optimize the
global objective.
IV. PAY TO CONTROL
Each driver as a self-interested unit chooses its waiting lo-
cation by maximizing the profit in Equation (1). To convince
the vehicles to relocate to another configuration, the service
provider needs to compensate for the difference between their
expected profit of the new location and their expected profit
for the waiting location from Equation (1). First, we pose the
problem between the drivers and the service provider as a
game. Then we provide an approximation algorithm to find
the optimal policy for the service provider.
A. Service Provider's Game
Let di be the incentive per unit distance offered to driver
m}) be the
i. Let Q = {q1, . . . , qm} (resp. Q(cid:48) = {q(cid:48)
configuration of the drivers before (resp. after) the incentive
pay. The game between the drivers and service provider
consists of the following:
1, . . . , q(cid:48)
• A set of m players and a service provider,
• An action set Ai for each driver i, which is the waiting
locations in the graph, i.e. Ai = V ∀i ∈ [m]. The action
set of the service provider is Q; and
• The profit function of the service provider is
i) + βD(Q(cid:48)),
diσc(qi, q(cid:48)
h(Q(cid:48)) =
(cid:88)
i∈m
where β ≥ 0 is a user-defined parameter that indicates
the importance of the service quality for the service
provider with respect to the incentive pay. For a small
value of β, the incentive pay is in the priority, thus the
service provider will offer the waiting locations close to
the driver's desired waiting location, however, for large
values of β, the service provider accepts high incentive
pay to relocate the drivers to the configuration with
minimum expected response time.
• The profit of driver i is the maximum of the expected
profit of the offered waiting location with incentive pay
and the expected profit of the waiting location from
Equation (1), i.e.,
max{(di − 1)σc(qi, q(cid:48)
i) + V(q(cid:48)
i)),
−σc(qi, u) + V(u, Bi − c(qi, u))}.
i, Bi − c(qi, q(cid:48)
max
u∈V
This is an instance of a leader-follower game [19]. The
service provider offers an incentive based on its utility and
the drivers as followers either take the offer or reject it.
The service provider is aware of the best action of the
drivers given any action taken by the service provider (i.e.,
incentive pay and the offered waiting location). The objective
is to find the optimal strategy for the service provider to
minimize a linear combination of the incentive pay and
the expected response time by relocating the drivers to the
desired configuration.
Driver i will accept the offer by the service provider to
relocate to q(cid:48)
i only if the offered incentives surpass the best-
expected profit of the driver. Since the profit functions of the
drivers are known to the service provider, then the minimum
di in which the drivers will accept the offer to move to
configuration Q(cid:48) is
maxu∈V −σc(qi, u) + Vi(u, B − c(qi, u))
di =
− Vi(q(cid:48)
c(qi, q(cid:48)
i)
i, Bi − c(qi, q(cid:48)
i))
σc(qi, q(cid:48)
i)
+ 1.
(7)
In the equation above, maxu∈V −σc(qi, u) + Vi(u, B −
c(qi, u)) is the maximum expected profit of the driver i by
i, Bi−c(qi, q(cid:48)
relocating to a new waiting location, and Vi(q(cid:48)
i))
is the expected profit of driver i by waiting at the location
q(cid:48)
i offered by the service provider. Knowing this minimum
di, the objective of the service provider becomes
c(q(cid:48)
h(Q(cid:48)) =
(cid:88)
σc(qi, q(cid:48)
(8)
i) + β
i, u)
pu min
i∈[m]
u∈V
−σc(qi, u) + Vi(u, B − c(qi, u))
i, Bi − c(qi, q(cid:48)
i)).
max
u∈V
Vi(q(cid:48)
i∈m
(cid:88)
(cid:88)
−(cid:88)
Observe that(cid:80)
i∈m
+
i∈m
i∈m maxu∈V −σc(qi, u)+Vi(u, B−c(qi, u))
is independent of the optimization parameters. Therefore, the
problem of minimizing the utility function of the service
provider h has the mobile facility location (MFL) problem
as a special case where Vi(v, Bi − c(u, v)) = 0 for all
u, v ∈ V and i ∈ [m]. The MFL is a well-known NP-hard
problem [20] where given a metric graph G = (F ∪D, E, c),
mapping µ : D → R+ and a subset Q ⊆ F ∪ D of size m.
m} ⊆ F
The objective is to find a subset Q(cid:48) = {q(cid:48)
u∈D µu minq(cid:48)∈Q(cid:48) c(u, q(cid:48)).
Remark IV.1 (Equilibrium). The optimal solution to the prob-
lem minQ(cid:48) h(Q(cid:48)) is the equilibrium of the leader-follower
game between the service provider and the drivers. Since
any other configuration will increase the cost function of the
service provider. In addition, By Equation (7), waiting in a
location other than the one suggested by the service provider
will decrease driver's expected profit.
minimizing(cid:80)
i) +(cid:80)
i∈[m] c(qi, q(cid:48)
1, . . . , q(cid:48)
B. Approximation Algorithm
We now propose a constant factor approximation for
the minimum pay-to-control problem, namely minimizing
σ maxu∈V σc(qi, u) − Vi(u, B −
Equation (8). Let wq(cid:48)
i,Ii = 1
i, Bi − c(qi, q(cid:48)
c(qi, u)) + V (q(cid:48)
i)), then the utility function of
the service provider becomes
i) − wq(cid:48)
h(Q(cid:48)) =
(cid:0)c(qi, q(cid:48)
(cid:1) + β
(cid:88)
(cid:88)
c(q(cid:48)
i,Ii
pa(u) min
i∈[m]
i, u).
u∈V
i∈[m]
The algorithm follows by a reduction from the minimum
pay-to-control problem to MFL.
we construct an MFL instance as follows:
Given an instance of the minimum pay-to-control problem
(i) A graph G = (Q∪ F ∪ V, E, c(cid:48)) where F is the set of
(ii) There is an edge between qi ∈ Q and q(cid:48) ∈ F with cost
(iii) There is an edge between q(cid:48) ∈ F and v ∈ V with cost
possible waiting locations for the drivers
c(cid:48)(qi, q(cid:48)) = c(qi, q(cid:48)) − wq(cid:48),Ii,
c(cid:48)(q(cid:48), v) = c(q(cid:48), v).
Fig. 4: Constructed MFL instance for optimizing the utility of the
service provider. An instance of the edges between the subsets is
shown with their respective costs.
(iv) The objective is to find a set of m vertices in F such
that minimizes
C(Q(cid:48)) =
(cid:88)
i∈m
(cid:88)
u∈V
c(cid:48)(qi, q(cid:48)
i) + β
pa(u) min
i∈[m]
c(cid:48)(q(cid:48)
i, u).
Figure 4 shows a schematic representation of the constructed
MFL instance. Suppose Q(cid:48) is a solution to the MFL instance,
we let Q(cid:48) be the solution of the minimum pay-to-control
problem and provide the following result on the cost of the
solution.
Lemma IV.2. Given an α-approximation algorithm for
the MFL problem,
the reduction above provides an α-
approximation for the minimum pay-to-control problem.
Proof. For any Q(cid:48) ⊆ F , by the construction of the MFL
instance, we have h(Q(cid:48)) = σC(Q(cid:48)). Therefore, given an
α-approximation algorithm for the MFL problem, and Q(cid:48)
obtained from the constructed MFL instance, we select
Q(cid:48) as a solution to the minimum pay-to-control problem.
Therefore,
h(Q(cid:48)) = σC(Q(cid:48)) ≤ ασ min
Q∗⊆F
C(Q∗) = α min
Q∗⊆F
h(Q∗).
By the result of Lemma IV.2, the 3 + o(1)-approximation
algorithm for the MFL problem in [20] applies to the
minimum pay-to-control problem.
V. SIMULATION RESULTS
We evaluate the performance of the two proposed indirect
controls on ride-sourcing data from Uber [21]. The data set
consists of the pick-up time and locations from April to
September 2014 in New York City, primarily Manhattan. The
jammed scenario occurs frequently with high demand in an
area and especially when surge price is applied for the high
demand area [2], [3].
To reduce the complexity of the large data set with 914
pick-up locations, we cluster the close pick-up locations into
125 clusters such that no two pick-up locations in a cluster
are farther than 500 meters apart. The drop-off location for
each ride is selected from the same set of clusters with equal
probability. Figure 5 shows the clustered pick-up locations
and the arrival rates for ride requests at each cluster is
(a) Ired = Q
(b) Ired = ∅
Fig. 6: Environment with three drivers and a pick-up location v. (a)
The red car has full information on the position of the other drivers,
therefore, Sred(v, qred) = 2 , (b) The positions of the other drivers
are not known to the driver of red car, thus, Sred(v, qred) = 0.
then the profit of a ride with pick-off at v and drop-off at w
for driver i located at u is σ(cid:48)c(w, v) − σc(u, v).
Now we define the expected profit of driver i located at
u and working for Bi period of time as follows:
(cid:88)
pd(wv)(cid:2)pi(v, u)(cid:0) max{0, σ(cid:48)c(w, v) (9)
− σc(u, v) + Vi(w, Bi − c(u, v) − c(v, w))}(cid:1)(cid:3),
v,w∈V
Vi(u, Bi) =
where Vi(u, 0) = 0 for all u ∈ V , drivers i ∈ [m] and
Ii ⊆ Q. Note that, Vi(w, Bi−c(u, v)−c(v, w)) represents the
expected profit of driver i after drop-off at w. The conditional
probability of the drop-off location is the common knowledge
of the drivers and the ride-sourcing companies based on prior
customer data.
Intuitively, the expected profit in Equation (9) represents
the total expected profit of servicing requests with pick-up
location at v and drop-offs at w only if the
σ(cid:48)c(w, v) − σc(u, v) + Vi(w, Bi − c(u, v) − c(v, w)) ≥ 0.
Observe that finding Vi(u, Bi) for all u ∈ V and a given Bi
is performed in polynomial time, however, since the drivers
are not going through the calculation of Vi(u, Bi) for all
u ∈ V , we assume that they have access to the expected
profit of the vertices by experience.
Since the calculation of the expected profit for drivers
for each time step is computationally expensive, we trained
a Random Forest Regressor [22] implemented by [23] to
approximate the values of V for each number of vehicles
in the system with training data over 10000 instances with
work-day Bi of 5 average length rides, fare σ(cid:48) = $1.06 per
mile and driving cost σ = $0.3 [24].
B. Partial Information Sharing
Figure 7 shows the percentage improvement in the ex-
pected response time for different number of vehicles using
the partial information sharing control method of Section III
in the two scenarios. Observe that as the number of drivers
increases, the average improvement in the expected response
time increases. However, with a large number of drivers
randomly placed in the environment, the expected response
time decreases and the possibility to further optimize it with
information sharing is limited. On the other hand, in the
jammed scenario where the drivers are concentrated in an
Fig. 5: The set of pick-up and drop off locations in Manhattan N.Y.
The bar at the location of each cluster represents the ride request
arrival rate.
represented with a bar. The performance of the proposed
control methods are evaluated in two scenarios: 1) the initial
location of the drivers are selected uniformly randomly, and
2) jammed scenario where the drivers are initialized at 20
closest locations to the Rockefeller center in Manhattan.
Observe that the proposed algorithms to find the controls
are applicable to various driver models for V. In the follow-
ing section, we propose a behavior model V for the drivers.
A. Drivers' model
A driver model is a process of evaluating the expected
profit of different locations at each time instance. Prior to
introducing our driver model, we provide the parameters in
the proposed model.
The pick-up probability of a location v as seen by driver i
is the probability that a ride-request at v is assigned to driver
i before any other ride request. Consider a configuration Q
of the drivers and a vertex u, where driver i is the kth
closest driver to the vertex location u. Since a request is
assigned to the closest available driver, then driver i can
expect to be assigned to the kth request arriving at u. Let
Si(v, u) denote the number of drivers closer to v than driver
i located at vertex u. Note that Si(v, u) is a function of
the information of driver i regarding the position of other
drivers. Figure 6 illustrates an instance with three vehicles
and a pick-up location. Given the full information of the
vehicle positions, Sred at pick-up location is 2 and Sred is
zero if Ired = ∅.
The probability that a request at v is assigned to driver i
located at u as expected by driver i before any other request
is approximated by
Si(v,u)+Si(w,u)+1(cid:88)
(cid:1)k(cid:0)
(cid:18)(cid:80)
(cid:1)Si(v,u)+Si(w,u)+1−k
k=Si(v,u)+1
λw
k
.
w∈V Si(w, u) + 1
(cid:19)
pi(v, u) = Πw∈V
(cid:0)
λv
λv + λw
λv + λw
Let σ(cid:48) be the fare per unit time of servicing a request,
Fig. 7: Improvement in expected response time by sharing location
information.
Fig. 8: The expected response time of a system of 20 drivers
executing 100 ride requests arriving over time under the partial
information control method. The drivers are initialized under the
jammed scenario.
area, the information sharing method improves the expected
response time by 10% on average. The information sharing
method shares information on the position of the other
drivers with 31.0% and 41.7% of the drivers in the random
initial configuration and jammed scenarios, respectively. The
results are the average of 1000 instances for different number
of drivers and scenarios. The boxes show the first, first and
third quartiles of each set of experiments. The expected
response time of the solution obtained from the LP-rounding
algorithm of Section III on this set of experiments is within
0.014% on average of the solution of the LP relaxation of the
information sharing problem. The maximum deviation from
the optimal solution of the LP relaxation is 0.42%.
Figure 8 shows the expected response time of a set of
20 drivers responding to 100 requests arriving over time
with the jammed initial configuration. In this experiment,
the maximum arrival rate on the vertices is 0.03 per minute,
therefore, we assumed that there exist enough time to relo-
cate between the ride request arrivals. Note that the partial
information method maintains a low expected response time
over the course of responding to 100 requests compared to
the same set of drivers with no control input on their waiting
locations. The results are an average of 100 experiments with
100 randomly generated requests for each experiment. The
lines represent the average and shaded areas represent the
first and third quartiles.
C. Pay to Control
In this section, we evaluate the performance of the PAY-
TO-CONTROL method for the two scenarios. Figure 9 il-
lustrates the improvement in the expected response (solid
lines) time and the total amount paid to the drivers (dashed
lines) to relocate for different β values in the random initial
configuration scenario. The shaded area represents the first
and third quartiles of 1000 random instances for each number
of vehicles. The amount paid is proportional to the average
cost of riding UberXL, i.e., $0.3 per minute [24]. For larger
β, the expected response time is more important than the
amount paid for relocation. Therefore, with a larger number
of vehicles, the PAY-TO-CONTROL method increases the
amount paid to the drivers to minimize the expected response
time. Notice that with β = 10 the expected response time
Fig. 9: The percentage improvement in expected response time and
the incentive pay. The solid lines represent the average improvement
in the expected response time and the dashed lines represent the
total amount paid to the drivers.
has improved by 25% for $1 per driver.
Next, we evaluate the performance of the pay-to-control
algorithm in the jammed scenario. Figure 10 shows that with
a larger number of vehicles concentrated in a small area,
the pay-to-control algorithm improves the expected response
time significantly with limited amounts paid to the drivers.
Notice that with β = 10 the expected response time has
improved by 70% for $2 per driver.
Figure 11 shows the expected response time and the
amount paid to a set of 20 drivers responding to 100 requests
arriving over time with the jammed initial configuration.
Notice that the PAY-TO-CONTROL method with β = 1
maintains a low expected response time over the course of
responding to 100 requests compared to the same set of
drivers with no control input on their waiting locations. The
total amount paid to the drivers over the course of responding
to 100 requests is $1.87 per request. The results are an
average of 100 experiments with 100 randomly generated
requests for each experiment. The lines represent the average
and shaded areas represent the first and third quartiles.
51015202530NumberofVehicles0510152025PercentageImprovementinResponseTimeRandominitialconfigurationJammedscenario020406080100NumberofRequests350400450500550600650ExpectedResponseTimePartialInformationShareNocontrol10203040NumberofDrivers05101520253035PercentageImprovementinResponseTimeβ=1.0β=10.0102030405060AmountPaidtoDrivers($) [4] W. Zhang, S. Guhathakurta, J. Fang, and G. Zhang, "The performance
and benefits of a shared autonomous vehicles based dynamic rideshar-
ing system: An agent-based simulation approach," in Transportation
Research Board 94th Annual Meeting, no. 15-2919, 2015.
[5] M. Hyland and H. S. Mahmassani, "Dynamic autonomous vehicle fleet
operations: Optimization-based strategies to assign AVs to immediate
traveler demand requests," Transportation Research Part C: Emerging
Technologies, vol. 92, pp. 278 -- 297, 2018.
[6] M. Chang, D. S. Hochbaum, Q. Spaen, and M. Velednitsky, "DIS-
PATCH: an optimal algorithm for online perfect bipartite matching
with i.i.d. arrivals," CoRR, vol. abs/1805.02014, 2018.
[7] M. Maciejewski, J. Bischoff, and K. Nagel, "An assignment-based
approach to efficient real-time city-scale taxi dispatching," IEEE
Intelligent Systems, vol. 31, no. 1, pp. 68 -- 77, 2016.
[8] Uber. Driving with Uber, wait less, earn more. [Online]. Available:
https://www.uber.com/info/get-trips-without-waiting/
[9] M. Pavone, S. L. Smith, E. Frazzoli, and D. Rus, "Robotic load
balancing for mobility-on-demand systems," The International Journal
of Robotics Research, vol. 31, no. 7, pp. 839 -- 854, 2012.
[10] M. Tsao, R. Iglesias, and M. Pavone, "Stochastic model predic-
tive control for autonomous mobility on demand," arXiv preprint
arXiv:1804.11074, 2018.
[11] G. C. Calafiore, C. Novara, F. Portigliotti, and A. Rizzo, "A flow
optimization approach for the rebalancing of mobility on demand
systems," in IEEE International Conference on Decision and Control,
2017, pp. 5684 -- 5689.
[12] D. B. Shmoys, "Approximation algorithms for facility location prob-
lems," in International Workshop on Approximation Algorithms for
Combinatorial Optimization. Springer, 2000, pp. 27 -- 32.
[13] V. Arya, N. Garg, R. Khandekar, A. Meyerson, K. Munagala, and
V. Pandit, "Local search heuristics for k-median and facility location
problems," SIAM Journal on computing, vol. 33, no. 3, pp. 544 -- 562,
2004.
[14] E. D. Demaine, M. Hajiaghayi, H. Mahini, A. S. Sayedi-Roshkhar,
S. Oveisgharan, and M. Zadimoghaddam, "Minimizing movement,"
ACM Transactions on Algorithms (TALG), vol. 5, no. 3, p. 30, 2009.
[15] A. Sadeghi and S. L. Smith, "Re-deployment algorithms for multiple
service robots to optimize task response," in IEEE International
Conference on Robotics and Automation, 2018, pp. 2356 -- 2363.
[16] S. Bandyapadhyay, A. Banik, S. Das, and H. Sarkar, "Voronoi game on
graphs," Theoretical Computer Science, vol. 562, pp. 270 -- 282, 2015.
[17] R. Salhab, J. Le Ny, and R. P. Malham´e, "A dynamic ride-sourcing
game with many drivers," in 55th Annual Allerton Conference on
Communication, Control, and Computing, 2017, pp. 770 -- 775.
[18] R. Schuler, "An algorithm for the satisfiability problem of formulas in
conjunctive normal form," Journal of Algorithms, vol. 54, no. 1, pp.
40 -- 44, 2005.
[19] T. Basar and G. J. Olsder, Dynamic noncooperative game theory.
Siam, 1999, vol. 23.
[23] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Van-
derplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and
E. Duchesnay, "Scikit-learn: Machine learning in Python," Journal
of Machine Learning Research, vol. 12, pp. 2825 -- 2830, 2011.
[24] Ride sharing driver. How much does uber cost? Uber fare estimator.
[Online]. Available: https://www.ridesharingdriver.com/
[20] S. Ahmadian, Z. Friggstad, and C. Swamy, "Local-search based
approximation algorithms for mobile facility location problems," in
Proceedings of the twenty-fourth annual ACM-SIAM symposium on
Discrete algorithms. SIAM, 2013, pp. 1607 -- 1621.
[21] (2014) Uber TLC FOIL response.
[Online]. Available: https:
//github.com/fivethirtyeight/uber-tlc-foil-response
[22] L. Breiman, "Random forests," Machine learning, vol. 45, no. 1, pp.
5 -- 32, 2001.
Fig. 10: The percentage improvement in the expected response
time and the incentive pay in the jammed scenario. The solid lines
represent the average improvement in the expected response time
and the dashed lines represent the total amount paid to the drivers.
Fig. 11: The expected response time and the paid amount to a
system of 20 drivers executing 100 ride requests arriving over time
under the PAY-TO-CONTROL method. The drivers are initialized
under the jammed scenario.
VI. CONCLUSION
This paper considered the problem of controlling self-
interested drivers in ride-sourcing applications. Two indirect
control methods were proposed and for each, a near-optimal
algorithm was presented. The extensive results show signif-
icant improvement in the expected response time on real-
world ride-sourcing data. In addition, we hope to extend the
results to capture vehicles with different capacities and ride-
sharing applications.
REFERENCES
[1] L. Rayle, S. Shaheen, N. Chan, D. Dai, and R. Cervero, "App-
based, on-demand ride services: Comparing taxi and ridesourcing
trips and user characteristics in san francisco university of california
transportation center (uctc)," University of California, Berkeley, United
States, 2014.
[2] N. Diakopoulos. How Uber surge pricing really works. [Online].
Available: https://www.washingtonpost.com/news/wonk/wp/2015/04/
17/how-uber-surge-pricing-really-works/
[3] A. Rosenblat and L. Stark, "Algorithmic labor and information asym-
metries: A case study of Ubers drivers," International Journal Of
Communication, 2016.
10203040NumberofDrivers01020304050607080PercentageImprovementinResponseTimeβ=1.0β=10.020406080AmountPaidtoDrivers($)020406080100NumberofRequests300400500600700ExpectedResponseTimePay-to-ControlNocontrolPaidamount05101520AmountPaidtoDrivers($) |
1705.01453 | 2 | 1705 | 2017-05-16T08:43:52 | Distributed Proportional-Fairness Control in MicroGrids via Blockchain Smart Contracts | [
"cs.MA"
] | Residential microgrids (MGs) may host a large number of Distributed Energy Resources (DERs). The strategy that maximizes the revenue for each individual DER is the one in which the DER operates at capacity, injecting all available power into the grid. However, when the DER penetration is high and the consumption low, this strategy may lead to power surplus that causes voltage increase over recommended limits. In order to create incentives for the DER to operate below capacity, we propose a proportional-fairness control strategy in which (i) a subset of DERs decrease their own power output, sacrificing the individual revenue, and (ii) the DERs in the subset are dynamically selected based on the record of their control history. The trustworthy implementation of the scheme is carried out through a custom-designed blockchain mechanism that maintains a distributed database trusted by all DERs. In particular, the blockchain is used to stipulate and store a smart contract that enforces proportional fairness. The simulation results verify the potential of the proposed framework. | cs.MA | cs | Distributed Proportional-Fairness Control in
MicroGrids via Blockchain Smart Contracts
Pietro Danzi, Marko Angjelichinoski, Cedomir Stefanovi´c, Petar Popovski
Department of Electronic Systems, Aalborg University, Denmark
Email: {pid,maa,cs,petarp}@es.aau.dk
7
1
0
2
y
a
M
6
1
]
A
M
.
s
c
[
2
v
3
5
4
1
0
.
5
0
7
1
:
v
i
X
r
a
Abstract-Residential microgrids (MGs) may host a large
number of Distributed Energy Resources (DERs). The strategy
that maximizes the revenue for each individual DER is the one in
which the DER operates at capacity, injecting all available power
into the grid. However, when the DER penetration is high and
the consumption low, this strategy may lead to power surplus
that causes voltage increase over recommended limits. In order
to create incentives for the DER to operate below capacity, we
propose a proportional-fairness control strategy in which (i) a
subset of DERs decrease their own power output, sacrificing
the individual revenue, and (ii) the DERs in the subset are
dynamically selected based on the record of their control history.
The trustworthy implementation of the scheme is carried out
through a custom-designed blockchain mechanism that maintains
a distributed database trusted by all DERs. In particular, the
blockchain is used to stipulate and store a smart contract that
enforces proportional fairness. The simulation results verify the
potential of the proposed framework.
I. INTRODUCTION
The deployment of Distributed Energy Resources (DERs)
in the residential low voltage (LV) microgrids (MGs) aims to
improve their self-sustainability and reduce the transmission
losses [1]. DERs based on renewable resources, such as solar
photovoltaic (PV), are traditionally operated at capacity to
inject all available power in the grid and thus maximize the
efficiency, regardless of the grid state. However, the variability
of the capacity may cause erratic voltage behavior on the dis-
tribution feeders. In particular, grids with high penetration of
PVs may experience voltage increase over the recommended
levels, e.g., during afternoons, when the production is high
and the household consumption is low [2]. An approach to
prevent grid instability due to dramatic voltage increases is
to control the power output of DERs, e.g., via active power
curtailment or reactive power adjustment [3]. In this respect,
[4] proposes a control strategy based on the principle of
proportional fairness, where all DERs equally contribute to
voltage regulation all the time; the strategy is executed by
the remote central authority such as the distribution system
operator (DSO). However, enforcing that all DERs participate
in voltage control and curtail their output power all the time
is characterized with a control complexity that increases with
the number of DERs and a reduced possibility for the owners
to operate an economic strategy. Moreover, existence of the
centralized authority involves the issue of Single Point of
Failure (SPoF).
In this paper, we propose a novel control scheme based on
the proportional fairness in which (i) only a subset of DERs
act as voltage regulators and curtail their individual power
outputs over control periods, and (ii) ensures that, in the long
term, DERs participation in voltage regulation is balanced. To
give the DERs the incentive to fairly participate in voltage
regulation, we introduce a principle based on exchange of
credits. Specifically, in order to join the regulating subset, DER
asks for a credit, which will be paid by the DERs that are
not in the regulating subset and therefore operate at their full
capacity. In turn, the decrease of credit status of the DERs that
have not participated in voltage regulation ultimately forces
them to participate in the voltage regulation in future.
In order to avoid existence of a central authority and,
thus, a SPoF,
the proposed protocol runs in a distributed
manner. The credit statuses of all DERs are tracked by the
use of a blockchain protocol, initially introduced with Bitcoin
cryptocurrency [5]. Its major advantage is the capability to
implement a distributed database that serves as record of
the system's state and of its history, trusted by all agents,
in this case DERs. All agents store identical copies of the
database. The database is hard to tamper with, since an agent
can add new record to it only if a proof-of-work (POW)
is obtained, where POW is the solution of a computational
puzzle that requires an investment of electrical energy to
run the computation. In particular, the blockchain is used
to memorize (i) the state of smart contracts [6], which are
computer programs that can receive, store and pay credit, i.e.,
cryptocurrency, and (ii) the credit history of agents. In the
proposed framework, a smart contract acts as a trustworthy
distrubuted control authority, realized via a custom-designed
blockchain mechanism operated by all DERs. Specifically,
DERs that are installed on the same distribution feeder stip-
ulate a smart contract among them, determining which units
will act as voltage regulator over control periods, based on
their available credit statuses and the economic strategy that
they individually adopt. In the long term, this ensures a fair
rotation in the participation of the DERs to the MG regulation.
The rest of the paper is organized as follows. Section II
introduces the system model and the proportional-fairness
control strategy assuming a centralized setup. Section III
introduces the main concepts of a blockchain protocol. Sec-
tion IV presents the distributed, blockchain-based architecture
fostering the proportional-fairness control. Section V contains
a case study based on a simulation of a power system and
an instance of blockchain protocol, verifying the proposed
framework. Section VI concludes this paper.
Fig. 1. The system model.
Fig. 2. Message exchanges between the DERs and the control authority.
max is the maximum tolerable apparent power.1 The
where s2
droop control law (1) clearly shows that, in VSC mode, the
active power of DER u is below capacity, leading to revenue
loss of the DER owner. Nevertheless, the presence of VSC
units is an imperative for voltage regulation of the MG.
II. SYSTEM MODEL
B. VSC election per MG based on proportional fairness
(cid:80)NMG
We consider a LV distribution grid (LVDG) composed of
NMG alternate-current MGs. A MG f hosts Uf DERs, Uf ≥
0, which are PV generators that supply residential loads, see
Fig. 1. The total number of DERs in the LVDG is Utotal =
f =1 Uf . All MGs jointly strive to maintain the LVDG grid
voltage amplitude within acceptable limits [7]. To do so, in
each MG there is a dedicated voltage regulator elected from
the local DERs. In the rest of this section, we first describe
the voltage regulation mechanism. Afterwards, we introduce
the regulator election strategy that each MG employs.
A. Voltage regulation
Assume MG f with Uf DERs indexed in the set U =
{1, 2, ..., Uf}. Each DER is connected to the MG via a power
electronic converter (PEC) that controls its output and that
supports dual mode capability [8], i.e., a PEC can operate in
current source converter (CSC) or voltage source converter
(VSC) mode. In CSC mode, DER u is operated at capacity,
outputting all available power gu using maximum power point
tracking algorithm, and is not capable of regulating the voltage
[7]. In VSC mode, DER u acts as the voltage regulator at the
expense of the reduced active output power pu, pu ≤ gu.
The voltage is regulated via the active output power using the
following droop control law [9]:
vu = vref − γ(gu − pu),
(1)
where vu and vref are the voltage at the output of the DER
and the reference voltage of the MG, respectively, while γ
is the droop parameter, chosen such that (i) vu is maintained
within tolerable limits vmin and vmax, and (ii) all VSCs in the
LVDG achieve proportional power sharing. The output reactive
power qu is determined via the active output power, subject
to a constraint on the apparent power, as follows:
qu =(cid:112)s2
max − p2
u,
(2)
We first formulate the centralized version of the VSC
election strategy, employed in each MG, and enabled by
an external control authority, e.g.,
the DSO, see Fig. 2.
The strategy runs periodically every T tc seconds, where the
interval between two consecutive VSC assignments is referred
to as control period, and is sufficiently long to allow reliable
exchange of messages among DERs and the authority. Let
µu(k) denote the operating mode of DER u in arbitrary control
period k. Hence, µu(k) is a binary variable:
µu(k) =
0
1
if DER u operates as CSC in period k,
if DER u operates as VSC in period k.
(cid:40)
(3)
Denote by µu the sequence of operating modes of DER u in
all periods up to interval K, i.e., µu = [µu(0), . . . , µu(K−1)].
Stacking µu, ∀u ∈ U, vertically, we form the Uf × K matrix
of control mode histories of all DERs in the MG up to period
K, denoted by M, where [M]u,k = µu(k). The k-th column
of M represents the operating modes of all DERs in the MG
in control period k, and is denoted by µ(k). The objective of
the VSC election strategy is to ensure fairness among DERs
in the MG, i.e., all DERs should equally participate in voltage
control over time. This can be expressed as follows:
M 1K =
K
Uf
1Uf ,
(4)
where 1Uf is the all-ones vector of length Uf . In addition, the
strategy should also ensure that only one DER per MG acts
as VSC in each control period, i.e.:
1T
Uf
µ(k) = 1,
(5)
with (·)T denoting the transpose operator.
DERs are given incentive to participate in voltage regulation
through acquisition of credit when they operate in VSC
mode. Denote by cu(k) the credit status of DER u in the
1We note that more sophisticated control schemes, based on both active
and reactive power adjustments, can be adopted for the voltage regulation.
However, they are beyond the scope of the paper.
Power plantMedium VoltageLow VoltageTransformer(PCC)𝑓1 𝑓2 𝑓3 𝑓4 𝑓5 𝑓6 Aggregated residential loadsDER𝑓7 Control authorityDERs...𝑤1(𝑘) 𝑤𝑈𝑓(𝑘) 1 𝑐1 𝑘 , µ1(𝑘) 𝑈𝑓 FeederFeeder...𝑓𝑁𝑀𝐺 𝑓1 µ𝑈𝑓(𝑘) 𝑐𝑈𝑓 𝑘 , (a)
(b)
(c)
(d)
Fig. 3. Different phases of blockchain protocol: (a) contract updates are propagated in the network, (b) contract updates are accumulated in the miners'
buffers, (c) a new block (red packet) that includes the updates is generated and propagated and (d) the new block is accepted by the entire network.
(cid:80)Uf
control period k and let c(k) = [c1(k), . . . , cUf (k)]. In the
proposed scheme, the total available credit is constant, i.e.,
u=1 cu(k) = C, ∀k, and the credit is only redistributed
among DERs, ultimately forcing them to act as VSCs.
The redistribution of the credit is performed in the following
way. In period k, DER u sends to the control authority the
credit demand wu(k) asked for performing the role of VSC
for the MG in the next period k + 1. The credit demand is
chosen in the interval [0, cu(k)], according to the individual
strategy adopted by DER u. For the sake of simplicity, in the
rest of the paper we assume that DER u ∈ U chooses wu(k),
∀k, randomly using uniform distribution:
wu(k) ∼ unif (0, cu(k)) .
(6)
The control authority stores the information sent by all DERs
in the MG in the vector w(k) = [w1(k), . . . , wUf (k)], and
chooses the DER with the lowest credit demand to operate as
VSC in the next control period:
u = argminuw(k).
(7)
Then, the control authority sends the messages about (i) the
control modes of all the DERs in the MG in the next control
period µ(k+1) = eu, where eu is a vector with 1 at position
u and 0 elsewhere, and (ii) the update of the total credit
(cid:16)
w(k)(cid:17)
c(k+1) = c(k) − w(k) +
1T
Uf
eu,
(8)
see Fig. 2. A careful inspection of (7) shows that the total
credit
is redistributed by (i) taking from the DERs their
respective credit demands and (ii) giving the total demanded
credit to the DER u that had the lowest demand and that will
take the VSC role in the next period.
The rule for choosing the credit demand (6) and the VSC
election rule (7) show that DERs with lower credit status have
higher chances of operating as VSC. On the other hand, low
credit status implies that the DER has previously infrequently
operated as VSC. In this way, the proportional fairness is
promoted among DERs in the MG. We also note that the
results provided in Section V verify that the objective (4)
becomes satisfied as the number of control period increases.
Finally,
To run the VSC election strategy, the LVDG requires a
communication network to interconnect the DERs with the
central authority to support the message exchange depicted in
Fig. 2. This can be implemented using wireless (e.g., cellular)
network, or wired network (e.g., power line communications).
the election strategy (7) and the credit update
rule (8) require presence of the central authority's database
that records the current credit status c(k) and that is trusted
by all agents. This need for the establishment of a trustful
relation among agents prevents the implementation of simple
decentralized control systems, such as token rings, in which
DERs hold the control in turns. Motivated by this insight, we
develop a decentralized control solution based on blockchain
protocol, which establishes a trustful distributed record both
of the credit status of all DERs and of the control history.
III. BLOCKCHAIN PROTOCOL
In this section, we describe a version of the blockchain pro-
tocol that realizes only the functionalities required to support
the proposed control strategy. An introduction to the general
variant of the blockchain can be found in [10].
A blockchain is a distributed database consisting of identical
copies that are stored in the memory of each agent. It is
organized as a concatenated list of blocks that can be expanded
by any agent with new blocks. Each block stores a set of smart
contracts updates. A smart contract is a computer program
that can be executed by an agent that has the blockchain
software. As the agent executes the program locally, whenever
it modifies the contract internal state, e.g., by exchanging
credit with it, the rest of the agents should be informed in
order to run the new version of the contract. The new state
is communicated through a smart contract update. To avoid
proliferation of different states of the contract, the blockchain
in parallel executes the update verification process via block
generation, ensuring that only verified updates are included in
the blockchain. We proceed by providing the details.
Fig. 3 illustrates the blockchain operation via an example
of a peer-to-peer network of interconnected agents. Every
agent is provided with a local copy of the blockchain, and
a buffer, named "mempool". When smart contract updates
MempoolBlockchain local copySmart contract updateBlock are produced, they are sent to the neighbors, Fig. 3(a). The
updates are temporary stored in mempools, Fig. 3(b), being
not considered valid yet. In parallel, agents are also generating
blocks, i.e., working to solve the computational puzzle (by
the blockchain protocol) and obtain POW. When an agent
obtains POW and generates a new block, it fills it with the
smart contract updates present in its mempool and transmits
the new block to its peers, see Fig. 3(c). Upon the reception
of the block, a neighbouring agent verifies that it contains a
valid POW, by checking the provided solution of the puzzle.
If the verification succeeds, the agent adds the block to its
blockchain, Fig. 3(d), and propagates it further to its neighbors.
The verification and propagation of successfully verified block
continues until all agents in the network are reached. Finally,
when all agents have received and verified the block, they have
the same updated version of the blockchain, see Fig. 3(d), and
thus, their copies of the smart contract have the same internal
state. The agents also remove from their mempools the smart
contract updates included in the received and verified blocks.
The smart contract can be abstracted as a virtual agent that
interacts with the actual agents according to the logic defined
in its program, and in this way regulating credit redistribution
among the actual agents through the contract updates. In the
context of the proposed framework, the smart contract updates
are produced by the MG control application, while the block
creation process is intrinsic to the blockchain and in charge
of keeping it consistent. These two process run independently,
where the latter ensures that the smart contract updates are
eventually stored in the blockchain.
A. Blockchain consistency
As depicted in Fig. 3(c)–(d), the blockchain consistency is
guaranteed through propagation of newly generated blocks
to all other agents. In order to avoid the generation of
uncontrolled amounts of blocks, a block generation requires
an investment of resources. In the seminal paper [5],
the
investment is the consumption of electrical power required to
solve a computational puzzle, known as POW. This process of
obtaining POW is named mining; the agent that first solves the
POW decides the content of the next block. The probability
pb to generate a new block before the other agents depends on
the computational resources allocated to solve the puzzle. The
difficulty of the puzzle is tuned to keep the block generation
rate constant over time, see [5] for details.
It may happen that two or more new valid blocks are created
concurrently by different agents, propagating in different sub-
graphs of the network. In this case, the network is split over
different, but valid versions of the blockchain. The contention
rule adopted by Bitcoin is "the longest chain wins", i.e., the
split is solved when some "newer" valid block is generated,
received and accepted by all agents, updating the blockchain
and making it consistent again [10].
B. The cost of mining
In standard blockchain realizations, the resource investment
in mining operation is compensated by attributing some credit,
i.e., cryptocurrency, to the miners, either by (i) increasing
the credit of the miner that generates the block, incrementing
the total amount of credit in the system, or (ii) letting the
miner demand a credit fee for each smart contract update
that it includes in a block. This stimulates miners to consume
electrical power to obtain POW, and thus reduces the risk that
a malicious miner may impose his version of the blockchain.
In this work, the miners are the DERs in the MG, i.e., the
blockchain is private, but they are not rewarded with credit
for mining. Nevertheless, a DER is stimulated to mine, in
order to ensure that the blockchain is consistent with its view
of the credit status and control history. Also, not providing
credit for mining eliminates the possibility of DERs being
more motivated in mining than in voltage regulation.
C. Peer-to-peer network and protocol messages
The blockchain is based on peer-to-peer networking, where
each agent connects to a set of randomly selected neighbors.2
The communication protocol adopts TCP connections to ex-
change two types of information, as shown in Fig. 3.
1) The smart contract updates, Fig. 3(a).
2) The blocks, Fig. 3(c).
D. Security issues of blockchain protocol
The blockchain protocol is founded on the assumption that
the resources required to generate blocks are well distributed
among agents, making it hard for a single agent
to con-
secutively generate blocks and and impose its view of the
blockchain that may potentially include false data [10]. This
assumption can be adopted in the scenario considered in the
paper, as it is expected that microcontrollers of the DERs have
similar hardware characteristics.
Security threats to blockchain may also come from peer-to-
peer networking. A prominent example is the eclipse attack
[11], in which all neighbors of an agent are under the control
of a malicious agent. In this case, the copy of the blockchain
of the attacked agent may be compromised. To avoid such
scenario, the communication peers are selected randomly.
IV. THE PROPOSED DISTRIBUTED CONTROL FRAMEWORK
The proposed solution is enabled by a blockchain protocol
(i.e., software) implemented by the agents,
i.e., DERs of
the distribution grid. The mining process is done by agents
themselves, where the agents do not cooperate and the POW
is of low difficulty determined by their limited computational
capabilities. The agents have access to a peer-to-peer network,
where the access is granted only to legitimate agents. We
assume that the rate at which POWs are obtained and new
blocks propagated through the network is sufficiently high
to prevent overflow of the mempools. Finally, the message
propagation delay is assumed negligible compared to the
control period duration T tc.
A smart contract is deployed for each grid feeder, playing
the role of the central authority (i.e., DSO), resulting in a
is required that
the communication graph formed in this way is
2It
connected, but the related details are beyond the scope of the paper.
(a) Credit sending.
(b) Credit receiving.
Fig. 4. Relationship among DERs in one feeder and their smart contract.
unique blockchain that stores NM G contracts. The relationship
of DERs in a feeder and their smart contract is depicted
in Fig. 4. During a control period, each DERs chooses its
credit demand for the next period via (6), and updates its
local copy of the smart contract by sending, i.e., transferring,
the demanded amount to it, see Fig. 4(a). These updates are
propagated to peers, Fig. 3(a)–(b), and the block generation
process, Fig. 3(c), ensures that all the copies of the smart
contract have the consistent knowledge of them, Fig. 3(d). The
credit sending is disabled during the final part of the period via
the contract locking, to ensure the consistency of local copies
at the control actuation instant, i.e., at the beginning of the
new period. At this point, based on its internal state, the smart
contract (i.e., each its copy) elects the (same) VSC for the
next control period using (7), and DERs are locally notified
(e.g., via reading the state of the smart contract) about their
operating mode for the next control period. In the next round,
the elected VSC updates its local smart contract to receive the
total demanded credit, see Fig. 4(b). This triggers a new round
of the update propagation, Fig. 3(a)–(b), which is eventually
stored in the blockchain after a block embedding it becomes
generated and propagated through the network, Fig. 3(c)–(d).
For the sake of completeness, in Fig. 5(a) we expose the
operating sequence of the framework related to a single control
period,
the contract updates
related to period k start in period k − 1:3
1) Credit sending: During period k − 1, each DER transfers
to the smart contract the credit demanded to operate as
VSC in the next period (represented by green arrows).
These updates are propagated through the network and
stored in the mempools, and gradually included in the
blockchain when new blocks are generated.
2) Contract lock and VSC election: At a predefined instant
before the end of period k − 1, all DERs modify the
state of the contract to lock it (purple arrow).4 The VSC
for control period k is elected via (7), uniquely over all
i.e., period k, assuming that
3In general, the framework can be operated such that the contract updates
for period k are started in a period k1, where k1 ≤ k − 1, allowing to solve
the eventual blockchain inconsistencies due to propagation and mining delays.
The related analysis is beyond the paper scope.
4Note that the contract is effectively locked only by the first contract update
that is included in the blockchain, which invalidates the consecutive locks.
(a) Blockchain-based scheme: the messages are exchanged among DERs.
(b) Centralized scheme: The messages are sent from DERs to the central
authority (uplink) and from the central authority to DERs (downlink).
Fig. 5. Sequence of the communication exchanges (a) in the blockchain-based
and (b) in the centralized scheme.
copies of the smart contract. All DERs set their control
mode accordingly.
3) Credit receiving: The VSC DER withdraws the credit
obtained for the control period k from the smart contract
(represented by the red arrow). To do this, it modifies
the state of the contract, communicates the state update
to the other agents, and waits for it to be included in a
newly mined block. The credit stored by the contract can
only be withdrawn by the DER operating as VSC.5
The figure also depicts the block generation process (repre-
sented by blue crosses) that is decoupled from the process
of contract updates. It may happen that some newly generated
blocks do not contain any updates and such blocks only verify
the consistency of the current state of the blockchain.
In the rest of this section, we outline several important
aspects of the framework.
A. Plug-and-play feature
The blockchain-enabled solution provides a smooth plug-
and-play since the joining DER just needs to start interacting
with the contract. Observe that it is likely that the joining DER
will be chosen as VSC in the next control periods, as its credit
status is zero.
B. Control availability and trustworthiness
The decentralized solution does not suffer from the SPoF,
providing an increased control availability with respect
to
the centralized architecture. It also avoids reliance on data
received by an external authority, which can be tampered
without being detected by the MG agents. On the other hand,
the major weakness of the blockchain-based architecture is the
modification of the database operated by a subset of agents
that are capable to generate new blocks faster, imposing their
version of the control history, as discussed in Section III-D.
C. Communication cost
Fig. 5 compares the communication exchanges in the
blockchain-based scheme, Fig. 5(a), with the exchanges in
5All contract updates are certified with public-key cryptography, cf. [5].
Update contract with credit demandsSmart contract:Collect 𝐰.Elect VSC.DER 1:Choose 𝑤#DER 𝑈%:Choose 𝑤&'. . .Update contract to receive total demanded creditSmart contractVSC DERxxxxx2) Contract lockNew blocks generated3) Credit receiving1) Credit sendingControl period k-1Control period kDecisionUplinkDownlinkActuationControl period k-1Control period k Fig. 6. Comparison of communication costs.
the centralized one, Fig. 5(b). We proceed by evaluating and
comparing the communication costs of both approaches, where
the communication cost is expressed via the average amount
of data exchanged per agent during a control period.
In the blockchain-based scheme, data is generated by the
processes of smart contract updates and blocks generation. The
number of blocks generated during a control period, denoted
by Nb, is determined by the protocol. The communication cost
for agent u with N peers is:
(a) Without control.
(b) With the control.
J bc = N Lu + J rc + pbNbN Lb + J rb,
(9)
Fig. 7. The voltage measured on the LV feeders on the dataset corresponding
to 9th of June, indicated in the Per-Unit (PU) base.
i.e., it comprises the costs of propagating its own transaction,
expressed through the message length Lu, the relaying of
other's transactions J rc, the transmission of its own blocks and
the relaying of other's blocks J rb. For the sake of comparison,
we provide a lower bound on J bc for the simplified case
where a DER has just a single peer and does not relay data6,
assuming that the computational power is equally distributed
among DERs, i.e., pb = 1/Utotal:
J bc
LB = Lu +
NbLb
Utotal
.
(10)
In the centralized architecture, we denote the length of
messages wu, cu, µu in bits as Lw, Lc, Lµ. In this case, the
communication cost per DER is equal to J c = Lw + Lc + Lµ,
i.e., it is a constant. We compare the communication costs
in Fig. 6, assuming that Lw, Lµ, Lc are 64 bits long, Lu
is 800 bits and Lb is 8000 bits. The block period is set
to 10 s, which together with T tc = 15 minute provides
Nb = 90 blocks for control period. The communication cost
in the blockchain-based scheme is slightly decreasing with
the number of DERs, i.e., Utotal, as this reduces the number
of blocks that they individually generate in a control period.
However, the centralized solution is clearly less demanding
due to its simpler communication architecture.
V. CASE STUDY
Inspired by the MG LV scenario in [2], we adopt the power
system composed by LV and MV MGs depicted in Fig. 1 as
a case study. The LV MG is connected to a medium voltage
(MV) microgrid in which a 6 MW solar PV power plant is
installed. The power exchange between the LV and MV MGs
takes place via a on-load tap changing transformer (OLTC),
that constitutes the point of common coupling (PCC). The LV
MG is composed by NM G = 7 feeders, each supporting the
consumption of 10 households. On each feeder, except feeder
f1, there are installed 4 PV systems (i.e., DERs) with the rated
power of 4 kW, resulting with a total of Utotal = 24 PVs. For
the modeling of PVs, OLTC, households consumption, and the
other grid components, we adopt Disc framework [2].
We simulated the system in the Disc framework both for
the case without control and the case in which the proposed
control scheme is applied. When no control is adopted for
the PVs, in the afternoon hours we observe an overvoltage
on the feeders, which increases with their distance from the
transformer, see Fig. 7(a). Fig. 7(b) shows the proposed control
strategy7 provides a reduction of the over-voltages with respect
to the scenario without control. One can also observe that
the slope of the voltage profile in afternoon/morning hours
is reduced, which is another benefit of the control strategy.
Specifically, smoothing the voltage profile curve in systems
dominated by PV power gives more time and flexibility to the
bulk generation, which is characterized with high inertia and
slow transient ramp-up/ramp-down response, to respond to the
the power supply variations. The proposed control strategy was
6We note that the scenario with a single neighbor should be avoided, as in
this case a malicious neighbor may corrupt or hide the information [11].
7We used the optimized value of the droop parameter γ = 0.005, obtained
via simulations.
020406080100Number of DERs102103104105106JJbcLBJc05101520Time [hrs]0.80.911.11.2Voltages [PU]f1f2f3f4f5f6f705101520Time [hrs]0.80.911.11.2Voltages [PU]f1f2f3f4f5f6f7 Fig. 8. The voltage measured on the LV feeders over the month of June
of Disc simulator (i) without control, depicted with purple boxes, and (ii)
with proposed control, blue boxes. The plot shows minimum, quartile and
maximum values.
Fig. 9. Evolution of the fraction of time spent by each DER as VSC in a
feeder with 4 DERs.
also simulated over a period of one month and the reduction
of over voltages was also verified, see Fig. 8. Observe that the
under-voltages are not reduced, as their control is not included
in the proposed strategy. Nevertheless, we note that a similar
credit system can be employed for the load prioritization
during low production periods.
The verification of the proportional fairness objective (4) in
a centralized setting was performed via MATLAB simulations,
where the initial credit was randomly distributed among DERs,
i.e., the credit status vector c was randomly initialized, and the
control period duration was set T tc = 15 minutes. We per-
formed 1000 simulation runs, corresponding to 250 hours of
LVDG operation, and obtained that the DERs spend the equal
fraction of time operating as VSCs (the detailed presentation
of these results is omitted due to space constraints).
Finally, we turn to the blockchain-related aspects of the
proposed framework. The blockchain software installed in
the DERs controller supported a private Ethereum blockchain
[6], which provides the possibility of writing complex smart
contracts, and is simulated using EthereumJS testrpc [12].
We deployed NM G contracts on the blockchain, where DERs
interacted with the one corresponding to their feeder, and
the initial credit was randomly distributed among DERs. The
scripts that interface DERs with the blockchain implemented
the functionalities described in Section IV. We monitored
the output of the scripts and observed that the contract was
effectively reproducing the fairness objective, see Fig. 9.
Clearly, after a transient period caused by the initial credit
distribution, DERs operate as VSCs for equal fractions of time.
VI. CONCLUSIONS
The current research efforts in MG control are oriented
towards distributed schemes, requiring development of novel
protocols to enforce the security and the information trust-
worthiness among control agents. The blockchain protocol has
interesting properties that can be used to this end, resulting in a
novel design of multi-agent control systems. In this paper, we
developed proportional fairness MG control and established
a comparison between the standard centralized architecture
and blockchain-based one, verifying that the blokchain-based
solution can reproduce the control objectives of the centralized
architecture. We note that further investigation should be
conducted with respect to more complex control schemes.
limits of
We also outline two potential
the private
blockchain architecture in the context of MG control: the
mining cost and the communication cost. Specifically, alter-
natives to the energy inefficient POW have to be found in
order to enable private blockchains for systems with limited
hardware capabilities, such as MG components. Secondly, the
communication cost of the blockchain protocol is significantly
higher than the cost of the centralized one. The design of
blockchain protocol tailored for MG applications is part of
our ongoing research.
ACKNOWLEDGMENT
The work presented in this paper was supported in part by
the EU, under grant agreement no. 607774 "ADVANTAGE".
REFERENCES
[1] L. E. Zubieta, "Are Microgrids the Future of Energy?: DC Microgrids
from Concept to Demonstration to Deployment," IEEE Electrific. Mag.,
vol. 4, no. 2, pp. 37–44, June 2016.
[2] R. Pedersen, C. Sloth, G. B. Andresen, and R. Wisniewski, "Disc: a
simulation framework for distribution system voltage control," in Proc.
of IEEE European Control Conference (ECC), 2015, pp. 1056–1063.
[3] R. Tonkoski, L. A. Lopes, and T. H. El-Fouly, "Coordinated active power
curtailment of grid connected PV inverters for overvoltage prevention,"
IEEE Trans. Sustain. Energy, vol. 2, no. 2, pp. 139–147, 2011.
[4] R. Pedersen, C. Sloth, and R. Wisniewski, "Coordination of electrical
distribution grid voltage control - a fairness approach," in Proc. IEEE
Conference on Control Applications (CCA), 2016, pp. 291–296.
[5] S. Nakamoto, "Bitcoin: A peer-to-peer electronic cash system," [Online].
Available: https://bitcoin.org/bitcoin.pdf, 2008, accessed: 2017-03-21.
[6] G. Wood, "Ethereum: A secure decentralised generalised transaction
ledger," [Online]. Available: http://gavwood.com/paper.pdf, 2014, ac-
cessed: 2017-03-21.
[7] J. Rocabert, A. Luna, F. Blaabjerg, and P. Rodriguez, "Control of power
converters in ac microgrids," IEEE Trans. Power Electron., vol. 27,
no. 11, pp. 4734–4749, 2012.
[8] T. Dragicevi´c, J. M. Guerrero, J. C. Vasquez, and D. Skrlec, "Supervi-
sory control of an adaptive-droop regulated dc microgrid with battery
management capability," IEEE Trans. Power Electron., vol. 29, no. 2,
pp. 695–706, 2014.
[9] J. M. Guerrero, J. C. Vasquez, J. Matas, L. G. De Vicuna, and
M. Castilla, "Hierarchical control of droop-controlled ac and dc mi-
crogridsa general approach toward standardization," IEEE Trans. Ind.
Electron., vol. 58, no. 1, pp. 158–172, 2011.
[10] A. Narayanan, J. Bonneau, E. Felten, A. Miller, and S. Goldfeder, Bit-
coin and Cryptocurrency Technologies: A Comprehensive Introduction.
Princeton University Press, 2016.
[11] E. Heilman, A. Kendler, A. Zohar, and S. Goldberg, "Eclipse attacks on
bitcoin's peer-to-peer network." in USENIX Security, 2015, pp. 129–144.
Available:
[12] "Ethereumjs
[Online].
testrpc,"
https://github.com/ethereumjs/testrpc, accessed: 2017-03-21.
012345678Feeder number0.911.11.2Voltage [PU]10203040506070k050100Relative periods [%]DER1DER2DER3DER4 |
1809.05897 | 1 | 1809 | 2018-09-16T15:43:11 | Systems of bounded rational agents with information-theoretic constraints | [
"cs.MA",
"cs.AI",
"cs.IT",
"cs.IT"
] | Specialization and hierarchical organization are important features of efficient collaboration in economical, artificial, and biological systems. Here, we investigate the hypothesis that both features can be explained by the fact that each entity of such a system is limited in a certain way. We propose an information-theoretic approach based on a Free Energy principle, in order to computationally analyze systems of bounded rational agents that deal with such limitations optimally. We find that specialization allows to focus on fewer tasks, thus leading to a more efficient execution, but in turn requires coordination in hierarchical structures of specialized experts and coordinating units. Our results suggest that hierarchical architectures of specialized units at lower levels that are coordinated by units at higher levels are optimal, given that each unit's information-processing capability is limited and conforms to constraints on complexity costs. | cs.MA | cs |
Systems of bounded rational agents with
information-theoretic constraints
Sebastian Gottwald1 and Daniel A. Braun1
1Institute of Neural Information Processing, Faculty of Engineering, Computer Science
and Psychology, University of Ulm.
Keywords: Bounded rationality, multi-agent systems, hierarchical structure, spe-
cialization, Free Energy principle
Abstract
Specialization and hierarchical organization are important features of efficient
collaboration in economical, artificial, and biological systems. Here, we investi-
gate the hypothesis that both features can be explained by the fact that each entity
of such a system is limited in a certain way. We propose an information-theoretic
approach based on a Free Energy principle, in order to computationally analyze
systems of bounded rational agents that deal with such limitations optimally. We
find that specialization allows to focus on fewer tasks, thus leading to a more ef-
ficient execution, but in turn requires coordination in hierarchical structures of
specialized experts and coordinating units. Our results suggest that hierarchical
architectures of specialized units at lower levels that are coordinated by units at
higher levels are optimal, given that each unit's information-processing capability
is limited and conforms to constraints on complexity costs.
1
Introduction
The question of how to combine a given set of individual entities in order to perform a
certain task efficiently is a long-lasting question shared by many disciplines, including
economics, neuroscience, and computer science. Even though the explicit nature of a
single individuum might differ between these fields, e.g. an employee of a company,
a neuron in a human brain, or a computer or processor as part of a cluster, they have
one important feature in common that usually prevents them from functioning isolated
by themselves: they are all limited. In fact, this was the driving idea that inspired Her-
bert A. Simons early work on decision-making within economic organizations (Simon,
1943, 1955), which earned him a Nobel prize in 1978. He suggested that a scientific
behavioral grounding of economics should be based on bounded rationality, which has
remained an active research topic until today (Russell and Subramanian, 1995; Lipman,
1995; Aumann, 1997; Kaelbling et al., 1998; DeCanio and Watkins, 1998; Gigerenzer
and Selten, 2001; Jones, 2003; Sims, 2003; Burns et al., 2013; Ortega and Braun, 2013;
Acerbi et al., 2014; Gershman et al., 2015). Subsequent studies in management theory
have been built upon Simons basic observation, because "if individual managers had
unlimited access to information that they could process costlessly and instantaneously,
there would be no role for organizations employing multiple managers" (Geanakoplos
and Milgrom, 1991). In neuroscience and biology, similar concepts have been used
to explore the evolution of specialization and modularity in nature (Kashtan and Alon,
2005; Wagner et al., 2007). In modern computer science, the terms parallel computing
and distributed computing denote two separate fields that share the concept of decen-
tralized computing (Radner, 1993), i.e. the combination of multiple processing units in
order to decrease the time of computationally expensive calculations.
Despite of their success, there are also shortcomings of most approaches to the or-
ganization of decision-making units based on bounded rationality: As (DeCanio and
Watkins, 1998) point out, existing agent-based methods (including their own) are not
using an overreaching optimization principle, but are tailored to the specific types of
calculations the agents are capable of, and therefore lack in generality. Moreover, it
is usually imposed as a separate assumption that there are two types of units, special-
ized operational units and coordinating non-operational units, which was expressed by
(Knight, 1921) as "workers do, and managers figure out what to do".
Here, we use a Free Energy optimization principle in order to study systems of
bounded rational agents, extending the work in (Ortega and Braun, 2011, 2013; Ge-
newein and Braun, 2013; Genewein et al., 2015) on decision-making, hierarchical in-
formation-processing, and abstraction in intelligent systems with limited information-
processing capacity, that has precursors in the economic and game-theoretic literature
(McKelvey and Palfrey, 1995; Ochs, 1995; Mattsson and Weibull, 2002; Wolpert, 2006;
Spiegler, 2011; Howes et al., 2009; Todorov, 2009; Still, 2009; Tishby and Polani, 2011;
Kappen et al., 2012; Edward et al., 2014; Lewis et al., 2014). Note that the Free En-
ergy optimization principle of information-theoretic bounded rationality is connected
to the Free Energy principle used in variational Bayes and Active Inference (Friston
et al., 2015a,b, 2017a,b), but has a conceptually distinct interpretation and some formal
differences (see Section 6.3 for a detailed comparison).
By generalizing the ideas in (Genewein and Braun, 2013; Genewein et al., 2015) on
two-step information-processing to an arbitrary number of steps, we arrive at a general
Free Energy principle that can be used to study systems of bounded rational agents. The
advantages of our approach can be summarized as follows:
(i) There is a unifying Free Energy principle that allows for a multi-scale problem
formulation for an arbitrary amount of agents distributed among the steps of gen-
eral multi-step processes (see Sections 3.3 and 4.2).
(ii) The computational nature of the optimization principle allows to explicitly cal-
culate and compare optimal performances of different agent architectures for a
given set of objectives and resource constraints (see Section 5).
(iii) The information-theoretic description implies the existance of the two types of
2
units mentioned above, non-operational units (selector nodes) that coordinate the
activities of operational units. Depending on their individual resource constraints,
the Free Energy principle assigns each unit to a region of specialization that is part
of an optimal partitioning of the underlying decision space (see Section 4.3).
In particular, we find that, for a wide range of objectives and resource limitations
(see Sections 5 and 5), hierarchical systems with specialized experts at lower levels and
coordinating units at higher levels generally outperform other structures.
2 Preliminaries
This section serves as an introduction to the terminology required for our framework
presented in Section 3 and 4.
X . Given a probability distribution p ∈ P
f : X → R is denoted by (cid:104)f(cid:105)p :=(cid:80)
Notation
We use curly letters, W, X , A, etc. to denote sets of finite cardinality, in particular the
underlying spaces of the corresponding random variables W , A, X, etc., whereas the
values of these random variables are denoted by small letters, i.e. w ∈ W, a ∈ A,
and x ∈ X , respectively. We denote the space of probability distributions on a given
set X by P
X , the expectation of a function
x p(x)f (x). If the underlying probability measure
is clear without ambiguity we just write (cid:104)f(cid:105).
For a function g with multiple arguments, e.g. for g : X × Y → R, (x, y)(cid:55)→ g(x, y),
we denote the function X → R, x(cid:55)→ g(x, y) for fixed y ∈ Y by g(·, y) (partial applica-
tion), i.e. the dot indicates the variable of the new function. Similarly, for fixed y ∈ Y,
we denote a conditional probability distribution on X with values p(xy) by p(·y). This
if F is a functional defined on functions of one variable, e.g. F [f ] :=(cid:80)
notation shows the dependencies clearly without giving up the original function names
and thus allows to write more complicated expressions in a concise form. For example,
x f (x) for all
functions f : X → R, then evaluating F on the function g in its first variable while
that the resulting value (which equals(cid:80)
keeping the second variable fixed, is simply denoted by F [g(·, y)]. Here, the dot indi-
cates on which argument of g the functional F is acting and at the same time it records
x g(x, y) in the case of the example) does not
depend on a particular x but on the fixed y.
2.1 Decision-making
Here, we consider (multi-task) decision-making as the process of observing a world
state w ∈ W, sampled from a given distribution ρ ∈ P
W, and choosing a corresponding
action a ∈ A drawn from a posterior policy P (·w) ∈ P
A. Assuming that the joint
distribution of W and A is given by p(a, w) := ρ(w)P (aw), then P is the conditional
probability distribution of A given W . Unless stated otherwise, the capital letter P
always denotes a posterior, while the small letter p denotes the joint distribution or a
marginal of the joint (i.e. a dependent variable).
3
(cid:88)
(cid:88)
a∈A
A decision-making unit is called agent. An agent is rational, if its posterior policy
P maximizes the expected utility
(cid:104)U(cid:105) =
w∈W
ρ(w)
P (aw) U (a, w)
(1)
for a given utility function U : W ×A → R. Note that the utility U may itself represent
an expected utility over consequences in the sense of von Neumann and Morgenstern
(1944), where W would serve as a context variable for different tasks. The posterior P
can be seen as a state-action policy that selects the best action a ∈ A with respect to a
utility function U given the state w ∈ W of the world.
2.2 Bounded rational agents
In the information-theoretic model of bounded rationality (Ortega and Braun, 2011,
2013; Genewein et al., 2015), an agent is bounded rational if its posterior P maximizes
(1) subject to the constraint
ρ(w) DKL(P (·w)(cid:107)q) (cid:54) D0 ,
(2)
(cid:10)DKL(P(cid:107)q)(cid:11) =
defined by DKL(p(cid:107)q) := (cid:80)
y∈Y
(cid:88)
w∈W
for a given bound D0 > 0 and a prior policy q ∈ P
A. Here, DKL(p(cid:107)q) denotes the
Kullback-Leibler (KL) divergence between two distributions p, q ∈ P
Y on a set Y,
p(y) log(p(y)/q(y)). Note that, for DKL(p(cid:107)q) to be
well-defined, p must be absolutely continuous with respect to q, so that q(y) = 0 implies
p(y) = 0. When p or q are conditional probabilities, then we treat DKL(p(cid:107)q) as a
function of the additional variables.
Given a world state w, the information-processing consists of transforming a prior q
to a world state specific posterior distribution P (·w). Since DKL(P (·w)(cid:107)q) measures
by how much P (·w) diverges from q, the upper bound D0 in (2) characterizes the
limitation of the agent's average information-processing capability: If D0 is close to
zero, the posterior must be close to the prior for all world states, which means that
A contains only little information about W , whereas if D0 is large, the posterior is
allowed to deviate from the prior by larger amounts and therefore A contains more
information about W . We use the KL-divergence as a proxy for any resource measure,
as any resource must be monotone in processed information, which is measured by the
KL-divergence between prior and posterior.
Technically, maximizing expected utility under the constraint (2) is the same as
minimizing expected complexity cost under the constraint of a minimal expected per-
formance, where complexity is given by the expected KL-divergence between prior and
posterior and performance by expected utility. Minimizing complexity means minimiz-
ing the number of bits required to generate the actions.
2.3 Free Energy principle
By the variational method of Lagrange multipliers, the above constrained optimization
problem is equivalent to the unconstrained problem
,
(3)
(cid:16)
max
P
(cid:104)U(cid:105) −
1
β
(cid:10)DKL(P(cid:107)q)(cid:11)(cid:17)
4
In the literature on
where β > 0 is chosen such that the constraint (2) is satisfied.
information-theoretic bounded rationality (Ortega and Braun, 2011, 2013), the objective
in (3) is known as the Free Energy F of the corresponding decision-making process. In
this form, the optimal posterior can be explicitly derived by determining the zeros of the
functional derivative of F with respect to P , yielding the Boltzmann-Gibbs distribution
(4)
q(a) eβ U (a,w) , Z(w) :=
q(a) eβ U (a,w) .
(cid:88)
1
P (aw) =
Z(w)
a∈A
Note how the Lagrange multiplier β (also known as inverse temperature) interpolates
between an agent with zero processing capability that always acts according to its prior
policy (β = 0) and a perfectly rational agent (β → ∞). Note that, plugging (4) back
into the Free Energy (3) gives
(cid:10) log Z(cid:11) .
P F[P ] =
max
1
β
2.4 Optimal prior
The performance of a given bounded rational agent crucially depends on the choice of
the prior policy q. Depending on D0 and the explicit form of the utility function, it
can be advantageous to a priori prefer certain actions over others. Therefore, optimal
bounded rational decision-making includes optimizing the prior in (3). In contrast to
(3), the modified optimization problem
(cid:16)
max
P,q
(cid:104)U(cid:105) −
1
β
(cid:10)DKL(P(cid:107)q)(cid:11)(cid:17)
(5)
(6)
(7)
does not have a closed form solution. However, since the objective is convex in (P, q),
a unique solution can be obtained iteratively by alternating between fixing one and op-
timizing the other variable (Csisz´ar and Tusn´ady, 1984), resulting in a Blahut-Arimoto
type algorithm (Arimoto, 1972; Blahut, 1972) that consists of alternating the equations
(cid:40)
q(a) = p(a) = (cid:80)
P (aw) = 1
Z(w) q(a) eβU (a,w) ,
w ρ(w)P (aw),
with Z(w) given by (4). In particular, the optimal prior policy is the marginal p of the
joint distribution of W and A. In this case, the average Kullback-Leibler divergence
between prior and posterior coincides with the mutual information between W and A,
I(W ; A) =
p(w, a) log
(cid:88)
(cid:88)
w∈W
a∈A
p(w, a)
ρ(w)p(a)
(cid:17)
(cid:16)
max
P
(cid:104)U(cid:105) −
1
β
=(cid:10)DKL(P, p)(cid:11) .
It follows that the modified optimization principle (6) is equivalent to
I(W ; A)
.
(8)
Due to its equivalence to rate distortion theory (Shannon, 1959) (with a negative
distortion measure given by the utility function), (8) is denoted as the rate distortion
case of bounded rationality in (Genewein and Braun, 2013).
5
2.5 Multi-step and multi-agent systems
When multiple random variables are involved in a decision-making process, such a
process constitutes a multi-step system (see Section 3). Consider the case of a prior
over A that is conditioned on an additional random variable X with values x ∈ X , i.e.
q(·x) ∈ P
A for all x ∈ X . Remember that we introduced a bounded rational agent as
a decision-making unit, that, after observing a world state w, transforms a single prior
policy over a choice space A to a posterior policy P (·w) ∈ P
A. Therefore, in the case
of a conditional prior, the collection of prior policies {q(·x)}x∈X can be considered as
a collection or ensemble of agents, or a multi-agent system, where for a given x ∈ X ,
the prior q(·x) is transformed to a posterior P (·x, w) ∈ P
A by exactly one agent.
Note that a single agent deciding about both, X and A, would be modelled by a prior
of the form q(x, a) with x ∈ X and a ∈ A, instead.
Hence, in order to combine multiple bounded rational agents, we are first splitting
the full decision-making process into multiple steps by introducing additional interme-
diate random variables (Section 3), which then will be used to assign one or more agents
to each of these steps (Section 4). In this view, we can regard a multi-agent decision-
making system as performing a sequence of successive decision steps until an ultimate
action is selected.
3 Multi-step bounded rational decision-making
3.1 Decision nodes
Let W and A denote the random variables describing the full decision-making process
for a given utility function U : W × A → R, as described in Section 2.
In order
to separate the full process into N > 1 steps, we introduce internal random variables
X1, . . . , XN−1, which represent the outputs of additional intermediate bounded rational
decision-making steps. For each k, let Xk denote the target space and xk ∈ Xk a partic-
ular value of Xk. We call a random variable that is part of a multi-step decision-making
system a (decision) node. For simplicity, we assume that all intermediate random vari-
ables are discrete (just like W and A).
Here, we are treating feed-forward architectures originating at X0 := W and ter-
k=0 according to the
minating in XN := A. This allows to label the variables {Xk}N
information flow, so that Xj potentially can only obtain information about Xi if i < j.
The canonical factorization
p(w, x1, . . . , xN−1, a) = ρ(w) p(x1w) p(x2x1, w)··· p(axN−1, . . . , x1, w)
of the joint probability distribution of {Xk}N
cies of each decision node.
k=0 therefore consists of the posterior poli-
3.2 Two types of nodes: inputs and prior selectors
A specific multi-step architecture is characterized by specifying the explicit depen-
dencies on the preceding variables for each node's prior and posterior, or better the
6
Figure 1. Example of a processing node that is part of a multi-step architecture with
N = 5, visualized as a directed graph. Here, X3 processes the output of X2 by trans-
forming a prior policy p(x3x1) to a posterior policy P (x3x2, x1). The prior of X3
being conditioned on the output of X1 (indicated by the dashed arrow), means that X1
determines which of the prior policies {p(·x1)}x1∈X1 is used by X3 to process a given
output of X2.
missing dependencies. For example, in a given multi-step system, the posterior of the
node X3 might depend explicitly on the outputs of X1 and X2 but not on W , so that
P (x3x2, x1, w) = P (x3x2, x1). If its prior has the form q(x3x1), then X3 has to pro-
cess the output of X2. Moreover, in this case, the actual prior policy q(·x1) ∈ P
X3 that
is used by X3 for decision-making is selected by X1 (see Figure 1).
In general, the inputs Xi, . . . , Xj that have to be processed by a particular node Xk,
are given by the variables in the posterior that are missing from the prior, and, if its prior
q is conditioned on the outputs of Xl, . . . , Xm, then these nodes select which of the prior
policies {q(·xl, . . . , xm)}xl∈Xl,...,xm∈Xm ⊂ P
Xk is used by Xk for decision-making, i.e.
for the transformation
q(xkxl, . . . , xm) −→ P (xkxl, . . . , xm, xi, . . . , xj) .
We denote the collection of input nodes of Xk by X k
selecting nodes of Xk by X k
is then given by
in (:= {Xi, . . . , Xj}) and the prior
sel (:= {Xl, . . . , Xm}). The joint distribution of X0, . . . , XN
(cid:0)xN
(cid:12)(cid:12)xN
sel, xN
in
(cid:1)
(9)
Specifying the sets X k
for all xk ∈ Xk and xk
in ∈ X k
in of selectors and inputs for each node in the system
then uniquely characterizes a particular multi-step decision-making system. Note that
we always have (X 1
sel, xk
sel and X k
sel ∈ X k
Decompositions of the form (9) are often visualized by directed acyclic graphs, so-
called DAGs (see e.g. Bishop, 2006, pp. 360). Here, in addition to the decomposition
of the joint in terms of posteriors, we have added the information about the prior depen-
dencies in terms of dashed arrows, as shown in Figure 1.
sel, X 1
in) = ({},{X0}).
p(x0, . . . , xN ) = ρ(w) P1
(cid:0)x1
(cid:12)(cid:12)x1
(cid:1)
sel, x1
in
··· PN
in (k = 1, . . . , N).
7
WX1X2X3X4Ap(x3x1)−→p(x3x2,x1) ,
sup
(10)
(cid:16)
1
βk
k=1
P1,q1,...,PN ,qN
(cid:104)U(cid:105) −
(cid:88)
N(cid:88)
(cid:10)DKL(P(cid:107)q)(cid:11) =
3.3 Multi-step Free Energy principle
If Pk and qk denote the posterior and prior of the k-th node of an N-step decision-
process, then the Free Energy principle takes the form
where, in addition to the expectation over inputs, the average of DKL(Pk(cid:107)qk) now also
includes the expectation with respect to Xsel,
(cid:10)DKL(Pk(cid:107)qk)(cid:11)(cid:17)
(cid:0)P (·xsel, xin)(cid:107)q(·xsel)(cid:1) .
Since the prior policies only appear in the KL-divergences, and moreover, there is
exactly one KL-divergence per prior, it follows as in 2.4, that for each k = 1, . . . , N the
(cid:88)
optimal prior is the marginal given for all xk ∈ Xk by
I(cid:0)X k
sel = xsel. Hence, the Free Energy principle can be simplified to
qk(xkxsel) = pk(xkxsel) :=
{x0,...,xN}\({xk}∪xsel)
p(xsel, xin) DKL
whenever X k
p(x0, . . . , xN ) ,
(cid:1)(cid:17)
(cid:12)(cid:12)X k
sel
N(cid:88)
p(xsel)
in; Xk
(cid:16)
(12)
(11)
xsel,xin
1
,
sup
P1,...,PN
(cid:104)U(cid:105) −
1
βk
k=1
where I(X; Y Z) denotes the conditional mutual information of two random variables
X, Y given a third random variable Z.
By optimizing (12) alternatingly, i.e. optimizing one posterior at a time while keep-
ing the others fixed, we obtain for each k = 1, . . . , N,
(cid:0)xkxsel, xin
(cid:1) =
Pk
(cid:1)
(cid:0)xkxsel
pk
Zk(xsel, xin)
(cid:104)
exp
βk Fk[P1, . . . , PN ](xk, xsel, xin)
,
(13)
(cid:105)
sel = xsel and X k
in = xin. Here, Zk(xsel, xin) denotes the normaliza-
whenever X k
tion constant and Fk[P1, . . . , PN ] denotes the (effective) utility function on which the
decision-making in Xk is based on. More precisely, given X = (Xk, X k
in),
sel, X k
it is the Free Energy of the subsequent nodes in the system, i.e.
for any value of
x := (xk, xsel, xin) we obtain for Fk := Fk[P1, . . . , PN ],
(cid:88)
1
p(x)
{x0,...,xN}\x
Fk(x) =
where
p(x0, . . . , xN )Fk,loc(x0, . . . , xN ) ,
(cid:88)
i>k
1
βi
log
in)
Pi(xixi
pi(xixi
sel, xi
sel)
(14)
.
Fk,loc(x0, . . . , xN ) := U (x0, xN ) −
in and xi
sel are collections of values of the random variables in X i
Here, xi
sel,
respectively. The final Blahut-Arimito-type algorithm consists of iterating (13), (11),
and (14) for each k = 1, . . . , N until convergence is achieved. Note that, since each op-
timization step is convex (marginal convexity), convergence is guaranteed but generally
not unique (Jain and Kar, 2017), so that, depending on the initialization, one might end
up in a local optimum.
in and X i
8
3.4 Example: two-step information-processing
The cases of serial and parallel information-processing studied in (Genewein and Braun,
2013), are special cases of multi-step decision-making systems introduced above. Both
cases are two-step processes (N = 2) involving the variables X0 = W , X1 = X, and
in) = ({},{X1}), and the parallel
X2 = A. The serial case is characterized by (X 2
in) = ({X1},{X0}). There is a third possible combination for N = 2,
case by (X 2
in) = ({},{X0, X1}). However, it can be shown that this case is
given by (X 2
equivalent to the (one-step) rate distoration case from Section 2, because if A has direct
world state access, then any extra input to the final node A = X2, that is not a prior
selector, contains redundant information.
sel, X 2
sel, X 2
sel, X 2
4 Systems of bounded rational agents
4.1 From multi-step to multi-agent systems
As explained in 2.5 above, a single random variable Xk that is part of an N-step
decision-making system can represent a single agent or a collection of multiple agents,
sel, i.e. whether Xk has multiple priors which are
depending on the cardinality of X k
sel or not. Therefore, an N-step bounded rational decision-
selected by the nodes in X k
making system with N > 1 represents a bounded rational multi-agent system (of depth
N).
sel of X k
For a given k ∈ {1, . . . , N}, each value x ∈ X k
sel corresponds to exactly one
agent in Xk. During decision-making, the agents that belong to the nodes in X k
sel are
sel agents in Xk is going to receive a given input xin (see 4.4
choosing which of the X k
below for a detailed example). This decision is based on how well the selected agent
x will perform on the input xin by transforming its prior policy pk(·x) into a posterior
(cid:88)
policy Pk(·x, xin), subject to the constraint
p(xinx) DKL
(cid:0)Pk(·x, xin)(cid:107)pk(·x)(cid:1) (cid:54) Dx ,
(cid:104)DKL(Pkpk)(cid:105)(x) :=
(15)
xin
where Dx > 0 is a given bound on the agent's information-processing capability. Sim-
ilarly to multi-step systems, this choice is based on the performance measured by the
Free energy of the subsequent agents.
4.2 Multi-agent Free Energy principle
In contrast to multi-step decision-making, the information-processing bounds are al-
lowed to be functions of the agents instead of just the nodes, resulting in an extra La-
grange multiplier for each agent in the Free Energy principle (10). As in (12), optimiz-
ing over the priors yields the simplified Free Energy principle
I(cid:0)X k
(cid:12)(cid:12)X k
(cid:1)(cid:17)
in; Xk
sel = xk
sel
,
(16)
(cid:16)
sup
P1,...,PN
(cid:104)U(cid:105) −
N(cid:88)
(cid:88)
k=1
sel∈X k
xk
sel
p(xk
βk(xk
sel)
sel)
9
which can be solved iteratively as explained in the previous section, the only difference
sel. Hence, for the posterior of
being that the Lagrange parameters βk now depend on xk
an agent that belongs to node k, we have
(cid:0)xkxsel, xin
(cid:1) =
Pk
(cid:1)
(cid:0)xkxsel
pk
Zk(xsel, xin)
where βk(xk
Fk is given by (14) except that now we have
sel) is chosen such that the constraint (15) is fulfilled for all x ∈ xk
sel, and
Fk,loc(x0, . . . , xN ) := U (x0, xN ) −
1
βi(xi
sel)
log
in)
Pi(xixi
pi(xixi
sel, xi
sel)
.
(18)
The resulting Blahut-Arimoto-type algorithm is summarized in Algorithm 1.
(cid:105)
exp
βk(xk
sel) Fk(xk, xsel, xin)
,
(17)
(cid:104)
(cid:88)
i>k
Algorithm 1 Blahut-Arimito-type algorithm for (16)
1: procedure GETMULTIAGENTSOLUTION(U, ρ,{(X k
initialize p(x0, . . . , xN ) ∀x0, . . . , xN,
repeat
sel, X k
in)}N
k=1, β, )
Fk(xk, xsel, xin) ← (14), (18) ∀xk, xsel, xin
Pk(xkxsel, xin) ← (17) ∀xk, xsel, xin
p(xkxsel) ← (11) ∀xk, xsel
p(x0, . . . , xN ) ← (9) ∀x0, . . . , xN
(cid:46) calc. effective utility
(cid:46) update posterior
(cid:46) update prior
(cid:46) update joint
p0 ← p
for k = 1, . . . , N do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14: end procedure
end for
error ← dist(p, p0)
until error <
return P1, . . . , PN
4.3 Specialization
Even though a given multi-agent architecture predetermines the underlying set of choices
for each agent, only a small part of such a set might be used by a given agent in the op-
timized system. For example, all agents in the final step potentially can perform any
action a ∈ A (see Figure 2 and the Example in 4.4 below). However, depending on
their indiviual information-processing capabilities, the optimization over the agents'
priors can result in a (soft) partitioning of the full action space A into multiple chunks,
where each of these chunks is given by the support of the prior of a given agent x,
supp(p(·x)) ⊂ A. Note that the resulting partitioning is not necessarily disjoint,
since agents might still be sharing a number of actions, depending on their available
information-processing resources. If the processing capability is low compared to the
amount of possible actions in the full space, and if there are enough agents at the same
level, then this partitioning allows each agent to focus on a smaller number of options
10
to choose from, provided that the coordinating agents have enough resources to decide
between the partitions reliably.
Therefore, the amount of prior adaptation of an agent, i.e. by how much its optimal
prior p deviates from a uniform prior p0 over all accessible choices, which is measured
by the KL-divergence DKL(p(cid:107)p0), determines its degree of specialization. More pre-
cisely, we define the specialization of an agent with prior p and choice space X by
S[p] :=
DKL(p(cid:107)p0)
log X
= 1 −
H[p]
log X
,
(19)
(cid:80)
where H[p] := −
x p(x) log p(x) denotes the Shannon entropy of p. By normalizing
with log X, we obtain a quantity between 0 and 1, since 0 (cid:54) H(p) (cid:54) log X. Here,
S[p] = 0 corresponds to H[p] = log X, which means that the agent is completely
unspecialized, whereas S[p] = 1 corresponds to H[p] = 0, which implies that p has
support on a single option x∗ ∈ X meaning that the agent deterministically performs
always the same action and therefore is fully specialized.
Figure 2. Example of a hierarchical architecture of 10 agents that are combined via
the 3-step decision-making system (N = 3) shown in the upper left corner (see 4.4 for
details). Here, every node -- and therefore every agent -- has access to the world states
(big circle). X1 consists of one agent that decides about which of the X1 = 3 agents in
X2 obtains a given world state as input. The selected agent in X2 selects which of the
X2 = 2 agents out of the X1 · X2 = 6 agents in A that are connected to it, obtains
the world state to perform the final decision about an action a ∈ A (grey circles on the
right). In our notation introduced below, this architecture is labelled by (1, 4)[1,3,(3,2)]
(see Section 5.1).
11
worldagentactioninputselectionWX1X2A 4.4 Example: Hierarchical multi-agent system with three levels
Consider the example of an architecture of 10 agents shown in Figure 2 that are com-
bined via the 3-step decision-making system given by
(X 2
sel, X 2
in) = ({X1},{W}),
(X 3
sel, X 3
in) = ({X1, X2},{W}),
(20)
as visualized in the upper left corner of Figure 2. The number of agents in each node
is given by the cardinality of the target space of the selecting node(s) (or equals one if
there are no selectors). Hence, X1 consists of one agent, X2 consists of X1 agents, and
A consists of X1 · X2 agents. For example, if we have X1 = 3 and X2 = 2, as in
Figure 2, then this results in a hierarchy of 1, 3 and 6 agents.
The joint probability of the system characterized by (20) is given by
p(w, x1, x2, a) = p(w)P1(x1w)P2(x2x1, w)P3(ax2, x1, w) ,
and the Free Energy by
(cid:88)
(cid:20)
F[P1, P2, P3] =
p(w, x1, x2, a)
U (a, w) −
1
β1
log
P1(x1w)
p1(x1)
where the priors p1, p2, and p3 are given by the marginals (11), i.e.
1
β3(x1, x2)
log
P3(ax2, x1, w)
p3(ax2, x1)
w,x1,x2,a
1
β2(x1)
−
log
p1(x1) =
p2(x2x1) =
p3(ax2, x1) =
P2(x2x1, w)
p2(x2x1) −
(cid:88)
(cid:88)
(cid:88)
w
w
ρ(w)P (x1w) ,
p(wx1)P2(x2x1, w) ,
(cid:21)
,
(cid:105)
.
By (13), the posteriors that iteratively solve the Free Energy principle are
p(wx1, x2)P3(ax2, x1, w).
w
exp(cid:2)β1F1(w, x1)(cid:3) ,
exp(cid:2)β2(x1)F2(w, x1, x2)(cid:3) ,
exp(cid:2)β3(x1, x2)U (a, w)(cid:3) ,
p1(x1)
Z(w)
p2(x2x1)
Z(w, x1)
p3(ax2, x1)
Z(w, x1, x2)
P1(x1w) =
P2(x2x1, w) =
P3(ax2, x1, w) =
where, by (14) and (18),
(cid:88)
x2,a
(cid:88)
a
F1(w, x1) :=
F2(w, x1, x2) :=
p(x2, ax1, w)
1
β2(x1)
log
U (a, w) −
1
−
β3(x1, x2)
(cid:105)
,
P2(x2x1, w)
p2(x2x1)
P3(ax2, x1, w)
p3(ax2, x1)
log
log
1
P3(ax2, x1, w)
p3(ax2, x1)
P3(ax2, x1, w)
U (a, w) −
β3(x1, x2)
(cid:104)
(cid:104)
12
Given a world state w ∈ W, the agent in X1 decides about which of the three agents
in X2 obtains w as an input. This narrows down the possible choices for the selected
agent in X2 to two out of the six agents in A. The selected agent performs the final
decision by choosing an action a ∈ A. Depending on its degree of specialization, which
is a result of his own and the coordinating agents' resources, this agent will choose his
action from a certain subset of the full space A.
5 Optimal Architectures
Here, we show how the above framework can be used to determine optimal architectures
of bounded rational agents. Summarizing the assumptions made in the derivations, the
multi-agent systems that we analyze must fulfill the following requirements:
(i) The information-flow is feed-forward: An agent in Xk can obtain information
directly from another agent that belongs to Xm only if m < k.
(ii) Intermediate agents cannot be endpoints of the decision-making process:
the
information-flow always starts with the processing of W and always ends with
a decision a ∈ A.
(iii) A single agent is not allowed to have multiple prior policies: Agents are the small-
est decision-making unit, in the sense that they transform a prior to a posterior
policy over a set of actions in one step.
The performance of the resulting architectures is measured with respect to the ex-
pected utility they are able to achieve under a given set of resource constraints. To
this end, we need to specify (1) the objective for the full decision-making process, (2)
the number N of decision-making steps in the system, (3) the maximal number n of
agents to be distributed among the nodes, and (4) the individual resource constraints
{D1, . . . , Dn} of those agents.
We illustrate the specifications (1) -- (4) with a toy example in Section 5.2 by show-
casing and explicitly explaining the differences in performance of several architectures.
Moreover, we provide a broad performance comparison in Section 5.3, where we sys-
tematically vary a set of objective functions and resource constraints, in order to deter-
mine which architectural features most affect the overall performance. For simplicity,
in all simulations we are limiting ourselves to architectures with N (cid:54) 3 nodes and
n (cid:54) 10 agents. In the following section, we start by describing how we characterize the
architectures conforming to the requirements (i) -- (iii).
5.1 Characterization of architectures
Type. In view of property (ii) above, we can label any N-step decision-making
process by a tuple (i, j), which we call the type of the architecture, where i characterizes
13
Figure 3. Overview of the resulting architectures for N (cid:54) 3, each of them being
labelled by its type.
the relation between the first N −1 variables W , X1, . . . , XN−1, and j determines how
these variables are connected to XN = A.
For example, for N (cid:54) 3, we obtain the types shown in Figure 3, where i ∈ {0, 1, 2}
and j ∈ {0, . . . , 5} represent the following relations:
i = 0 : (X 2
i = 1 : (X 2
i = 2 : (X 2
j = 0 : (X 3
j = 1 : (X 3
j = 2 : (X 3
sel, X 2
sel, X 2
sel, X 2
sel, X 3
sel, X 3
sel, X 3
sel, X 3
sel, X 3
sel, X 3
in) = ({},{X1})
in) = ({X1},{W})
in) = ({},{W})
in) = ({},{X2})
in) = ({},{X1, X2})
in) = ({X1},{X2})
in) = ({X2},{X1})
in) = ({X1, X2},{W})
in) = ({X2},{W}) .
j = 3 : (X 3
j = 4 : (X 3
j = 5 : (X 3
For example, the architecture shown in Figure 2 has the type (1, 4). Correspondingly,
the two-step cases are labelled by (i, ) for i ∈ {0, 1, 2}, and the one-step rate distoration
case by (−1, ). Note that not every combination of i ∈ {0, 1, 2} and j ∈ {0, . . . , 5}
describes a unique system, e.g. (2, 3) is equivalent to (2, 2) when replacing X1 by
14
(−1,)WA(0,)WXA(1,)WXA(2,)WXA(0,0)WX1X2A(0,1)WX1X2A(0,2)WX1X2A(0,3)WX1X2A(0,4)WX1X2A(0,5)WX1X2A(1,0)WX1X2A(1,1)WX1X2A(1,2)WX1X2A(1,3)WX1X2A(1,4)WX1X2A(1,5)WX1X2A(2,1)WX1X2A(2,2)WX1X2A(2,3)WX1X2A(2,4)WX1X2A X2. Moreover, as mentioned above, (2, ) is equivalent to (−1, ), and similarly, (0, 1) is
equivalent to (0, ).
Figure 4. Visualization of exemplary 3-step multi-agent architectures specified by their
types and shapes.
Shape. After the number of nodes has been fixed, the remaining property that charac-
terizes a given architecture is the number of agents per node. For most architectures
there are multiple possibilities to distribute a given amount of agents among the nodes,
even when neglecting individual differences in resource constraints. We call such a dis-
tribution a shape, denoted by [n1, n2, . . . ], where nk denotes the number of agents in
node k. Note that, not all architectures will be able to use the full amount of available
agents, most immanently the one-step rate distortion case (1 agent), or the two-step
serial-case (2 agents). For these systems, we always use the agents with the highest
available resources in our simulations.
For example, for N (cid:54) 3 the resulting shapes for a maximum of n = 10 agents are
as follows:
• [1] for (−1, ), [1, 1] for (0, ), and [1, 9] for (1, ),
15
(0,3)[1,1,8](1,0)[1,8,1](1,2)[1,4,4](1,5)[1,3,6](2,1)[1,1,1]worldagentactioninputselection(2,4)[1,1,(2,4)]W118W181W144W136W111W118 Figure 5. Utility function for Example 5.2 (left) and posterior policy of a single agent
with an information bound of 4.6 bit (right). The set of phone calls W is partitioned into
three separate regions, corresponding to three different topics about which customers
might have complaints or questions. Each of these can be divided into two subcate-
gories of four customer calls each. For each phone call there is exactly one answer that
achieves the best result (U = 1). Moreover, the responses that belong to one subcat-
egory of calls are also suitable for the other calls in that particular subcategory, albeit
slightly less effective (U = 0.85) than the optimal answers. Similarly, the responses
that belong to the same topic of calls are still a lot better (U = 0.7) than responses to
other topics (U = 0).
• [1, 1, 1] for (0, 0) and (2, 1),
• [1, 1, 8] for (0, 2), (0, 3), (0, 5), (2, 2),
• [1, 1, (2, 4)] and [1, 1, (4, 2)] for (0, 4) and (2, 4),
• [1, 8, 1] for (1, 0) and (1, 1),
• [1, 4, 4] for (1, 2),
• [1, 2, 7], [1, 3, 6], [1, 4, 5], [1, 5, 4], [1, 6, 3], [1, 7, 2] for (1, 3) and (1, 5),
• [1, 2, (2, 3)] and [1, 3, (3, 2)] for (1, 4),
where a tuple inside the shape means that two different nodes are deciding about the
agents in that spot, e.g.
[1, 1, (2, 4)] means that there are 8 agents in the last node,
labeled by the values (x1, x2) ∈ X1 × X2 with X1 = 2 and X2 = 4. In Figure 4, we
visualize one example architecture for each of the above 3-step shapes, except for the
shapes of type (1, 4) of which one example is shown in Figure 2.
Together, the type (i, . . . ) and shape [n1, . . . ] uniquely characterize a given multi-
agent architecture, denoted by (i, . . . )[n1,... ].
5.2 Example: Callcenter
Consider the operation of a company's callcenter as a decision-making process, where
customer calls (world states) must be answered with an appropriate response (action)
in order to achieve high customer satisfaction (utility). The utility function shown in
Figure 5 on the left can be viewed as a simplistic model for a real-world callcenter of a
big company such as a communication service provider. In this simplification, there are
16
04812162024responsesa04812162024phonecallswUtilityfunctionU04812162024responsesa04812162024Posteriorofasingleagentwith4.6bit0.00.20.40.60.81.0 24 possible customer calls that belong to three separate topics, for example questions
related to telephone, internet, or television, which can be further subdivided into two
subcategories, for example consisting of questions concerning the contract or problems
with the hardware. See the description of Figure 5 for the explicit utility values.
Handling all possible phone calls perfectly by always choosing the corresponding
response with maximum utility requires log2(24) ≈ 4.6 bit (see Figure 5). However, in
practice a single agent is usually not capable of knowing the optimal answers to every
single type of question. For our example this means that the callcenter only has access
to agents with information-processing capability less than 4.6 bit. It is then required
to organize the agents in a way so that each agent only has to deal with a fraction of
the customer calls. This is often realized by first passing the phone call through several
filters in order to forward it to a specialized agent. Arranging these selector or filter units
in a strict hierarchy then corresponds to architectures of the form of (1, 4) or (1, 5) (see
below for a comparison of these two), where at each stage a single operator selects how
a call is forwarded. In contrast, architectures of the form of (2, 4) allow for multiple
independent filters working in parallel, for example realized by multiple trained neural
networks, where each is responsible for a particular feature of the call (for example,
one node deciding about the language of the call, and another node deciding about the
topic). In the following we do not discriminate between human and artificial decision-
makers, since both can qualify equally well as information-processing units.
Assume that there are n = 10 bounded rational agents available. Considering the
given utility function, the architectures (1, 4)[1,3,(3,2)] (shown in Figure 2) and (1, 5)[1,3,6]
(shown in Figure 4) might be obvious choices as they represent the hierarchical structure
of the utility function. With an information bound of 1.6 (≈ log2(3)) bit for the first
agent and 0.1 bit for the rest, the optimal prior policies for (1, 5)[1,3,6] obtained by our
Free Energy principle are shown in Figure 6. We can see that, for this architecture, the
choice x1 of the agent at the first step corresponds to the general topic of the phone call,
the decisions x2 of the three agents at the second stage correspond to the subcategory
on which one of the six agents at the final stage is specialized to, who then makes the
decision about the final response a by picking one of the four actions in the support of
its prior.
Figure 6. Optimal prior policies for each agent of the architecture (1, 5)[1,3,6] with an
information bound of (D1, D2, . . . , D10) = (1.6, 0.1, . . . , 0.1).
17
0.00.2p(x1)012choicesxi0.00.20.4p(x2x1=0)0123450.00.20.4p(x2x1=1)0123450.00.20.4p(x2x1=2)0123450.00.2p(ax2=0)051015200.00.2p(ax2=1)051015200.00.2p(ax2=2)051015200.00.2p(ax2=3)051015200.00.2p(ax2=4)051015200.00.2p(ax2=5)05101520 Figure 7. Performance comparison under two different information bounds.
We can see in Figure 7 on the left that a hierarchical structure as in (1, 5)[1,3,6] or
(1, 4)[1,3,(3,2)] is indeed superior when comparing with the architecture (2, 4)[1,1,(2,4)],
because there is no good selector for the second filter. We have also added two architec-
tures to the comparison that have a bottleneck of the information flow at either end of the
decision-making process, (0, 3)[1,1,8] and (1, 0)[1,8,1] (see Figure 4 for a visualization),
which are performing considerably worse than the others: in (0, 3)[1,1,8] the first agent is
the only one who has direct contact to the customer and passes the filtered information
on to everybody else, whereas in (1, 0)[1,8,1] the customer talks to multiple agents, but
these cannot take any decisions but pass on the information to a final decision node
who has to select from all possible options. Interestingly, as can be seen on the right
side of Figure 7, when changing the resource bounds such that the first agent only has
D1 = 1 bit instead of 1.6 and the second agent has D2 = 0.5 bit instead of 0.1, then the
strictly hierarchical architectures (1, 5)[1,3,6] and (1, 4)[1,3,(3,2)] are outperformed by the
architecture (2, 4)[1,1,(2,4)], because their first agent is not able to perfectly distinguish
between the three topics anymore. This is an ideal situation for (2, 4)[1,1,(2,4)], since
here the total information-processing for filtering the phone calls is split up efficiently
between the first two agents in the system.
Note that (1, 4) and (1, 5) do not necessarily perform identically (as can be seen on
the right in Figure 7), even though the structure of the utility function might suggest that
it is ideal for (1, 5)[1,3,6] to always have the optimal priors shown in Figure 6. However,
this crucially depends on the given information-processing bounds. In Figure 8, we
illustrate the difference between the two types in more detail, by showing the processed
information that can actually be achieved per agent in the respective architecture for an
information bound of D = (0.4, 2.6, 2.6, 2.6, 0.4, . . . , 0.4). When the first agent in the
hierarchy has low capacity, then the rigid structure of (1, 4) is penalized because the
agents at the second stage cannot compensate the errors of the first agent, irrespectively
Figure 8. Demonstration of the difference between the two architectures (1, 4)[1,3,(3,2)]
and (1, 5)[1,3,6] for an information bound of D = (0.4, 2.6, 2.6, 2.6, 0.4, . . . , 0.4).
18
(0,3)[1,1,8](1,0)[1,8,1](1,4)[1,3,(3,2)](1,5)[1,3,6](2,4)[1,1,(2,4)]0.00.20.40.60.8hUi(D1,D2,...,D10)=(1.6,0.1,...,0.1)[bit](0,3)[1,1,8](1,0)[1,8,1](1,4)[1,3,(3,2)](1,5)[1,3,6](2,4)[1,1,(2,4)]0.00.20.40.60.8(D1,D2,D3,...,D10)=(1.0,0.5,0.1,...,0.1)[bit](1,4)[1,3,(3,2)](1,5)[1,3,6]0.00.20.40.60.81.0hUi12345678910agenti0.00.51.01.52.02.5DKL(Pikpi)[bit](1,4)[1,3,(3,2)]0123456789agenti0.00.51.01.52.02.5DKL(Pikpi)[bit](1,5)[1,3,6)]04812162024responsesa04812162024phonecallsw(1,4)[1,3,(3,2)]04812162024responsesa04812162024phonecallsw(1,5)[1,3,6]PerformanceProcessedinformationEffectivepolicyp(aw) of their capacity. In contrast, for (1, 5), the connection between the second stage and
the executing stage can be changed freely, which leads to ignoring the first agent and
letting the three agents in the second stage determine the distribution of phone calls
completely. In this sense, (1, 5) is more robust to errors in the first filter than (1, 4).
5.3 Systematic performance comparison
In this section, we move away from an explicit toy example to a broad performance
comparison of all architectures for N (cid:54) 3, averaged over multiple types of utility func-
tions and a large number of resource constraints (as defined below). In Section 6.1,
this is supplemented with an analysis of the architectural features that best explain the
performances.
Figure 9. Utility functions on which the performances are measured.
Objectives. We compare all possible architectures for twelve different utility func-
tions, {Uk}12
k=1, defined on a world and action space of W = A = 20 elements, and
we assume the same cardinality for the range of all hidden variables. Note that the car-
dinality of the target set X for selector nodes X ∈ Xsel is given by the number of agents
it decides about. In particular, we consider three kinds of utility functions (one-to-one,
many-to-one, one-to-many) that we vary in a 2×2 paradigm, where the first dimension
19
0510152005101520worldstatesU1(diagisosm)0510152005101520U2(diagisomm)0510152005101520U3(diagnonisosm)0510152005101520U4(diagnonisomm)0510152005101520worldstatesU5(samwisosm)0510152005101520U6(samwisomm)0510152005101520U7(samwnonisosm)0510152005101520U8(samwnonisomm)05101520actions05101520worldstatesU9(maswisosm)05101520actions05101520U10(maswisomm)05101520actions05101520U11(maswnonisosm)05101520actions05101520U12(maswnonisomm) Figure 10. Proportion of conditions where the given architectures had the highest per-
formance, for all conditions, and separately for each of the three different schemes of
resource constraints.
is the number of maximum utility peaks (single, multiple) and the second dimension is
the range of utility values (binary, multi-valued). The utility functions are visualized in
Figure 9, where the three kinds of functions correspond to the three rows of the plot. A
one-to-one scenario applies to a needle-in-a-haystack situation where each world state
affords only a unique action, and vice versa each optimal action allows to uniquely iden-
tify the world state, for example an absolute identification task. A many-to-one scenario
allows for abstractions in the world states, for example in categorization when multiple
instances are judged to belong to the same class (e.g. vegetables are boiled, fruit is eaten
raw). A one-to-many scenario allows for abstractions in the action space, for example
in hierarchical motor control when a grasp action can be performed in many different
ways.
Resource limitations. We are considering three schemes of resource constraints:
(i) Same constraints for all agents.
(ii) Same constraints for all agents but one, which has a higher limit than the other
agents.
(iii) Same constraints for all but two agents, which can have a different limit and have
higher limits than all the other agents.
For (i), we compare 20 sets of constraints {D0, D1, . . .} with Di equally spaced in
the range between 0 and 3 bits, for (ii) we compare 39 sets in the same range but the
high resource agent having 1, 2 and 3 bits, and for (iii) we allow 89 sets with similar
constraints than in (ii) but additional combinations for the second high-resource agent.
Simulation results. The performance of an architecture is given by its expected util-
ity with respect to a given objective and a given information bound as defined above.
20
(−1,),[1](0,2),[1,1,8](0,4),[1,1,(2,4)](0,4),[1,1,(4,2)](1,3),[1,2,7](1,3),[1,4,5](1,4),[1,2,(2,3)](1,4),[1,3,(3,2)](1,5),[1,2,7](1,5),[1,3,6](1,5),[1,4,5](1,5),[1,5,4](1,5),[1,6,3](1,5),[1,7,2](1,),[1,9](2,2),[1,1,8](2,4),[1,1,(2,4)]0.00.10.20.30.40.5WinningrateAllconditions(0,4),[1,1,(2,4)](0,4),[1,1,(4,2)](1,3),[1,2,7](1,3),[1,4,5](1,4),[1,2,(2,3)](1,4),[1,3,(3,2)](1,5),[1,2,7](1,5),[1,3,6](1,5),[1,4,5](1,5),[1,5,4](1,5),[1,6,3](1,5),[1,7,2](1,),[1,9](2,2),[1,1,8](2,4),[1,1,(2,4)]0.00.10.20.30.4WinningrateSameconstraintsforallagents(−1,),[1](0,2),[1,1,8](0,4),[1,1,(2,4)](0,4),[1,1,(4,2)](1,3),[1,4,5](1,4),[1,2,(2,3)](1,4),[1,3,(3,2)](1,5),[1,2,7](1,5),[1,3,6](1,5),[1,4,5](1,5),[1,6,3](1,5),[1,7,2](1,),[1,9](2,2),[1,1,8](2,4),[1,1,(2,4)]0.00.10.20.30.40.5WinningrateSameconstraintsforallagentsbutone(−1,),[1](0,4),[1,1,(2,4)](0,4),[1,1,(4,2)](1,3),[1,4,5](1,4),[1,2,(2,3)](1,4),[1,3,(3,2)](1,5),[1,2,7](1,5),[1,3,6](1,5),[1,4,5](1,5),[1,5,4](1,5),[1,6,3](1,5),[1,7,2](1,),[1,9](2,2),[1,1,8](2,4),[1,1,(2,4)]0.00.20.40.6WinningrateSameconstraintsforallagentsbuttwoWins Figure 11. Architecture performances averaged over all conditions (first row), averaged
over all information bounds for each utility function (second row), and averaged over
all objectives for each information bound (last three rows).
21
(−1,),[1](0,),[1,1](1,),[1,9](0,0),[1,1,1](0,2),[1,1,8](0,3),[1,1,8](0,4),[1,1,(2,4)](0,4),[1,1,(4,2)](0,5),[1,1,8](1,0),[1,8,1](1,1),[1,8,1](1,2),[1,4,4](1,3),[1,2,7](1,3),[1,3,6](1,3),[1,4,5](1,3),[1,5,4](1,3),[1,6,3](1,3),[1,7,2](1,4),[1,2,(2,3)](1,4),[1,3,(3,2)](1,5),[1,2,7](1,5),[1,3,6](1,5),[1,4,5](1,5),[1,5,4](1,5),[1,6,3](1,5),[1,7,2](2,1),[1,1,1](2,2),[1,1,8](2,4),[1,1,(2,4)]0.00.10.20.30.40.50.60.70.80.9AveragescoreAverageoverallconditionsU1U2U3U4U5U6U7U8U9U10U11U120.60.70.80.91.0AveragescoreAverageoverallconstraints -- foreachutilityfunction(2,4)[1,1,(2,4)](1,5)[1,3,6](1,4)[1,3,(3,2)]{0.1,...,0.1}{0.2,...,0.2}{0.3,...,0.3}{0.4,...,0.4}{0.5,...,0.5}{0.6,...,0.6}{0.7,...,0.7}{0.8,...,0.8}{0.9,...,0.9}{1.0,...,1.0}{1.1,...,1.1}{1.2,...,1.2}{1.3,...,1.3}{1.4,...,1.4}{1.5,...,1.5}{1.6,...,1.6}{1.7,...,1.7}{1.8,...,1.8}{1.9,...,1.9}{2.0,...,2.0}0.10.20.30.40.50.60.70.80.9AverageexpectedutilityAverageoverallutilityfunctions -- sameconstraintforallagents(2,4)[1,1,(2,4)](1,5)[1,3,6](1,4)[1,3,(3,2)]{1.0,0.1,...,0.1}{1.0,0.3,...,0.3}{1.0,0.4,...,0.4}{1.0,0.6,...,0.6}{1.0,0.7,...,0.7}{1.0,0.9,...,0.9}{2.0,0.1,...,0.1}{2.0,0.3,...,0.3}{2.0,0.4,...,0.4}{2.0,0.6,...,0.6}{2.0,0.7,...,0.7}{2.0,0.9,...,0.9}{2.0,1.0,...,1.0}{2.0,1.2,...,1.2}{2.0,1.3,...,1.3}{2.0,1.4,...,1.4}{2.0,1.6,...,1.6}{2.0,1.7,...,1.7}{2.0,1.9,...,1.9}{3.0,0.1,...,0.1}{3.0,0.3,...,0.3}{3.0,0.4,...,0.4}{3.0,0.6,...,0.6}{3.0,0.7,...,0.7}{3.0,0.9,...,0.9}{3.0,1.0,...,1.0}{3.0,1.2,...,1.2}{3.0,1.3,...,1.3}{3.0,1.4,...,1.4}{3.0,1.6,...,1.6}{3.0,1.7,...,1.7}{3.0,1.9,...,1.9}{3.0,2.0,...,2.0}{3.0,2.2,...,2.2}{3.0,2.3,...,2.3}{3.0,2.5,...,2.5}{3.0,2.6,...,2.6}{3.0,2.7,...,2.7}{3.0,2.9,...,2.9}0.10.20.30.40.50.60.70.80.9AverageexpectedutilityAverageoverallutilityfunctions -- sameconstraintforallagentsbutone(2,4)[1,1,(2,4)](1,5)[1,3,6](1,4)[1,3,(3,2)]{1.0,1.0,0.1,...,0.1}{1.0,1.0,0.3,...,0.3}{1.0,1.0,0.4,...,0.4}{1.0,1.0,0.6,...,0.6}{1.0,1.0,0.7,...,0.7}{1.0,1.0,0.9,...,0.9}{1.0,2.0,0.1,...,0.1}{1.0,2.0,0.3,...,0.3}{1.0,2.0,0.4,...,0.4}{1.0,2.0,0.6,...,0.6}{1.0,2.0,0.7,...,0.7}{1.0,2.0,0.9,...,0.9}{1.0,3.0,0.1,...,0.1}{1.0,3.0,0.3,...,0.3}{1.0,3.0,0.4,...,0.4}{1.0,3.0,0.6,...,0.6}{1.0,3.0,0.7,...,0.7}{1.0,3.0,0.9,...,0.9}{2.0,1.0,0.1,...,0.1}{2.0,1.0,0.3,...,0.3}{2.0,1.0,0.4,...,0.4}{2.0,1.0,0.6,...,0.6}{2.0,1.0,0.7,...,0.7}{2.0,1.0,0.9,...,0.9}{2.0,2.0,0.1,...,0.1}{2.0,2.0,0.3,...,0.3}{2.0,2.0,0.4,...,0.4}{2.0,2.0,0.6,...,0.6}{2.0,2.0,0.7,...,0.7}{2.0,2.0,0.9,...,0.9}{2.0,2.0,1.0,...,1.0}{2.0,2.0,1.2,...,1.2}{2.0,2.0,1.3,...,1.3}{2.0,2.0,1.4,...,1.4}{2.0,2.0,1.6,...,1.6}{2.0,2.0,1.7,...,1.7}{2.0,2.0,1.9,...,1.9}{2.0,3.0,0.1,...,0.1}{2.0,3.0,0.3,...,0.3}{2.0,3.0,0.4,...,0.4}{2.0,3.0,0.6,...,0.6}{2.0,3.0,0.7,...,0.7}{2.0,3.0,0.9,...,0.9}{2.0,3.0,1.0,...,1.0}{2.0,3.0,1.2,...,1.2}{2.0,3.0,1.3,...,1.3}{2.0,3.0,1.4,...,1.4}{2.0,3.0,1.6,...,1.6}{2.0,3.0,1.7,...,1.7}{2.0,3.0,1.9,...,1.9}{3.0,1.0,0.1,...,0.1}{3.0,1.0,0.3,...,0.3}{3.0,1.0,0.4,...,0.4}{3.0,1.0,0.6,...,0.6}{3.0,1.0,0.7,...,0.7}{3.0,1.0,0.9,...,0.9}{3.0,2.0,0.1,...,0.1}{3.0,2.0,0.3,...,0.3}{3.0,2.0,0.4,...,0.4}{3.0,2.0,0.6,...,0.6}{3.0,2.0,0.7,...,0.7}{3.0,2.0,0.9,...,0.9}{3.0,2.0,1.0,...,1.0}{3.0,2.0,1.2,...,1.2}{3.0,2.0,1.3,...,1.3}{3.0,2.0,1.4,...,1.4}{3.0,2.0,1.6,...,1.6}{3.0,2.0,1.7,...,1.7}{3.0,2.0,1.9,...,1.9}{3.0,3.0,0.1,...,0.1}{3.0,3.0,0.3,...,0.3}{3.0,3.0,0.4,...,0.4}{3.0,3.0,0.6,...,0.6}{3.0,3.0,0.7,...,0.7}{3.0,3.0,0.9,...,0.9}{3.0,3.0,1.0,...,1.0}{3.0,3.0,1.2,...,1.2}{3.0,3.0,1.3,...,1.3}{3.0,3.0,1.4,...,1.4}{3.0,3.0,1.6,...,1.6}{3.0,3.0,1.7,...,1.7}{3.0,3.0,1.9,...,1.9}{3.0,3.0,2.0,...,2.0}{3.0,3.0,2.2,...,2.2}{3.0,3.0,2.3,...,2.3}{3.0,3.0,2.5,...,2.5}{3.0,3.0,2.6,...,2.6}{3.0,3.0,2.7,...,2.7}{3.0,3.0,2.9,...,2.9}Information-processingconstraints[bits]0.10.20.30.40.50.60.70.80.9AverageexpectedutilityAverageoverallutilityfunctions -- sameconstraintforallagentsbuttwo(2,4)[1,1,(2,4)](1,5)[1,3,6](1,4)[1,3,(3,2)] In Figure 10, we show which of the architectures won at least one condition, together
with the proportion of conditions won by each of these architectures. We can see that
(2, 4)[1,1,(2,4)] overall outperforms all the other systems (see Figure 4 for a visualiza-
tion). In the case when all agents have the same resource constraints, the architecture
(1, 4)[1,3,(3,2)] is a strong second winner, however this is not the case if one or two agents
have more resources than the rest. It is not surprising that in these situations the parallel
case with one high-resource agent distributing the work among the low resource agents,
and even the case of a single agent that does everything by himself, are both performing
well.
A closer look on the achieved expected utilities however, shows that there are several
architectures that are almost equally well performing for many conditions. In order to
increase comparability between the different utility functions, we measure performance
in terms of a relative score, which, for a given utility function and resource constraint, is
given by the architectures' expected utility divided by the maximum expected utility of
all architectures. The score averaged over all conditions is shown for each architecture
in Figure 11 in the top row. We can see that the best architectures are pretty close to each
other. As expected, the architecture that won the most conditions also has the highest
overall performance, however there are multiple architectures that are very close. The
top three architectures are
(2, 4)[1,1,(2,4)], (1, 5)[1,3,6], (1, 4)[1,3,(3,2)],
(21)
which have been visualized above (Figure 2 and 4).
A better understanding of their performances under different resource constraints
can be gathered from the remaining graphs in Figure 11. In the second row we can
see that the top three overall architectures also perform best for almost all utility func-
tions when averaged over the information bounds. The last three graphs in Figure 11
show the expected utility of each architecture averaged over all utility functions for each
information bound. We can see how the expected utility increases with higher informa-
tion bounds, for some architectures more than for others. The top three architecures
perform differently for most of the bounds, with spans of bounds where each of them
clearly outperforms the others.
6 Discussion
6.1 Analysis of the simulations
There are plenty of factors that influence the performance of each of the given architec-
tures. Here, we attempt to unfold the features that determine their performances in the
clearest way. To this end, we compare the architectures with respect to the following
quantities:
Average specialization of operational agents: the specialization (19) averaged over
all agents in the final stage of the architecture.
Hierarchical: boolean value that specifies whether an architecture is hierarchical
or not, meaning that consecutive nodes are occupied by an increasing amount of
agents.
22
Agents with direct w-access: the number of agents with direct world state access.
operational agents with direct w-access: the number of agents in the last node of
the architecture.
Number of w-bottlenecks: the total number of nodes that are missing direct access
to the world state.
Figure 12. Proposed features to explain the architectures' performances (see 6.1).
As can be seen from Figure 12, we found that these architectural features explain the
differences in performance quite well. More precisely, the architectures can be roughly
grouped into three different categories, indicated by slightly different color saturations
in Figure 12): The poorest performing group consists of architectures that have between
one and two w-bottlenecks, and therefore have only few agents with direct w-access,
in particular none of their operational agents has direct w-access. Moreover, in this
group, most architectures are not hierarchical at all, and their operational agents have
low specialization, with two exceptions that both have two w-bottlenecks.
23
0.000.250.500.751.00Scores0.00.20.40.60.8Averagespecializationofexecutingagents0.000.250.500.751.00Hierarchical0.02.55.07.510.0Agentswithdirectw-access02468Executingagentswithdirectw-access(2,4),[1,1,(2,4)](1,5),[1,3,6](1,4),[1,3,(3,2)](1,5),[1,2,7](1,5),[1,4,5](1,4),[1,2,(2,3)](1,5),[1,5,4](1,5),[1,6,3](1,),[1,9](1,5),[1,7,2](0,4),[1,1,(4,2)](2,2),[1,1,8](0,4),[1,1,(2,4)](-1,),[1](0,5),[1,1,8](1,3),[1,2,7](1,3),[1,3,6](1,3),[1,4,5](1,2),[1,4,4](1,3),[1,5,4](1,3),[1,6,3](0,2),[1,1,8](1,3),[1,7,2](2,1),[1,1,1](1,1),[1,8,1](0,3),[1,1,8](1,0),[1,8,1](0,),[1,1](0,0),[1,1,1]0.00.51.01.52.0Numberofw-bottlenecks The architectures with medium performance have maximally one w-bottleneck and
many of them are hierarchical. Here, those systems that have operational units with
high specialization are missing direct w-access, and the systems that have operational
units with direct w-access have low specialization.
All architectures in the top group have many agents with direct world-state access
and they have no w-bottlenecks. Interestingly, the best six architectures are all strictly
hierarchical. Moreover, the order of performance is almost in direct accordance with
the average specialization of the operational agents.
Overall we can say that, it is best to have as many operational units as needed
to discriminate the actions well, as long as the coordinating agents have enough re-
sources to discriminate between them properly. The architecture (1, 4)[1,1,(2,4)] has eight
operational agents, which are managed by two coordinating units, which need maxi-
mally two bits (for choosing among four agents) and one bit (for choosing among two
agents) in order to perform well. Both of the other top three architectures, (1, 5)[1,3,6]
and (1, 4)[1,3,(3,2)], have six operational agents, which are managed by three coordinat-
ing units, so that each of them needs maximally one bit. But compared to (1, 4)[1,1,(2,4)],
there are less agents to spare for the operational stage. Hence, if the operational units
have low resources, it is always a trade-off between the number of operational units and
the resources of the coordinating ones.
Another way to see why the architecture (1, 4)[1,1,(2,4)] overall outperforms all the
other high-ranked systems, might be its lower average choice-per-agent ratio, i.e. the
average number options for the decision of each agent in the system. In (1, 4)[1,1,(2,4)],
the second agent also directly observes the world state, and moreover, the choice space
of eight agents at the operational stage is split into two and four choices. Therefore,
10 = 2.6 choices per agent on average, whereas for (1, 5)[1,3,6] and
there are only 2+4+20
(1, 4)[1,3,(3,2)], there are 3+6+20
10 = 2.9.
6.2 Limitations of our analysis
The analysis presented above only provides a rough explanation of the differences in
performance. Which architecture is optimal, depends a lot on the actual information
bounds of each agent. In all of our conditions, we assumed that most agents have the
same processing capabilities, which is why there is a certain bias towards architectures
that perform well under this assumption (low variance in choice-per-agent ratio across
the agents).
Due to the large amount of Lagrange parameters in the Free Energy principle (16),
the data generation was done by running the Blahut-Arimoto-type algorithm for 10.000
different combinations of parameters for each of the architectures, for each type of
the different types of resource limitations, (i) -- (iii) in 5, and for each of the utility
functions defined in 5. For a given information bound, the corresponding parameters
were determined by looking for the points with the highest Free Energy that still respect
the bound.A better approach would be to enhance the global parameter search by a more
fine-grained local search. Another possibility is to use an evolutionary algorithm, where
each population is given by multiple sets of parameters and the information constraints
are built in by a method similar to (Chehouri et al., 2016). This works well but requires
significantly more time to process.
24
Since the Blahut-Arimoto type algorithm is not guaranteed to converge to a global
maximum, the resulting values for the expected utility and mutual information for a
given set of parameters can depend on the initialization of the algorithm. In practice, this
variation is small enough, so that it influences the average performance over multiple
conditions only by a negligable amount. However, direct comparisons of architectures
for a given information bound and utility function should be repeated multiple times to
make sure that the results are stable.
6.3 Relation to Variational Bayes and Active Inference
Above, we determined the architectures that achieve the highest expected utility under a
given resource constraint. These constraints are fulfilled by tuning the Lagrange multi-
pliers in the Free Energy principle. If the Lagrange multipliers themselves are fixed, for
instance as exchange rates between information and utility (Ortega and Braun, 2010),
or inverse temperatures in thermodynamics (Ortega and Braun, 2013), then the Free
Energy itself would be an appropriate performance measure. This is done, for example
in Bayesian model selection, which is also known as structure learning and represents
an important problem in Bayesian inference and machine learning. The Bayesian ap-
proach for evaluating different Bayesian network structures, in order to find the relation
of a given set of hidden variables that best explains a dataset D, consists in comparing
the marginal likelihood or evidence p(DS) of the structures S (Friedman and Koller,
2003). This can be seen to be analogous to a performance comparison of different
decision-making architectures measured by the Free Energy. In the simple case of one
observable Y and one hidden variable X, we have
(cid:88)
x∈X
p(yS) =
p(xS) p(yx, S)
∀y ∈ Y ,
where the likelihood p(yx, S) is assumed to be known. Given a prior p(xS) and, for
simplicity, a single observed datapoint y ∈ Y , the posterior distribution of X can be
inferred by using Bayes' rule,
p(xy, S) =
p(xS) p(yx, S)
p(yS)
∀x ∈ X .
(22)
As has been noted before (Ortega and Braun, 2013), when comparing (22) with the
Boltzmann equation (4) we can see that (22) is equivalent to the posterior P of a
bounded rational decision-maker with choice space X , prior policy p(xS), Lagrange
parameter β = 1, and utility function given by U (x) := log p(yx, S). Since the
marginal likelihood p(yS) is the normalization constant in (22), it follows immedi-
ately from (5) that log p(yS) is the optimal Free Energy Fvar[P = p(·y, S)] of this
25
decision-maker, where
Fvar[P ] :=
=
=
(cid:88)
(cid:88)
(cid:88)
x
x
x
P (x) log p(yx, S) −
P (x) log
P (x) log
p(xy, S)
P (x)
p(x, yS)
P (x)
(cid:88)
x
P (x) log
P (x)
p(xS)
+ log p(yS)
.
(23)
In Bayesian statistics, Fvar is known as the variational Free Energy, and the given
decomposition is often referred to in terms of the difference between accuracy (ex-
pected log-likelihood) and complexity (KL-divergence between prior and posterior).
It is used in the variational characterization of Bayes' rule, i.e. the approximation of
the exact Bayesian posterior p(·y, S) given by (22) in terms of a simpler -- for exam-
ple a parametrized -- distribution q by minimizing the KL-divergence between q and
p(·y, S). Since DKL(q(cid:107)p(·y, S)) = −Fvar[q] + log p(y, S), this is equivalent to the
maximization of Fvar.
The same is true for multiple hidden variables. For example, let S be the 3-step
architecture of type (1, 4) from Section 4.4 with W = Y and hidden variables X1, X2,
and X3 = A. Setting β1 = β2 = β3 = 1 and U (a, x1, x2, y) = log p(ya, x1, x2, S), we
obtain
F2(y, x1, x2) = log p(yx1, x2, S) , F1(y, x1) = log p(yx1, S) ,
and
Z(y, x1, x2) = p(yx1, x2, S) , Z(y, x1) = p(yx1, S) , Z(y) = p(yS) .
Note that, even though so far we always assumed that the utility function only depends
on the world states and actions, the equations in Sections 3, 4, and 4.4 are also valid
in the general case of U depending on all the variables in the system. The total Free
Energy for a given y ∈ Y then takes the form
(cid:17)
(cid:88)
(cid:88)
x1,x2
=
x1
(cid:16)
(cid:16)
p(x1, x2y, S)
log p(yx1, x2, S) − log
p(x1y, S)
log p(yx1, S) − log
p(x2x1, y, S)
(cid:17)
p(x2x1, S) − log
= log p(yS) .
p(x1y, S)
p(x1S)
p(x1y, S)
p(x1S)
Hence, also in this case, the logarithm of the marginal likelihood is given by the Free
Energy of the corresponding decision-making system. Choosing the multi-step archi-
tecture with the highest Free Energy is then analogous to Bayesian model selection with
the marginal likelihood or Bayesian model evidence as performance measure.
Another interesting interpretation of (23) is that here the hidden variable X can be
thought of as an action causing observed outcomes y. This is close to the framework
of Active Inference (Friston et al., 2015b, 2017b), where actions directy cause transi-
tions of hidden states, which generate outcomes that are observed by the actor. More
precisely, there the real-world process generating observable outcomes is distinguished
26
from an internal generative model describing the beliefs about the external generative
process (e.g. a Markov decision process). Observations are generated from transitions
of hidden states, which depend on the decision-maker's actions. Decision-making is
given by the optimization of a variational Free Energy analogous to (23), where the
log-likelihood is given by the generative model, which describes beliefs about the hid-
den and control states of the generative process. This way utilities are absorbed into a
(desired) prior (Ortega and Braun, 2015). There are several differences to our approach.
First, the structure of the Free Energy principle of bounded rationality originates from
the maximization of a given pre-defined external utility function under information
constraints, whereas the Free Energy principle of Active Inference aims to minimize
surprise or Bayesian model evidence, effectively minimizing the divergence between
approximate and true posterior. Second, in Active Inference, utility is transformed into
preferences in terms of prior beliefs, while in bounded rationality prior policies over
actions can be part of the optimization process, which results in specialization and ab-
straction. In constrast, Active Inference compounds utilities and priors into a single
desired prior which is fixed and does not allow to separately optimize utility and action
priors.
7 Conclusion
In this work, we have presented an information-theoretic framework to study systems
of decision-making units with limited information-processing capabilities. It is based
on an overreaching Free Energy optimization principle, which, on the one hand, allows
to compute the optimal performances of explicit architectures, and on the other hand,
produces optimal partitions of the involved choice spaces into regions of specialization.
In order to combine a given set of bounded rational agents, first the full decision-making
process is split into multiple decision steps by introducing intermediate decision vari-
ables, and then a given set of agents is distributed among these variables. We have
argued that this leads to two types of agents, non-operational units that distribute the
work among subordinates, and operational units that are doing the actual work in the
sense of choosing a particular action that either serves as an input for another agent
in the system, or represents the final decision of the full process. This "vertical" spe-
cialization is enhanced by optimizing over the agents' prior policies, which leads to
an optimal soft partitioning of the underlying choice space of each step in the system,
resulting in a "horizontal" specialization as well.
In order to illustrate the proposed framework, we have simulated and analyzed the
performances under a number of different resource constraints and tasks for all possible
3-step architectures whose information flow starts by observing a given world state and
ends with the selection of a final decision. Even though the relative architecture perfor-
mances depend crucially on the explict information-processing constraints, the overall
best performing architectures tend to be hierarchical systems of non-operational "man-
ager" units at higher hierarchical levels and operational "worker" units at the lowest
level.
Our approach is based on earlier work on information-theoretic bounded rationality
27
(Ortega and Braun, 2011, 2013; Genewein and Braun, 2013; Genewein et al., 2015)
(see also the references therein).
In particular, the N-step decision-making systems
introduced in Section 3 generalize the two-step processes studied in (Genewein and
Braun, 2013; Genewein et al., 2015). According to Simon (Simon, 1979), there are
three different bounded rational procedures that can transform intractable into tractable
decision problems: (i) Looking for satisfactory choices instead of optimal ones, (ii)
replacing global goals with tangible subgoals, and (iii) dividing the decision-making
task among many specialists. From this point of view, the decision-making process of
a single agent, given by the one-step case of information-theoretic bounded rationality
(Ortega and Braun, 2011, 2013) described in Section 2, corresponds to (i), while the
bounded rational multi-step and multi-agent decision-making processes introduced in
Section 3 and 4, can be attributed to (ii) and (iii).
The main advantage of a purely information-theoretic treatment is its universality.
To our knowledge this work is the first systematic theory-guided approach to the organi-
zation of agents with limited resources in the generality of information theory. In other
approaches, more specific methods are used instead, that are tailored to each particular
focus of study. In particular, bounded rationality has usually a very specific meaning,
often being implemented by simply restricting the cardinality of the choice space. For
example, in management theory the well-known results by Graicunas from the 1930s
(Graicunas, 1933) suggest that managers must have a limited span of control in order
to be efficient. By counting the number of possible relationships between managers
and their subordinates, he concludes that there is an explicit upper bound of five or
six subordinates. Of course, there are many cases of successful companies today that
disagree with Graicunas' claim, e.g. Apple's CEO has 17 managers that are reporting
directly to him. However, current management experts think that the optimal number
is somewhere between 5 and 12. The idea of restricting the cardinality of the space of
decision-making is also studied for operational agents. For example in (Camacho and
Persky, 1988), Camacho and Persky explore the hierarchical organization of specialized
producers with a focus on production. Even though their treatment is more abstract and
more general than many preceeding studies, their take on bounded rationality is very
explicit and based on the assumption that the number of elementary parts that form a
product, as well as the number of possibilities of each part, are larger than a single
individual can handle. Similarly, in most game theoretic approaches that are based on
automaton theory (Neyman, 1985; Abreu and Rubinstein, 1988; Hern´andez and Solan,
2016), the boundedness of an agent's rationality is expressed by a bound on the number
of states of the automaton. Most of these non-information theoretic treatments consider
cases when there is a hard upper bound on the number of options, but they usually lack a
probabilistic description of the behaviour in cases when the number of options is larger
than the given bound.
The work by Geanakoplos and Milgrom (1991) uses "information" to describe the
limited attention of managers in a firm. But here, the term is used more informally, and
not in the classical information-theoretical sense. However, one of their results suggests
that "firms with more prior information about parameters [...] will employ less able
managers, or give their managers wider spans of control" (Geanakoplos and Milgrom,
1991, p. 207). This observation is in line with information-theoretic bounded rationality,
since by optimizing over priors in the Free Energy principle, the required processing-
28
information is decreased compared to the case of non-optimal priors, so that less able
agents can perform a given task, or similarly, an agent with a higher information bound
can have a larger choice space.
In neuroscience, the variational Bayes approach explained in Section 6.3 has been
proposed as a theoretical framework to understand brain function in terms of Active
Inference (Friston, 2009, 2010; Friston et al., 2015a,b, 2017a,b), where perception is
modelled as variational Bayesian inference over hidden causes of observations. There,
a processing node (usually a neuron) is limited in the sense that it can only linearly
combine a set of input signals into a single output signal. Decision-making is mod-
elled by approximating Bayes' rule in terms of these basic operations, and then tuning
the weights of the resulting linear transformations in order to optimize the Free En-
ergy (23). Hence, there, the Free Energy serves as a tool to computationally simplify
Bayesian inference on the neuronal level, whereas our Free Energy principle is a tool
to computationally trade off expected utility and processing costs, providing an abstract
probabilistic description of the best possible choices when the information-processing
capability is limited.
In the general setting of approximate Bayesian inference, there are many interesting
algorithms and belief update schemes, for example belief propagation in terms of mes-
sage passing on factor graphs (see e.g. Yedidia et al., 2005). These algorithms make use
of the notion of the Markov boundary (minimal Markov blanket) of a node X, which
consists of the nodes that share a common factor with X (so-called neighbours). Con-
ditioned on its Markov boundary a given random variable is independent of all other
variables in the system, which allows to approximate marginal probabilities in terms of
local messages between neighbours. These approximations are generally only exact on
tree-like factor graphs without loops (M´ezard and Montanari, 2009, Thm. 14.1). This
raises the interesting question of whether such algorithms could also be applied to our
setting. First, it should be noted that variational Bayesian inference constitutes only
a subclass of problems that can be expressed by utility optimization with information
constraints. In this subclass, all random variables have to appear either in utility func-
tions, that is they have to be given as log-likelihoods, or they have to appear in marginal
distributions that are kept fixed -- see for example the definition of the utility in the in-
ference example above where U (a, x1, x2, y) = log p(ya, x1, x2, S) compared to the
utility functions of the form U (w, a) used throughout the paper that leave all interme-
diate random variables X1, . . . , XN−1 unspecified. Second, while it may be possible to
exploit the notion of Markov blankets by recursively computing free energies between
the nodes in a similar fashion to message-passing, there can also be contributions from
outside the Markov boundary, for example when the action node has to take an expec-
tation over possible world states that lie outside the Markov boundary. Finally, it may
be interesting to study whether message passing algorithms can be extended to deal
with our general problem setting and at least to approximately generate the same kind
of solutions as Blahut-Arimoto, even though in general we do not have tree-structured
graphs.
There are plenty of other possible extensions of the basic framework introduced in
this work. Marschak and Reichelstein (1998) study multi-agent systems in terms of
communication cost minimization, while ignoring the actual decision-making process.
One could combine our model with the information bottleneck method (Tishby et al.,
29
1999) and explicitly include communication costs in order to study more general agent
architectures, in particular systems with non-directed information flow. Moreover, we
have seen in our simulations that specialization of operational agents is an important
feature shared among all of the best performing architectures. In the biological liter-
ature, specialization is often paired with modularity. For example Kashtan and Alon
(2005) and Wagner et al. (2007) show that modular networks are an evolutionary con-
sequence of modularly varying goals. Similarly, it would be interesting to study the
effects of changing environments on specialization, abstraction, and optimal network
architectures of systems of bounded rational agents.
Acknowledgement
This study was funded by the European Research Council (ERC-StG-2015-ERC Start-
ing Grant, Project ID: 678082, "BRISC: Bounded Rationality in Sensorimotor Coordi-
nation").
Appendix
7.1 Proof of (13)
The Free Energy functional F that is optimized in the Free Energy principle (12) is
given by
F[P1, . . . , PN ] =
p(x) F0,loc(x) ,
(cid:1)
where x := (x0, . . . , xN ), and for all k ∈ {0, . . . , n}
p(x) = ρ(x0) P1
Fk,loc(x) = U (x0, xN ) −
By writing
F0,loc(x) = Fk,loc(x) −
1
βk
log
where x<k := (x0, . . . , xk−1), and
Rk(x<k) :=
(cid:88)
i<k
(cid:88)
x
(cid:0)x1
(cid:12)(cid:12)x1
(cid:88)
sel, x1
in
1
βi
i>k
(cid:0)xN
(cid:12)(cid:12)xN
(cid:1) ,
··· PN
Pi(xixi
log
pi(xixi
sel, xi
sel)
sel, xN
in
in)
.
Pk(xkxk
pk(xkxk
in)
sel, xk
sel) − Rk(x<k) ,
in)
Pi(xixi
pi(xixi
sel, xi
sel)
,
1
βi
log
30
we obtain for any k ∈ {0, . . . , n},
sel, xk
F[P1, . . . , PN ] =
p(xk
(cid:88)
xk
sel,xk
in
1
βk
−
DKL
(cid:20)(cid:88)
(cid:0)Pk(·xk
in)
xk
Pk(xkxk
sel, xk
(cid:88)
sel)(cid:1)(cid:21)
(cid:13)(cid:13)pk(xkxk
in)
xc
−
sel, xk
in)
p(xcx)Fk,loc(x)
(cid:88)
p(x<k)Rk(x<k)
x<k
utility Fk(x) =(cid:80)
with x = (xk, xk
In this form, we can see that
optimizing for Pk yields the Boltzmann distribution (13) with respect to the effective
sel, xk
in) and xc := (x0, . . . , xN ) \ x.
xc p(xcx)Fk,loc(x) as defined in (14).
References
Abreu, D. and Rubinstein, A. (1988). The structure of nash equilibrium in repeated
games with finite automata. Econometrica, 56(6):1259 -- 1281.
Acerbi, L., Vijayakumar, S., and Wolpert, D. M. (2014). On the origins of suboptimality
in human probabilistic inference. PLOS Computational Biology, 10(6):1 -- 23.
Arimoto, S. (1972). An algorithm for computing the capacity of arbitrary discrete
memoryless channels. IEEE Transactions on Information Theory, 18(1):14 -- 20.
Aumann, R. J. (1997). Rationality and bounded rationality. Games and Economic
Behavior, 21(1):2 -- 14.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning (Information Science
and Statistics). Springer-Verlag New York, Inc., Secaucus, NJ, USA.
Blahut, R. E. (1972). Computation of channel capacity and rate-distortion functions.
IEEE Transactions on Information Theory, 18(4):460 -- 473.
Burns, E., Ruml, W., and Do, M. B. (2013). Heuristic search when time matters. Journal
of Artificial Intelligence Research, 47(1):697 -- 740.
Camacho, A. and Persky, J. J. (1988). The internal organization of complex teams:
Bounded rationality and the logic of hierarchies. Journal of Economic Behavior &
Organization, 9(4):367 -- 380.
Chehouri, A., Younes, R., Perron, J., and Ilinca, A. (2016). A constraint-handling tech-
nique for genetic algorithms using a violation factor. Journal of Computer Sciences,
12(7):350 -- 362.
Csisz´ar, I. and Tusn´ady, G. (1984). Information geometry and alternating minimization
procedures. Statistics and Decisions, Supplement Issue, 1:205 -- 237.
DeCanio, S. J. and Watkins, W. E. (1998). Information processing and organizational
structure. Journal of Economic Behavior & Organization, 36(3):275 -- 294.
31
Edward, V., Noah, G., L., G. T., and B., T. J. (2014). One and done? optimal decisions
from very few samples. Cognitive Science, 38(4):599 -- 637.
Friedman, N. and Koller, D. (2003). Being bayesian about network structure. a bayesian
approach to structure discovery in bayesian networks. Machine Learning, 50(1):95 --
125.
Friston, K. J. (2009). The free-energy principle: a rough guide to the brain? Trends in
Cognitive Sciences, 13(7):293 -- 301.
Friston, K. J. (2010). The free-energy principle: a unified brain theory? Nature Reviews
Neuroscience, 11.
Friston, K. J., Levin, M., Sengupta, B., and Pezzulo, G. (2015a). Knowing one's place:
a free-energy approach to pattern regulation. Journal of The Royal Society Interface,
12(105).
Friston, K. J., Lin, M., Frith, C., and Pezzulo, G. (2017a). Active inference, curiosity
and insight. Neural Computation, 29(10):2633 -- 2683.
Friston, K. J., Parr, T., and de Vries, B. (2017b). The graphical brain: Belief propagation
and active inference. Network Neuroscience, 1(4):381 -- 414.
Friston, K. J., Rigoli, F., Ognibene, D., Mathys, C., Fitzgerald, T., and Pezzulo, G.
(2015b). Active inference and epistemic value. Cognitive Neuroscience, 6(4):187 --
214. PMID: 25689102.
Geanakoplos, J. and Milgrom, P. (1991). A theory of hierarchies based on limited man-
agerial attention. Journal of the Japanese and International Economies, 5(3):205 --
225.
Genewein, T. and Braun, D. A. (2013). Abstraction in decision-makers with limited
information processing capabilities. NIPS workshop on Planning with Information
Constraints.
Genewein, T., Leibfried, F., Grau-Moya, J., and Braun, D. A. (2015). Bounded ratio-
nality, abstraction, and hierarchical decision-making: An information-theoretic opti-
mality principle. Frontiers in Robotics and AI, 2:27.
Gershman, S. J., Horvitz, E. J., and Tenenbaum, J. B. (2015). Computational rationality:
A converging paradigm for intelligence in brains, minds, and machines. Science,
349(6245):273 -- 278.
Gigerenzer, G. and Selten, R. (2001). Bounded Rationality: The Adaptive Toolbox.
MIT Press.
Graicunas, V. A. (1933). Relationship in organization. Bulletin of the International
Management Institute, (7):39 -- 42.
Hern´andez, P. and Solan, E. (2016). Bounded computational capacity equilibrium. Jour-
nal of Economic Theory, 163:342 -- 364.
32
Howes, A., Lewis, R. L., and Vera, A. (2009). Rational adaptation under task and pro-
cessing constraints: Implications for testing theories of cognition and action. Psy-
chological Review, 116(4):717 -- 751.
Jain, P. and Kar, P. (2017). Non-convex Optimization for Machine Learning. Now
Foundations and Trends.
Jones, B. D. (2003). Bounded rationality and political science: Lessons from public
administration and public policy. Journal of Public Administration Research and
Theory: J-PART, 13(4):395 -- 412.
Kaelbling, L. P., Littman, M. L., and Cassandra, A. R. (1998). Planning and acting in
partially observable stochastic domains. Artificial Intelligence, 101(1):99 -- 134.
Kappen, H. J., G´omez, V., and Opper, M. (2012). Optimal control as a graphical model
inference problem. Machine Learning, 87(2):159 -- 182.
Kashtan, N. and Alon, U. (2005). Spontaneous evolution of modularity and network
motifs. Proceedings of the National Academy of Sciences, 102(39):13773 -- 13778.
Knight, F. (1921). Risk, Uncertainty and Profit. Cambridge, MA: Houghton Mifflin.
Lewis, R. L., Howes, A., and Singh, S. (2014). Computational rationality: Linking
mechanism and behavior through bounded utility maximization. Topics in Cognitive
Science, 6(2):279 -- 311.
Lipman, B. L. (1995). Information processing and bounded rationality: A survey. The
Canadian Journal of Economics / Revue canadienne d'Economique, 28(1):42 -- 67.
Marschak, T. and Reichelstein, S. (1998). Network mechanisms, informational effi-
ciency, and hierarchies. Journal of Economic Theory, 79(1):106 -- 141.
Mattsson, L.-G. and Weibull, J. W. (2002). Probabilistic choice and procedurally
bounded rationality. Games and Economic Behavior, 41(1):61 -- 78.
McKelvey, R. D. and Palfrey, T. R. (1995). Quantal response equilibria for normal form
games. Games and Economic Behavior, 10(1):6 -- 38.
M´ezard, M. and Montanari, A. (2009). Information, Physics, and Computation. Oxford
Graduate Texts.
Neyman, A. (1985). Bounded complexity justifies cooperation in the finitely repeated
prisoners' dilemma. Economics Letters, 19(3):227 -- 229.
Ochs, J. (1995). Games with unique, mixed strategy equilibria: An experimental study.
Games and Economic Behavior, 10(1):202 -- 217.
Ortega, P. A. and Braun, D. A. (2010). A conversion between utility and information.
The Third Conference on Artificial General Intelligence, pages 115 -- 120.
33
Ortega, P. A. and Braun, D. A. (2011). Information, Utility and Bounded Rationality,
pages 269 -- 274. Springer Berlin Heidelberg, Berlin, Heidelberg.
Ortega, P. A. and Braun, D. A. (2013). Thermodynamics as a theory of decision-making
with information-processing costs. Proceedings of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences, 469(2153).
Ortega, P. A. and Braun, D. A. (2015). What is epistemic value in free energy models
of learning and acting? a bounded rationality perspective. Cognitive Neuroscience,
6(4):215 -- 216. PMID: 25990838.
Radner, R. (1993). The organization of decentralized information processing. Econo-
metrica, 61(5):1109 -- 1146.
Russell, S. J. and Subramanian, D. (1995). Provably bounded-optimal agents. Journal
of Artificial Intelligence Research, 2(1):575 -- 609.
Shannon, C. E. (1959). Coding theorems for a discrete source with a fidelity criterion.
IRE International Convention Record, 7:142 -- 163.
Simon, H. A. (1943). A Theory Of Administrative Decision. PhD thesis, University of
Chicago.
Simon, H. A. (1955). A behavioral model of rational choice. The Quarterly Journal of
Economics, 69(1):99 -- 118.
Simon, H. A. (1979). Rational decision making in business organizations. The Ameri-
can Economic Review, 69(4):493 -- 513.
Sims, C. A. (2003).
Implications of rational inattention. Journal of Monetary Eco-
nomics, 50(3):665 -- 690. Swiss National Bank/Study Center Gerzensee Conference
on Monetary Policy under Incomplete Information.
Spiegler, R. (2011). Bounded Rationality and Industrial Organization. Oxford Univer-
sity Press: Oxford.
Still, S. (2009).
Information-theoretic approach to interactive learning. EPL (Euro-
physics Letters), 85(2):28005.
Tishby, N., Pereira, F. C., and Bialek, W. (1999). The information bottleneck method.
pages 368 -- 377.
Tishby, N. and Polani, D. (2011).
Perception-Action Cycle: Models, Architectures, and Hardware. Springer.
Information theory of decisions and actions.
In
Todorov, E. (2009). Efficient computation of optimal actions. Proceedings of the Na-
tional Academy of Sciences, 106(28):11478 -- 11483.
von Neumann, J. and Morgenstern, O. (1944). Theory of Games and Economic Behav-
ior. Princeton University Press.
34
Wagner, G. P., Pavlicev, M., and Cheverud, J. M. (2007). The road to modularity. Nature
Reviews Genetics, 8:921.
Wolpert, D. H. (2006). Information Theory -- The Bridge Connecting Bounded Rational
Game Theory and Statistical Physics, pages 262 -- 290. Springer Berlin Heidelberg,
Berlin, Heidelberg.
Yedidia, J. S., Freeman, W. T., and Weiss, Y. (2005). Constructing free-energy ap-
proximations and generalized belief propagation algorithms. IEEE Transactions on
Information Theory, 51(7):2282 -- 2312.
35
|
1503.00071 | 1 | 1503 | 2015-02-28T06:06:33 | Crowd Congestion and Stampede Management through Multi Robotic Agents | [
"cs.MA"
] | Crowd management is a complex, challenging and crucial task. Lack of appropriate management of crowd has, in past, led to many unfortunate stampedes with significant loss of life. To increase the crowd management efficiency, we deploy automated real time detection of stampede prone areas. Then, we use robotic agents to aid the crowd management police in controlling the crowd in these stampede prone areas. While doing so, we aim for minimum interference by robotic agents in our environment. Thereby not disturbing the ambiance and aesthetics of the place. We evaluate the effectiveness of our model in dealing with difficult scenarios like emergency evacuation and presence of localized congestion. Lastly, we simulate a multi agent system based on our model and use it to illustrate the utility of robotic agents for detecting and reducing congestion. | cs.MA | cs | Crowd Congestion and Stampede Management
through Multi Robotic Agents
Garima Ahuja and Kamalakar Karlapalem
Center for Data Engineering,
International Institute of Information Technology Hyderabad, India
{[email protected],[email protected]}
Abstract. Crowd management is a complex, challenging and crucial
task. Lack of appropriate management of crowd has, in past, led to
many unfortunate stampedes with significant loss of life. To increase the
crowd management efficiency, we deploy automated real time detection
of stampede prone areas. Then, we use robotic agents to aid the crowd
management police in controlling the crowd in these stampede prone ar-
eas. While doing so, we aim for minimum interference by robotic agents
in our environment. Thereby not disturbing the ambience and aesthetics
of the place. We evaluate the effectiveness of our model in dealing with
difficult scenarios like emergency evacuation and presence of localized
congestion. Lastly, we simulate a multi agent system based on our model
and use it to illustrate the utility of robotic agents for detecting and
reducing congestion.
Keywords: Crowd Management, Stampede, Congestion, Localized Con-
gestion, Emergency Evacuation
1
Introduction
Managing large crowds is a difficult task. Ineffective execution of this task can
potentially lead to stampedes. History suggests that crowd management police
appointed to prevent such incidents have not been very successful. During a coro-
nation festival in Russia, a police force of 1,800 men failed to control the crowd,
leading to loss of 1,389 lives by trampling [4]. Crowd management is a team
task, it requires strategic communication to figure out where to lead the crowd.
And due to perennial availability of agents, communication between agents is
more reliable than communication between humans. Moreover, these incidents
are not entirely unexpected, they repeatedly occur at similar occasions like re-
ligious gatherings [1,3], music concerts, sports tournaments, etc [2,5]. Wherever
large crowds gather in a relatively small area, risk of stampede is high. And that
is why police officers in large numbers are assigned at these places to ensure
safe management of crowd. To help the police officers in crowd management, we
propose a multi agent based solution. More specifically, in this paper, we deal
with the problem of congestion detection and congestion control in large crowds.
We assume a two dimensional space on ground where humans can move in any
5
1
0
2
b
e
F
8
2
]
A
M
.
s
c
[
1
v
1
7
0
0
0
.
3
0
5
1
:
v
i
X
r
a
2
Garima Ahuja and Kamalakar Karlapalem
arbitrary direction. Given such a scenario, we propose automated strategies to
detect congested areas of the field and an automated strategy for congestion
reduction. As demonstrated in our simulations, the agents we propose can suc-
cessfully detect and control congestion for obedient crowds. The agents can also
successfully detect presence of in-obedient crowds and request police intervention
in such scenarios.
The organization of the paper is as follows. In section 2, we describe our
crowd management model. Following which, we present our congestion detection
strategies and address congestion control in sections 3 and 4 respectively. Then,
we illustrate the results in section 5. In section 6, we discuss the existing work in
crowd dynamics modeling and crowd management. Finally we present conclusion
in section 7.
2 Crowd Management Agents
We have two kinds of agents for crowd management: Congestion Detecting
Agents (CDAs) and Congestion Controlling Agents (CCAs). Congestion detect-
ing agents inspect parts of the field and report presence or absence of congestion
in them. Congestion controlling agents take pro-active measures to reduce con-
gestion in a congested area.
Congestion Detecting Agents (CDAs) are computing machines present off-
site. They receive location coordinates of humans present inside the field. (We
do not delve into the details of how we receive the location coordinates, possible
ways to obtain positional information are GPS or RFID). CDAs then use this
information to check for congestion.
Congestion Controlling Agents (CCAs) on the other hand are flying robots
(like droids). They have a spotlight and a speaker attached to them to allow them
to illuminate the areas below them and to announce instructions. CCAs receive
instructions from CDAs and execute them to achieve congestion reduction.
Our CDAs never enter the site. This helps in accomplishing minimum in-
terference by robotic agents in our human environment. We do not send any
robotic agent inside the field unless there is a stampede risk; which is when,
it becomes absolutely necessary to do so. In the following sections, we present
how we achieve congestion detection and congestion control using the agents
described.
3 Congestion Detection
For computational ease and efficiency, we divide our field into smaller rectangular
areas (henceforth called grids). The division allows multiple CDAs to parallely
inspect grids.
We fork as many CDAs as there are grids to parallely inspect the grids and
check for presence of congestion in them. For the grids for which presence of
congestion is reported, we find the extent and boundaries of the congested area
using the breadth first search algorithm Algorithm 1.
Title Suppressed Due to Excessive Length
3
Algorithm 1 Given a grid id as input, return grid ids of the neighbouring
congested grids.
1: Given : gridId
2: list congestedGridIds ← emptyList
3: list toBeCheckedGridIds ← emptyList
4: toBeCheckedGridIds.add(gridId)
5: while toBeCheckedGridIds (cid:54)= empty do
6:
7:
8:
currentGridId ← toBeCheckedGridIds.getF irstElement()
/*
Strategies for checkForCongestion() function are described in subsections of sec-
tion 3.
*/
if checkF orCongestion(currentGridId) == true then
//congestion present
congestedGridIds.add(currentGridId)
list neighbouringGridIds ← getF ourOverlappingGridIds(currentGridId)
for each neighbouringGridId in neighbouringGridIds do
/*
One of the overlapping grids of a child grid would be the parent congested
grid itself, gridId of which would be present in congestedGridIds. Since we
have already checked the parent congested grid, we do not add it to the
toBeCheckedGridIds list.
*/
if neighbouringGridId not present in congestedGridIds then
toBeCheckedGridIds.add(neighbouringGridId)
end if
end for
end if
/*
The first element of the toBeCheckedGridIds list is currentGridId. We finished
inspecting the grid corresponding to the currentGridId, so we remove it from the
list.
*/
toBeCheckedGridIds.removeF irstElement()
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27: end while
28: return congestedGrids
4
Garima Ahuja and Kamalakar Karlapalem
The CCAs need to know the boundaries of the congested areas to resolve
congestions. We discuss the details of congestion control in section 4.
In the following subsections, we describe the different strategies that could
be used for congestion detection in a grid.
3.1 Naive Strategy
We can estimate the capacity of a grid based on its area. A CDA would then
count the number of humans present inside the grid and report presence of
congestion if this number is more than the capacity of the grid. Otherwise, the
CDA would report absence of congestion.
3.2 Free Flow Strategy
CDA counts the number of humans who entered or left the grid in one unit time.
If the number of humans who left the grid is more than the number of humans
who entered, we say that there is free flow in the grid and report absence of
congestion; otherwise we report presence of congestion.
3.3 Trapped Humans Strategy
We report presence of congestion in the grid when there are one or more trapped
humans present. A human is trapped when he has more than minN neighbours
in his neighbourhood where the neighbourhood is defined as a circle of radius r
around a human. Figure 1 shows one such scenario. This strategy is not computa-
tionally efficient as it involves inspecting neighbourhood of every human present
in the grid. However, search space can be decreased using Locality Sensitive
Hashing (LSH), thereby improving the performance.
Fig. 1. Trapped humans are represented by filled circles.
3.4 A Macro Micro Strategy
The strategies suggested so far use only the location coordinates of humans
present inside the field for congestion detection. We can however use a human’s
Title Suppressed Due to Excessive Length
5
previous and current positions to determine the direction he/she is moving in. We
can also record the time stamps as and when we receive the location coordinates
and use it to determine his/her speed.
Another important piece of information we have not considered in the other
strategies is the human behaviour. Humans tend to visit places in groups of
families or friends [13]. In a particular group, all the humans tend to be close
to each other and move in the same direction. Their direction of movement also
depends on their inherent tendency to avoid collisions.
We now propose a macro micro strategy for congestion detection which uses
all the information we have not been utilizing till now. We divide the humans
into groups where a group is defined as a connected group of humans moving in
the same direction. The congestion detection then proceeds as follows:
(i) At the macro or inter-group level, we count the number of groups that are
moving towards a fixed point. We report presence of congestion if this number is
greater than or equal to three. (Our simulations show that when the number of
conflicting directions is two, the human tendency to avoid collisions is ensuring
hassle free movement.)
(ii) At the micro or intra-group level, we further divide the groups into con-
nected sub-groups of humans moving with similar speeds. We report presence
of congestion if speed of a sub-group is more than the speed of the sub-group
moving in front of it (front is along the direction of movement of the group).
(iii) If both the conditions stated above do not hold, we report absence of
congestion.
4 Congestion Control
Our simulation results (presented in the next section) show that the macro micro
strategy performs better than the other strategies for congestion detection. We
therefore use it and, based on it, devise a strategy for congestion control.
Recall that the CDA not just reports presence or absence of congestion in
a grid but also finds out the boundaries of the congested area. Thus, the CDA
knows where to lead the humans present in the congested area, i.e. out of the
congestion boundary. The CDA calls a CCA and then they work together to
achieve congestion control. One of the following cases will arise:
(i) There are no conflicting directions (intersecting directions) but the speed
of a sub-group is greater than the speed of the sub-group moving in front of it
(front is along the direction of movement of the group). Recall that the CCA has
a spotlight attached to it to illuminate the area below. The CCA starts moving
above the sub-group while shining its spotlight on the sub-group and playing a
pre-recorded message requesting the humans to slow down.
(ii) There are conflicting directions but at least one of the group has space in
front to proceed (see Figure 2). The CCA then positions itself above that group
and leads the group in its target direction while requesting the other groups to
wait for it to return. The group being led is requested to follow the spotlight and
maintain the speed of the spotlight. After the group is led out of the congested
6
Garima Ahuja and Kamalakar Karlapalem
area, the new scenario could conform to case (ii) or case (iii) and is handled
accordingly recursively till congestion is resolved.
Fig. 2. When there is space in front of at least one of the groups, that group is led
while the others wait.
(iii) There are conflicting directions and none of the groups have space in front
to proceed (see Figure 3). Using the pre-recorded messages and the spotlight, the
CCA leads one of the group through a semi circular path around the congested
area while requesting the other groups to wait for it to return. The center of the
semi circular path is given by (1) and (2) and the radius is given by (3).
centerx = {(ax + bx)/2dist(a, b) ≥ dist(i, j)∀i, j ∈ humansInCongestedArea}
(1)
centery = {(ay + by)/2dist(a, b) ≥ dist(i, j)∀i, j ∈ humansInCongestedArea}
(2)
radius = {dist(a, b)/2dist(a, b) ≥ dist(i, j)∀i, j ∈ humansInCongestedArea}
(3)
Fig. 3. When there is no space in front of any of the groups, one of the group is led
through a semi circular path around the congested area while the other groups wait.
After the group is led out of the congested area, the new scenario could con-
form to case (ii) or case (iii) and is handled accordingly recursively till congestion
is resolved.
Title Suppressed Due to Excessive Length
7
Notice in Figure 3 that the groups finally end up moving in the direction
they originally intended to. Only difference being that the CCAs made them
move around the congested area while they were trying to move through the
congested area.
Also, note that the isVacant() function used in the congestion boundary de-
tection algorithm (presented in section 3) would use results from naive strategy,
not macro micro strategy. The reason being that the CCAs need to know where
there is empty space so that they can lead the conflicting groups accordingly.
Macro Micro strategy would report absence of congestion for densely populated
grids too as long as there are no conflicting directions present and therefore it
cannot be used for the isVacant() function.
The environment we are dealing with is non-deterministic. The outcome of
the congestion control strategy would depend on the obedience of the crowd.
Therefore, it is crucial to take feedback from the environment and act accord-
ingly. To that end, the CDA keeps recalculating humans’ speeds and directions
to calculate the percentage of disobedience. A human is disobedient if he/she
does not have the speed or direction he/she is being instructed to have. If the
percentage of disobedience is high, the CDA requests the police in charge to take
over and manage the crowd.
5 Results
We use Helbing and Molnar’s social force model [9] for crowd simulation. The
motion of each human is governed by the summation of all the forces exerted on
and by the human. The following forces help us in modeling group behavior:
(i) Intent: A human exerts force in the direction he wants to move in.
(ii) Cohesion: A human exerts force so as to remain close to his group.
(iii) Coherency: A human exerts force so as to walk in the same direction
as his group.
(iv) Momentum: Inertial force.
(v) Avoidance: A human exerts force to avoid colliding with other humans
and obstacles.
The first scenario we consider is an emergency evacuation scenario. To sim-
ulate that, we add a force towards the exit door to the above forces. In an
evacuation scenario, we do not want our congestion controlling agents to inter-
rupt the evacuation if the humans are moving towards the exit door in a calm
manner. So we want absence of congestion to be reported if the movement of
the crowd is not unruly. Figure 4 shows such a scenario where the humans are
moving in an orderly manner towards the exit gate while avoiding obstacles.
Naive strategy, free flow strategy and trapped humans strategy all fail in the
evacuation scenario because they cannot distinguish between a calm crowd and
an unruly one. However, macro micro strategy recognizes the gradually changing
direction as one direction and reports absence of congestion if speed conflicts are
absent. Therefore, we will not unnecessarily interrupt evacuation if the crowd
is calm (speed conflicts absent). However as soon as the crowd becomes unruly,
8
Garima Ahuja and Kamalakar Karlapalem
speed conflicts will be detected and remedial measures will be taken. Hence, our
crowd managing model is able to successfully handle the emergency evacuation
scenario without any unnecessary interruptions.
Fig. 4. Outputs of different congestion detection strategies when humans are moving
calmly towards the golden exit gate while avoiding obstacles. Direction of the compass
represents direction of the human. Red compasses represent congestion.
We now consider a localized congestion scenario i.e. a scenario where con-
gestion is localized to a relatively small area (in this case a grid). To simulate
this scenario, to the above listed forces we add a force directed towards a fixed
point. Results of various congestion detection strategies for localized congestion
scenario is shown in Figure 5. Naive strategy, free flow strategy and trapped hu-
mans strategy again fail in this scenario because the grid is sparsely populated
and number of humans entering the grid is zero. However, macro micro strat-
egy is able to identify the root cause behind congestion which is the conflicting
directions and reports congestion successfully.
Fig. 5. Outputs of different congestion detection strategies for the localized conges-
tion scenario. Direction of the arrow represents direction of the human. Red humans
represent congestion.
Title Suppressed Due to Excessive Length
9
Now that we have shown that the macro micro strategy works better than
other strategies for congestion detection, we test our congestion control strategy
for the above stated localized congestion scenario. The results are presented in
Figure 6. Figure 6(a) shows the present scenario. Figure 6(b) shows the blue
group being led on a semi circular path. Figure 6(c) shows the brown group
being led on a straight path. Figure 6(d) shows the scenario after congestion is
successfully resolved.
Fig. 6. Output of macro micro congestion control for the localized congestion scenario.
The present scenario is shown in (a), (b) shows the blue group being led on a semi
circular path, (c) shows the brown group being led on a straight path and (d) shows
the scenario after congestion is successfully resolved.
Finally, we simulate humans on a central park background and test our con-
gestion detection and congestion control strategies. The results are presented in
Figure 7.
Fig. 7. Situation before(a) and after(b) congestion control. Red dots represent humans
experiencing congestion.
The time taken to resolve congestion for different magnitudes of congestion
is tabulated in Table 1.
10
Garima Ahuja and Kamalakar Karlapalem
Table 1. Time taken to resolve congestion
No.
Number of hu-
mans in the con-
gested area
1
2
3
4
5
6
6
12
19
28
34
41
Number of con-
flicting
groups
in the congested
area
3
3
3
5
6
8
Maximum walk-
ing speed of a
human in metres
per second
1.4
1.2
1.4
1.3
1.3
1.2
Time taken to re-
solve
congestion
in seconds
2.2
10.9
24.3
97.6
134.0
247.8
Remember that we pointed out that the groups finally end up moving in the
direction they originally intended to. Only difference being that the CCAs made
them move around the congested area while they were trying to move through
the congested area. Since we are not making the humans move in a direction they
did not intend to, we cannot assure complete absence of congestion. Direction
conflicts will inevitably arise unless we force the humans to move in a direction
they do not want to move in (see Figure 7(b)). Nevertheless, our congestion
control strategy resolves congestion as soon as it is detected and keeps on doing
so to ensure hassle free movement of the crowd.
6 Related Work
There has been a lot of research in trying to model crowd dynamics and be-
haviour. There are cellular automata based models [6,7,11] and agent-based
models [14,15,17]. Helbing and Molnar [9] model is a social force model that
simulates motion of each individual in the crowd under the influence of the
forces exerted on the individual by other individuals and inanimate objects. An
individual can exert two types of forces, physical and social. The social forces
represent the intention of a human to move in a chosen direction and avoid colli-
sion. The resultant motion of the human is governed by sum of all forces exerted
by the human and on the human. We used this model to simulate human agents
in our experiments.
These crowd behaviour models help in increasing our understanding and pro-
vide a framework to work towards crowd management models. Helbing, Johans-
son and Al-Abideen studied the dynamics of crowd disasters [8]. They analyzed
the videos of crowd disasters during Hajj and studied crowd response and be-
haviour. Helbing and Mukerji analyzed the Love Parade disaster [10]. Sime [16]
proposed that early warning would help in efficient crowd management during
disasters. Kolli and Karlapalem [12] presented crowd management strategies in
long queues, a one-dimensional and unidirectional scenario. However, the prob-
lem of crowd management in a two-dimensional scenario using robotic agents
proposed in this paper has not been studied.
Title Suppressed Due to Excessive Length
11
7 Conclusion
In this paper, we presented a multi agent based solution to crowd management.
Our results show that the agents are able to detect the presence of congestion as
well as take remedial actions even in difficult scenarios like localized congestion
and emergency evacuation. Thus the robotic agents help the police force in
ensuring a congestion free movement of the crowd. Future work may look into
how to handle congestion in a three dimensional scenario where say there are
stairways present. Such a scenario is more dangerous as the chances of stumbling
and falling increase and therefore it needs to be addressed.
References
1. Allahabad
stampede
kills
36
kumbh
mela
pilgrims.
http://in.reuters.com/article/2013/02/11/kumbh-mela-stampede-allahabad-
update-idINDEE91907I20130211.
2. Hillsborough disaster and its aftermath. http://www.bbc.com/news/uk-19545126.
3. A history of hajj tragedies. http://www.theguardian.com/world/2006/jan/13/saudiarabia.
4. Khodynka tragedy. http://worldhistoryproject.org/1896/5/18/khodynka-tragedy.
5. The who concert disaster. http://en.wikipedia.org/wiki/The Who concert disaster.
6. J. Dijkstra, H. J. Timmermans, and A. Jessurun. A multi-agent cellular automata
system for visualising simulated pedestrian activity. In Theory and Practical Issues
on Cellular Automata, pages 29–36. Springer, 2001.
7. T. Hamagami and H. Hirata. Method of crowd simulation by using multiagent on
cellular automata. In Intelligent Agent Technology, 2003. IAT 2003. IEEE/WIC
International Conference on, pages 46–52. IEEE, 2003.
8. D. Helbing, A. Johansson, and H. Z. Al-Abideen. Dynamics of crowd disasters: An
empirical study. Physical review E, 75(4):046109, 2007.
9. D. Helbing and P. Molnar. Social force model for pedestrian dynamics. Physical
review E, 51(5):4282, 1995.
10. D. Helbing and P. Mukerji. Crowd disasters as systemic failures: analysis of the
love parade disaster. EPJ Data Science, 1(1):1–40, 2012.
11. A. Kirchner and A. Schadschneider. Simulation of evacuation processes using a
bionics-inspired cellular automaton model for pedestrian dynamics. Physica A:
Statistical Mechanics and its Applications, 312(1):260–276, 2002.
12. S. Kolli and K. Karlapalem. Mama: multi-agent management of crowds to avoid
In Proceedings of the 2013 international conference
stampedes in long queues.
on Autonomous agents and multi-agent systems, pages 1203–1204. International
Foundation for Autonomous Agents and Multiagent Systems, 2013.
13. R. Kota, V. Bansal, and K. Karlapalem. System issues in crowd simulation using
massively multi-agent systems. In Workshop on Massively Multi Agent Systems,
pages 251–257, 2006.
14. Y. Murakami, K. Minami, T. Kawasoe, and T. Ishida. Multi-agent simulation
for crisis management. In Knowledge Media Networking, 2002. Proceedings. IEEE
Workshop on, pages 135–139. IEEE, 2002.
15. T. Osaragi. Modeling of pedestrian behavior and its applications to spatial evalu-
ation. In Proceedings of the Third International Joint Conference on Autonomous
Agents and Multiagent Systems-Volume 2, pages 836–843. IEEE Computer Society,
2004.
12
Garima Ahuja and Kamalakar Karlapalem
16. J. D. Sime. Crowd facilities, management and communications in disasters. Facil-
ities, 17(9/10):313–324, 1999.
17. M. C. Toyama, A. L. Bazzan, and R. Da Silva. An agent-based simulation of
pedestrian dynamics: from lane formation to auditorium evacuation. In Proceedings
of the fifth international joint conference on Autonomous agents and multiagent
systems, pages 108–110. ACM, 2006.
|
1108.3462 | 1 | 1108 | 2011-08-17T12:13:44 | A Multiagent Simulation for Traffic Flow Management with Evolutionary Optimization | [
"cs.MA",
"nlin.AO"
] | A traffic flow is one of the main transportation issues in nowadays industrialized agglomerations. Configuration of traffic lights is among the key aspects in traffic flow management. This paper proposes an evolutionary optimization tool that utilizes multiagent simulator in order to obtain accurate model. Even though more detailed studies are still necessary, a preliminary research gives an expectation for promising results. | cs.MA | cs |
A Multiagent Simulation for Traffic Flow
Management with Evolutionary Optimization
Patryk Filipiak
Institute of Computer Science
University of Wroclaw, Poland
[email protected]
October 11, 2018
Abstract
A traffic flow is one of the main transportation issues in nowadays
industrialized agglomerations. Configuration of traffic lights is among the
key aspects in traffic flow management. This paper proposes an evolution-
ary optimization tool that utilizes multiagent simulator in order to obtain
accurate model. Even though more detailed studies are still necessary, a
preliminary research gives an expectation for promising results.
1 Traffic flow management
Early models of traffic flow (called macroscope models) treated vehicles in the
collective manner basing on the analogy to particles in a fluid [7]. Later on, more
precise (mezoscope) models were being created consecutively (mostly based on
gas kinetics) [5]. Currently, a multiagent simulation mechanism provides much
more efficient microscope model where each vehicle can be regarded separately
allowing for highly detailed analysis including collision avoidance [4], traffic
virtualization [2], interactions with pedestrians [3], etc.
Traffic flow management varies from tracing main roads average capacity
across certain area (where macroscope models give satisfactory results) to very
low-level manipulations including re-arrangement of lanes, modifying traffic
lights configuration, planning bridge locations, etc. where microscope models
are most suitable [9].
2 Multiagent traffic flow simulator
The agent-based traffic flow simulator described in [9] is the universal microscope
model. The environment for agents in this model comprises of a system of city
streets defined in XML files using the following entities illustrated in Figure 1:
1
Figure 1: Elements of environment description presented on the part of a sample
junction.
• road -- defines the one-way part of a street, i.e. an existing path leading
from one junction to another. For the sake of simplicity, roads at the
frontiers of a model are considered dead-end even though they play roles
of ingoing or outgoing roads.
• queue -- defines a single lane represented as a first-in-first-out list of ve-
hicles located on this lane one after another. Note that there has to be at
least one queue for each road at each junction.
• trajectory -- corresponds to a trajectory of vehicles crossing the junction
from a given ingoing road to a given outgoing road. Each trajectory stores
the information about other trajectories with possibile collisions with it.
• track -- an abstract pair queue-trajectory binding three entities, namely
queue at ingoing road, queue at outgoing road and trajectory from one to
another.
Each vehicle in the environment is an agent that is aware of its current
speed and exact location within a model (particularly, its relative position on
the present road). What is more, a vehicle is assigned another agent who drives
it in order to achive the driver's own aim which is to get from point A to point
B chosing the best possible way.
Driver agents are equipped with a simple perception mechanism that allows
them to acquire information about a vehicle driven by them (e.g. current lane,
driving speed or acceleration) and the current state of environment in their
proximity, particularly: other vehicles (their behaviour, distance to them, etc.)
and traffic lights.
For the sake of efficiency, described model includes the following simplifica-
tions that should be taken into account: there are traffic lights at each junction,
there are no roundabouts and the presence of pedestrians is ignored.
2
Simulation process in described model is the sequence of N ∈ N iterations
executed every τ > 0 miliseconds (by default N = 1000 and τ = 200). Each
iteration includes the following operations:
1. Apply changes in the environment (if there are any at this time period)
including the presence of colliding tracks and the state of traffic lights.
2. Refresh all agents' local information about the state of environment.
3. Refresh locations of all moving vehicles according to intentions of their
drivers considering current speed and acceleration of vehicles.
4. Remove agents that achived their aims and create new ones if possible.
Vehicle agents collect some essential information during simulation, i.e. exact
trace; number, locations and durations of stops; average speed etc. for the sake
of further analysis.
Figure 3 depicts arrangement of main streets in Wroclaw, Poland that can
be easily modelled using described agent-based simulator.
3 Optimization of traffic lights
Traffic lights are commonly used as an effective solution for car flow routine
regulation in virtually all agglomerations. Modification of traffic lights configu-
ration is the key aspect in optimization of traffic flow for the model presented
in the previous section.
In many real-world situations, traffic lights configuration settings are based
on combination of averaged theoretical data and designer's assumptions. A
multiagent simulation approach can be used instead in order to provide more
accurate data.
A cyclic sequence of consecutive changes of lights including detailed informa-
tion about duration of time when each light is on combined with dependencies
amongst all traffic lights within a given area (typically a crossroad) is referred
to as traffic lights programme. The usual representation of such programme is
a graphical timeline presented in Figure 2.
Figure 2: A sample junction with two pairs of traffic lights (left) and a corre-
sponding traffic lights programme (right).
3
Figure 3: Arrangement of main streets in Wroclaw (Poland) that can be easily
modelled with the agent-based simulator.
The merit of underlying optimization process lies in realistic accuracy pro-
vided by the agent-based simulator. Performance of contemporary hardware is
sufficiently enough for launching simulations containing even millions of agents
in place of less accurate statistical estimations that were used in previous years.
Thus, the values of a function that is being optimized are in fact results of
agent-based simulation.
3.1 Constraints
Any traffic lights programme is obliged to satisfy the following constraints:
• A change of lights cannot violate the sequence (green, yellow, red, red+yellow)
repeated continuously.
• Duration of green light must be greater or equal to a given minimal time
length whereas durations of transient lights (i.e. yellow and red+yellow)
are constant and cannot undergo any optimization.
• No collision possibilities are allowed, i.e. red light must be lit on at one
pair of traffic lights until green light turns to yellow on another pair and
remains yellow for a given time period; similar condition must be satisfied
during a change from green to red light.
3.2 Criteria
The aim is to optimize a traffic lights programme by maximizing a car flow
subject to constraints described above. Typical optimization criteria in such
case would be: maximizing the number of agents that achived their aims during
simulation, minimizing the average time of a ride, maximizing average speed of
vehicles, etc. Since most of the criteria correspond to each other an aggregation
of at least two of them provides a fair evaluation function.
4
Figure 4: A sample programme of i-th traffic lights represented in binary chro-
mosome form (up) and its corresponding timeline form (down).
3.3 Representation
Assume that T > 0 is a time length of one traffic lights cycle at a junction. Time
range [0, T ] can be discretized into T /τ intervals of τ > 0 miliseconds each as it
was mentioned in the previous section. Let tmin be the minimal and
tmax =
T
τ
− k · tmin − C,
(C = const)
(1)
be the maximal feasible time length of green light duration. Note that tmax is
determined by the time length of green light duration for k tracks colliding with
current track and some constant value reserved for transient lights.
As a result, any programme of M > 0 traffic lights can be well-defined by
the set of pairs
{(start1, t1), (start2, t2), . . . , (startM , tM )},
(2)
where starti represents the number of time interval when i-th traffic light turns
to green and remains for the next ti time intervals (i = 1, . . . , M ). Duration of
red light can be computed therefore as T − ti − C.
Evolutionary algorithms provide an efficient optimization mechanism for the
problem presented above. Let n > 0 be the lowest integer value that satisfies
condition T /τ ≤ 2n, hence
0 < tmin ≤ tmax ≤ 2n.
(3)
Any traffic lights programme expressed in the form (2) can be encoded as a
2M N -element binary chromosome containing M pairs of N -bit values (Li, Ri)
such that
starti = Li mod T /τ,
ti = tmin + [Ri mod (tmax − tmin + 1)]
(4)
(5)
5
for i = 1, . . . , M .
3.4 Algorithm
Population-Based Incremental Learning (PBIL) [1] can be used as an evolu-
tionary optimization tool for the stated problem thanks to its simplicity and
promising results in a broad range of applications.
It is also worth noticing
that even more complex evolutionary algorithms could be applied instead with-
out a significant decrease of performance due to considerably time consuming
evaluation process that is suggested in this paper.
Algorithm 1 presents the pseudocode of PBIL. It is so called Estimation of
Distribution Algorithm (EDA) [6] based on the idea of optimizing a probability
model for solution rather than implicit set of solutions. Later on, individuals
are randomly generated according to obtained probability distribution.
The requirement for collision avoidance is not satisfied by the traffic lights
programme definition itself, hence the presence of infeasible individuals in popu-
lation might be expected. Figure 5 presents a simple conflicts resolving method
for this case. If a collision possibility occurres between a pair of trajectories
then corresponding traffic lights programme can be modified in order to ensure
arbitrary set C ′ distance between green light durations. If any of modified green
light durations becomes shorter than tmin then it is automaticaly expanded to
tmin in the only possible way. Finally, if a conflict still occurres, a new random
individual is selected.
Algorithm 1 Standard PBIL algorithm with parameters 0 < θ1, θ2, θ3 < 1 on a
population of size individuals each of which is represented as d-element binary
vector.
p ← InitialProbabilityVector()
P ← RandomPopulation(p, size)
PopulationEvaluation(P )
while not TerminationCondition(P ) do
xi ← BestIndividual(P )
for k ← 1 to d do
pk ← pk · (1 − θ1) + xik · θ1
end for
for k ← 1 to d do
if UniformRandom(0, 1) ¡ θ2 then
pk ← pk · (1 − θ3) + BinaryRandom(0.5) · θ3
end if
end for
P ← RandomPopulation(p, size)
PopulationEvaluation(P )
end while
6
4 Conclusions and Perspectives
Even though the results are partial at the moment and more detailed studies
are still necessary, a preliminary research gives an expectation for promising
results. An improvement of agent-based traffic flow model could be a next step.
Also a modification of evolutionary mechanism may improve the performance of
optimization process. Replacing PBIL with Bayesian Optimization Algorithm
(BOA) [8] could ensure obtaining much more detailed model. Unlike other
evolutionary algorithms, BOA not only searches for optimal solution but also
gives some information about its structure. Lots of dependencies (e.g. among
traffic lanes and even adjacent junctions) could be modelled in terms of Bayesian
networks that are utilized in BOA. What is more, such dependency model could
include the fact that more effort should be put to traffic flow management inside
city centers than suburbs because of greater amount of junctions, narrow and/or
one-way street, etc.
As it is clearly seen, the effectiveness of agent-based approach applied for
traffic flow managements provides a broad range of ideas.
5 Acknowledgements
The multiagent simulator developed by Krzysztof Sroka within confines of his
Master's dissertation [9] was extensively used during research process for this
paper.
Figure 5: Simple conflicts resolving method.
7
References
[1] Baluja, S., Population-Based Incremental Learning: A Method for Integrating Ge-
netic Search Based Function Optimization and Competitive Learning, Technical
Report (Pittsburgh, PA: Carnegie Mellon University), 1994.
[2] Berg van den, J., Sewall, J., Lin, M., Manocha, D., Virtualized Traffic: Recon-
structing Traffic Flows from Discrete Spatio-Temporal Data in Proceedings of IEEE
VR, 2009.
[3] Godara, A., Lassarre, S., Banos, A., Simulating Pedestrian-Vehicle Interaction
in an Urban Network Using Cellular Automata and Multi-Agent Models. Lecture
Notes in Computer Science, Traffic and Granular Flow05, Part IV, 2007
[4] Hoffmann, G., M., Tomlin, C., J., Decentralized cooperative collision avoidance for
acceleration constrained vehicles. 47th IEEE Conference on Decision and Control
(CDC), 2008.
[5] Hoogendoorn, S., P., Bovy, P., H., L., State-of-the-art of Vehicular Traffic Flow
Modelling, Special Issue on Road Traffic Modelling and Control of the Journal of
Systems and Control Engineering, 2001.
[6] Larranaga, P., Lozano, J. A., Estimation of distribution algorithms: A new tool
for evolutionary computation. Kluwer Academic Publishers, Boston, 2002.
[7] Lighthill, M., H., Whitham, G., B., On kinematic waves II: a theory of traffic flow
on long, crowded roads. Proceedings of the Royal Society of London series A, 1955.
[8] Pelikan, M., Goldberg, D. E., Cantu-Paz, E., Linkage problem, distribution estima-
tion, and Bayesian networks. IlliGAL Report No. 98013: Illinois Genetic Algorithm
Laboratory, University of Illinois at Urbana-Champaign, 1998.
[9] Sroka, K., A traffic lights programme optimization, Master's Thesis, Institute of
Computer Science, University of Wroclaw, 2009.
8
|
1912.10944 | 2 | 1912 | 2019-12-26T14:47:34 | A Survey of Deep Reinforcement Learning in Video Games | [
"cs.MA",
"cs.AI",
"cs.LG"
] | Deep reinforcement learning (DRL) has made great achievements since proposed. Generally, DRL agents receive high-dimensional inputs at each step, and make actions according to deep-neural-network-based policies. This learning mechanism updates the policy to maximize the return with an end-to-end method. In this paper, we survey the progress of DRL methods, including value-based, policy gradient, and model-based algorithms, and compare their main techniques and properties. Besides, DRL plays an important role in game artificial intelligence (AI). We also take a review of the achievements of DRL in various video games, including classical Arcade games, first-person perspective games and multi-agent real-time strategy games, from 2D to 3D, and from single-agent to multi-agent. A large number of video game AIs with DRL have achieved super-human performance, while there are still some challenges in this domain. Therefore, we also discuss some key points when applying DRL methods to this field, including exploration-exploitation, sample efficiency, generalization and transfer, multi-agent learning, imperfect information, and delayed spare rewards, as well as some research directions. | cs.MA | cs | A Survey of Deep Reinforcement Learning in Video
Games
Kun Shao, Zhentao Tang, Yuanheng Zhu, Member, IEEE, Nannan Li, and Dongbin Zhao, Fellow, IEEE
1
9
1
0
2
c
e
D
6
2
]
A
M
.
s
c
[
2
v
4
4
9
0
1
.
2
1
9
1
:
v
i
X
r
a
Abstract -- Deep reinforcement learning (DRL) has made great
achievements since proposed. Generally, DRL agents receive
high-dimensional inputs at each step, and make actions according
to deep-neural-network-based policies. This learning mechanism
updates the policy to maximize the return with an end-to-
end method. In this paper, we survey the progress of DRL
methods, including value-based, policy gradient, and model-based
algorithms, and compare their main techniques and properties.
Besides, DRL plays an important role in game artificial
in-
telligence (AI). We also take a review of the achievements of
DRL in various video games, including classical Arcade games,
first-person perspective games and multi-agent real-time strategy
games, from 2D to 3D, and from single-agent to multi-agent.
A large number of video game AIs with DRL have achieved
super-human performance, while there are still some challenges
in this domain. Therefore, we also discuss some key points
when applying DRL methods to this field, including exploration-
exploitation, sample efficiency, generalization and transfer, multi-
agent learning,
imperfect information, and delayed spare re-
wards, as well as some research directions.
Index Terms -- reinforcement
learning, deep learning, deep
reinforcement learning, game AI, video games.
I. INTRODUCTION
A RTIFICIAL intelligence (AI) in video games is a long-
standing research area. It studies how to use AI tech-
nologies to achieve human-level performance when playing
games. More generally, it studies the complex interactions be-
tween agents and game environments. Various games provide
interesting and complex problems for agents to solve, making
video games perfect environments for AI research. These
virtual environments are safe and controllable. In addition,
these game environments provide infinite supply of useful data
for machine learning algorithms, and they are much faster than
real-time. These characteristics make games the unique and
favorite domain for AI research. On the other side, AI has
been helping games to become better in the way we play,
understand and design them [1].
Broadly speaking, game AI involves the perception and
the decision-making in game environments. With these com-
ponents,
there are some crucial challenges and proposed
solutions. The first challenge is that the state space of the
game is very large, especially in strategic games. With the rise
K. Shao, Z. Tang, Y. Zhu, N. Li, and D. Zhao are with the State Key
Laboratory of Management and Control for Complex Systems, Institute of Au-
tomation, Chinese Academy of Sciences. Beijing 100190, China. They are also
with the University of Chinese Academy of Sciences, Beijing, China (e-mail:
[email protected]; [email protected]; [email protected];
[email protected], [email protected]).
This work is supported by National Natural Science Foundation of
China (NSFC) under Grants No.61573353, No.61603382, No.6180337, and
No.61533017.
of representation learning, the whole system has successfully
modeled large-scale state space with deep neural networks.
The second challenge is that learning proper policies to make
decisions in dynamic unknown environment is difficult. For
this problem, data-driven methods, such as supervised learn-
ing and reinforcement learning (RL), are feasible solutions.
The third challenge is that the vast majority of game AI is
developed in a specified virtual environment. How to transfer
the AI's ability among different games is a core challenge. A
more general learning system is also necessary.
For a long time, solving these challenges with reinforcement
learning is widely used in game AI. And in the last few
years, deep learning (DL) has achieved remarkable perfor-
mance in computer vision and natural language processing
[2]. The combination, deep reinforcement learning (DRL),
teaches agents to make decisions in high-dimensional state
space in an end-to-end framework, and dramatically improves
the generalization and scalability of traditional RL algorithms.
Especially, DRL has made great progress in video games,
including Atari, ViZDoom, StarCraft, Dota2, and so on. There
are some related works to introduce these achievements in this
field. Zhao et al. [3] and Tang et al. [4] survey the development
of DRL research, and focus on AlphaGo and AlphaGo Zero.
Justesen et al. [5] reviews DL-based methods in video game
play,
including supervised learning, unsupervised learning,
reinforcement learning, evolutionary approaches, and some
hybrid approaches. Arulkumaran et al. [6] make a brief intro-
duction of DRL, covering central algorithms and presenting
a range of visual RL domains. Li [7] gives an overview of
recent achievements of DRL, and discusses core elements,
important mechanisms, and various applications. In this paper,
we focus on DRL-based game AI, from 2D to 3D, and from
single-agent to multi-agent. The main contributions include
the comprehensive and detailed comparisons of various DRL
methods, their techniques, properties, and the impressive and
diverse performances in these given video games.
The organization of the remaining paper is arranged as
follows. In Section II, we introduce the background of DL and
RL. In Section III, we focus on recent DRL methods, including
value-based, policy gradient, and model-based DRL methods.
After that, we make a brief introduction of research platforms
and competitions, and present performances of DRL methods
in classical single-agent Arcade games, first-person perspective
games, and multi-agent real-time strategy games. In Section
V, we discuss some key points and research directions in this
field. In the end, we draw a conclusion of this survey.
2
Fig. 1. The framework diagram of the typical DRL for video games. The deep learning model takes input from video games API, and extract meaningful
features automatically. DRL agents produces actions based on these features, and make the environments transfer to next state.
II. BACKGROUND
B. Reinforcement learning
Generally speaking, training an agent to make decisions
with high-dimensional inputs is difficult. With the development
of deep learning, researchers take deep neural networks as
function approximations, and use plenty of samples to opti-
mize policies successfully. The framework diagram of typical
DRL for video games is depicted in Fig. 1.
A. Deep learning
Deep learning comes from artificial neural networks, and
is used to learn data representation. It
is inspired by the
theory of brain development, and can be learned in supervised
learning, unsupervised learning and semi-supervised learning.
Although the term deep learning is introduced in 1986 [8],
deep learning has a winter time because of lacking data and
incapable computation hardware. However, with more and
more large-scale datasets being released, and capable hardware
being available, a big revolution happens in DL [9].
Convolutional neural network (CNN) [10] is a class of
deep neural networks, which is widely applied to computer
vision. CNN is inspired by biological processes, and is
shift invariant based on shared-weights architecture. Recurrent
Neural Network (RNN) is another kind of deep nerial net-
work, especially for natural language processing. As a special
kind of RNN, Long Short Term Memory (LSTM) [11] is
capable of learning long-term dependencies. Deep learning
architectures have been applied into many fields, and have
achieved significant successes, such as speech recognition, im-
age classification and segmentation, semantic comprehension,
and machine translation [2]. DL-based methods with efficient
parallel distributed computing resources can break the limit
of traditional machine learning methods. This method inspires
scientists and researchers to achieve more and more state-of-
the-art performance in respective fields.
Reinforcement learning is a kind of machine learning meth-
ods where agents learn the optimal policy by trial and error
[12]. By interacting with the environment, RL can be suc-
cessfully applied to sequential decision-making tasks. Consid-
ering a discounted episodic Markov decision process (MDP)
(S, A, γ, P, r), the agent chooses an action at according to
the policy π(atst) at state st. The environment receives the
action, produces a reward rt+1 and transfers to the next state
st+1 according to the transition probability P (st+1st, at).
This transition probability is unknown in RL domain. The
process continues until the agent reaches a terminal state or a
maximum time step. The objective is to maximize the expected
discounted cumulative rewards
∞(cid:88)
Eπ[Rt] = Eπ[
γirt+i],
(1)
where γ ∈ (0, 1] is the discount factor.
i=0
Reinforcement learning can be devided into off-policy and
on-policy methods. Off-policy RL algorithms mean that the
behavior policy used for selecting actions is different from the
learning policy. On the contrary, behavior policy is the same
with the learning policy in on-policy RL algorithms. Besides,
reinforcement learning can also be devided into value-based
and policy-based methods. In value-based RL, agents update
the value function to learn suitable policy, while policy-based
RL agents learn the policy directly.
Q-learning is a typical off-policy value-based method. The
update rule of Q-learning is
δt = rt+1 + γ arg max
Q(st+1, a) − Q(st, at),
a
Q(st, at) ← Q(st, at) + αδt.
(2a)
(2b)
δt is the temporal difference (TD) error, and α is the learning
rate.
VideoGamesAPIValue-based Policy gradient Model-based etc…StatesFeaturesEnvironmentsRL AgentsActionsRewardsCNN, LSTM, etc… 3
Fig. 2. The network architectures of typical DRL methods, with increased complexity and performance. (a): DQN network; (b)Dueling DQN network; (c):
DRQN network; (d): Actor-critic network; (e): Reactor network.
Policy gradient [13] parameterizes the policy and updates
parameters θ. In its general form, the objective function of
policy gradient is defined as
J(θ) = Eπ[
log πθ(atst)R].
(3)
∞(cid:88)
∞(cid:88)
R is the total accumulated return.
t=0
Actor-critic [12] reinforcement learning improves the policy
gradient with an value-based critic
J(θ) = Eπ[
Ψt log πθ(atst)].
(4)
t=0
Ψt is the critic, which can be the state-action value function
Qπ(st, at), the advantage function Aπ(st, at) = Qπ(st, at) −
V π(st) or the TD error rt + V π(st+1) − V π(st).
III. DEEP REINFORCEMENT LEARNING
DRL makes a combination of DL and RL, achieving rapid
developments since proposed. This section will
introduce
various DRL methods, including value-based methods, policy
gradient methods, and model-based methods.
A. Value-based DRL methods
Deep Q-network (DQN) [14] is the most famous DRL
model which learns policies directly from high-dimensional
inputs. It receives raw pixels, and outputs a value function to
estimate future rewards, as shown in Fig. 2(a). DQN uses the
experience replay method to break the sample correlation, and
stabilizes the learning process with a target Q-network. The
loss function at iteration i is
Li(θi) = E(s,a,r,s(cid:48))∼U (D)[(yDQN
i
− Q(s, a; θi))2],
with
yDQN
i
= r + γ max
a(cid:48) Q(s(cid:48), a(cid:48); θi
−).
(5)
(6)
DQN bridges the gap between high-dimensional visual
inputs and actions. After that, researchers have improved DQN
in different aspects. Double DQN [15] introduces double Q-
learning to reduce observed overestimations, and it leads to
much better performance. Prioritized experience replay [16]
helps prioritize experience to replay important
transitions
i(cid:80)
more frequently. The sample probability of transition i as
P (i) = pα
, where pi is the priority of transition i. Dueling
DQN [17] uses the dueling neural network architecture for
model-free DRL. It includes two separate estimators: one for
state value function V (s; θ, β) and the other for advantage
function A(s, a; θ, α), as shown in Fig. 2(b).
k pα
k
Q(s, a : θ, α, β) = V (s; θ, β) + A(s, a; θ, α).
(7)
Pop-Art [18] is proposed to adapt to different and non-
stationary target magnitudes, which successfully replaces the
clipping of rewards as done in DQN to handle various mag-
nitudes of targets. Fast reward propagation [19] is a novel
training algorithm for reinforcement learning, which combines
the strength of DQN, and exploits longer state-transitions in
experience replays by tightening the optimization via con-
straints. This novel
technique makes DRL more practical
by drastically reducing training time. Gorila [20] is the first
massively distributed architecture for DRL. This architecture
uses four main components: parallel actors; parallel learners;
a distributed neural network to represent the value function
or behavior policy; and a distributed store of experience. To
address the limited memory and imperfect game information
at each decision point, Deep Recurrent Q-Network (DRQN)
[21] replaces the first fully-connected layer with a recurrent
neural network in DQN, as shown in Fig. 2(c).
Generally, DQN learns rich domain representations and
approximates the value function with deep neural networks,
while batch RL algorithms with linear representations are
more stable and require less hyperparameter tuning. The Least
Squares DQN (LS-DQN) [22] combines DQN's rich feature
representations with the stability of a linear least squares
method. In order to reduce approximation error variance in
DQNs target values, averaged-DQN [23] averages previous
Q-values estimates,
leading to a more stable training and
improved performance. Deep Q-learning from Demonstrations
(DQfD) [24] combines DQN with human demonstrations,
which improves the sample efficiency greatly. DQV [25] uses
TD learning to train a Value neural network, and uses this
(a)(b)(c)(d)(e) it
network to train a second Quality-value network to esti-
mate state-action values. DQV learns significantly faster and
better than double-DQN. Researchers have proposed several
improvements to DQN. However,
is unclear which of
these are complementary and how much can be combined.
Rainbow [26] combines with main extensions to DQN, and
gives each component's contribution to overall performance.
RUDDER [27] is a novel reinforcement learning approach
for finite MDPs with delayed rewards, which is also a return
decomposition method, RUDDER is exponentially faster on
tasks with different lengths of reward delays. Ape-X DQfD
[28] uses a new transformed Bellman operator to process
rewards of varying densities and scales, and applies human
demonstrations to ease the exploration problem to guide agents
towards rewarding states. Additional, it proposes an auxiliary
temporal consistency loss to train stably extending the effective
planning horizon by an order of magnitude. Soft DQN [29]
is an entropy-regularized versions of Q-learning, with better
robustness and generalization .
Distributional DRL learns the value distribution, in contrast
to common RL that models the expectation of return, or value.
C51 [30] focuses on the distribution of value, and designs
distributional DQN algorithm to learn approximate value dis-
tributions. QR-DQN [31] methods close a number of gaps
between theoretical and algorithmic results. Distributional
reinforcement learning with Quantile regression in which the
distribution over returns is modeled explicitly instead of only
estimating the mean. Implicit Quantile Networks (IQN) [32] is
a flexible, applicable, and state-of-the-art distributional DQN.
IQN approximates the full Quantile function for the return
distribution with Quantile regression, and provides a fully
integrated distributional RL agent without prior assumptions
on the parameterization of the return distribution. Furthermore,
IQN allows to expand the class of control policies to a wide
range of risk-sensitive policies connected to distortion risk
measures.
B. Policy gradient DRL methods
Policy gradient DRL optimizes the parameterized policy
directly. Actor-critic architecture computes the policy gradi-
ent using a value-based critic function to estimate expected
future reward, as shown in Fig. 2(d). Asynchronous DRL
is an efficient framework for DRL that uses asynchronous
gradient descent to optimize the policy [33]. Asynchronous
advantage actor-critic (A3C) trains several agents on multiple
environments, showing a stabilizing effect on training. The
objective function of the actor is demonstrated as
J(θ) = Eπ[
Aθ,θv (st, at) log πθ(atst) + βHθ(π(st))],
(8)
where Hθ(π(st)) is an entropy term used to encourage explo-
ration.
t=0
GA3C [34] is a hybrid CPU/GPU version of A3C, which
achieves a significant speed up compared to the original CPU
implementation. UNsupervised REinforcement and Auxiliary
Learning (UNREAL) [35] learns separate policies for max-
imizing many other pseudo-reward functions simultaneously,
∞(cid:88)
4
including value function replay, reward prediction, and pixel
control. This agent drastically improves both data efficiency
and robustness to hyperparameter settings. PAAC [36] is a
novel framework for efficient parallelization of DRL, where
multiple actors learn the policy on a single machine. Pol-
icy gradient methods are efficient
techniques for policies
improvement, while they are usually on-policy and unable
to take advantage of off-policy data. The new method is
referred as PGQ [37], which combines policy gradient with Q-
learning. PGQ establishes an equivalency between regularized
policy gradient techniques and advantage function learning
algorithms. Retrace(λ) [38] takes the best of the importance
sampling, off-policy Q(λ), and tree-backup(λ), resulting in
low variance, safety, and efficiency. It makes a combination
of dueling DRQN architecture and actor-critic architecture,
as shown in Fig. 2(e). Reactor [39] is a sample-efficient
and numerical efficient reinforcement learning agent based
on a multi-step return off-policy actor-critic architecture. The
network outputs a target policy, an action-value Q-function,
and an estimated behavioral policy. The critic is trained with
the off-policy multi-step Retrace method and the actor is
trained by a β-leave-one-out policy gradient. Importance-
Weighted Actor Learner Architecture (IMPALA) [40] is a new
distributed DRL, which can scale to thousands of machine.
IMPALA uses a single reinforcement learning agent with a
single set of parameters to solve a mass of tasks. This method
achieves stable learning by combining decoupled acting and
learning with a novel V-trace off-policy correction method,
which is critical for achieving learning stability.
1) Trust region method: Trust Region Policy Optimization
(TRPO) [41] is proposed for optimizing control policies,
with guaranteed monotonic improvement. TRPO computes
an ascent direction to improve on policy gradient, which
can ensure a small change in the policy distribution. The
constrained optimization problem of TRPO in each epoch is
maximizeθ Es∼ρθ(cid:48) ,a∼πθ(cid:48) [
Aθ(cid:48)(s, a)],
s.t. Es∼ρθ(cid:48) [DKL(πθ(cid:48)(·s))] ≤ δKL.
πθ(as)
πθ(cid:48)(as)
(9a)
(9b)
This algorithm is effective for optimizing large nonlinear
policies. Proximal policy optimization (PPO) [42] samples
data by interaction with the environment, and optimizes the
objective function with stochastic gradient ascent
rt(θ) =
,
(10a)
πθ(atst)
πθold(atst)
(10b)
L(θ) = Et[min(rt(θ) At, clip(rt(θ), 1 − , 1 + ) At].
rt(θ) denotes the probability ratio. This objective function
clips the probability ratio to modify the surrogate objective.
PPO has some benefits over TRPO, and is much simpler to
implement, with better sample complexity. Actor-critic with
experience replay (ACER) [43] introduces several innovations,
including stochastic dueling network, truncated importance
sampling, and a new trust region method, which is stable and
sample efficient. Actor-critic using Kronecker-Factored Trust
Region (ACKTR) [44] bases on natural policy gradient, and
uses Kronecker-factored approximate curvature (K-FAC) with
trust region to optimize the actor and the critic. ACKTR is
sample efficient compared with other actor-critic methods.
2) Deterministic policy: Apart from stochastic policy, deep
deterministic policy gradient (DDPG) [45] is a kind of deter-
ministic policy gradient method which adapts the success of
DQN to continuous control. The update rule of DDPG is
Q(st, at) = r(st, at) + γQ(st+1, πθ(st+1)).
(11)
DDPG is an actor-critic, off-policy algorithm, and is able
to learn reasonable policies on various tasks. Distributed
Distributional DDPG (D4PG) [46] is a distributional update to
DDPG, combined with the use of multiple distributed workers
all writing into the same replay table. This method has a much
better performance on a number of difficult continuous control
problems.
3) Entropy-regularized policy gradient: Soft Actor Critic
(SAC) is an off-policy policy gradient method, which estab-
lishes a bridge between DDPG and stochastic policy optimiza-
tion. SAC incorporates the clipped double-Q trick, and the
objective function of maximum entropy DRL is
T(cid:88)
J(π) =
E(st,at)∼ρπ [r(st, at) + αH(π(.st))],
(12)
t=0
SAC uses an entropy regularization in its objective function.
It trains the policy to maximize a trade-off between entropy
and expected return. The entropy is a measure of randomness
in the policy. This mechanism is similar to the trade-off
between exploration and exploitation. Increasing entropy can
encourage more exploration, and accelerate learning process.
Moreover, it can also prevent the learning policy from con-
verging to a poor local optimum.
C. Model-based DRL methods
is a differentiable, recursive,
Combining model-free reinforcement learning with on-line
planning is a promising approach to solve the sample effi-
ciency problem. TreeQN [47] is proposed to address these
challenges. It
tree-structured
model that serves as a drop-in replacement for any value
function network in DRL with discrete actions. TreeQN dy-
namically constructs a tree by recursively applying a transition
model in a learned abstract state space and then aggregating
predicted rewards and state-values using a tree backup to esti-
mate Q-values. ATreeC is an actor-critic variant that augments
TreeQN with a softmax layer to form a stochastic policy
network. Both approaches are trained end-to-end, such that
the learned model is optimized for its actual use in the plan-
ner. TreeQN and ATreeC outperform n-step DQN and value
prediction networks on multiple Atari games. Vezhnevets et
al. [48] presents STRategic Attentive Writer (STRAW) neural
network architecture to build implicit plans. STRAW purely
interacts with an environment, and is an end-to-end method.
STRAW model can learn temporally abstracted high-level
macro-actions, which enables both economic computation and
structured exploration. STRAW employs temporally extended
planning strategies and achieves strong improvements on Atari
games. The world model [49] uses an unsupervised manner
to train a generative recurrent neural network, which can
5
model RL environments through compressed spatiotemporal
representations. It feeds extracted features into simple and
compact policies, achieving impressive results in several en-
vironments. Value propagation (VProp) [50] bases on value
iteration, and is an efficient differentiable planning module. It
can successfully be trained to learn to plan using reinforcement
learning. As a general framework of AlphaZero, MuZero
[54] combines MCTS with a learned model, and predicts
the reward, the action-selection policy, and the value function
to make planning. It extends model-based RL to a range of
logically complex and visually complex domains, and achieves
superhuman performance.
A general review of various DRL methods from 2017 to
2019 is presented in Table I.
IV. DRL IN VIDEO GAMES
Playing video games like human experts is challenging for
computers. With the development of DRL, agents are able
to play various games end-to-end. Here we focus on game
research platforms and competitions, and impressive progress
in various video games, from 2D to 3D, and from single-agent
to multi-agent, as shown in Fig. 3.
A. Game research platforms
Platforms and competitions make great contributions to
the development of game AI, and help to evaluate agents'
intelligence, as presented in Table II. Most platforms can be
described by two major categories: General Platforms and
Specific Platforms.
General Platforms: Arcade Learning Environment (ALE)
[55] is the pioneer evaluation platform for DRL algorithms,
which provides an interface to plenty of Atari 2600 games.
ALE presents both game images and signals, such as player
scores, which makes it a suitable testbed. To promote the
progress of DRL research, OpenAI integrates a collection of
reinforcement learning tasks into a platform called Gym [56],
which mainly contains Algorithmic, Atari, Classical Control,
Board games, 2D and 3D robots. After that, OpenAI Universe
[57] is a platform for measuring and training agents' general
intelligence across a large supply of games. Gym Retro [58]
is a wrapper for video game emulator with a unified interface
as Gym, and makes Gym easy to be extended with a large
collection of video games, not only Atari but also NEC,
Nintendo, and Sega, for RL research. The OpenAI Retro
contest aims at exploring the development of DRL that can
generalize from previous experience. OpenAI bases on the
Sonic the HedgehogT M video game, and presents a new DRL
benchmark [59]. This benchmark can help to measure the
performance of few-shot
learning and transfer learning in
reinforcement learning. General Video Game Playing [60] is
intended to design an agent to play multiple video games
without human intervention. The General Video Game AI
(GVGAI) [61] competition is proposed to provide a easy-to-
use and open-source platform for evaluating AI methods, in-
cluding DRL. DeepMind Lab [62] is a first-person perspective
learning environment, and provides multiple complicated tasks
in partially observed, large-scale, and visually diverse worlds.
A GENERAL REVIEW OF RECENT DRL METHODS FROM 2017 TO 2018.
TABLE I
Main Techniques
Networks
Category
6
DRL Algorithms
DQN [14]
Double DQN [15]
Dueling DQN [17]
Prioritized DQN [16]
Bootstrapped DQN [51]
Gorila [20]
LS-DQN [22]
Averaged-DQN [23]
DQfD [24]
DQN with Pop-Art [18]
Soft DQN [29]
DQV [25]
Rainbow [26]
RUDDER [27]
experience replay, target Q-network
double Q-learning
dueling neural network architecture
prioritized experience replay
combine deep exploration with DNNs
massively distributed architecture
combine least-squares updates in DRL
averaging learned Q-values estimates
learn from the demonstration data
adaptive normalization with Pop-Art
KL penalty and entropy bonus
training a Quality-value network
integrate six extensions to DQN
return decomposition
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN
CNN-LSTM
CNN
CNN
CNN
CNN
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
value-based, off-policy
CNN
CNN
CNN-LSTM policy gradient, on-policy
CNN-LSTM policy gradient, on-policy
CNN-LSTM policy gradient, on-policy
CNN-LSTM policy gradient, off-policy
CNN-LSTM policy gradient, on-policy
policy gradient, off-policy
CNN-LSTM policy gradient, on-policy
CNN-LSTM policy gradient, off-policy
policy gradient, on-policy
CNN-LSTM policy gradient, off-policy
CNN-LSTM policy gradient , on-policy
policy gradient , on-policy
policy gradient, off-policy
CNN-LSTM policy gradient, on-policy
CNN-LSTM policy gradient, on-policy
model-based, on-policy
model-based, on-policy
model-based, on-policy
model-based, off-policy
CNN-LSTM
CNN
CNN
CNN
CNN
CNN
Ape-X DQfD [28]
transformed Bellman operator, temporal consistency loss
C51 [30]
QR-DQN [31]
IQN [32]
A3C [33]
GA3C [34]
PPO [42]
ACER [43]
ACKTR [44]
Soft Actor-Critic [52]
UNREAL [35]
Reactor [39]
PAAC [36]
DDPG [45]
TRPO [41]
D4PG [46]
PGQ [37]
IMPALA [40]
FiGAR-A3C [53]
TreeQN/ATreeC [47]
STRAW [48]
World model [49]
MuZero [54]
distributional Bellman optimality
distributional RL with Quantile regression
an implicit representation of the return distribution
asynchronous gradient descent
hybrid CPU/GPU version
clipped surrogate objective, adaptive KL penalty coefficient
experience replay, truncated importance sampling
K-FAC with trust region
entropy regularization
unsupervised auxiliary tasks
Retrace(λ), β-leave-one-out policy gradient estimate
parallel framework for A3C
DQN with deterministic policy gradient
incorporate a KL divergence constraint
distributed distributional DDPG
combine policy gradient and Q-learning
importance-weighted actor learner architecture
fine grained action repetition
on-line planning, tree-structured model
macro-actions, planning strategies
mixture density network, variational autoencoder
representation function, dynamics function, and prediction function
A LIST OF GAME AI COMPETITIONS SUITABLE FOR DRL RESEARCH.
TABLE II
Competition Name
ViZDoom AI competition
Time
2016, 2017, 2018
StarCraft AI competitions (AIIDE, CIG, SSCAIT)
2010 -- 2019
microRTS competition
The GVGAI competition -- learning track
Microsoft Malmo collaborative AI challenge
The multi-agent RL in Malmo competition
The OpenAI Retro contest
NeurIPS Pommerman competition
Unity Obstacle Tower Challenge
NeurIPS MineRL competition
2017, 2018, 2019
2017, 2018
2017
2018
2018
2018
2019
2019
Unity ML-Agents Toolkit [63] is a new toolkit for creating
and interacting with simulation environments. This platform
has sensory, physical, cognitive, and social complexity, and
enables fast and distributed simulation, and flexible control.
Specific Platforms: Malmo [64] is a research platform for
AI experiments, which is built on top of Minecraft. It is a
first-person 3D environment, and can be used for multi-agent
research in Microsoft Malmo collaborative AI challenge 2017
and the multi-agent RL in MalmO competition 2018. TORCS
[65] is a racing car simulator which has both low-level and
visual features for the self-driving car with DRL. ViZDoom
[66] is a first-person shooter game platform, and encourages
DRL agent to utilize the visual information to perform naviga-
tion and shooting tasks in a semi-realistic 3D world. ViZDoom
AI competition has attracted plenty of researchers to develop
their DRL-based Doom agents since 2016. As far as we
7
Fig. 3. The diagram of various video games AI, from 2D to 3D, and from single-agent to multi-agent.
know, real-time strategy (RTS) games are very challenging for
reinforcement learning method. Facebook proposes TorchCraft
for StarCraft I [67], and DeepMind releases StarCraft II
learning environment [68]. They expect researchers to propose
powerful DRL agents to achieve high-level performance in
RTS games and annual StarCraft AI competitions. CoinRun
[69] provides a metric for an agent's ability to transfer its
experience to novel situations. This new training environment
strikes a desirable balance in complexity: the environment is
much simpler than traditional platform games, but it still poses
a worthy generalization challenge for DRL algorithms. Google
Research Football is a new environment based on open-source
game Gameplay Football for DRL research.
B. Atari games
ALE is an evaluation platform that aims at building agents
with general intelligence across hundreds of Atari 2600 games.
As the most popular testbed for DRL research, a large num-
ber of DRL methods have achieved outstanding performance
consecutively. Machado et al. [70] takes a review at
the
ALE in DRL research community, proposes diverse evaluation
methodologies and some key concerns. In this section, we
will introduce the main achievements in the ALE domain,
including the extremely difficult Montezuma's Revenge.
As the milestone in this domain, DQN is able to surpass
the performances of previous algorithms, and achieves human-
level performance across 49 games [14]. Averaged-DQN ex-
amines the source of value function estimation errors, and
demonstrates significantly improved stability and performance
on the ALE benchmark [23]. UNREAL significantly outper-
forms the previous best performance on Atari, averaging 880%
expert human performance [35]. PAAC achieves sufficiently
good performance on ALE after a few hours of training [36].
DQfD has better initial performance than DQN on most Atari
games, and receives more average rewards than DQN on 27
of 42. In addition, DQfD learns faster than DQN even when
given poor demonstration data [24]. Noisy DQN replaces the
MEAN AND MEDIAN SCORES ACROSS 57 ATARI GAMES OF TYPICAL DRL
METHODS, MEASURED AS PERCENTAGES OF HUMAN BASELINE.
TABLE III
Methods
DQN [14]
C51 [30]
UNREAL [35]
QR-DQN [30]
IQN [32]
Rainbow [26]
Ape-X DQN [71]
Ape-X DQfD ∗ [28]
Mean
228%
701%
880%
915%
1019%
1189%
1695%
2346%
Median
79%
178%
250%
211%
218%
230%
434%
702%
year
2015
2017
2017
2017
2018
2018
2018
2018
Note: ∗ means this method is measured across 42 Atari games.
conventional exploration heuristics with NoisyNet, and yields
substantially higher scores in ALE domain. As a distributional
DRL method, C51 obtains a new series of impressive results,
and demonstrates the importance of the value distribution in
approximated RL [30]. Rainbow provides improvements in
terms of sample efficiency and final performance. The authors
also show the contribution of each component to overall per-
formance [26]. QR-DQN algorithm significantly outperforms
recent improvements on DQN, including the related C51 [30].
IQN shows substantial gains on the Atari benchmark over
QR-DQN, and even halves the distance between QR-DQN
and Rainbow [32]. Ape-X DQN substantially improves the
performance on the ALE, achieving better final score in less
wall-clock training time [71]. When tested on a set of 42 Atari
games, the Ape-X DQfD algorithm exceeds the performance
of an average human on 40 games using a common set of
hyperparameters. Mean and median scores across multiple
Atari games of typical DRL methods that achieve state-of-
the-art performance consecutively are presented in Table III.
Montezuma's Revenge is one of the most difficult Atari
video games. It is a goal-directed behavior learning environ-
ment with long horizons and sparse reward feedback signals.
2D3DSingle-agentMulti-agentMinecraftQuake III Arena CTFViZDoomTORCSMontezuma's RevengeALEDM LabStarCraftDota2DimensionsNumber of agents Players must navigate through a number of different rooms,
avoid obstacles and traps, climb ladders up and down, and then
pick up the key to open new rooms. It requires a long sequence
of actions before reaching the goal and receiving a reward,
and is difficult to explore an optimal policy to tackle tasks.
Efficient exploration is considered as a crucial factor to learn
in a delayed feedback environment. Then, Ostrovski et al. [72]
provide an improved version of count-based exploration with
PixelCNN as a supplement for pseudo-count, also reveals the
importance of Monte Carlo return for effective exploration.
In addition to improve the exploration efficiency, learning
from human data is also a proper method to reach better
performance in this problem. Le et al. [73] leverage imitation
learning from expert interaction and hierarchical reinforcement
learning at different
levels. This method learns obviously
faster than original hierarchical reinforcement learning, and
also significantly more efficiently than conventional imitation
learning. Other than gameplay, the demonstration is also a
valuable kind of sample for agent to learn. DQfD utilizes
a small set of demonstration data to speed up the learning
process [24]. It combines prioritized replay mechanism with
temporal difference updates and supervised classification, and
finally achieves a better and impressive result. Further, Aytar
et al. [74] only use YouTube video as a demonstration sample
and invests a transformed Bellman operator for learning from
human demonstrations. Interestingly, these two works both
claim being the first to solve the entire first level of Mon-
tezuma's Revenge. Go-explore [75] makes further progress,
and achieves scores over 400,000 on average. Go-Explore sep-
arates learning into exploration and robustification. It reliably
solves the whole game, and generalizes well.
C. First-person perspective games
Different from Atari games, agents in first-person perspec-
tive video games can only receive observations from their own
perspectives, resulting from imperfect information inputs. In
RL domain, this is a POMDP problem which requires efficient
exploration and memory.
1) ViZDoom: First-person shooter (FPS) games play an
important role in game AI research. Doom is a classical
FPS game, and ViZDoom is presented as a novel testbed
for DRL [66]. Agents learn from visual inputs, and interact
with the ViZDoom environment in a first-person perspective.
Wu et al. [76] propose a method that combines A3C and
curriculum learning. The agent learns to navigate and attack
via playing against built-in agents progressively. Parisotto et
al. [77] develop Neural Map, which is a memory system
with an adaptable write operator. Neural Map uses a spatially
structured 2D memory image to store the environment's in-
formation. This method surpasses other DRL memories on
several challenging ViZDoom maze tasks and shows a capable
generalization ability. Shao et al. [78] show that ACKTR can
successfully teach agents to battle in ViZDoom environment,
and significantly outperform A2C agents by a significant
margin.
2) TORCS: TORCS is a racing game where actions are ac-
celeration, braking and steering. This game has more realistic
8
graphics than Atari games, but also requires agents to learn the
dynamic of the car. FIGAR-DDPG can successfully complete
the race task and finish 20 laps of the circuit, with a 10× total
reward against that obtained by DDPG, and much smoother
policies [53]. Normalized Actor-Critic (NAC) normalizes the
Q-function effectively, and learns an initial policy network
from demonstration and refine the policy in a real environment
[79]. NAC is robust to suboptimal demonstration data, learns
robustly and outperforms existing baselines when evaluated
on TORCS. Mazumder et al. [80] incorporate state-action
permissibility (SAP) and DDPG, and applies it to tackle the
lane keeping problem in TORCS. The proposed method can
speedup DRL training remarkably for this task. In [81], a
two-stage approach is proposed for the vision-based vehicle
lateral control problem which includes an multi-task learning
perception stage and an RL control stage. By exerting the
correlation between multiple learning task,
the perception
module can robustly extract track features. Additionally, the
RL agent learns by maximizing the geometry-based reward
and performs better than the LQR and MPC controllers. Zhu
et al. [82] use DRL to train a CNN to perceive driving data
from images of first-person view, and learns a controller to get
driving commands, showing a promising performance.
3) Minecraft: Minecraft is a sandbox construction game,
where players can build creative creations, structures, and
artwork across various game modes. Recently, it becomes a
popular platform for game AI research, with 3D infinitely
varied data. Project Malmo is an experimentation platform [83]
that builts on the Minecraft for AI research. It supports a large
number of scenarios, including navigation, problem solving
tasks, and survival to collaboration. Xiong et al. [84] propose
a novel Q-learning approach with state-action abstraction and
warm start using human reasoning to learn effective policies in
the Microsoft Malmo collaborative AI challenge. The ability to
transfer knowledge from source task to target task in Minecraft
is one of the major challenges. Tessler et al. [85] provides a
DRL agent which can transfer knowledge by learning reusable
skills, and then incorporated into hierarchical DRL network
(H-DRLN). H-DRLN exhibits superior performance and low
learning sample complexity compared to regular DQN in
Minecraft, and the potential to transfer knowledge between
related Minecraft tasks without any additional learning. To
solve the partial or non-Markovian observations problems,
Jin et al. [86] propose a new DRL algorithm based on
counterfactual regret minimization that iteratively updates an
approximation to a cumulative clipped advantage function. On
the challenging Minecraft first-person navigation benchmarks,
this algorithm can substantially outperform strong baseline
methods.
4) DeepMind lab: DeepMind lab is a 3D first-person
game platform extended from OpenArena, which is based
on Quake3. Comparable to other first-person game platforms,
DeepMind lab has considerably richer visuals and more re-
alistic physics, making it a significantly complex platform.
On a challenging suite of DeepMind lab tasks, the UNREAL
agent leads to a mean speedup in learning of 10× over A3C
and averaging 87% expert human performance. As learning
agents become more powerful, continual learning has made
quick progress recently. To test continual learning capabilities,
Mankowitz et al. [87] consider an implicit sequence of tasks
with sparse rewards in DeepMind lab. The novel agent archi-
tecture called Unicorn, demonstrates strong continual learning
and outperforms several baseline agents on the proposed
domain. Schmitt et al. [88] present a method which uses
teacher agents to kickstart the training of a new student agent.
On a multi-task and challenging DMLab-30 suite, kickstarted
training improves new agents' sample efficiency to a great
extend, and surpasses the final performance by 42%. Jaderberg
et al. [89] focus on Quake III Arena Capture the Flag, which is
a popular 3D first-person multiplayer video game, and demon-
strates that DRL agents can achieve human-level performance
with only pixels and game points as input. The agent uses
population based training to optimize the policy. This method
trains a large number of agents concurrently from thousands
of parallel matches, where agents plays cooperatively in teams
and against each other on randomly generated environments.
In an evaluation, the trained agents exceed the winrate of self-
play baseline and high-level human players both as teammates
and opponents, and are proved far stronger than existing DRL
agents.
D. Real-time strategy games
Real-time strategy games are very popular among players,
and have become popular platforms for AI research.
1) StarCraft: In StarCraft, players need to perform actions
according to real-time game states, and defeat the enemies.
Generally speaking, designing an AI bot have many chal-
lenges, including multi-agent collaboration, spatial and tem-
poral reasoning, adversarial planning, and opponent model-
ing. Currently, most bots are based on human experiences
and replays, with limited flexibility and intelligence. DRL is
proved to be a promising direction for StarCraft AI, especially
in micromanagement, build order, mini-games and full-games
[90].
Recently, micromanagement is widely studied as the first
step to solve StarCraft AI. Usunier et al. [91] introduce the
greedy MDP with episodic zero-order optimization (GMEZO)
algorithm to tackle micromanagement scenarios, which per-
forms better than DQN and policy gradient. BiCNet [92] is
a multi-agent deep reinforcement learning method to play
StarCraft combat games. It bases on actor-critic reinforcement
learning, and uses bi-directional neural networks to learn
collaboration. BiCNet successfully learns some cooperative
strategies, and is adaptable to various tasks, showing better
performances than GMEZO. In aforementioned works, re-
searchers mainly develops centralized methods to play mi-
cromanagement. Foerster et al. [93] focus on decentralized
control for micromanagement, and propose a multi-agent
actor-critic method. To stabilize experience replay and solve
nonstationarity, they use fingerprints and importance sampling,
which can improve the final performance. Shao et al. [94]
follow decentralized micromanagement task, and propose pa-
rameter sharing multi-agent gradient descent SARSA(λ) (PS-
MAGDS) method. To resue the knowledge between various
micromanagement scenarios, they also combine curriculum
9
transfer learning to this method. This improves the sample
efficiency, and outperforms GMEZO and BiCNet in large-scale
scenarios. Kong et al. [95] bases on master-slave architecture,
and proposes master-slave multi-agent reinforcement learning
(MS-MARL). MS-MARL includes composed action represen-
tation, independent reasoning, and learnable communication.
This method has better performance than other methods in
tasks. Rashid et al. [96] focus on sev-
micromanagement
eral challenging StarCraft II micromanagement
tasks, and
use centralized training and decentralized execution to learn
cooperative behaviors. This eventually outperforms state-of-
the-art multi-agent deep reinforcement learning methods.
Researchers also use DRL methods to optimize the build
order in StarCraft. Tang et al. [97] put forward neural network
fitted Q-learning (NNFQ) and convolutional neural network
fitted Q-learning (CNNFQ) to build units in simple StarCraft
maps. These models are able to find effective production
sequences, and eventually defeat enemies. In [68], researchers
present baseline results of several main DRL agents in the
StarCraft II domain. The fully convolutional advantage actor-
critic (FullyConv-A2C) agents achieve a beginner-level
in
StarCraft II mini-games. Zambaldi et al. [98] introduce the
relational DRL to StarCraft, which iteratively reasons about
the relations between entities with self-attention, and uses
it to guide a model-free RL policy. This method improves
sample efficiency, generalization ability, and interpretability of
conventional DRL approaches. Relational DRL agent achieves
impressive performance on SC2LE mini-games. Sun et al. [99]
develop the DRL based agent TStarBot, which uses flat action
structure. This agent defeats the built-in AI agents from level
1 to level 10 in a full game firstly. Lee et al. [100] focus
on StarCraft II AI, and present a novel modular architecture,
which splits responsibilities between multiple modules. Each
module controls one aspect of the game, and two modules are
trained with self-play DRL methods. This method defeats the
built-in bot in "Harder" level. Pang et al. [101] investigate
a two-level hierarchical RL approach for StarCraft II. The
macro-action is automatically extracted from expert's data, and
the other is a flexible and scaleable hierarchical architecture.
More recently, DeepMind proposes AlphaStar, and defeats
professional players for the first time.
2) MOBA and Dota2: MOBA (Multiplayer Online Battle
Arena) is originated from RTS games, which has two teams,
and each team consists of five players. To beat the opponent,
five players in a team must cooperate together, kill enemies,
upgrade heros, and eventually destroy the opponent base. Since
MOBA research is still in a primary stage, there are fewer
works than conventional RTS games. Most works on MOBA
concentrate on dataset analysis and case study. However, due
to a series of breakthroughs that DRL achieves in game AI,
researchers start to pay more attention to MOBA recently.
King of Glory (a simplified mobile version of Dota) is the most
popular mobile-end MOBA game in China. Jiang et al. [102]
apply Monte-Carlo Tree Search and deep neural networks to
this game. The experimental results indicate that MCTS-based
DRL method is efficient and can be used in 1v1 MOBA
scenario. Most impressive works on MOBA are proposed by
OpenAI. Their results prove that DRL method with self-play
can not only be successful in a 1v1 and 2v2 Dota2 scenarios
[103], but also in 5v5 [104] [105]. The model architecture is
simple, using a LSTM layer as the core component of neural
network. Under the support of massively distributed cloud
computing and PPO optimization algorithm, OpenAI Five can
master the critical abilities of team fighting, searching forest,
focusing, chasing, and diversion for team victory, and defeat
human champion OG with 2:0. Their works truly open a new
door to MOBA research with DRL method.
V. CHALLENGES IN GAMES WITH DRL
Since DRL has achieved large progress in some video
games, it is considered as one of most promising ways to
realize the artificial general intelligence. However, there are
still some challenges should be conquered towards goal. In this
secition, we discuss some crucial challenges for DRL in video
games, such as tradeoff between exploration and exploitation,
low sample efficiency, dilemma in generalization and overfit-
ing, multi-agent learning, incomplete information and delayed
sparse rewards. Though there are some proposed approaches
have been tried to solve these problems, as presented in Fig.
4, there are still some limitations should be broken.
A. Exploration-exploitation
Exploration can help to obtain more diversity samples,
while exploitation is the way to learn the high reward policy
with valuable samples. The trade-off between exploration and
exploitation remains a major challenge for RL. Common meth-
ods for exploration require a large amount of data, and can not
tackle temporally-extended exploration. Most model-free RL
algorithms are not computationally tractable in complicated
environments.
Parametric noise can help exploration to a large extend in
the training process [106] [107]. Besides, randomized value
functions become an effective approach for efficient explo-
ration. Combining exploration with deep neural networks can
help to learn much faster, which greatly improves the learning
speed and final performance in most games [51].
A simple generalization of popular count-based ap-
proach can reach satisfactory performance on various high-
dimensional DRL benchmarks [108]. This method maps states
to hash codes, and counts their occurrences via a hash table.
Then, according to the classic count-based method, we can use
these counts to compute a reward bonus. On many challenging
tasks,
these simple hash functions can achieve impressive
performance. This exploration strategy provides a simple and
powerful baseline to solve MDPs requiring considerable ex-
ploration.
B. Sample efficiency
DRL algorithms usually take millions of samples to achieve
human-level performance. While humans can quickly master
highly rewarding actions of an environment. Most model-
free DRL algorithms are data inefficient, especially for a
environment with high dimension and large explore space.
They have to interact with environment in a large time cost
10
for seek out high reward experiences in a complex sample
space, which limits their applicability to many scenarios. In
order to reduce the exploration dimension of environment and
ease the expenditure of time on interaction, some solutions
can be used for improving data efficiency, such as hierarchy
and demonstration.
Hierarchical reinforcement learning (HRL) allows agents to
decompose the task into several simple subtasks, which can
speed up training and improve sample efficiency. Temporal
abstraction is key to scaling up learning, while creating
such abstractions autonomously has remained challenging. The
option-critic architecture has the ability to learn the internal
policies and the options' termination conditions, without any
additional rewards or subgoals [109]. FeUdal Networks (FuNs)
include a Manager module and a Worker module [110]. The
Manager sets abstract goals at high-level. The Worker receives
these goals, and generates actions in the environment. FuN
dramatically outperforms baseline agents on tasks that involve
long-term credit assignment or memorization. Representation
learning methods can also be used to guide the option discov-
ery process in HRL domain [111].
Demonstration is a proper technique to improve sample
efficiency. Current approaches that learn from demonstration
use supervised learning on expert data and use reinforcement
learning to improve the performance. This method is difficult
to jointly optimize divergent losses, and is very sensitive to
noisy demonstrations. Leveraging data from previous control
of the system can greatly accelerate the learning process even
with small amounts of demonstration data [24]. Goals defined
with human preferences can effectively solve complicated RL
tasks without the reward function, while greatly reducing the
cost of human oversight [112].
C. Generalization and Transfer
The ability to transfer knowledge across multiple environ-
ments is considered as a critical aspect of intelligent agents.
With the purpose of promoting the performance of generaliza-
tion in multiple environments, multi-task learning and policy
distillation have been focus on these situations.
Multi-task learning with shared neural network parameters
can solve the generalization problem, and efficiency can be
improved through transfer across related tasks. Hybrid reward
architecture takes a decomposed reward function as input and
learns a separate value function for each component [113].
The whole value function is much smoother, which can be
easily approximated with a low-dimensional representation,
and learns more effectively. IMPALA shows the effectiveness
for multi-task reinforcement learning, using less data and ex-
hibiting positive transfer between tasks [40] . PopArt-IMPALA
combines PopArt's adaptive normalization with IMPALA, and
allows a more efficient use of parallel data generation, showing
impressive performance on multi-task domain [114].
To successfully learn complex tasks with DRL, we usually
need large task-specific networks and extensive training to
achieve good performance. Distral shares a distilled policy
which can learn common knowledge across multiple tasks
[115]. Each worker is trained to solve individual task and to be
close to the shared policy, while the shared policy is trained by
distillation. This approach shows efficient transfer on complex
tasks, with more robust and more stable performance. Mix &
Match is a training framework that is designed to encourage
effective and rapid learning in DRL agents [116]. It allows to
automatically form a curriculum over agent, and progressively
trains more complex agents from simpler agents.
D. Multi-agent learning
Multi-agent
learning is very important
in video games,
such as StarCraft. In a cooperative multi-agent setting, curse-
of-dimensionality, communication, and credit assignment are
major challenges.
Team learning uses a single learner to learn joint solutions
in multi-agent system, while concurrent learning uses multiple
learners for each agent. Recently,
the centralised training
of decentralised policies is becoming a standard paradigm
for multi-agent training. Multi-agent DDPG considers other
agents' action policy and can successfully learn complex
multi-agent coordination behavior [117]. Counterfactual multi-
agent policy gradients uses a centralized critic to estimate
the action-value function and decentralized actors to opti-
mize each agents' policies, with a counterfactual advantage
function to address the multi-agent credit assignment problem
[118] . In addition, communication protocols is important
to share information to solve multi-agent tasks. Reinforced
Inter-Agent Learning (RIAL) and Differentiable Inter-Agent
Learning (DIAL) use deep reinforcement learning to learn end-
to-end communication protocols in complex environments.
Analogously, CommNet is able to learn continuous communi-
cation between multiple agents.
E. Imperfect information
In partially observable and first-perspective games, DRL
agents need to tackle imperfect information to learn a suitable
policy. Making decisions in these environments is challenging
for DRL agents.
A critical component of enabling effective learning in these
environment is the use of memory. DRL agents have used
some simple memory architectures, such as several past frames
or an LSTM layer. But these architectures are limited to only
remember transitory information. Model-free episode control
learns difficult sequential decision-making tasks much faster,
and achieves a higher overall reward [119]. Differentiable
neural computer uses a neural network to read from and
write to an external memory matrix [120]. This method can
solve complex, structured tasks which can not access to neural
networks without external read and write memory. Neural
episodic control
inserts recent state representations paired
with corresponding value functions into the appropriate neural
dictionary, and learns significantly faster than other baseline
agents [121].
F. Delayed spare rewards
The sparse and delayed reward is very common in many
games, and is also one of the reasons that reduce sample
efficiency in reinforcement learning.
11
In many scenarios, researchers use curiosity as an intrinsic
reward to encourage agents to explore environment and learn
useful skills. Curiosity can be formulated as the error that
the agent predicts its own actions' consequence in a visual
space [122]. This can scale to high-dimensional continuous
state spaces. Moreover, it leaves out the aspects of environment
that cannot affect agents. Curiosity search for DRL encourages
intra-life exploration by rewarding agents for visiting as many
different states as possible within each episode [123].
VI. CONCLUSION AND DISCUSSION
Game AI with deep reinforcement learning is a challenging
and promising direction. Recent progress in this domain has
promote the development of artificial intelligence research. In
this paper, we review the achievements of deep reinforcement
learning in video games. Different DRL methods and their
successful applications are introduced. These DRL agents
achieve human-level or super-human performances in various
games, from 2D perfect information to 3D imperfect infor-
mation, and from single-agent to multi-agent. In addition to
these achievements, there are still some major problems when
applying DRL methods to this field, especially in 3D imperfect
information multi-agent video game. A high-level game AI
requires to explore more efficient and robust DRL techniques,
and needs novel frameworks to be implemented in complex
environment. These challenges have not been fully investigated
and could be opened for further study in the future.
ACKNOWLEDGMENT
The authors would like to thank Qichao Zhang, Dong Li and
Weifan Li for the helpful comments and discussions about this
work.
REFERENCES
[1] N. Y. Georgios and T. Julian, Artificial Intelligence and Games. New
[2] Y. Lecun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521,
York: Springer, 2018.
no. 7553, pp. 436 -- 444, 2015.
[3] D. Zhao, K. Shao, Y. Zhu, D. Li, Y. Chen, H. Wang, D. Liu, T. Zhou,
and C. Wang, "Review of deep reinforcement learning and discussions
on the development of computer Go," Control Theory and Applications,
vol. 33, no. 6, pp. 701 -- 717, 2016.
[4] Z. Tang, K. Shao, D. Zhao, and Y. Zhu, "Recent progress of deep
learning: from AlphaGo to AlphaGo Zero," Control
reinforcement
Theory and Applications, vol. 34, no. 12, pp. 1529 -- 1546, 2017.
[5] J. Niels, B. Philip, T. Julian, and R. Sebastian, "Deep learning for video
game playing," CoRR, vol. abs/1708.07902, 2017.
[6] A. Kailash, P. D. Marc, B. Miles, and A. B. Anil, "Deep reinforcement
learning: A brief survey," IEEE Signal Processing Magazine, vol. 34,
pp. 26 -- 38, 2017.
[7] L. Yuxi, "Deep reinforcement learning: An overview," CoRR, vol.
[8] R. Dechter, "Learning while searching in constraint-satisfaction-
[9] J. Schmidhuber, "Deep learning in neural networks," Neural Networks,
abs/1701.07274, 2017.
problems," pp. 178 -- 183, 1986.
vol. 61, pp. 85 -- 117, 2015.
[10] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification
with deep convolutional neural networks," in International Conference
on Neural Information Processing Systems, 2012, pp. 1097 -- 1105.
[11] S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural
Computation, vol. 9, no. 8, pp. 1735 -- 1780, 1997.
[12] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction.
MIT Press, 1998.
[13] J. W. Ronald, "Simple statistical gradient-following algorithms for
connectionist reinforcement learning," Machine Learning, vol. 8, pp.
229 -- 256, 1992.
[14] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G.
Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, and G. Os-
trovski, "Human-level control through deep reinforcement learning,"
Nature, vol. 518, no. 7540, p. 529, 2015.
[15] v. H. Hado, G. Arthur, and S. David, "Deep reinforcement learning
with double Q-learning," in AAAI Conference on Artificial Intelligence,
2016.
[16] S. Tom, Q. John, A. Ioannis, and S. David, "Prioritized experience
replay," in International Conference on Learning Representations,
2016.
[17] W. Ziyu, S. Tom, H. Matteo, v. H. Hado, L. Marc, and d. F. Nando,
"Dueling network architectures for deep reinforcement learning," in
International Conference on Machine Learning, 2016.
[18] P. v. H. Hado, G. Arthur, H. Matteo, M. Volodymyr, and S. David,
"Learning values across many orders of magnitude," in Advances in
Neural Information Processing Systems, 2016.
[19] S. H. Frank, L. Yang, G. S. Alexander, and P. Jian, "Learning to play
in a day: faster deep reinforcement learning by optimality tightening,"
in International Conference on Learning Representations, 2017.
[20] "Massively parallel methods for deep reinforcement
learning," in
International Conference on Machine Learning Workshop on Deep
Learning, 2015.
[21] J. H. Matthew and S. Peter, "Deep recurrent Q-learning for partially
observable MDPs," CoRR, vol. abs/1507.06527, 2015.
[22] L. Nir, Z. Tom, J. M. Daniel, T. Aviv, and M. Shie, "Shallow updates
for deep reinforcement learning," in Advances in Neural Information
Processing Systems, 2017.
[23] A. Oron, B. Nir, and S. Nahum, "Averaged-DQN: variance reduction
and stabilization for deep reinforcement learning," in International
Conference on Machine Learning, 2017.
[24] "Deep Q-learning from demonstrations," in AAAI Conference on Arti-
ficial Intelligence, 2018.
[25] M. Sabatelli, G. Louppe, P. Geurts, and M. Wiering, "Deep quality-
value (dqv) learning." abs/1810.00368, 2018.
[26] "Rainbow: combining improvements in deep reinforcement learning,"
in AAAI Conference on Artificial Intelligence, 2018.
[27] A. A.-M. Jose, G. Michael, W. Michael, U. Thomas, and H. Sepp,
"RUDDER: return decomposition for delayed rewards," CoRR, vol.
abs/1806.07857, 2018.
[28] "Observe and look further: achieving consistent performance on Atari,"
CoRR, vol. abs/1805.11593, 2018.
[29] S. John, A. Pieter, and C. Xi, "Equivalence between policy gradients
and soft Q-learning," CoRR, vol. abs/1704.06440, 2017.
[30] G. B. Marc, D. Will, and M. Remi, "A distributional perspective
on reinforcement learning," in International Conference on Machine
Learning, 2017.
[31] D. Will, R. Mark, G. B. Marc, and M. R´emi, "Distributional rein-
forcement learning with quantile regression," in AAAI Conference on
Artificial Intelligence, 2018.
[32] D. Will, O. Georg, S. David, and M. R´emi, "Implicit quantile networks
for distributional reinforcement learning," CoRR, vol. abs/1806.06923,
2018.
[33] "Asynchronous methods for deep reinforcement learning," in Interna-
tional Conference on Machine Learnin, 2016.
[34] B. Mohammad, F. Iuri, T. Stephen, C. Jason, and K. Jan, "Reinforce-
ment learning through asynchronous advantage actor-critic on a GPU,"
in International Conference on Learning Representations, 2017.
[35] J. Max, M. Volodymyr, C. Wojciech, S. Tom, Z. L. Joel, S. David, and
K. Koray, "Reinforcement learning with unsupervised auxiliary tasks,"
in International Conference on Learning Representations, 2017.
[36] "Efficient parallel methods for deep reinforcement learning," CoRR,
vol. abs/1705.04862, 2017.
[37] O. Brendan, M. R´emi, K. Koray, and M. Volodymyr, "PGQ: combin-
ing policy gradient and Q-learning," in International Conference on
Learning Representations, 2017.
[38] M. R´emi, S. Tom, H. Anna, and G. B. Marc, "Safe and efficient
off-policy reinforcement learning," in Advances in Neural Information
Processing Systems, 2016.
[39] G. Audrunas, G. A. Mohammad, G. B. Marc, and R. Munos,
"The Reactor: a sample-efficient actor-critic architecture," CoRR, vol.
abs/1704.04651, 2017.
[40] "IMPALA: scalable distributed deep-RL with importance weighted
actor-learner architectures," CoRR, vol. abs/1802.01561, 2018.
12
[41] S. John, L. Sergey, A. Pieter, I. J. Michael, and M. Philipp, "Trust
region policy optimization," in International Conference on Machine
Learning, 2015.
[42] S. John, W. Filip, D. Prafulla, R. Alec, and K. Oleg, "Proximal policy
optimization algorithms," CoRR, vol. abs/1707.06347, 2017.
[43] W. Ziyu, B. Victor, H. Nicolas, M. Volodymyr, M. R´emi, K. Koray,
and d. F. Nando, "Sample efficient actor-critic with experience replay,"
in International Conference on Learning Representations, 2017.
[44] W. Yuhuai, M. Elman, L. Shun, B. G. Roger, and B. Jimmy, "Scalable
trust-region method for deep reinforcement learning using Kronecker-
factored approximation," in Advances in Neural Information Processing
Systems, 2017.
[45] P. L. Timothy, J. H. Jonathan, P. Alexander, H. Nicolas, E. Tom,
T. Yuval, S. David, and W. Daan, "Continuous control with deep
reinforcement learning," CoRR, vol. abs/1509.02971, 2015.
[46] "Distributed distributional deterministic policy gradients," CoRR, vol.
abs/1804.08617, 2018.
[47] F. Gregory, R. Tim, I. Maximilian, and W. Shimon, "TreeQN and
ATreeC: differentiable tree planning for deep reinforcement learning,"
in International Conference on Learning Representations, 2018.
[48] V. Alexander, M. Volodymyr, O. Simon, G. Alex, V. Oriol, A. John,
and K. Koray, "Strategic attentive writer for learning macro-actions,"
in Advances in Neural Information Processing Systems, 2016.
[49] D. Ha and J. Schmidhuber, "Recurrent world models facilitate policy
evolution," Neural Information Processing Systems, 2018.
[50] N. Nantas, S. Gabriel, L. Zeming, K. Pushmeet, H. S. T. Philip, and
U. Nicolas, "Value propagation networks," CoRR, vol. abs/1805.11199,
2018.
[51] O. Ian, B. Charles, P. Alexander, and V. R. Benjamin, "Deep ex-
ploration via bootstrapped DQN," in Advances in Neural Information
Processing Systems, 2016.
[52] T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, "Soft actor-critic: Off-
policy maximum entropy deep reinforcement learning with a stochastic
actor," international conference on machine learning, pp. 1856 -- 1865,
2018.
[53] S. Sahil, S. L. Aravind, and R. Balaraman, "Learning to repeat:
learning," in
fine grained action repetition for deep reinforcement
International Conference on Learning Representations, 2017.
[54] S. Julian, A. Ioannis, and H. Thomas, "Mastering atari, go, chess and
shogi by planning with a learned model," abs/1911.08265, 2019.
[55] G. B. Marc, N. Yavar, V. Joel, and H. B. Michael, "The Arcade learning
environment: An evaluation platform for general agents," J. Artif. Intell.
Res., vol. 47, pp. 253 -- 279, 2013.
[56] B. Greg, C. Vicki, P. Ludwig, S. Jonas, S. John, T. Jie, and Z. Wojciech,
"OpenAI Gym," CoRR, vol. abs/1606.01540, 2016.
[57] "OpenAI Universe github," https://github.com/openai/universe, 2016.
[58] "OpenAI Retro github," https://github.com/openai/retro, 2018.
[59] N. Alex, P. Vicki, H. Christopher, K. Oleg, and S. John, "Gotta
learn fast: a new benchmark for generalization in RL," CoRR, vol.
abs/1804.03720, 2018.
[60] "General video game AI: a multi-track framework for evaluat-
ing agents, games and content generation algorithms," CoRR, vol.
abs/1802.10363, 2018.
[61] R. T. Ruben, B. Philip, T. Julian, L. Jialin, and P.-L. Diego, "Deep
learning for general video game AI," CoRR, vol.
reinforcement
abs/1806.02448, 2018.
[62] "DeepMind Lab," CoRR, vol. abs/1612.03801, 2016.
[63] J. Arthur, B. Vincent-Pierre, V. Esh, G. Yuan, H. Hunter, M. Marwan,
and L. Danny, "Unity: a general platform for intelligent agentst," CoRR,
vol. abs/1809.02627, 2018.
[64] "OpenAI Malmo github," https://github.com/Microsoft/malmo, 2017.
[65] B. Wymann, E. Espi´e, C. Guionneau, C. Dimitrakakis, R. Coulom, and
A. Sumner, "Torcs, the open racing car simulator," Software available
at http://torcs. sourceforge. net, vol. 4, p. 6, 2000.
[66] M. Kempka, M. Wydmuch, G. Runc, J. Toczek, and W. Jakowski,
"ViZDoom: a Doom-based AI research platform for visual reinforce-
ment learning," in IEEE Conference on Computational Intelligence and
Games, 2017, pp. 1 -- 8.
[67] S. Gabriel, N. Nantas, A. Alex, C. Soumith, L. Timoth´ee, L. Zeming,
R. Florian, and U. Nicolas, "TorchCraft: a library for machine learning
research on real-time strategy games," CoRR, vol. abs/1611.00625,
2016.
[68] "StarCraft II: a new challenge for reinforcement learning," CoRR, vol.
abs/1708.04782, 2017.
[69] C. Karl, K. Oleg, H. Chris, K. Taehoon, and S. John, "Quantifying
generalization in reinforcement learning," CoRR, vol. abs/1812.02341,
2018.
[70] C. M. Marlos, G. B. Marc, T. Erik, V. Joel, J. H. Matthew, and
B. Michael, "Revisiting the Arcade learning environment: evaluation
protocols and open problems for general agents," Journal of Artificial
Intelligence Research, vol. 61, pp. 523 -- 562, 2018.
[71] H. Dan, Q. John, B. David, B.-M. Gabriel, H. Matteo, v. H. Hado, and
S. David, "Distributed prioritized experience replay," in International
Conference on Learning Representations, 2018.
[72] O. Georg, G. B. Marc, O. Aaron, and M. Remi, "Count-based ex-
ploration with neural density models," in International Conference on
Machine Learning, 2017.
[73] M. L. Hoang, J. Nan, A. Alekh, D. Miroslav, Y. Yisong, and D. Hal,
"Hierarchical imitation and reinforcement learning," in International
Conference on Machine Learning, 2018.
[74] Y. Aytar, T. Pfaff, D. Budden, T. L. Paine, Z. Wang, and N. D. Freitas,
"Playing hard exploration games by watching youtube," 2018.
[75] "Montezuma's revenge solved by go-explore, a new algorithm for hard-
exploration problems," https://eng.uber.com/go-explore/, 2018.
[76] Y. Wu and Y. Tian, "Training agent for first-person shooter game
with actor-critic curriculum learning," in International Conference on
Learning Representations, 2017.
[77] P. Emilio and S. Ruslan, "Neural map: structured memory for deep
reinforcement learning," CoRR, vol. abs/1702.08360, 2017.
[78] K. Shao, D. Zhao, N. Li, and Y. Zhu, "Learning battles in ViZDoom via
deep reinforcement learning," in IEEE Conference on Computational
Intelligence and Games, 2018.
[79] G. Yang, X. Huazhe, L. Ji, Y. Fisher, L. Sergey, and D. Trevor,
"Reinforcement learning from imperfect demonstrations," CoRR, vol.
abs/1802.05313, 2018.
[80] M. Sahisnu, L. Bing, W. Shuai, Z. Yingxuan, L. Lifeng, and L. Jian,
"Action permissibility in deep reinforcement learning and application
to autonomous driving," in ACM SIGKDD Conference on Knowledge
Discovery and Data Mining, 2018.
[81] D. Li, D. Zhao, Q. Zhang, and Y. Chen, "Reinforcement learning
and deep learning based lateral control for autonomous driving," IEEE
Computational Intelligence Magazine, 2018.
[82] Y. Zhu and D. Zhao, "Driving control with deep and reinforcement
learning in the open racing car simulator," in International Conference
on Neural Information Processing, 2018.
[83] J. Matthew, H. Katja, H. Tim, and B. David, "The Malmo platform for
artificial intelligence experimentation," in International Joint Confer-
ences on Artificial Intelligence, 2016.
[84] X. Yanhai, C. Haipeng, Z. Mengchen, and A. Bo, "HogRider: cham-
pion agent of Microsoft Malmo collaborative AI challenge," in AAAI
Conference on Artificial Intelligence, 2018.
[85] T. Chen, G. Shahar, Z. Tom, J. M. Daniel, and M. Shie, "A deep
hierarchical approach to lifelong learning in Minecraft," in AAAI
Conference on Artificial Intelligence, 2017.
[86] H. J. Peter, L. Sergey, and K. Kurt, "Regret minimization for partially
observable deep reinforcement learning," in International Conference
on Learning Representations, 2018.
[87] "Unicorn: continual learning with a universal, off-policy agent," CoRR,
vol. abs/1802.08294, 2018.
[88] "Kickstarting
deep
abs/1803.03835, 2018.
reinforcement
learning,"
CoRR,
vol.
[89] "Human-level
performance
with population-based deep reinforcement
abs/1807.01281, 2018.
in first-person multiplayer
games
learning," CoRR, vol.
[90] Z. Tang, K. Shao, Y. Zhu, D. Li, D. Zhao, and T. Huang, "A review
of computational intelligence for StarCraft AI," in IEEE Symposium
Series on Computational Intelligence (SSCI), 2018.
[91] N. Usunier, G. Synnaeve, Z. Lin, and S. Chintala, "Episodic ex-
ploration for deep deterministic policies: an application to StarCraft
tasks," in International Conference on Learning
micromanagement
Representations, 2017.
[92] P. Peng, Q. Yuan, Y. Wen, Y. Yang, Z. Tang, H. Long, and J. Wang,
"Multiagent bidirectionally-coordinated nets for learning to play Star-
Craft combat games," 2017.
[93] J. Foerster, N. Nardelli, G. Farquhar, T. Afouras, P. H. S. Torr,
P. Kohli, and S. Whiteson, "Stabilising experience replay for deep
multi-agent reinforcement learning," in International Conference on
Machine Learning, 2017.
[94] K. Shao, Y. Zhu, and D. Zhao, "StarCraft micromanagement with
learning," IEEE
Intelligence,
reinforcement
Transactions on Emerging Topics in Computational
DOI:10.1109/TETCI.2018.2823329, 2018.
learning and curriculum transfer
13
[95] K. Xiangyu, X. Bo, L. Fangchen, and W. Yizhou, "Revisiting the
master-slave architecture in multi-agent deep reinforcement learning,"
CoRR, vol. abs/1712.07305, 2017.
[96] R. Tabish, S. Mikayel, S. d. W. Christian, F. Gregory, N. F. Jakob, and
W. Shimon, "QMIX: monotonic value function factorisation for deep
multi-agent reinforcement learning," CoRR, vol. abs/1803.11485, 2018.
[97] T. Zhentao, Z. Dongbin, Z. Yuanheng, and G. Ping, "Reinforcement
learning for build-order production in StarCraft II," in International
Conference on Information Science and Technology, 2018.
[98] V. Zambaldi, D. Raposo, A. Santoro, V. Bapst, Y. Li, I. Babuschkin,
K. Tuyls, D. Reichert, T. Lillicrap, E. Lockhart et al., "Relational deep
reinforcement learning," CoRR, vol. abs/806.01830, 2018.
[99] P. Sun, X. Sun, L. Han, J. Xiong, Q. Wang, B. Li, Y. Zheng, J. Liu,
Y. Liu, H. Liu, and T. Zhang, "TStarBots: defeating the cheating level
builtin AI in StarCraft II in the full game," CoRR, vol. abs/1809.07193,
2018.
[100] L. Dennis, T. Haoran, O. Z. Jeffrey, X. Huazhe, D. Trevor, and
A. Pieter, "Modular architecture for starcraft ii with deep reinforcement
learning," CoRR, p. abs/1811.03555, 2018.
[101] P. Zhen-Jia, L. Ruo-Ze, M. Zhou-Yu, Z. Yi, Y. Yang, and L. Tong,
"On reinforcement learning for full-length game of starcraft," CoRR,
p. abs/1809.09095, 2018.
[102] R. J. Daniel, E. Emmanuel, and L. Hao, "Feedback-based tree search
for reinforcement learning," in International Conference on Machine
Learning, 2018.
[103] "OpenAI Dota 1v1," https://blog.openai.com/dota-2/, 2017.
[104] "OpenAI Dota Five," https://blog.openai.com/openai-five/, 2018.
[105] B. Christopher, B. Greg, and C. Brooke, "Dota 2 with large scale deep
reinforcement learning," abs/1912.06680, 2019.
[106] "Noisy networks for exploration," CoRR, vol. abs/1706.10295, 2017.
[107] "Parameter space noise for exploration," CoRR, vol. abs/1706.01905,
2017.
[108] T. Haoran, H. Rein, F. Davis, S. Adam, C. Xi, D. Yan, S. John, D. T.
Filip, and A. Pieter, "Exploration: a study of count-based exploration
for deep reinforcement learning," in Advances in Neural Information
Processing Systems, 2017.
[109] B. Pierre-Luc, H. Jean, and P. Doina, "The option-critic architecture,"
in AAAI Conference on Artificial Intelligence, 2017.
[110] S. V. Alexander, O. Simon, S. Tom, H. Nicolas, J. Max, S. David, and
K. Koray, "FeUdal networks for hierarchical reinforcement learning,"
in International Conference on Machine Learning, 2017.
[111] C. M. Marlos, R. Clemens, G. Xiaoxiao, L. Miao, T. Gerald, and
C. Murray, "Eigenoption discovery through the deep successor rep-
resentation," CoRR, vol. abs/1710.11089, 2017.
[112] F. C. Paul, L. Jan, B. B. Tom, M. Miljan, L. Shane, and A. Dario,
"Deep reinforcement learning from human preferences," in Advances
in Neural Information Processing Systems, 2017.
[113] v. S. Harm, F. Mehdi, L. Romain, R. Joshua, B. Tavian, and T. Jeffrey,
"Hybrid reward architecture for reinforcement learning," in Advances
in Neural Information Processing Systems, 2017.
[114] H. Matteo, S. Hubert, E. Lasse, C. Wojciech, S. Simon, and v. H.
Hado, "Multi-task deep reinforcement learning with popart," CoRR,
vol. abs/1809.04474, 2018.
[115] "Distral: robust multitask reinforcement
learning," in Advances in
Neural Information Processing Systems, 2017.
[116] "Mix& Match - agent curricula for reinforcement learning," CoRR, vol.
abs/1806.01780, 2018.
[117] L. Ryan, W. Yi, T. Aviv, H. Jean, A. Pieter, and M. Igor, "Multi-agent
actor-critic for mixed cooperative-competitive environments," 2017, pp.
6382 -- 6393.
[118] J. Foerster, G. Farquhar, T. Afouras, N. Nardelli, and S. Whiteson,
"Counterfactual multi-agent policy gradients," in AAAI Conference on
Artificial Intelligence, 2018.
[119] B. Charles, U. Benigno, P. Alexander, L. Yazhe, R. Avraham, Z. L. Joel,
W. R. Jack, W. Daan, and H. Demis, "Model-free episodic control,"
CoRR, vol. abs/1606.04460, 2016.
[120] "Hybrid computing using a neural network with dynamic external
memory," Nature, vol. 538, pp. 471 -- 476, 2016.
[121] "Neural episodic control," in International Conference on Machine
Learning, 2017.
[122] P. Deepak, A. Pulkit, A. E. Alexei, and D. Trevor, "Curiosity-driven
exploration by self-supervised prediction," in IEEE Conference on
Computer Vision and Pattern Recognition Workshops, 2017, pp. 488 --
489.
[123] S. Christopher and C. Jeff, "Deep curiosity search: intra-life exploration
learning
improves performance on challenging deep reinforcement
problems," CoRR, vol. abs/1806.00553, 2018.
|
1305.2386 | 1 | 1305 | 2013-05-10T16:47:02 | Disappointment in Social Choice Protocols | [
"cs.MA"
] | Social choice theory is a theoretical framework for analysis of combining individual preferences, interests, or welfare to reach a collective decision or social welfare in some sense. We introduce a new criterion for social choice protocols called social disappointment. Social disappointment happens when the outcome of a voting system occurs for those alternatives which are at the end of at least half of individual preference profiles. Here we introduce some protocols that prevent social disappointment and prove an impossibility theorem based on this key concept. | cs.MA | cs |
Disappointment in Social Choice Protocols
Mohammad Ali Javidian, Rasoul Ramezanian1
Department of Mathematics Sharif University of Technology, Tehran, Iran
Complex and Multi Agent Systems Lab
[email protected], [email protected]
Abstract
Social choice theory is a theoretical framework for analysis of combining individual
preferences, interests, or welfares to reach a collective decision or social welfare in some
sense. We introduce a new criterion for social choice protocols called "social disappoint-
ment". Social disappointment happens when the outcome of a voting system occurs for
those alternatives which are at the end of at least half of individual preference pro-
files. Here we introduce some protocols that prevent social disappointment and prove
an impossibility theorem based on this key concept.
1
Introduction
In social sciences, we are facing two kinds of social choices: voting which is used to make a
political decision and market mechanism as a tool to make an economic decision[1]. Here,
we are merely concerned with the former.
The theory underlying voting systems is known as social choice theory and is concerned
with the design and analysis of methods for collective decision making[27]. Voting pro-
cedures are among the most important methods for collective decision making.
In this
paper, our attention is on voting procedures. Voting procedures focus on the aggregation
of individuals' preferences to produce collective decisions. In practice, a voting procedure
is characterized by ballot responses and the way ballots are tallied to determine winners.
Voters are assumed to have clear preferences over candidates (alternatives) and attempt
to maximize satisfaction with the election outcome by their ballot responses. Voting pro-
cedures are formalized by social choice functions, which map ballot response profiles into
election outcomes(see[6], page:175).
We use a broad class of social choice functions such as Condorcet method, Plurality rule,
Hare system, Borda count, Sequential Pairwise Voting with a Fixed Agenda (Seq. Pairs),
and Dictatorship. Condorcet method is typically attributed to the Marquis de Condorcet
(1743-1794); However, it dates back to Ramon Llull in the thirteenth century([28] p. 6).
Hare procedure was introduced by Thomas Hare in 1861, and is also known by names
such as the "single transferable vote system"(STV) or "instant runoff voting" ([28] p. 7).
Jean Charles Chevalier de Borda (1733-99) in 1781 [5] introduced an aggregation procedure
known as Borda count. Interestingly, recent historical work by McLean, Urken (1993) [17],
and Pukelsheim (unpublished) reveals that Bordas system had been explicitly described in
1433 by Nicholas of Cusa (1401-64), a Renaissance scholar interested in the question of how
1Corresponding Author
German kings should be elected ([30] p. 9). For more details and examples, see [28] sec.
1.3.
There are five desirable properties that relate to such procedures: Always-A-Winner
Condition (AAW), Condorcet Winner Criterion (CWC), Pareto Condition, Monotonicity
(Mono), Independence of Irrelevant Alternatives (IIA). A social choice procedure is said to
satisfy AAWcondition if every sequence of individual preference lists produces at least one
winner. An alternative 'x' is said to be a Condorcet winner if it is the unique winner in
Condorcets method. A social choice procedure is said to satisfy CWC provided that-if there
is a Condorcet winner-then it alone is the social choice. A social choice procedure is said to
satisfy the Pareto condition (or just Pareto) if the following holds for every pair 'x' and 'y'
of alternatives: If everyone prefers 'x' to 'y', then y is not a social choice. A social choice
procedure is said to be monotone provided that the following holds for every alternative 'x':
If 'x' is the social choice (or tied for such) and someone changes his or her preference list by
moving 'x' up one spot, then 'x' should still be the social choice (or tied for such). A social
choice procedure has IIA condition the social choice set includes 'x' but not 'y', and one or
more voters change their preferences, but no one changes his or her mind about whether 'x'
is preferred to 'y' or 'y' to 'x', then the social choice set should not change so as to include
'y'. The condition of "independence of irrelevant alternatives" was first used by Arrow [1]
in 1951. For more details of these properties see [28], sec. 1.4.
In this paper, we propose a new property for social choice procedures called social
disappointment in voting which is a situation that happens when the winner is the least
favorable candidate for at least half of the voters. See the definition and an example of
social disappointment in section 2.
The rest of the paper is organized as follows.
In section 2 we introduce the concept
of social disappointment and the Least Public Resentment procedure (L.P.R) which is a
procedure that prevents social disappointment in voting. In section 3 in imitation of Tay-
lor's work (see[29] also[28] pp. 28-31) we present an impossibility theorem based on social
disappointment in voting.
Notice:
In this paper, we will follow the notation and basic results of Taylor and
Pacelli(see[28], chapter:1).
2 Social Disappointment
We start explaining the social disappointment property by the following example.
Example 2.1 Consider the following situation in which there are four Dutchmen, three
Germans, and two Frenchmen who have to decide which drink will be served for lunch (only
a single drink will be served to all).
4
3
2
Milk
Beer Wine
Wine Wine Beer
Beer Milk Milk
Now, which drink should be served based on these individual preferences? Milk could be
chosen since it has the most agents ranking it first. Milk is the winner according to the
plurality rule, which only considers how often each alternative is ranked in first place. How-
ever, a majority of agents will be dissatisfied with this choice as they prefer any other drink
to Milk(see[7], pp. 3,4).
Definition 2.1 Social disappointment in voting happens when the outcome of a voting
system (for 3 or more alternatives) occurs for those alternatives which are at the end of at
least half of individual preference profiles.
Now the question is for which protocols social disappointment(S.D) may hap- pen? The
answer is given in the following table (where a 'yes' indicates that social disappointment
may happen).
Plurality Borda Hare Seq. Pairs Dictator Condorcet
S.D
Yes
Yes
Yes
Yes
Yes
Yes
We prove the seven claims in Table 1.
Claim 2.1 The Plurality rule does not prevent social disappointment.
Proof. See Example 2.1. ⊣
Claim 2.2 The Borda count does not prevent social disappointment.
Proof. Consider the three alternatives 'a', 'b', and 'c' and the following sequence of two
preference lists:
Voters Voters
3and4
1and2
a
b
c
c
b
a
Alternatives 'a', 'b' and 'c' are the social choice when the Borda count procedure is used.
Although 'a' is the social choice (also 'c'), but it is at the bottom of individual preference
lists and so social disappointment has taken place. ⊣
Claim 2.3 The Hare procedure does not prevent social disappointment.
Proof. Consider the three alternatives 'a', 'b', and 'c' and the following sequence of ten
preference lists grouped into voting blocks of size four, three, and two:
Voters Voters Voters
8-10
1-4
a
b
c
5-7
c
b
a
Alternative 'a' is the social choice when the Hare system is used. Although 'a' is the social
choice, but it is at the bottom of individual preference lists and so social disappointment
has taken place. ⊣
Claim 2.4 Sequential pairwise voting with a fixed agenda does not prevent social disap-
pointment.
Proof. Consider the three alternatives 'b', 'c', and a and suppose that this ordering of
the alternatives is also the agenda. Consider the following sequence of four preference lists
grouped into voting blocks of size two, one, and one:
Alternatives 'a', 'b' are the social choice when the Hare system is used. Although 'a'
is the social choice, but it is at the bottom of individual preference lists and so social
disappointment has taken place. ⊣
Claim 2.5 A dictatorship does not prevent social disappointment.
Proof. Consider the three alternatives 'a', 'b', and 'c' and the following three preference
lists:
Voters Voter Voter
1and2
b
c
a
4
c
b
a
4
c
b
a
Voters Voter Voter
1and2
3
b
c
a
3
c
b
a
a
b
c
a
b
c
Assume that Voter 1 is the dictator. Then 'a' is the social choice, but obviously social
disappointment has happened. ⊣
Claim 2.6 If in Condorcet method more than half of voters put 'a' at the bottom of indi-
vidual preference lists then for sure 'a' would not be the social choice and in this case social
disappointment would not occur. But if the number of voters is even and precisely half of
voters put 'a' in the end of their lists, one of these two possibilities happens:
• 1. Not all voters in the other half put 'a' at the top of their lists, which in this case,
'a' definitely does not hold the social choice and social disappointment occures.
• 2. All the voters in the other half also put 'a' at the top of their lists, which in this
case 'a' is definitely in the set of social choice and therefore we face with the social
disappointment.
Proof. It is concluded from the definitions. ⊣
Remark 2.1 Regarding case 2 in claim 2.6 if we have only three alternatives, the set of
social choice certainly have more than one member.
2.1 The least public resentment procedure (L.P.R):
We observed that none of the famous procedures listed in Table 1 prevents social disap-
pointment. We introduce a procedure which prevents social disappointment in voting, called
The least public resentment procedure (L.P.R). In this procedure we begin by deleting the
alternative or alternatives occurring at bottom of the most lists. At this stage we have lists
that are at least one alternative shorter than the lists we started with. Now, we simply
repeat this procedure of deleting the least public resentment alternative or alternatives.
The alternative(s) deleted last is declared as the winner.
Example 2.2 Consider Example 2.1, we decide which alternative occurs at the bottom of
the most lists and delete it from all the lists. Milk is deleted from each list leaving the
following:
4
3
2
Wine Beer Wine
Beer Wine Beer
Now, Beer occurs at the bottom of six of the nine lists, and thus is eliminated. Hence, Wine
is the social choice when the L.P.R is used.
Which properties does this procedure satisfy? The answer is given in the following table:
AAW CWC Pareto Mono
Yes
Yes
No
L.P.R
No
IIA Non S.D
No
Yes
Proposition 2.1 The L.P.R procedure satisfies AAW, Pareto, and nonexistence of social
disappointment (Non S.D) criterion but does not satisfy CWC, Monotonicity, and IIA.
Proof. For this procedure, the description makes it clear that there is at least one winner
for every profile. So L.P.R satisfies AAW condition.
Suppose that there is a winner that is in the end of at least half of preference profiles.
The L.P.R procedure would delete this alternative from profile lists in the first stage, so the
social disappointment for this alternative could not occure.
Consider the three alternatives 'a', 'b', and 'c' and the following sequence of seven
preference lists grouped into voting blocks of size two, two, two, and one:
Voters Voters Voters Voter
1and2
3and4
5and6
a
b
c
a
c
b
b
c
a
7
c
b
a
The L.P.R procedure produces 'b' as the social choice. However, 'a' is clearly the Condorcet
winner, defeating each of the other alternatives in one-on-one competitions. Since the
Condorcet winner is not the social choice in this situation, we have that the L.P.R procedure
does not satisfy the Condorcet winner criterion.
L.P.R procedure satisfies the Pareto condition. Because if in all lists 'b' has occurred
down below 'a', therefore at some point 'b' would be gone but 'a' stands still due to the
fact that 'b' socially is more resentful than 'a'. So based on social choice procedure in this
system 'b' would be eliminated in early stages or at most in comparison with 'a'.
Consider the three alternatives 'a', 'b', and 'c' and the following sequence of seven
preference lists grouped into voting blocks of size two, two, two, and one:
Voters Voters Voter
1and2
3and4
b
a
c
c
a
b
5
b
c
a
We delete the alternatives which have taken place more than the other alternatives at the
end of the votes. In this case, that would be alternatives 'c' and 'b' with the two last places
in votes for each as compared to one for 'a'. But now 'a' is the only alternative left, and so
it is the social choice when the L.P.R procedure is used.
Now suppose that the single voter on the most right changes his or her list by in-
terchanging 'a' with the alternative that is right above 'a' on this list. This apparently
favorable-to-'a'-change yields the following sequence of preference lists:
Voters Voters Voter
1and2
3and4
b
a
c
c
a
b
If we apply the L.P.R procedure again, we delete the alternatives which have taken place
more than the other alternatives at the end of the votes. In this case, 'c' is that alternative.
But the reader can now easily check that with 'c' so eliminated, alternative 'b' is at bottom
of two of the fifth lists. Alternative 'a' is deleted and so 'b' is the social choice. This change
in social choice from 'a' to 'c' shows that the L.P.R procedure does not satisfy monotonicity.
Consider the three alternatives 'a', 'b', and 'c' and the following sequence of four pref-
erence lists grouped into voting blocks of size one, one, and two:
5
b
a
c
b
c
a
Voter Voter Voters
3and4
1
b
a
c
2
a
c
b
Alternative 'b' is the social choice when the L.P.R procedure is used.
In particular, 'b'
is a winner and 'a' is a non-winner. Now suppose that Voter 4 changes his or her list by
interchanging the alternatives 'a' and 'c'. The lists then become:
Voter1 Voter2 Voter3 Voter4
b
a
c
a
c
b
b
c
a
b
a
c
Notice that we still have 'b' over 'a' in Voter 4s list. However, L.P.R procedure now has
'a' and 'b' tied for the win. Thus, although no one changed his or her mind about whether
'a' is preferred to 'b' or 'b' to 'a', the alternative 'a' went from being a non-winner to
being a winner. This shows that independence of irrelevant alternatives fails for the L.P.R
procedure. ⊣
3 A Glimpse of Impossibility
Taylor proved in [29] also [28] pp. 28-31 that there is no social choice procedure for three
or more alternatives that satisfies the always-a-winner criterion, independence of irrelevant
alternatives, and the Condorcet winner criterion. Now we prove an impossibility theorem
based on social disappointment concept.
Theorem 3.1 There is no social choice procedure for four or more alternatives that satisfies
the nonexistence of S.D. criterion, and the Condorcet winner criterion.
Proof. We assume that we have a social choice procedure that satisfies the Condorcet
winner criterion. We then show that if this procedure is applied to the profile that consists
of Condorcets voting paradox [10], then it produces a winner which will lead to social
disappointment. We prove this claim for when we have four alternatives.
Assume that we have a social choice procedure that satisfies the Condorcet winner
criterion. Consider the following profile:
b b
d d d
c
c
c
a
a
a
a
b b
a
b d d d
c
c
c
a
b
Alternative 'd' is the unique social choice when the Condorcets method is used. Although
'd' is the social choice, but it is at the bottom of individual preference lists and so social
disappointment has taken place. ⊣
Corollary 3.1 There is no social choice procedure for four or more alternatives that sat-
isfies:
• a) the nonexistence of S.D. criterion, always a winner, and the Condorcet winner
criterion.
• b) the nonexistence of S.D. criterion, monotonicity, and the Condorcet winner crite-
rion.
• c) the nonexistence of S.D. criterion, Independence of Irrelevant Alternatives, and the
Condorcet winner criterion.
• d) the nonexistence of S.D. criterion, Pareto, and the Condorcet winner criterion.
Proof. It is obvious considering Theorem 3.1. ⊣
3.1 Condorcet with an amendment and Seq. Pairs with an amendment
procedure
Before considering the rest of the possible cases, we introduce following two procedures and
investigate their properties which we mentioned in this article.
Condorcet with an amendment procedure
This protocol is done the same as Condorcets method, with the difference that in the end
we remove those alternatives with the social disappointment from the set of social choice.
Seq. Pairs with an amendment procedure
This protocol is done the same as Seq. Pairs, with the difference that in the end we remove
those alternatives with the social disappointment from the set of social choice.
Which properties do the procedures satisfy? The answer is given in the following table:
Condorcet with an amendment
Seq. Pairs with an amendment
AAW CWC Pareto Mono
Yes
Yes
No
Yes
No
No
Yes
Yes
IIA Non S.D
Yes
No
Yes
Yes
Here we will only prove three items of the table above. We leave the rest for the reader
(The proof will be easy to comprehend according to the given definitions and [28] sections
1.5,1.6 ).
Proposition 3.1 Condorcet with an amendment and Seq. Pairs with an amendment pro-
cedures do not satisfy the Condorcet winner criterion. Furthermore Condorcet with an
amendment procedure does not satisfy the always a winner criterion.
Proof. Consider the following profile:
b b
d d d
c
c
a
a
c
a
b b
a
a
b d d d
c
c
c
a
b
Alternative 'd' is the unique social choice when the Condorcets method is used. There-
fore there is no winner (NW) when the Condorcet with an amendment procedure is used.
Consider Seq. Pairs voting with fixed agenda a,b,c,d. Alternatives 'c', 'd' are the social
choices when the Seq. Pairs voting with this particular fixed agenda is used. Thus alterna-
tive 'c' is the social choice when Seq. Pairs with an amendment is used, so Condorcet with
an amendment and Seq. Pairs with an amendment procedures do not satisfy the Condorcet
winner criterion. Furthermore according to what was said Condorcet with an amendment
procedure does not satisfy the always a winner criterion.
⊣
Remark 3.1 In light of Remark 2.1 we understand that set of social choice would be either
all three alternatives or two alternatives which one would be free from social disappointment.
Anyway, set of social choice would include the alternative with no social disappointment.
Considering this fact the following proposition will be prove.
Proposition 3.2 There exist some social choice procedures for three alternatives that sat-
isfy:
• a) the nonexistence of S.D. criterion, and the Condorcet winner criterion.
• b) the nonexistence of S.D. criterion, the Condorcet winner, and always a winner
criterion.
• c) the nonexistence of S.D. criterion, the Condorcet winner, and Pareto criterion.
• d) the nonexistence of S.D. criterion, the Condorcet winner, and monotonicity crite-
rion.
• e) the nonexistence of S.D. criterion, the Condorcet winner, and independence of
irrelevant alternatives criterion.
• f ) the Pareto criterion, the Condorcet winner, and independence of irrelevant alter-
natives criterion.
• g) the nonexistence of S.D. criterion, monotonicity, and independence of irrelevant
alternatives criterion.
Proof. For (b) consider Seq. Pairs with an amendment and for the rest Condorcet's method
with an amendment. ⊣
3.2 The Least Unpopular (L.U) and The Least Unpopular Reselection
procedure(L.U.R)
To investigate the remaining cases we introduce and investigate The Least Unpopular and
The Least Unpopular Reselection procedures.
The Least Unpopular procedure (L.U)
The social choice in this procedure is (are) the alternative(s) that appears lesser than the
others at the bottom of individual preference lists. This protocol satisfies AAW, monotonic-
ity, and nonexistence of S.D. criterion, but does not satisfy CWC, Pareto, and IIA criterion.
Proposition 3.3 The Least Unpopular procedure does not satisfy CWC, IIA, and Pareto
criterion.
Proof. Consider the four alternatives 'a', 'b', 'c', and 'd' and the following profile:
Voters Voter Voter
1and2
a
b
c
d
3
c
a
b
d
4
d
a
b
c
Alternatives 'a', 'b' are the social choices when the Least Unpopular procedure is used.
Thus, alternative 'b' is in the set of social choice even though everyone prefers 'a' to 'b'.
This show that Pareto fails. Now consider the three alternatives 'a', 'b', 'c' and following
profile:
Voters Voter
1and2
3
b
c
a
a
b
c
Alternative 'b' is the social choice when the Least Unpopular procedure is used. However,
'a' is clearly the Condorcet winner, defeating each of the other alternatives in one-on-one
competitions. Since the Condorcet winner is not the social choice in this situation, we have
that the Least Unpopular procedure does not satisfy the Condorcet winner criterion. On
the other hand 'b' is a non-winner. Now suppose that voter 3 changes his or her list by
interchanging the alternatives 'a' and 'c'. The lists then become:
Voters Voter
1and2
3
b
a
c
a
b
c
Notice that we still have 'b' over 'a' in Voter 4s list. However, Least Unpopular procedure
now has 'a' and 'b' tied for the win. Thus, although no one changed his or her mind about
whether 'a' is preferred to 'b' or 'b' to 'a', the alternative 'a' went from being a non-winner
to being a winner. This shows that independence of irrelevant alternatives fails for the
Least Unpopular procedure. ⊣
Proposition 3.4 There are some social choice procedures for three or more alternatives
that satisfy:
a) the nonexistence of S.D. criterion, monotonicity, and always a winner criterion.
b) the nonexistence of S.D. criterion, Pareto, and always a winner criterion.
Proof. The Least Unpopular procedure is one of them. ⊣ Now we introduce and investigate
The Least Unpopular Reselection procedure.
The Least Unpopular Reselection (L.U.R)
First of all we choose a set of alternatives appearing lesser than the others at the bottom of
individual preference lists. If this set has only one member, it would be the social choice.
Otherwise we remove the remaining alternatives and run the L.U. procedure for the set
obtained from the first stage, and keep doing this until there will be no continuing. The
obtained set in the last repetition would be the social choice. This protocol satisfies AAW,
monotonicity, Pareto, and nonexistence of S.D. criterion, but does not satisfy CWC, and
IIA criterion.
Proposition 3.5 The Least Unpopular Reselection procedure does not satisfy CWC and
IIA criterion.
Proof. Perform as we did in the proof of Proposition 3.3. ⊣
Proposition 3.6 There are some social choice procedures for three or more alternatives
that satisfy the nonexistence of S.D. criterion, Pareto, and monotonicity criterion.
Proof. The L.U.R procedure is one of them. ⊣
4 Conclusion and Future Direction
Here in the following table we summarize what we came to conclusion in the previous
sections:
AAW CWC Pareto Mono
Condorcet
Plurality
Borda count
Hare system
Seq. Pairs
Dictator
L.P.R
Condorcet with an amendment
Seq. Pairs with an amendment
L.U
L.U.R
no
yes
yes
yes
yes
yes
yes
no
yes
yes
yes
yes
no
no
no
no
no
no
no
no
no
no
yes
yes
yes
yes
no
yes
yes
yes
yes
no
yes
yes
yes
yes
no
yes
yes
yes
yes
yes
yes
yes
IIA Non S.D
yes
no
no
no
no
no
no
yes
no
no
no
no
no
no
no
no
no
yes
yes
yes
yes
yes
Finally we must note that the well-known procedures that satisfy the condition IIA are
Condorcet extension which means that they choose the Condorcet winner whenever one
exists. Since CWC is not compatible with the nonexistence of S.D. criterion, we dont
know whether there are some social choice procedures for three alternatives that satisfy the
nonexistence of S.D. criterion, always a winner criterion, and independence of irrelevant
alternatives criterion.
This question can be a motivation for future work.
History and Related Works
Kenneth Arrow showed that it is impossible to design a voting rule that satisfies some very
natural properties (Arrow, 1950)[2]. This seminal work is thus named Arrows impossibility
theorem, and is broadly regarded as the beginning of modern Social Choice Theory, which
is an active research direction in Economics[32].
In terms of the literature, there are, as one might expect, an abundance of treatments
of Arrows theorem. Two of the most notable book-length treatments are Arrow (1963)[2]
and Kelly (1978)[14]. Textbook coverage (with proofs) is also widely available, including
chapters in Kelly (1987)[15], Saari (1995)[21], and Taylor(1995)[28].
The 1970s seem to have been the heyday for social choice research in the second half
of the twentieth century. For references, Kelly (1991)[16] is remarkable. Books on social
choice (from a number of different perspectives) include Sen (1970)[24], Fishburn (1973)[13],
Feldman (1980)[11], Riker (1982)[19], Schofield (1985)[23], Campbell (1992)[8], Shepsle and
Bonchek (1997)[25], Austen-Smith and Banks (2000)[3], and Arrow, Sen, and Suzumura,
(2002)[4]. Books on different aspects of voting include Straffin (1980)[26], Nurmi (1987)[18],
Saari (1994)[20], Felsenthal and Machover (1998)[12], Taylor and Zwicker (1999)[31], and
Saari (2001)[22]. An important recent survey is Brams and Fishburn (2002)[6](see[28], pp.
19,20).
Originating in economics and political science, social choice theory has since found its
place as one of the fundamental tools for the study of multi-agent systems. The reasons for
this development are clear: if we view a multi-agent system as a society of autonomous soft-
ware agents, each of which has different objectives, is endowed with different capabilities,
and possesses different information, then we require clearly defined and well-understood
mechanisms for aggregating their views so as to be able to make collective decisions in
such a multi-agent system(see[7], page:3).
In fact, a burgeoning areaComputational So-
cial Choiceaims to address problems in computational aspects of information/preference
representation and aggregation in multi-agent scenarios(see[32],[9]and[7]).
References
[1] Arrow, K. A difficulty in the concept of social welfare. The Journal of Political Economy
58 (1950), 328-46.
[2] Arrow, K. Social Choice and Individual Values. John Wiley and Sons, (1951). Second
edition (1963).
[3] Austen-Smith, D. and J. Banks, Positive political theory I: collective preferences. Uni-
versity of Michigan Press: Ann Arbor, (2000).
[4] Arrow, K. and A. Sen and K. Suzumura (Eds.), Handbook of social choice and welfare,
vol. I, North-Holland, New York, (2002).
[5] Borda J-C de (1953) M´emoire sur les ´elections au Scrutin. In: Histoire de l'Acad´emie
Royale des Sciences, (1781). (Translated by Grazia A de as Mathematical derivation
of an election system. Isis 44:42-51)
[6] Brams, S. and P. Fishburn, Voting procedures. In the Handbook of Social Choice and
Welfare. Arrow, Sen, and Suzumura, eds. (2002), 175-236.
[7] Brandt, F., Conitzer, V., and Endriss, U. Computational Social Choice. MIT Press
(2012).
[8] Campbell, D. Equity, efficiency, and social choice. Clarendon Press, Oxford, (1992).
[9] Chevaleyre, Y., Endriss, U., Lang, J., and Maudet, N. (2007), A Short Introduction
to Computational Social Choice, in Proceedings of the 33rd Conference on Current
Trends in Theory and Practice of Computer Science (SOFSEM-2007).
[10] Condorcet, Marquis De. Essai sur l'application de l'analyse `a la probabilit´e des d´ecisions
rendues `a la pluralit´e des voix, Paris, (1785).
[11] Feldman, A. Welfare economics and social choice theory. Kluwer: Nijhoff, (1980).
[12] Felsenthal, D. and M. Machover, The measurement of voting power: theory and prac-
tice, problems and paradoxes. Edward Elgar: Cheltenham, UK, (1998).
[13] Fishburn, P. The theory of social choice. Princeton University Press: Princeton, NJ,
(1973).
[14] Kelly, J. Arrow impossibility theorems. Academic Press: New York, (1978).
[15] Kelly, J. Social choice theory: an introduction. Springer-Verlag: New York, (1987).
[16] Kelly, J. Social choice bibliography. Social Choice and Welfare 8 (1991), 97-169.
[17] McLean, I. and Urken, A. B. General introduction, in: I. McLean and A. Urken (eds),
Classics of Social Choice, The University of Michigan Press(1995).
[18] Nurmi, H. Comparing voting systems. D. Reidel Publishing Company: Dordrecht,
Holland, 1987.
[19] Riker,W. Liberalism against populism: a confrontation between the theory of democ-
racy and the theory of social choice. W. H. Freeman: San Francisco, (1982).
[20] Saari, D. The geometry of voting. Springer-Verlag: New York, (1994).
[21] Saari, D. Basic geometry of voting. Springer-Verlag: New York, (1995).
[22] Saari, D. Choatic elections: a mathematician looks at voting. The American Mathe-
matical Society, (2001).
[23] Schofield, N. Social choice and democracy. Springer Verlag: Berlin, (1985).
[24] Sen, A. Collective choice and social welfare. Holden Day: San Francisco, (1970).
[25] Shepsle, K. and M. Bonchek, Analyzing politics: rationality, behavior, and institutions.
Norton: New York and London, (1997).
[26] Straffin, P. Topics in the theory of voting. Birkhauser: Boston, (1980).
[27] Suzumura, K. Handbook of Social Choice and Welfare, Volume 1, Edited by K.J Arrow,
A.K. Sen and K Suzumura. Elsevier Science B.V.(2002), 1-32.
[28] Taylor, A. Mathematics and politics: strategy, voting, power, and proof. Springer-
Verlag: New York, (1995). Second edition (2008).
[29] Taylor, Alan. A glimpse of impossibility, Perspectives on Political Science 26:2326
(1997).
[30] Taylor, Alan. Social Choice and the Mathematics of Manipulation, Cambridge, Eng-
land: Cambridge University Press (2005).
[31] Taylor, A. and W. Zwicker, Simple games: desirability relations, trading, and pseu-
doweightings. Princeton University Press: Princeton, NJ, (1999).
[32] Xia, Lirong. Computational Voting Theory: Game-Theoretic and Combinatorial As-
pects. Dissertation submitted in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in the Department of Computer Science in the Graduate School
of Duke University (2011).
|
1909.00893 | 1 | 1909 | 2019-09-03T00:04:24 | A Predictive Deep Learning Approach to Output Regulation: The Case of Collaborative Pursuit Evasion | [
"cs.MA"
] | In this paper, we consider the problem of controlling an underactuated system in unknown, and potentially adversarial environments. The emphasis will be on autonomous aerial vehicles, modelled by Dubins dynamics. The proposed control law is based on a variable integrator via online prediction for target tracking. To showcase the efficacy of our method, we analyze a pursuit evasion game between multiple autonomous agents. To obviate the need for perfect knowledge of the evader's future strategy, we use a deep neural network that is trained to approximate the behavior of the evader based on measurements gathered online during the pursuit. | cs.MA | cs | A Predictive Deep Learning Approach to Output Regulation: The Case of
Collaborative Pursuit Evasion
S. Shivam1, A. Kanellopoulos2, K. G. Vamvoudakis2, and Y. Wardi1
9
1
0
2
p
e
S
3
]
A
M
.
s
c
[
1
v
3
9
8
0
0
.
9
0
9
1
:
v
i
X
r
a
Abstract -- In this paper, we consider the problem of con-
trolling an underactuated system in unknown, and potentially
adversarial environments. The emphasis will be on autonomous
aerial vehicles, modelled by Dubins dynamics. The proposed
control
law is based on a variable integrator via online
prediction for target tracking. To showcase the efficacy of our
method, we analyze a pursuit evasion game between multiple
autonomous agents. To obviate the need for perfect knowledge
of the evader's future strategy, we use a deep neural network
that is trained to approximate the behavior of the evader based
on measurements gathered online during the pursuit.
I. INTRODUCTION
Output tracking in dynamical systems, such as robots,
flight control, economics, biology, cyber-physical systems,
is the practice of designing decision makers which ensure
that a system's output tracks a given signal [1], [2].
Well-known existing methods for nonlinear output reg-
ulation and tracking include control techniques based on
nonlinear inversions [3], high-gain observers [4], and the
framework of model predictive control (MPC) [5], [6].
Recently a new approach has been proposed, based on
the Newton-Raphson flow for solving algebraic equations
[7]. Subsequently it has been tested on various applications
including controlling an inverted pendulum, and position
control of platoons of mobile robotic vehicles [8], [9]. While
perhaps not as general as the aforementioned established
techniques, it seems to hold out promise of efficient com-
putations and large domains of stability.
The successful deployment of complex control systems
in real world applications increasingly depends on their
ability to operate on highly unstructured -- even adversarial
-- settings, where a-priori knowledge of the evolution of
the environment is impossible to acquire. Moreover, due to
the increasing interconnection between the physical and the
cyber domains, control systems become more intertwined
with human operators, making model-based solutions fragile
to unpredictable. Towards that, methods that augment low-
level control
techniques with intelligent decision making
mechanisms have been extensively investigated in [10]. Ma-
chine learning [11], [12], offers a suitable framework to
1S. Shivam, and Y. Wardi are with the School of Electrical and Computer
Engineering, Georgia Institute of Technology, Atlanta, GA, 30332, USA.
e-mail: ([email protected], [email protected]).
2A. Kanellopoulos, and K. G. Vamvoudakis are with the Daniel
Guggenheim School of Aerospace Engineering, Georgia Institute of Tech-
nology, Atlanta, GA, 30332, USA. e-mail: ([email protected], kyri-
[email protected]).
This work was supported in part, by ONR Minerva under grant No.
N00014 ´ 18 ´ 1 ´ 2160, and by NSF under grants No. SaTC-1801611
and CPS-1851588.
allow control systems to autonomously adapt by leveraging
data gathered from their environment. To enable data-driven
solutions for autonomy,
learning algorithms use artificial
neural networks (NNs); classes of functions that, due to
properties that stem from their neurobiological analogy,
offer adaptive data representations and prediction based on
external observations.
NNs have been used extensively in control applications
[13], both in open-loop and closed-loop fashion. In closed-
loop applications, NNs have been utilized as dynamics ap-
proximators, or in the framework of reinforcement learning,
in enabling online solution of the Hamilton-Jacobi-Bellman
equation [14]. However, the applicability of NNs in open-
loop control objectives is broader, due to their ability to
operate as classifiers, or as nonlinear function approximators
[15].
The authors of [13] introduced NN structures for system
identification as well as adaptive control. Extending the
identification capabilities of learning algorithms, the authors
of [16] introduce a robustification term that guarantees
asymptotic estimation of the state and the state derivative.
Furthermore, reinforcement learning has received increas-
ing attention since the development of methods that solve
optimal control problems for continuous time control sys-
tems online without the knowledge of the dynamics [17].
Prediction has been in the forefront of research conducted
on machine learning. Learning-based attack prediction was
employed both in [18] and [19] in the context of cyber-
security, and [20] utilized NNs to solve a pursuit evasion
game by constructing both the evader's and the pursuer's
strategies offline using pre-computed trajectories. Recently,
authors of this paper have applied NN for on-line model
construction in a control application [21].
This paper applies an NN technique to the pursuit-evasion
problem investigated in [22], which is more challenging
than the problem addressed in [21]. The strategies of both
pursuers and evader are based on respective games. In Ref.
[22], the pursuers know the game of the evader ahead of time,
and an MPC technique is used to determine their trajectories.
In this paper the pursuers do not have an a-priori knowledge
of the evader's game or its structure, and they employ an NN
in real time to identify its input-output mapping. We use our
tracking-control technique [7] rather than MPC, and obtain
similar results to [22]. Furthermore, the input to the system
has a lesser dimension that its output, and hence the control
is underactuated. We demonstrate a way of overcoming this
limitation, which may have a broad scope in applications.
The rest of the paper is structured as follows. Section II
Fig. 1.
Basic control system scheme.
describes our proposed control technique and some prelim-
inary results on NN, and it formulates the pursuers-evader
problem. Section III describes results on model-based and
learning-based strategies. Simulation results are presented
in Section IV. Finally, Section V concludes the paper and
discusses directions for future research.
II. PRELIMINARIES AND PROBLEM FORMULATION
A. Tracking Control Technique
This subsection recounts results published in our previous
work in which prediction-based output tracking was used for
fully-actuated systems [7] -- [9]. Consider a system as shown
in Figure 1 with rptq P Rm, yptq P Rm, uptq P Rm, and
eptq :" rptq ´ yptq. The objective of the controller is to
ensure that
tÑ8rptq ´ yptq ă ε,
lim
(1)
for a given (small) ε P R`.
´Bg
¯´1`
r ´ yptq
To illustrate the basic idea underscoring the controller, let
us first assume that (i) The plant subsystem is a memoryless
nonlinearity of the form
yptq " gpuptqq,
(2)
for a continuously-differentiable function g : Rm Ñ Rm,
and (ii) the target reference trptq : t P r0,8qu is a constant,
rptq " r for a given r P Rm.1 These assumptions will be
relaxed later. In this case, the tracking controller is defined
by the following equation,
9uptq "
puptqq
,
Bu
(3)
assuming that the Jacobian matrix BgBupuptqq is nonsingular at
every point uptq computed by the controller via (3). Observe
that (3) defines the Newton-Raphson flow for solving the
algebraic equation r ´ gpuq " 0, and hence (see [7], [8])
rptq ´
the controller converges in the sense that limtÑ8
yptq
" 0. Next, suppose that the reference target is time-
dependent, while keeping the assumption that the plant is
a memoryless nonlinearity. Suppose that trptqu is bounded,
continuous, piecewise-continuously differentiable, and t 9rptqu
is bounded. Define
`
η :" lim sup
tÑ8
9rptq,
(4)
then (see [8]), with the controller defined by (3), we have
that
tÑ8rptq ´ yptq ď η.
lim
(5)
1Henceforth we will use the notation txptqu For a generic signal
txptq, t P r0,8qu, to distinguish it from its value at a particular point
t, xptq.
Note that Eqs. (2) and (3) together define the closed-loop
system. Observe that the plant-equation (2) is an algebraic
equation while the controller equation (3) is a differential
equation, hence the closed-loop system represents a dynam-
ical system. Its stability, in the sense that typtqu is bounded
whenever trptqu and t 9rptqu are bounded, is guaranteed by
(5) as long as the control trajectory tuptqu does not pass
through a point uptq where the Jacobian matrix BgBupuptqq is
singular.
Finally, let us dispense with the assumption that the plant
subsystem is a memoryless nonlinearity. Instead, suppose
that it is a dynamical system modeled by the following two
equations,
`
x ` 1
9xptq " fpxptq, uptqq, xp0q :" x0
yptq " hpxptqq,
(6)
(7)
where the state variable xptq is in Rn, and the functions
f : Rn Rm Ñ Rn and h : Rn Ñ Rm satisfy the following
assumption.
Assumption 1. (i). The function f : Rn Rm Ñ Rn is
continuously differentiable, and for every compact set Γ Ă
Rm there exists K P R` such that, for every x P Rn and
u P Γ, fpx, uq ď K
. (ii). The function h :
Rn Ñ Rm is continuously differentiable.
l
This assumption ensures that whenever the control signal
tuptqu is bounded and continuous, the state equation (6) has
a unique solution xptq on the interval t P r0,8q.
In this setting, yptq is no longer a function of uptq, but
rather of xptq which is a function of tupτq : τ ă tu.
Therefore (2) is no longer valid, and hence the controller
cannot be defined by (3). To get around this conundrum
we pull the feedback not from the output yptq but from a
predicted value thereof. Specifically, fix the look-ahead time
T P R`, and suppose that at time t the system computes
a prediction of ypt ` Tq, denoted by ypt ` Tq. Suppose
also that ypt ` Tq is a function of pxptq, uptqq, hence can
be written as ypt ` Tq " gpxptq, uptqq, where the function
g : Rn Rm Ñ Rm is continuously differentiable.
Now the feedback law is defined by the following equa-
Bu
. (8)
pxptq, uptqq
rpt ` Tq ´ gpxptq, uptqq
tion,
9uptq "
The state equation (6) and control equation (8) together
define the closed-loop system. This system can be viewed
as an pn ` mq-dimensional dynamical system with the state
variable pxptqT, uptqTqT P Rn`m and input rptq P Rm. We
are concerned with a variant of Bounded-Input-Bounded-
State (BIBS) stability whereby if trptqu and t 9rptqu are
bounded, txptqu is bounded as well. Such stability no-longer
can be taken for granted as in the case where the plant is a
memoryless nonlinearity.
We remark that a larger T means larger prediction errors,
and these translate into larger asymptotic tracking errors. On
the other hand, an analysis of various second-order systems
in [7] reveals that they all were unstable if T is too small, and
¯´1`
´Bg
r(t) e(t) � Controller u(t) Plant y(t) � � +.- stable if T is large enough. It can be seen that, a requirement
for a restricted prediction error can stand in contradiction
with the stability requirement. This issue was resolved by
speeding up the controller in the following manner. Consider
α ą 1, and modify (8) by multiplying its right hand side by
α, resulting in the following control equation:
9uptq " α
rpt ` Tq ´ gpxptq, uptqq
¯´1`
pxptq, uptqq
´Bg
.
Bu
It was verified in [7] -- [9], that regardless of the value of
T P R`, a large-enough α stabilizes the closed-loop system.2
Furthermore, if the closed-loop system is stable then the
following bound holds,
lim sup
tÑ8
rptq ´ yptq ď η
α
,
(9)
where η is defined by (4). Thus, a large gain α can stabilize
the closed-loop system and reduce the asymptotic tracking
error.
B. Problem Formulation
In an attempt to broaden the application scope of the
control algorithm, underactuated systems such as the fixed-
wing aircraft are explored, which are widely used in the
domain of aerospace engineering. The behavior of a fixed
wing aircraft at constant elevation can be approximated by
a planar Dubins vehicle with 3 states [23] @t ě 0,
1ptq " V p cos θpptq,
9zp
2ptq " V p sin θpptq,
9zp
9θpptq " uptq,
it
„
«
ff
1ptq, zp
where pzp
2ptqqT denotes the planar position of the
vehicle, θpptq its heading and uptq the angular acceleration,
constrained as, (cid:107)u(cid:107) ď umax. The input saturation enforces
a minimum turning radius equal to V0{umax. For testing
the efficacy of the controller for the underactuated system,
henceforth referred to as the pursuer,
is tasked with
tracking an evading vehicle, modeled as a single integrator,
with dynamics as follows:
1ptq
ze
2ptq
ze
d
dt
2ptqqJ denote the planar position of the
1ptq, ze
where pze
evader, and V e is its speed. We consider two cases; one
where the evader is agnostic to the pursuer and follows
a known trajectory and the other where the the evader is
adversarial
in nature and its trajectory is not known to
the pursuer. The next section will provide two solutions
for the problem of estimating the evader's trajectory based,
respectively, on a model-based approach and a learning-
based approach.
V e cos θe
V e sin θe
"
,
2This statement seems to have a broad scope, and does not require the
plant to be a minimum-phase system.
III. PREDICTIVE FRAMEWORK
1ptq, zp
A. Model-Based Pursuit Evasion
The considered system is underactuated because the pur-
2ptqqJ, is two-dimensional while it is
suer's position, pzp
controlled by an one-dimensional variable, uptq. This raises
a problem since the application of the proposed tracking
technique requires the control variable and system's output
to have the same dimension. To get around this difficulty,
we define a suitable function F : R2 Ñ R` and set
gpxptq, uptqq :"
Fpyppτq ´ yepτqqdτ where yppτq and
yepτq are the predicted position of the pursuer and the evader
at time τ; we apply the Newton-Raphson flow to the equation
gpxptq, uptqq " 0. The modified controller becomes
ş
´Bg
¯´1`
t`T
t
gpxptq, uptqq
, t ě 0.
9uptq " ´α
pxptq, uptqq
Bu
(10)
the modified algorithm works
Since gpx, uq is a scalar,
similar to the base case.
Assume general nonlinear system dynamics as in (6) with
output described in (7). The predicted state trajectory is
computed by holding the input to a constant value over
the prediction horizon, given by the following differential
equation:
pτq ` Bf
Bu
pτq " Bf
Bξ
9ξpτq " fpξpτq, uptqq, τ P rt, t ` Ts,
pξpτq, uptqqBξ
Bu
(11)
with the initial condition ξptq " xptq as shown in [7]. The
predicted output at τ is yppτq " hpξpτqq. Furthermore, by
taking the partial derivative of (11) with respect to u(t), we
obtain
9Bξ
Bu
(12)
with the initial condition BξBuptq " 0. The above is a
differential equation in BξBupτq; τ P rt, t ` Ts and (11) and
(12) can be solved numerically. Finally, the values of gpx, uq
and BgBupx, uq can be substituted in (10) to get the control law.
In the next section, results are presented for an agnostic
as well as an adversarial pursuer- evader system. However,
as mentioned above, in the adversarial problem formulation,
the trajectory of the evader is not known in advance, which
can be overcome in two ways.
pξpτq, uptqq,
In the first approach, the pursuer(s) use game theory to
predict the approximate direction of evasion. As mentioned
in [24], in the case of single pursuer, the evader's optimal
strategy is to move along the line joining the evader and
pursuer's position, if the pursuer is far enough. When the
distance between the pursuer and the evader reduces to the
turning radius of the pursuer, the evader switches strategies
and enters into the non-holonomic constraint region of the
pursuer. This can be represented as follows:
θE "
arctan
arctan
2ptq´zp
ze
1ptq´zp
ze
2ptq
1ptq
2ptq´zp
ze
1ptq´zp
ze
2ptq
1ptq
,
d ą RP ,
π{2, d ď RP .
(13)
$''''&''''%
Here θE is the expected evasion angle of the evader and d
is the distance between the pursuer and evader,
If there are multiple pursuers, it is assumed that the evader
follows the same strategy by considering only the closest
pursuer. It is notable that this will not provide the pursuers a
correct prediction of the evader's motion as they do not know
about the goal seeking behavior mentioned above. However,
it gives a good enough approximation of the pursuer's motion
that the algorithm can be used for tracking.
The second approach involves learning the evader's be-
havior over time using NN. The pursuers take their positions
and the position of the evader as input and the NN gives the
estimated evasion direction as the output after training.
To showcase the efficacy of our method, we consider a
pursuit evasion problem, involving multiple pursuing agents.
Such problems are typically formulated as zero-sum dif-
ferential games [24]. Due to the difficulty of solving the
underlying Hamilton-Jacobi-Isaacs (HJI) equations [25] of
this problem, we shall utilize the method described in II-A
to approximate the desired behavior. Furthermore, we show
that augmenting the controller with learning structures in
order to tackle the pursuit evasion problem without explicit
knowledge of the evader's behavior is straightforward.
In order to formulate the pursuit evasion problem, we
define a global state space system consisting of the dynamics
of the pursuers and the evader. For ease of exposition, the
analysis will focus on the 2-pursuer, 1-evader problem, since
extending the results to multiple pursuers is straightforward.
The global state dynamics become,
» -- -- -- -- -- -- -- -- -- -- -- -- --
"
» -- -- -- -- -- -- -- -- -- -- -- -- --
fiffiffiffiffiffiffiffiffiffiffiffiffifl
1 ptq
p1
z
2 ptq
p1
z
θp1ptq
1 ptq
p2
z
2 ptq
p2
z
θp2ptq
1ptq
ze
2ptq
ze
d
dt
fiffiffiffiffiffiffiffiffiffiffiffiffifl ,
V p1 cos θp1
V p1 sin θp1
up
1
V p1 cos θp2
V p2 sin θp2
up
2
V e cos θe
V e sin θe
(14)
where the subscripts indicate the autonomous agent. For
compactness, we denote the global state vector as xptq P R8,
the pursuers' control vector uptq P R2, and the nonlinear
mapping described by the right-hand side of (14). Thus,
given the initial states of the agents x0 P R8, the evolu-
tion of the pursuit evasion game is described by 9xptq "
fpxptq, u, ueq, xp0q " x0, t ě 0.
Subsequently, this zero-sum game can be described as a
e´γtLpxqdt
β1pd2
e´γt
minimax optimization problem through the cost index,
Jpx, u, ueq "
:"
a
1 ´ zeq2 ` pzi
pzi
dt,
(15)
where di "
2q2, i P tp1, p2u is the
distance between the i-th pursuer and the evader, β1, β2 P
d2
1d2
2
1 ` d2
d2
2q ` β2
1 ` d2
2 ´ ze
0
2
ż 8
ż 8
0
R` are user defined contants, and γ P R` is a discount
factor. The first term ensures that the pursuers remain close
to the evader, while the second term encourages cooperation
between the agents. The cost decreases exponentially to
ensure that the integral has a finite value in the absence of
equilibrium points.
Let V pxq : R8 Ñ R be a smooth function quantifying the
value of the game when specific policies are followed starting
from state xptq. Then, we can define the corresponding
`
Hamiltonian of the game as,
H
The optimal feedback policies u‹pxq, u‹
(16)
epxq of this game
" Lpxq ` BV
Bx
fpx, u, ueq ` γV.
BV
Bx
x, u, ue,
are known to constitute a saddle point [25] such that,
T
u
u‹pxq " arg min
epxq " arg max
u‹
`
ue
x, u‹, u‹
e,
H
Hpx, u, ueq,
Hpx, u, ueq.
‹
" 0.
BV
Bx
(17)
(18)
(19)
Under the optimal policies (17),(18), the HJI equation is
satisfied,
Evaluating the optimal pursuit policies, yields the singular
optimal solutions described by, Vθp1u1 " Vθp2 u2 " 0, where
Vxi is the partial derivative of the value function with respect
to the state xi, calculated by solving (19). To obviate the
need for bang-bang control, as is derived by (17) and (18)
we shall employ the predictive tracking technique described
in Section II-A to derive approximate, easy to implement,
feedback controllers for the pursuing autonomous agents.
Furthermore, by augmenting the predictive controller with
learning mechanisms, the approximate controllers will have
epxq, the evader's policy.
no need for explicit knowledge of u‹
The following theorem presents bounds on the optimality
loss induced by the use of the look-ahead controller approx-
imation.
Theorem 1. Let the pursuit evasion game evolve according to
the dynamics given by (14), where the evader is optimal with
respect to (15) and the pursuers utilize the learning-based
predictive tracking strategy given (10). Then, the tracking
error of the pursuers and the optimality loss due to the use
of the predictive controller are bounded if D ¯∆ P R`, such
that, ∆pxptq, uptq, uptqeq ď ¯∆, @t ě 0, where ∆px, u, ueq "
1 ´
Vxe vepcos ue ´ cos u‹
2 ´ u2q, with Vξ denoting the partial derivative
u1q ` Vθp2pu‹
of the game value with respect to the state component ξptq.
Proof: Consider the Hamiltonian function when the ap-
proximate controller, denoted uptq and the NN-based pre-
diction of the evader's policy, ueptq are used,
eq ` Vye vepsin ue ´ sin u‹
eq ` Vθp1pu‹
Hpx, u, ueq " Lpxq `
fpx, u, ueq ` γV.
(20)
T
`BV
Bx
Taking into account the nonlinear dynamics of the system
(14), one can rewrite (20) in terms of the optimal Hamil-
tonian as,Hpx, u, ueq " Hpx, u‹, u‹
eq ` ∆pu, ueq, where
Hpx, u‹, u‹
eq " 0 is the HJI equation that is obtained after
T
`BVBx
substituting (17) and (18) in (16). Now, take the orbital
derivative of the value function along the trajectories using
fpx, u, ueq.
the approximate controllers as,
Substituting (20) yields 9V " ´Lpxq ´ γV ` ∆px, u, ueq.
Thus, since Lpxq ą 0, @x P R8zt0u,
9V "
9V ă ´γV ` ∆px, u, ueq ñ 9V ă ´γV ` ¯∆.
Hence for V ě ¯∆{γ, we have 9V ď 0. Thus tx P
R8 V pxq ď ¯∆{γu is a forward invariant set, which implies
that the tracking error and the optimality loss over any finite
horizon is bounded.
Remark 1. Note that we do not use optimal control or MPC
to solve the pursuit evasion problem. Instead, the controller
is governed by (10), which is simple to implement and has
l
low computational complexity.
B. Deep Learning-Based Pursuit Evasion
A deep NN, consisting of L ą 2 hidden layers, describes
a nonlinear mapping between its input space Rn and output
space Rp. Each layer receives the output of the previous
layer as an input and, subsequently, feeds its own output to
the next layer. Each layer's output consists of the weighted
sum of its input alongside a bias term, filtered through an
application-specific activation function [11].
Specifically, let Rnl be the input space of a specific layer,
the corresponding output space. Then the layer's
and Rpl
output is,
X1
. . . Xnl
where X1 "
gathered from training data or from the output of previous
layers, vij P R is a collection of nl weights for each layer,
vi0 P R the bias term and σ : Rnl Ñ R is the layer's
activation function. We note that it is typical to write the
output of layer compactly, with slight abuse of notation, as,
"
(21)
P Rpnl`1qpl
Y1
l Ñ Rnl is the activation function of the previous
Y " σpW Tσ1pXqq,
"
P Rpl, W "
. . . Ypl
where Y "
and σ1 : Rn1
layer, taking as input the vector X "
ıT
X1T
‰
"
‰
vij
1
.
It is known [26], that two-layer NNs possess the universal
approximation property, according to which, any smooth
function can be approximated arbitrarily close by an NN
of two or more layers. Let S Ă Rn be a simply connected
compact set and consider the nonlinear function κ : S Ñ Rp.
Given any b ě 0, there exists a NN such structure such that,
W Tσ1pxq
where }} ď b. We note that,
function of the output layer σpq is taken to be linear.
` , @x P S,
typically,
κpxq " σ
the activation
`
Evaluating the weight matrix W in a network is the main
concern of the area of machine learning. In this work, we em-
ploy the gradient descent based backpropagation algorithm.
Given a collection of Nd training data, stored in the tuple
nlÿ
j"1
Yipxq " σ
"
vijXj ` vi0
‰T P Rnl is the input vector,
, i " 1, 2, . . . , pl,
txk, κkuk, where xk P Rn, κk P Rp, @k " 1, . . . , Nd, we
denote the output errors as rk " κpxkq´κk. Then, the update
equation for the weights at each optimization iteration tk is
given by,
BprT
krkq
Bwij
1 , δzpi
1 , δzp1
, @tk P N,
wijptk ` 1q " wijptkq ´ η
1 , δzpN
2 , . . . , δzpN
this presupposes that
2 q, where, pδzpi
(22)
where η P R` denotes the learning rate. We note that the
update index tk need not correspond to the sample index
k, since different update schedules leverage the gathered
data in different ways [26]. It can be seen that in order
for the proposed method to compute the pursuers' control
inputs, an accurate prediction of the future state of the evader
is required. However,
the pursuers
themselves have access to the evader's future decisions; an
assumption that is, in most cases, invalid. Thus, we augment
the pursuers' controllers with a NN structure, that learns to
predict the actions of the evader, based on past recorded data.
Initially, we assume that the evader's strategy is com-
puted by a feedback algorithm, given her relative posi-
tion to the pursuers. This way, the unknown function we
wish to approximate is f : R2N Ñ R2, with, ue "
fpδzp1
2 q denote
the distance of pursuer i to the evader in the X and Y axes,
respectively. In order to train the network, we let the pursuers
gather data regarding the fleet's position with respect to
the evader, as well as her behavior over a predefined time
window Tl ą 0.
Remark 2. Increasing the time window Tl will allow the
pursuers to gather more training data for the predictive
network. However,
this will not only increase the com-
putational complexity of the learning procedure, but will
make the pursuers more inert
to sudden changes in the
evader's behavior. Simulation results corroborate our choice
`
l
of training parameters.
‰
Subsequently, we denote by uepxq, the current prediction
W T σ1pχq
function for the evader's strategy, i.e., uepxq " σ
,
where χ "
P R2N , W
denotes the current weight estimate of the NNs output
layer, and σ1pq is the current estimate of the hidden layers,
parametrized by appropriate hidden weights.
Remark 3. While the learning algorithm for the evader's
behavior operates throughout the duration of the pursuit, thus
making the approximation weights time-varying, we suppress
their explicit dependence on time since the process is open-
loop, in the sense that the system is learning in batches,
l
rather that in a continuous fashion.
δxN δyN
"
δy1
δz1
. . .
IV. SIMULATION RESULTS
This section presents results for the problems briefly
described in the previous section. First, the agnostic evader
case is considered followed by the adversarial case. For
the second case, single and multiple pursuer systems are
considered separately. The controller is implemented on a
Dubins vehicle. For the purpose of tracking, we define the
system output to be yi "
‰T, i P tp1, p2, eu.
"
zi
1
zi
2
A. Single Pursuer - Agnostic Target
In this subsection, the controller is tested on a Dubins
vehicle with the task of pursuing an agnostic target moving
along a known trajectory. Since the vehicle has a constant
speed and an input saturation is enforced, it has an inherent
minimum turning radius. For this simulation, we set V p "
2 rad/s and then
2 m/s and the input saturation is first set to π
to 2π rad/s. The evader moves along two semicircular curves
with a constant speed which is less than V p.
As a consequence, when the pursuer catches up to the
evader, it overshoots and has to go around a full circle to
again start tracking. Naturally, lower turning radius translates
to better tracking as the vehicle can make "tighter" turns.
This can be seen when comparing the trajectories of the
vehicle in Figure 2 with Figure 4. For the same trajectory
of the evader, the tracking performance is far better in the
second case. Once the pursuer catches up to the target, the
maximum tracking error in the first case is approximately
4 meters and only 1 meter in the second case, shown in
Figures 3 and 5. This is consistent with the fact that the
ratio of the turning radii is 4 : 1.
B. Single Pursuer - Adversarial Evader
The pursuer is again modelled as a Dubins vehicle, while
the evader is modelled as a single integrator with a maximum
velocity less than the speed of the pursuer. Hence, while the
pursuer is faster, the evader is more agile, and can instantly
change its direction of motion. In this and subsequent cases,
the evader is considered adversarial in nature and uses game
theory to choose evasion direction.
Let ypptq and yeptq be the position vector of the pursuer
and evader respectively at time t. First, the pursuer makes
an estimate of the optimal evasion direction based on the
relative position of the evader and itself at time t using
(13). Assuming this direction of evasion to be fixed over
the prediction window from t to t ` T gives the predicted
position of the evader at all time instances in this interval,
Algorithm 1 Deep Learning-Based and Predictive Pursuit
Evasion
Inputs: XPiptq, @it1, . . . , Nu, XEptq and evasion strategy
approximation weights W .
Output: uPiptq, @it1, . . . , Nu.
1: Compute pδxi, δyiq, i P t1, . . . , Nu.
2: Predict evader's future behavior via (21).
3: Train NN as in (22).
4: Predict evader's future state as XEpt ` Tq " XEptq `
5: Propagate pursuer dynamics to get XPpt ` Tq.
6: Computed current Newton flow parameters using (23).
7: Computed control dynamics 9uPiptq from (3).
8: Propagate actual system evolution using (14).
9: Append current distances pδxi, δyiq to a stack of previ-
rVE cos θE VE sin θEsTT .
ous observations.
10: Update evader prediction network through (22).
Fig. 2. Agnostic evader with a large turning radius.
Fig. 3. Evolution of an agnostic evader tracking error with a large turning
radius.
Fig. 4. Agnostic evader with a small turning radius.
Fig. 5. Evolution of the agnostic evader tracking error with a small turning
radius.
denoted as yepτq, τ P rt, t ` Ts. Next, the pursuer estimates
its own predicted position if its input is kept constant, called
yppτq, τ P rt, t ` Ts. Finally, gptq is set as yept ` Tq ´
yppt ` Tq2 and the value of BgBupxptq, uptqq (xptq being the
ensemble vector of the states of the pursuer and the evader)
is used to compute the input differential equation (10).
Figures 6 shows the trajectories of the pursuer and the
evader, with the goal for the evader set to to point p150, 60q.
It can be observed that the evader moves towards the goal
while the pursuer is far away and starts evasive maneuvers
when it gets close to it, by entering its non-holonomic region.
Figure 7 displays the tracking error, defined as the distance
between the pursuer and the evader, which is almost periodic.
This is because the evader's maneuver forcing the pursuer to
circle back. The peak tracking error after the pursuer catches
up is slightly more than twice the turning radius, as expected.
of the tracking error of the model-based algorithm vis-`a-
vis the NN-based control. Figure 12 showcases the quality
of the performance of the proposed algorithm based on the
game theoretic cost metric. From these figures, it can be seen
that the NN structure offers fast predictive capabilities to the
controller; hence the overall performance is comparable to
the model based control.
Fig. 6.
Trajectories for a single pursuer-evader system.
Fig. 7. Evolution of the tracking error for a single pursuer-evader system.
C. Multiple Pursuers - Adversarial Evader
While the previous section had only one pursuer, this
simulation considers the case of two pursuers and a single
evader. Having multiple pursuers means there must be coop-
eration between them in order to optimally utilize resources.
Thus, a pursuer can no longer make decisions solely based
on the position of the evader relative to itself. The positions
of the rest of the pursuers must also be factored in. Thus
we redefine the expression for gpx, uq to include these
parameters as shown below for the case of two pursuers.
Let dp be the distance between the two pursuers, and let
*
2pτq
1pτqd2
2pτqq ` β2
gpxptq, uptqq :"
d2
2pτq
1pτq ` d2
d2
dτ, @t ě 0.
(23)
"
β1pd2
` β3e
1pτq ` d2
´γdppτq
ż
t`T
t
The first term ensures that the pursuers remain close to
the evader, while the second term encourages cooperation
between agents. The last term is added to repel pursuers
apart if they come close to each other, as having multiple
pursuers in close vicinity of each other is sub-optimal.
Figure 8 shows the trajectories of the pursuers and the
evader when the goal for the evader is set to the point
p15,´1q. In this case, the pursuers close in on the evader
and trap it away from its goal due to their cooperative
behavior. The evader is forced to continuously perform
evasive maneuvers as the other pursuer closes in when the
first has to make a turn. This can be seen more clearly in
the tracking error plot given in Figure 9. After catching up
with the evader, it can be seen that when one pursuer is at its
maximum distance, the other is at its minimum. The results
achieved show good coordination between the pursuers and
low tracking error and are qualitatively comparable to [22].
Lastly, we present the results under the learning-based
prediction. In Figure 11, we present a comparative result
Fig. 8.
Trajectories for the two pursuer-single evader system.
Fig. 9. Evolution of the tracking error for the two pursuer-single evader
system.
V. CONCLUSION AND FUTURE WORK
This work extends the framework of prediction-based
nonlinear tracking in the context of pursuit evasion games.
We present results for vehicle pursuit of agnostic targets,
modeled as moving along known trajectories, as well as
adversarial target tracking, where the evader evolves accord-
ing to game-theoretic principles. Furthermore, to obviate the
need for explicit knowledge of the evader's strategy, we em-
ploy learning algorithms alongside the predictive controller.
The overall algorithm is shown to produce comparable results
to those in the literature, while it precludes the need for
solving an optimal control problem.
Future work will focus on developing robustness guaran-
tees will allow for more realistic scenarios, where noise and
external disturbances are taken into consideration.
REFERENCES
[1] S. Devasia, D. Chen, and B. Paden, "Nonlinear inversion-based output
tracking," IEEE Transactions on Automatic Control, vol. 41, no. 7, pp.
930 -- 942, 1996.
[2] P. Martin, S. Devasia, and B. Paden, "A different look at output
tracking: control of a vtol aircraft," Automatica, vol. 32, no. 1, pp.
101 -- 107, 1996.
[3] A. Isidori and C. Byrnes, "Output regulation of nonlinear systems,"
IEEE Trans. Automat. Control, vol. 35, pp. 131 -- 140, 1990.
[4] H. Khalil, "On the design of robust servomechanisms for minimum
phase nonlinear systems," Proc. 37th IEEE Conf. Decision and Con-
trol, Tampa, FL, pp. 3075 -- 3080, 1998.
and Design of Hybrid Systems (2018 ADHS)l, Oxford, the UK, July
11-13, 2018.
[9] S. Shivam, I. Buckley, Y. Wardi, C. Seatzu, and M. Egerstedt, "Track-
ing control by the newton-raphson flow: Applications to autonomous
vehicles," in European Control Conference, Naples, Italy, June 25-28,
2018.
[10] G. Saridis, "Intelligent robotic control," IEEE Transactions on Auto-
matic Control, vol. 28, no. 5, pp. 547 -- 557, 1983.
[11] S. S. Haykin, Neural networks and learning machines. Pearson Upper
Saddle River, 2009, vol. 3.
[12] D. Vrabie, K. G. Vamvoudakis, and F. L. Lewis, Optimal adaptive
control and differential games by reinforcement learning principles.
IET, 2013, vol. 2.
[13] K. S. Narendra and K. Parthasarathy, "Identification and control of
dynamical systems using neural networks," IEEE Transactions on
neural networks, vol. 1, no. 1, pp. 4 -- 27, 1990.
[14] K. G. Vamvoudakis and F. L. Lewis, "Online actor -- critic algorithm to
solve the continuous-time infinite horizon optimal control problem,"
Automatica, vol. 46, no. 5, pp. 878 -- 888, 2010.
[15] C. M. Bishop et al., Neural networks for pattern recognition. Oxford
university press, 1995.
[16] S. Bhasin, R. Kamalapurkar, M. Johnson, K. G. Vamvoudakis, F. L.
Lewis, and W. E. Dixon, "A novel actor -- critic -- identifier architecture
for approximate optimal control of uncertain nonlinear systems,"
Automatica, vol. 49, no. 1, pp. 82 -- 92, 2013.
[17] K. G. Vamvoudakis, "Q-learning for continuous-time linear systems:
A model-free infinite horizon optimal control approach," Systems &
Control Letters, vol. 100, pp. 14 -- 20, 2017.
[18] B. G. Weber and M. Mateas, "A data mining approach to strategy
prediction," in 2009 IEEE Symposium on Computational Intelligence
and Games.
IEEE, 2009, pp. 140 -- 147.
[19] T. Alpcan and T. Bas¸ar, Network security: A decision and game-
theoretic approach. Cambridge University Press, 2010.
[20] H. J. Pesch, I. Gabler, S. Miesbach, and M. H. Breitner, "Synthesis
of optimal strategies for differential games by neural networks," in
New Trends in Dynamic Games and Applications.
Springer, 1995,
pp. 111 -- 141.
[21] A. Kanellopoulos, K. Vamvoudakis, and Y. Wardi, "Predictive learning
via lookahead simulation," in AIAA Scitech 2019 Forum, San Diego,
California, January 7-11, 2019.
[22] S. A. Quintero, D. A. Copp, and J. P. Hespanha, "Robust uav coordina-
tion for target tracking using output-feedback model predictive control
with moving horizon estimation," in American Control Conference,
Chicago, Illinois, July 1-3, 2015.
[23] S. M. LaValle, Planning algorithms.
Cambridge university press,
2006.
[24] R. Isaacs, Differential games: a mathematical theory with applications
to warfare and pursuit, control and optimization. Courier Corpora-
tion, 1999.
[25] T. Basar and G. J. Olsder, Dynamic noncooperative game theory.
Siam, 1999, vol. 23.
[26] F. Lewis, S. Jagannathan, and A. Yesildirak, Neural network control
of robot manipulators and non-linear systems. CRC Press, 1998.
Fig. 10. Trajectories for two pursuers-single evader system with learning.
Fig. 11.
learning.
Evolution of the tracking error for the systems with and without
Fig. 12. Total cost for the system with and without learning.
[5] F. Allgower and A. Zheng, Nonlinear model predictive control.
Birkhauser, 2012, vol. 26.
[6] J. Rawlings, D. Mayne, and M. Diehl, Model Predictive Control:
Nob Hill, LLC,
Theory, Computation, and Design, 2nd Edition.
2017.
[7] Y. Wardi, C. Seatzu, M. Egerstedt, and I. Buckley, "Performance reg-
ulation and tracking via lookahead simulation: Preliminary results and
validation," in 56th IEEE Conf. on Decision and Control, Melbourne,
Australia, December 12-15, 2017.
[8] Y. Wardi, C. Seatzu, and M. Egerstedt, "Tracking control via variable-
gain integrator and lookahead simulation: Application to leader-
follower multiagent networks," in Sixth IFAC Conference on Analysis
|
1905.04083 | 1 | 1905 | 2019-05-10T11:48:27 | ES-CTC: A Deep Neuroevolution Model for Cooperative Intelligent Freeway Traffic Control | [
"cs.MA"
] | Cooperative intelligent freeway traffic control is an important application in intelligent transportation systems, which is expected to improve the mobility of freeway networks. In this paper, we propose a deep neuroevolution model, called ES-CTC, to achieve a cooperative control scheme of ramp metering, differential variable speed limits and lane change control agents for improving freeway traffic. In this model, the graph convolutional networks are used to learn more meaningful spatial pattern from traffic sensors, a knowledge sharing layer is designed for communication between different agents. The proposed neural networks structure allows different agents share knowledge with each other and execute action asynchronously. In order to address the delayed reward and action asynchronism issues, the evolutionary strategy is utilized to train the agents under stochastic traffic demands. The experimental results on a simulated freeway section indicate that ES-CTC is a viable approach and outperforms several existing methods | cs.MA | cs |
ES-CTC: A Deep Neuroevolution Model for Cooperative Intelligent Freeway
Traffic Control
1School of Mechanical Engineering, Beijing Institute of Tachnology, China
Yuankai Wu1 , Huachun Tan2∗ , Zhuxi Jiang3 and Bin Ran4
2∗School of Transportation Engineering, Southeast University, China
3Momenta, China
[email protected], 2∗[email protected], [email protected], [email protected],
4College of Engineering, University of Wisconsin-Madison, USA
Abstract
Cooperative intelligent freeway traffic control is an
important application in intelligent transportation
systems, which is expected to improve the mobil-
ity of freeway networks. In this paper, we propose
a deep neuroevolution model, called ES-CTC, to
achieve a cooperative control scheme of ramp me-
tering, differential variable speed limits and lane
change control agents for improving freeway traf-
fic. In this model, the graph convolutional networks
are used to learn more meaningful spatial pattern
from traffic sensors, a knowledge sharing layer
is designed for communication between different
agents. The proposed neural networks structure al-
lows different agents share knowledge with each
other and execute action asynchronously. In order
to address the delayed reward and action asynchro-
nism issues, the evolutionary strategy is utilized to
train the agents under stochastic traffic demands.
The experimental results on a simulated freeway
section indicate that ES-CTC is a viable approach
and outperforms several existing methods.
1 Introduction
The ongoing drastic expansion of car ownership and travel
demand have led to increasing freeway congestion, with ad-
verse effects on the economy. To relieve freeway congestion,
numerous freeway traffic control approaches, e.g. dynamic
routing, variable speed limit (VSL), ramp metering (RM),
lane change control (LCC) etc., are studied. From a system-
atic viewpoint, using one management approach alone cannot
fully optimize the freeway traffic in practice. The mainlane
flow, on-ramp flow, routing behaviors and lane changing be-
haviors need to be regulated in a coordinated manner in order
to improve the freeway condition. This is the motivation for
investigating the coordination of different traffic control ap-
proaches.
There is a large volume of published studies describing the
cooperative traffic control: Hedgy et.al [2005] developed a
predictive coordinated control approach for the coordination
of VSL and RM. Carlson et.al [2010] formulated coordinated
VSL and RM control as an optimal control problem using
second-order traffic flow model. Recently, the coordination
of RM, VSL and LCC under connected autonomous vehicle
environment was studied [Roncoli et al., 2015]. Two limita-
tions worth noting in respect of the studies mentioned above
are: 1) The control model are highly dependent on the in-
tegrated traffic flow models, which are inevitably inconsis-
tent with the real-world traffic breakdown. 2) The success
of proactive approaches are based on robustness and reliabil-
ity of the short-term traffic prediction model. The accurate
and reliable short-term traffic prediction is not an easy task
because the evolution of traffic state is related to many fac-
tors [Wu et al., 2018b].
Recently, the advent of deep reinforcement learning (DRL)
has lead to potential applications of reinforcement learning
(RL) techniques to tackle challenging control problems in in-
telligent transportation systems. DRL has given promising
results in RM [Belletti et al., 2018], traffic light control [Wei
et al., 2018], differential VSL control [Wu et al., 2018c], fleet
management [Lin et al., 2018] and hybrid electric vehicle en-
ergy management [Wu et al., 2018a]. The utilization of deep
learning algorithms within RL allows a well-trained traffic
control agent achieves a proactive control scheme, and opti-
mizes the transportation benefits. The success of DRL on one
specific traffic control approach hold great promise for ap-
plication of DRL on coordination of different traffic control
approaches.
However, the coordination of different traffic control ap-
proaches within one DRL framework is not an easy task. The
first challenge is due to the difference between the control cy-
cle of different agents. In many situations, the agents change
actions asynchronously, a somewhat different situation from
that familiar from popular multi-agent DRL frameworks [Fo-
erster et al., 2016; Lowe et al., 2017]. For example, the agents
controlling on-ramp flow should decide whether to change
traffic light phase every few seconds. While the control cy-
cle for VSL agents are always above 1 minute because a fre-
quently change speed limit will unstablilize the traffic flow.
The second challenge stems from the difficulties in defin-
ing a representative reward signal for different traffic control
agents. The aim of traffic management would be to reduce
travel time and increase traffic flow. However, the average
travel time and total flow cannot be computed until all the
vehicles have completed their routes, which causes the issue
of delayed rewards [Van der Pol and Oliehoek, 2016]. The
delayed rewards would cause further credit assignment prob-
lems in multi-agent DRL [Foerster et al., 2017].
The third challenge lies in the modeling of the traffic state.
Traditional, the traffic state collected from sensors are mod-
eled as images and/or vectors, and is directly taken as an input
for a convolutional neural networks (CNN) [Wei et al., 2018]
or fully connected neural networks (FC) [Li et al., 2016].
However, sensors on the road network contain complex spa-
tial correlations and exhibits graph structure. There have been
numerous studies reported that the graph convolutional net-
work (GCN) is more suitable for modeling spatial correlation
of traffic sensors than CNN and FCN in traffic prediction [Li
et al., 2018; Lv et al., 2018].
To tackle those challenges, we propose a deep neuroevolu-
tion [Salimans et al., 2017] based multi-agent framework for
cooperative traffic control (ES-CTC). The main contributions
of this paper can be summarized as follows:
1. We find that the deep neuroevolution approach is a per-
fect match for cooperative traffic control. In deep neu-
roevolution approach like evolutional strategies (ES), the
only feedback signal for different agents is the final re-
turn of an episode. As a result, the problem of delayed
reward is readily solved with ES.
2. We proposed a novel structure named knowledge shar-
ing graph convolutional nets (KS-GCN) to generate con-
trol actions from state collected from traffic sensors.
GCN is used as the building block for the proposed
structure, which can fully capture the spatial depen-
dency between different sensors. The structure allows
communication and knowledge-sharing between differ-
ent agents. Based on the knowledge sharing layer, the
neural agent can coordinate with other agents by execut-
ing action in its own control circle.
3. The travel demands for training the neural networks
are modeled as a stochastic distribution, leading to the
changes in system dynamics of the environment. The ex-
periments show that the proposed approach works well
under stochastic travel demands.
2 Problem Statement
The freeway section considered in this paper is given in Fig-
ure 1. The freeway section in Figure 1 is composed by mul-
tiple lanes and it presents an on-ramp and an off-ramp. As it
may be seen in the figure, the interference between vehicles
is appearing in the merging area between inflow of on-ramp
and outflow of mainstream. The conflicts cause further speed
reductions in the merging area, contributing to the creation of
a generalised bottleneck.
Following the statement in [Roncoli et al., 2015], we con-
sider that the freeway flow with a high ratio of connected au-
tonomous vehicle (CAV). Therefore the differential VSL and
LCC can be successfully implemented. More specifically, the
following control agents are considered in this paper:
• Ramp-metering agent: The agent is to regulate the in-
flow from on-ramp to mainstream by change the phase
of the traffic light in on-ramp.
Figure 1: The freeway section has an on-ramp and an off-ramp.
There is a recurrent bottleneck caused by conflicts between inflow in
on-ramp and outflow in mainlane. There are several traffic detectors
and controllers in this freeway section. The controllers include VSL
signs, traffic lights for RM and road side units for LCC.
• Differential VSL (DVSL) agent: The DVSL agent
aims at regulating the outflow of controlled area to pre-
vent the capacity drop at bottlenecks. The conflicts be-
tween vehicles occur mostly in the right lanes. There-
fore different speed limits among lanes might be more
effective. The DVSL strategy can be implemented un-
der CAV environment. The DVSL signs can send speed
limit orders to the vehicles in the corresponding lane,
the vehicles are forced to drive under the received speed
limit.
• LCC agent: The LCC is used to regulate the lateral
flows for each lane. The implementation of LCC agent
is more challenging than RM and DVSL agents. In this
paper, we only considered to use a road-side unit (RSU)
to send "keep lane" orders to the vehicle in left 2 lanes
of the merge area. The reason is that the lateral inflow
from left lanes to right lanes will cause severe conges-
tion when traffic breakdown occured in the merge area
of the right lanes.
Each control agent executes its own action according to its
own control cycle. We denote by T R the control cycle for
RM agent, T D for DVSL agent and T L for LCC agent. The
main goal of these agents is to reduce congestion and promote
the freeway capacity in a coordinated manner.
3 The KS-GCN Model Description
Figure 2 presents the architecture of KS-GCN, which is com-
prised of several GCN layers, traffic state inputs for DVSL,
RM, LCC, several knowledge sharing layers, DVSL, RM and
LCC actuators respectively.
3.1 Framework
The function of KS-GCN is to generate coordinated actions
for the DVSL agent, RM agent and LCC agent given observed
traffic state from correlated sensors/detectors on the targeted
freeway section. Each agent only receives states from its
mostly related sensors. Each sensor collects P traffic vari-
ables (e.g., velocity, occupancy rate) in one cycle and is de-
noted as a vector xT−T c ∈ RP . The sensor network can be
represented as a weighted undirected graph G = (V,E, W),
where V is a set of nodes V = N, E is a set of edges,
j
self-connections. IN is the identity matrix. ¯Dii =(cid:80)
where ¯W = W + IN is the adjacency matrix that added
¯Wij.
U(l) ∈ Rf (l)×f (l+1), b(l) ∈ Rf (l+1) are the layer-specific
trainable weight matrix and bias. H(l) ∈ RN×f (l), N is the
number of graph signal, f (l) is the number of feature in l-th
layer, and αl() is the activation in l-th layer.
In KS-GCN,
there are 3 stacked GCNs, which are used to learn features
from traffic states for RM, DVSL and LCC agents respec-
tively.
On top of the GCN, we further use a knowledge sharing
layer to learn the sharing features for each agent. After L
layers of GCN, the last output matrix HL is of size N × f (L).
We use a simple FC layer for knowledge sharing, the output
matrix is reshaped as a vector hL ∈ RN f (L). The sharing
feature s can be obtained by:
s = αks(UkshL + bks),
(3)
Uks ∈ RK×N f (L) and bks ∈ RK are trainable weights for
the knowledge sharing layer. K is the dimension of the shar-
ing knowledge. Each agent shares its own knowledge with
the other agents for generating specific action. The sharing
process is done by concatenation:
zRM = concat(hL,RM , sDV SL, sLCC),
zDV SL = concat(hL,DV SL, sRM , sLCC),
zLCC = concat(hL,LCC, sRM , sDV SL).
(4)
Here, z is the final vectorized feature for generating control
action, concat is the concatenation layer.
3.3 Action Design
In this subsection, we introduce the action representation of
different agents. The action for RM is represented by the
phase of traffic light in the on-ramp. It is defined as aRM = 1:
change the light to green phase (the vehicles in on-ramp is
allowed to enter the freeway), and aRM = 0: change the
light to red phase. The action for RM agent can be generated
by a FC layer with softmax activation:
aRM = argmax(sof tmax(URM zRM + bRM ))
(5)
where URM ∈ R2×f RM , and bRM ∈ R2 are the trainable
weights. argmax is used to find the index with maximum
value.
A similar action design can be applied to LCC agent. The
action of LCC agent is defined as aLCC = 1: allow lane
change in left 2 lanes, and aLCC = 0: forbidden lane change
in left 2 lanes. The generation process of aLCC is:
aLCC = argmax(sof tmax(ULCCzLCC + bLCC)).
(6)
The action aDV SL interacts the speed limit of all lanes
in the controlled area. Therefore aDV SL ∈ Rc, where c
is the number of lane at the controlled section. Consider-
ing the real world implementation and the driver compli-
ance issue, the elements of aDV SL is set as discrete values
∈ [0, 1,··· , M ]. And the speed limits V ∈ Rc is
aDV SL
i
equal to V0 + jaDV SL, where V0 is the minimum value of
Figure 2: The architecture of KS-GCN
W ∈ RN×N is a weighted adjacency matrix. The KS-GCN
learns functions that map graph signals to traffic control sig-
nals asynchronously:
[XR
T−T R , XD
T−T D , XL
[XR
T−T R , XD
T−T D , XL
[XR
T−T R , XD
T−T D , XL
T−T L , WR, WD, WL] → aRM,
if T = iT R
T−T L , WR, WD, WL] → aDVSL,
if T = iT D
T−T L , WR, WD, WL] → aLCC,
if T = iT L
(1)
where XR ∈ RN R×P , XD ∈ RN D×P and XL ∈ RN L×P
are graph sensor signals that related to RM, DVSL and LCC
agents respectively.The 3 agents can share sensors, therefore
N R + N D + N L ≥ N. WR, WD and WL are RM, DVSL
and LCC similarity matrices derived from W. i is an integer.
KS-GCN asynchronously updates the control signals every
control cycle. The control cycles of RM (T R), DVSL (T D)
and LCC (T L) can be different from each other.
3.2 Network Structure
We use the GCN architecture proposed in [Kipf and Welling,
2016] to learn the spatial dependence between traffic signals
on the graph. The layer-wise propagation rule of the specific
GCN is:
H(l+1) = αl( ¯D− 1
2 ¯W ¯D− 1
2 H(l)U(l) +
(2)
)
b(l)
.
.
.
b(l)
the speed limit, j is the integer multiples, the maximum value
of speed limits is V0 + jM. It is not feasible for a neural
networks to generate explicit discrete speed limits for mul-
tiple lanes because the total number of actions for a c-lane
freeway section will be as large as M c. The neural networks
with limited size is difficult or impossible to handle such a
large action space. Follow the work in [Wu et al., 2018c], the
action generation process for the DVSL agent is defined as:
aDV SL = int((M +1)sigmoid(UDV SLzDV SL +bDV SL)),
(7)
The activation of FC layer for DVSL agent is sigmoid func-
tion. The outputs of the FC layer are then multiplied with
M + 1. The discrete action aDV SL is obtained by the integer
parts of the scaled outputs.
4 Evolutionary Strategy for Optimization
In this section our aim is to propose an efficient and effective
optimization algorithm for coopetative traffic control using
KS-GCN based on evolutionary strategy (ES). Finding an op-
timal coopetative control policy for a given freeway section
in section 2 can be seen as an optimization problem to search
for a trainable parameter set θ for KS-GCN that maximize the
0 rt of the freeway section. rt is the
total outflow F (θ) =(cid:80)T
instantaneous outflow of the freeway section.
The parameters θ of KS-GCN can be directed updated by
using the final return Fj(θ + σj) of parallel workers in ES,
therefore we proposed to use ES as the optimization algo-
rithm for KS-GCN. Another objective of the freeway control
agents is to achieve an optimal control scheme under stochas-
tic traffic demand. This also can be easily done via ES. In
simulation, the traffic demand is modeled as a random pro-
cess. In each episode, a new traffic demand is set by sampling
data from the random process, then several parallel workers
are used to run on simulations with the same traffic demand,
finally the parameters θ is updated by the final returns of
these parallel workers. We find that this stochastic training
approach guarantees the generalization of the agents.
Another core challenge is how to balance exploration and
exploitation using ES. The total outflow as the reward func-
tion is sometimes deceptive, e.g, the agents that achieved high
outflow for a specific traffic demand might perform badly un-
der another traffic demand sampled from the same random
process. Without adequate exploration, the agents might fail
to discover effective traffic control strategies. In this paper,
we exploit the novelty-seeking (NS) proposed in [Conti et
al., 2018] for exploration.
In NS, the novelty of one pol-
icy is characterized by a behavior vector b(πθ) that describes
its behavior. For CTC, we define traffic demand D specific
b(πθ, D) as:
b(πθ, D) = avg(aRM
D ), avg(
aDV SL
D
M + 1
D )
), avg(aLCC
(8)
D
and aLCC
D , aDV SL
where aRM
are vectors that contain all
time RM, DVSL and LCC actions under demand D. The
original work of NS use a set of parameters to calculate the
novelty. Because the traffic demand changes every episode,
D
Learning rate α,
noise
random demand procees PD,
standard deviation
balance parameter
Algorithm 1 ES-CTC
Input:
σ,
w
1: For t = 1,2,... do
2:
3:
4:
5:
6:
7:
8:
9:
10: End For
End For
Set θt+1 = θt + α 1
nσ
Decrease w
Sample traffic demand Dt from PD
Compute bahavior vector bt(θt, Dt)
For each work j = 1,2,...,n do
Sample j ∼ N (0, I)
Compute returns Fj and novelties Nj using Eq (10)
(cid:80)n
j=1(1 − w)Fj + wNj
calculating demand specific behavior vectors for a set of pa-
rameters will be very time-consuming. In this paper, the nov-
elty of a parallel worker is directly defined as the distance
between its behavior vector and the one of unperturbed agent
on demand D:
Nj(b(πθ+σj , D), b(πθ, D)) = (cid:107)b(πθ+σj , D) − b(πθ, D)(cid:107)2,
(9)
The parameter update rule for ES-CTC is then expressed as
follows:
θt+1 = θt + α
1
nσ
(1 − w)Fj + wNj
(10)
n(cid:88)
j=1
where n is the number of parallel workers, α is the learning
rate. 0 ≤ w ≤ 1 is the parameter to balance between ex-
ploration and exploitation. In this work, we slowly decrease
w every episode. Algorithm 1 summarizes the optimization
procedure of ES-CTC
5 Experiments
In this section, we mainly conducted experiments on a sim-
ulated freeway section built by SUMO to evaluate the effec-
tiveness of ES-CTC.
5.1 The simulated freeway section
The open source software SUMO is selected for the exper-
iments. The software supports set the speed limits for each
lane, set traffic phase for traffic light and forbidden lane
changing using its API -- the Traffic Control Interface (TraCI)
package. A 874.51m freeway section with on- and off- ramps
of I405 north bound in California, USA is selected. The
original speed limits for the mainlane of this section are
65mile/h, for the on- and off- ramps are 50mile/h. The
freeway section in SUMO and each agents' control area can
be found in Figure 3. The travel demand of this freeway
can be categorized into 3 routes: 1) From mainlane to main-
lane (M2M), 2) From mainlane to off-ramp (M2Off), and
3) From on-ramp to mainline (On2M). Based on observa-
tion from recorded traffic flow from sensors of PeMS1, the
hourly demand of these 3 routes is modeled as Poisson distri-
bution with average value 5427, 1809 and 1153 respectively.
1http://pems.dot.ca.gov
The depart lane of the vehicles are randomly set according to
uniform distribution. Passenger car with a length 3.5m and
truck/bus with a length 8m are selected as vehicle types in
the simulated traffic stream. The type of vehicles are selected
randomly according to probability [0.85, 0.15]. Each round
simulation lasts for 1 hour.
Figure 3: The freeway section in SUMO
We place sensors in the upstream of DVSL controlled area,
DVSL controlled area, on-ramp and merge area to detect the
traffic state. The sensors on off-ramp and downstream area
are used to calculate the outflow of the freeway section. The
outflow can be used to compute the final return for the agents.
The traffic speed and occupancy rate collected from these
sensors are used as inputs for the KS-GCN. Specifically, the
on-ramp and upstream of merge area are used for RM agent.
The sensors in the upstream of DVSL controlled, DVSL con-
trolled area and upstream of merge area are used for DVSL
agent. The sensors in the merge area are used for LCC agent.
The sizes of XR, XD and XL are 8× 2, 22× 2 and 12× 2 re-
spectively. The element wij of similarity matrix W for input
states is given by:
exp
0.9
1
if
if
−loc(i)−loc(j)
10
if
(i, j) ∈ S
i = j
wij =
(i, j) ∈ D
(11)
where loc() denotes the location of the sensor. (i, j) ∈ D
means that sensor i and sensor j belong to different freeway
sections. (i, j) ∈ S denotes that sensor i and sensor j are
in the same freeway section. The control cycle T R, T D and
T L of RM, DVSL and LCC agents are set to 3, 60 and 30
seconds respectively. The speed limits set for DVSL agent is
[10mph, 15mph,··· , 75mph].
5.2 Benchmarks
We compare ES-CTC with the following baseline methods,
which include numerous DRL based traffic control models:
• No control: The baseline without any DVSL, RM and
LCC control.
• DQN-RM A modified version of DQN based traffic
light control for RM. The state input of the neural net-
works is the vectorization of XR. The agent is modeled
as a neural networks with two hidden FC layers.
• TRPO-RM The actor and critic of the agent are mod-
eled as neural networks with two hidden FC layers.
• DDPG-DVSL A DRL based DVSL control model
whose actor and critic of the agent are modeled as a neu-
ral networks with two hidden FC layers.
The traffic state XR is used as the state variable for DQN-
RM and TRPO-RM. The traffic state XD is used as the state
variable for DDPG-DVSL. The neural networks of DQN-
RM, actor and critic of DDPG-DVSL and TRPO-RM have
2 hidden FC layers, which contain 30 hidden neurons and 20
hidden neuron respectively. The agents of ES-CTC are built
upon 2 layer GCNs, the numbers of feature in 1st and 2nd
are 5 and 3 respectively, the dimensions of sharing feature
are set as 8. The reward signal of DQN-RM, TRPO-RM and
DDPG-DVSL is the outflow rt of the freeway section at time
point t. Their discount factors are set to 0.9. The return Fj
for ES-CTC is the total outflow of the freeway section.
5.3 Performance Comparisons
Scenario 1
We first evaluate all models on a simple case, they are con-
stantly optimized on a same demand profile. The DRL based
DQN-RM, TRPO-RM and DDPG-DVSL are trained with the
demand with 2000 episodes. The number of parallel work-
ers n for ES-CTC is set to 50. To make the comparison fair,
we update the parameters of ES-CTC 40 times therefore all
models are learned with same number of simulation. In this
scenario, we can observe whether the compared models can
converge to a stable and optimal point by the training process
of all models. The evolution of the overall outflow of each
algorithm during training can be seen in Figure 4.
(a) Evolution of total outflow of DQN-RM, TRPO-RM
and DDPG-DVSL
(b) Evolution of total outflow of the ES-CTC model
Figure 4: Evolution of total outflow of the models over iterations of
the algorithms.
We discover that the DQN-RM, TRPO-RM and DDPG-
DVSL fail to converge to a stable value. Several oscillations
can be observed from Figure 5(a). The outflow are related to
many other factors such as the inflow of on-ramp and outflow
of off-ramp, which could not fully controlled by the agents.
Moreover, the vehicle can be computed as a out vehicle only
when it has leaved the freeway section, there could be a delay
between the control effects of the agents on the vehicle and
computation of reward signal. These issues make the DRL
based approaches difficult to converge.
It is observed that
ES-CTC is more stable from Figure 5(c). ES-CTC reaches a
relatively high outflow after 25 round generation and achieves
the highest max outflow with 6609 vehicles. Another advan-
tage of ES-CTC models is that they are significantly faster
Models
ES-CTC
DQN-RM
TRPO-RM
DDPG-DVSL
Outflow
6725.1
6567.2
6563.9
6642.3
T DS
0.7949
0.7819
0.7839
0.7904
IL
0.0656
0.0471
0.0437
0.0502
Table 1: The average evaluation metrics on 100 stochastic traffic
demands
than DRL models due to their higher parallelization capabil-
ity. The results indicate that deep neuroevolution model is
more suitable for cooperative traffic control compared with
DRL models. The total number of outflow only reaches 6289
when no control strategy is implemented. The maximum out-
flows of all DRL models and ES-CTC are significantly higher
than 6289. The maximum outflows for DQN-RM, TRPO-RM
and DDPG-DVSL are 6577, 6570 and 6588 respectively. It
shows that the traffic control strategies can promote the ca-
pacity of the freeway.
Scenario 2
In the second case, the DQN-RM, TRPO-RM, DDPG-DVSL
and ES-CTC are trained and evaluated on stochastic traffic
demand. The DRL based DQN-RM, TRPO-RM and DDPG-
DVSL are trained with the demand with 3000 episodes. They
are trained with a new traffic demand in each episode. The
number of parallel workers n for ES-CTC is set to 100. In or-
der to guarantee all models consume similar wall-clock time,
we evolved the ES-CTC model with 200 generations. After
training, we compare the average outflow of all models on
100 stochastic demands. The traditional performance metric
used in the RL problems is the average total return achieved
by the model in an episode. In order to obtain more represen-
tative metrics independent of reward shaping for traffic con-
trol, we also compute the average traffic demand satisfaction
degree T DS and average improvement level IL, which are
defined as
Fi
T DS =
Di
Fi − F N
i
IL =
.
F N
i
(12)
Here Di is the total demand of the ith episode, F N
is the to-
i
tal outflow of ith episode without any traffic control agents.
The evaluation results of 4 models are given in Table 1. We
can find ES-CTC achieves relatively higher average outflow,
T DS and IL than three DRL benchmarks on 100 stochastic
traffic demands. The ES based optimization strategy, graph
convolutional structure and coordination between different
agents are the keys to its success.
The RM, DVSL and LCC actions of ES-CTC obtained
from one simulation are presented in Figure 5. The most in-
teresting one is the speed limits produced by DVSL agent.
The DVSL agent has learned to always set a maximum speed
limit for the leftest lane.
it automatically set the left lanes
as overtaking lanes. The agents mainly adjusts inflow to the
bottleneck by adjusting the speed limits of the right lanes,
on-ramp vehicles and vehicles' lane change behaviors. As
stated before, the conflicts between vehicles occur mostly in
the right lanes. Therefore it is not necessary to decrease the
speed limits of left 2 lanes (lane 4 and lane 5).
(a) The RM action produced
by ES-CTC in the first 6
minute
(b) The variable speed limits
produced by ES-CTC
(c) The LCC action pro-
duced by ES-CTC
Figure 5: Visualization of RM, DVSL and LCC actions produced by
ES-CTC.
6 Conclusion
In this paper we have proposed a deep neuroevolutional
model for cooperative freeway traffic control.
In order to
learn the spatial dependence between traffic sensors, the neu-
ral networks structure of the model are built upon graph con-
volutional layer. Our structure allows several traffic control
agents with different control cycles work cooperatively to im-
prove the freeway traffic efficiency. Our solution outperforms
the state-of-the-art DRL based solutions in terms of improve-
ments in freeway capacity.
Several interesting questions stem from our paper both the-
oretically and practically, that we plan to study in the future.
We aim to extend the approach to large freeway networks and
a broader set of dynamic events such as adverse weather and
traffic incidents in the future. Another interesting direction
we plan to study is the incorporation of more advanced traffic
control strategies. In this paper, the most basic graph con-
volutional network architecture and evolutionary strategy are
used. We believe that a more systemic research of architec-
tures and optimization strategies may provide improvements
in control performance.
Acknowledgments
The work was supported by national natural science foun-
dation of China (61620106002). Any opinions expressed in
this paper are solely those of the authors and do not repre-
sent those of the sponsors. The authors would like to thank
experienced anonymous reviewers for their constructive and
[Lowe et al., 2017] Ryan Lowe, Yi Wu, Aviv Tamar, Jean
Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-
agent actor-critic for mixed cooperative-competitive envi-
ronments. In Advances in Neural Information Processing
Systems, pages 6379 -- 6390, 2017.
[Lv et al., 2018] Zhongjian Lv,
Jiajie Xu, Kai Zheng,
Hongzhi Yin, Pengpeng Zhao, and Xiaofang Zhou. Lc-
rnn: A deep learning model for traffic speed prediction. In
IJCAI, pages 3470 -- 3476, 2018.
[Roncoli et al., 2015] Claudio Roncoli, Markos Papageor-
giou, and Ioannis Papamichail. Traffic flow optimisa-
tion in presence of vehicle automation and communica-
tion systems -- part ii: Optimal control for multi-lane mo-
torways. Transportation Research Part C: Emerging Tech-
nologies, 57:260 -- 275, 2015.
[Salimans et al., 2017] Tim Salimans,
Jonathan Ho,
Xi Chen, Szymon Sidor, and Ilya Sutskever. Evolution
strategies as a scalable alternative to reinforcement
learning. arXiv preprint arXiv:1703.03864, 2017.
[Van der Pol and Oliehoek, 2016] Elise Van der Pol and
Frans A Oliehoek. Coordinated deep reinforcement learn-
ers for traffic light control. Proceedings of Learning, Infer-
ence and Control of Multi-Agent Systems (at NIPS 2016),
2016.
[Wei et al., 2018] Hua Wei, Guanjie Zheng, Huaxiu Yao, and
Zhenhui Li.
Intellilight: A reinforcement learning ap-
proach for intelligent traffic light control. In Proceedings
of the 24th ACM SIGKDD International Conference on
Knowledge Discovery & Data Mining, pages 2496 -- 2505.
ACM, 2018.
[Wu et al., 2018a] Jingda Wu, Hongwen He, Jiankun Peng,
Yuecheng Li, and Zhanjiang Li. Continuous reinforcement
learning of energy management with deep q network for a
power split hybrid electric bus. Applied Energy, 222:799 --
811, 2018.
[Wu et al., 2018b] Yuankai Wu, Huachun Tan, Lingqiao
Qin, Bin Ran, and Zhuxi Jiang. A hybrid deep learning
based traffic flow prediction method and its understanding.
Transportation Research Part C: Emerging Technologies,
90:166 -- 180, 2018.
[Wu et al., 2018c] Yuankai Wu, Huachun Tan, and Bin Ran.
Differential variable speed limits control for freeway re-
current bottlenecks via deep reinforcement learning. arXiv
preprint arXiv:1810.10952, 2018.
valuable suggestions for improving the overall quality of this
paper.
References
[Belletti et al., 2018] Francois Belletti, Daniel Haziza,
Gabriel Gomes, and Alexandre M Bayen. Expert level
control of
ramp metering based on multi-task deep
reinforcement learning. IEEE Transactions on Intelligent
Transportation Systems, 19(4):1198 -- 1207, 2018.
[Carlson et al., 2010] Rodrigo C Carlson,
Ioannis Pa-
pamichail, Markos Papageorgiou, and Albert Messmer.
Optimal motorway traffic flow control involving variable
speed limits and ramp metering. Transportation Science,
44(2):238 -- 253, 2010.
[Conti et al., 2018] Edoardo Conti, Vashisht Madhavan, Fe-
lipe Petroski Such, Joel Lehman, Kenneth Stanley, and Jeff
Clune.
Improving exploration in evolution strategies for
deep reinforcement learning via a population of novelty-
seeking agents. In Advances in Neural Information Pro-
cessing Systems, pages 5032 -- 5043, 2018.
[Foerster et al., 2016] Jakob Foerster,
Ioannis Alexandros
Assael, Nando de Freitas, and Shimon Whiteson. Learn-
ing to communicate with deep multi-agent reinforcement
learning. In Advances in Neural Information Processing
Systems, pages 2137 -- 2145, 2016.
[Foerster et al., 2017] Jakob Foerster, Gregory Farquhar,
Triantafyllos Afouras, Nantas Nardelli, and Shimon
Whiteson. Counterfactual multi-agent policy gradients.
arXiv preprint arXiv:1705.08926, 2017.
[Hegyi et al., 2005] Andreas Hegyi, Bart De Schutter, and
Hans Hellendoorn. Model predictive control for optimal
coordination of ramp metering and variable speed limits.
Transportation Research Part C: Emerging Technologies,
13(3):185 -- 209, 2005.
[Hellinga and Mandelzys, 2011] Bruce
and
Michael Mandelzys. Impact of driver compliance on the
safety and operational impacts of freeway variable speed
Journal of Transportation Engineering,
limit systems.
137(4):260 -- 268, 2011.
Hellinga
[Kipf and Welling, 2016] Thomas N Kipf and Max Welling.
Semi-supervised classification with graph convolutional
networks. arXiv preprint arXiv:1609.02907, 2016.
[Li et al., 2016] Li Li, Yisheng Lv, and Fei-Yue Wang.
Traffic signal timing via deep reinforcement learning.
IEEE/CAA Journal of Automatica Sinica, 3(3):247 -- 254,
2016.
[Li et al., 2018] Yaguang Li, Rose Yu, Cyrus Shahabi, and
Yan Liu. Diffusion convolutional recurrent neural net-
work: Data-driven traffic forecasting. 2018.
[Lin et al., 2018] Kaixiang Lin, Renyu Zhao, Zhe Xu, and
Jiayu Zhou. Efficient large-scale fleet management via
multi-agent deep reinforcement learning. arXiv preprint
arXiv:1802.06444, 2018.
|
1703.01931 | 1 | 1703 | 2017-03-06T15:44:10 | Context-Based Concurrent Experience Sharing in Multiagent Systems | [
"cs.MA"
] | One of the key challenges for multi-agent learning is scalability. In this paper, we introduce a technique for speeding up multi-agent learning by exploiting concurrent and incremental experience sharing. This solution adaptively identifies opportunities to transfer experiences between agents and allows for the rapid acquisition of appropriate policies in large-scale, stochastic, homogeneous multi-agent systems. We introduce an online, distributed, supervisor-directed transfer technique for constructing high-level characterizations of an agent's dynamic learning environment---called contexts---which are used to identify groups of agents operating under approximately similar dynamics within a short temporal window. A set of supervisory agents computes contextual information for groups of subordinate agents, thereby identifying candidates for experience sharing. Our method uses a tiered architecture to propagate, with low communication overhead, state, action, and reward data amongst the members of each dynamically-identified information-sharing group. We applied this method to a large-scale distributed task allocation problem with hundreds of information-sharing agents operating in an unknown, non-stationary environment. We demonstrate that our approach results in significant performance gains, that it is robust to noise-corrupted or suboptimal context features, and that communication costs scale linearly with the supervisor-to-subordinate ratio. | cs.MA | cs |
Context-Based Concurrent Experience Sharing in
Multiagent Systems
Dan Garant
University of Massachusetts
[email protected]
Amherst, MA 01002
Victor Lesser
University of Massachusetts
[email protected]
Amherst, MA 01002
Bruno C. da Silva
Federal University of Rio
Grande do Sul. Porto Alegre,
Brazil 91540-000
[email protected]
Chongjie Zhang
Tsinghua University
Beijing, China
[email protected]
ABSTRACT
One of the key challenges for multi-agent learning is scalability.
In this paper, we introduce a technique for speeding up multi-
agent learning by exploiting concurrent and incremental experi-
ence sharing. This solution adaptively identifies opportunities to
transfer experiences between agents and allows for the rapid ac-
quisition of appropriate policies in large-scale, stochastic, homo-
geneous multi-agent systems. We introduce an online, distributed,
supervisor-directed transfer technique for constructing high-level
characterizations of an agent's dynamic learning environment --
called contexts -- which are used to identify groups of agents oper-
ating under approximately similar dynamics within a short tempo-
ral window. A set of supervisory agents computes contextual infor-
mation for groups of subordinate agents, thereby identifying candi-
dates for experience sharing. Our method uses a tiered architecture
to propagate, with low communication overhead, state, action, and
reward data amongst the members of each dynamically-identified
information-sharing group. We applied this method to a large-scale
distributed task allocation problem with hundreds of information-
sharing agents operating in an unknown, non-stationary environ-
ment. We demonstrate that our approach1 results in significant per-
formance gains, that it is robust to noise-corrupted or suboptimal
context features, and that communication costs scale linearly with
the supervisor-to-subordinate ratio.
1.
INTRODUCTION
In large-scale multi-agent systems consisting of hundreds to
thousands of reinforcement-learning agents, convergence to a near-
optimal joint policy, when possible, may require a large number
of samples. These systems, however, may contain groups of agents
working on nearly identical local tasks or under approximately sim-
ilar environmental dynamics. Identifying such groups may prove
useful in cooperative domains, due to the opportunity of exploiting
shared information. Information sharing has been extensively stud-
ied in single-agent settings with the goal of transferring knowledge
from a source task to novel tasks [26, 12, 2]. Applying this idea
to the multi-agent setting (MAS), it is apparent that experiences
may be transferred not only across similar tasks, but also between
concurrently-learning agents in a shared environment. This paper
focuses on the problem of online transfer of experiences between
1A more complete presentation of our approach, as well as addi-
tional experiments, can be found in [5].
such agents -- with an emphasis on the adaptive discovery of groups
of agents where experience sharing is possible and beneficial.
In multi-agent settings, agents need to interact and learn con-
currently. The environment, from each agent's perspective, is
non-stationary due to the presence of other concurrently-adapting
agents. Since the observations made by one agent are conditioned
on the behaviors of its neighbors, it is not clear when they can be
usefully exchanged and reused by other agents -- which may be op-
erating under different local environments and may be interacting
with different types of neighbors. Experience exchange is, there-
fore, not straightforward in non-stationary MAS. As an example,
consider the task allocation problem depicted in Figure 1. Agents
Figure 1: Sample task allocation network.
are represented as nodes and may receive tasks from the environ-
ment or from other agents. Each agent can choose to fulfill a given
task or to forward it to a neighboring agent. Assume that agents
have partial knowledge of the system: they do not know the global
structure of the network nor have access to state or policy informa-
tion of other agents. This results in a non-stationary problem where
it may be inappropriate to transfer information between some pairs
of agents. Agent D, for instance, receives a large number of tasks
from the environment and may need to forward them to a neigh-
bor; agent C receives tasks from a neighbor and may need to direct
them away from its heavily-loaded neighbor. Agents A and B, on
CBDA302792773EEEE1923111810 the other hand, undergo similar task-forwarding patterns with re-
spect to their neighbors. Experience transfer, then, may be appro-
priate between agents A and B (said to be contextually compatible
agents), but not between agents D and C.
To address the information-sharing problem in non-stationary
MAS we propose modeling contexts as inherently dynamic local
characterizations of the environment under which agents operate.
They are defined over short timescales during which policies and
models are approximately static.
In Sections 5 and 6 we intro-
duce and motivate a context-similarity measure grounded in the
comparison of abstract representations of environment dynamics,
rather than policies or Q-values, and advocate the use of supervi-
sory agents (which periodically collect data from subordinates) as
a way of identifying contextually-compatible agent groups where
experiences may be shared. As will be further discussed in the
following sections, contextual modeling is made possible through
the commonly-studied property of interaction sparsity [29], or what
Simon [24] referred to as nearly-decomposable systems. This is ap-
parent in many domains, such as distributed task allocation, disaster
planning [9, 19] and sensing networks [31], in which agents inter-
act strongly with only a small group of closely-related partners.
Although other transfer mechanisms are possible, to our knowl-
edge no other methods exist that address the particular setting and
scale presented in this paper. We believe this is the first algorithm
that allows experience sharing in a concurrent and interacting MAS
with ∼1000 agents while undergoing low communication and com-
putational overhead. We evaluate our method on a large-scale dis-
tributed problem and demonstrate that context-based transfer yields
significant performance gains. We further show 1) that the time
complexity of our method scales with the number of agents within
each supervisory group, not with the total number of agents in the
network; 2) that our method is robust to noise-corrupted or sub-
optimal context features; 3) that communication costs scale lin-
early with the supervisor-to-subordinate ratio; and 4) that sparse
lossy compression schemes may be deployed and provide signifi-
cant improvements in communication costs, while inducing negli-
gible negative impact on system-wide performance.
2. RELATED WORK
In this section we discuss related work that also aims at expe-
rience sharing in MAS. Kretchmar et al.
introduced a technique
for agents to periodically exchange Q-values to accelerate learning
[11]. They assumed, however, that agents operate on independent
copies of an environment and do not interact. Boutsioukis et al. re-
laxed this assumption via a method for using Q-values of a source
task to bias the initial policy of target tasks [1]. They assumed
that learning on the source task had to be completed before trans-
fer was made possible, and required the use of inter-task mappings.
We do not assume that such mappings are needed and instead in-
fer when observations may be transfered by identifying groups of
agents that operate under similar local contexts. Taylor et al. in-
troduced a transfer method that allowed for source and target tasks
to be learned in parallel [25]. It implicitly assumed that all agents
experienced tasks with similar state values -- which may not be true
if they operate in contexts with different transition dynamics. More
recently, Mnih et al. introduced a technique for accelerating deep
learning algorithms via asynchronous sharing of policy gradients
[16]. This allowed for independent agents to cooperate in solving a
complex task, but required that agents did not interact while doing
so. In Section 5 we extend the discussion presented here and intro-
duce related techniques relevant to the problem of characterizing
local contexts in order to identify sharing opportunities.
3. SETTING
Multi-agent decision-making problems are often framed in the
context of Markov games [20]. Markov games model n agents
operating in an environment described by a joint state S. A state
transition function specifies the conditional probability of the en-
vironment transitioning to state S(cid:48), given that it was in S and
that agents executed a particular joint action (a1, . . . , an); i.e.,
P (S(cid:48)S, a1, . . . , an). In Markov games, each agent i holds a par-
ticular reward function Ri(riS, a1, . . . , an), which we consider
here to be a conditional distribution over rewards.
In cooperative environments, individual reward functions may
be identical -- each agent's individual performance perfectly aligns
with the system's performance. In this paper we consider the more
general case of decomposable reward functions, which arise in
structured settings such as Network Distributed-Partially Observ-
able Markov decision processes [18] or factored multi-agent MDPs
[6]. We also assume that the global state S is decomposable into
(potentially overlapping) components si, each of which represents
the portions of the state that agent i can directly observe. This
arises in systems where it may be infeasible for agents to learn
over the full joint space or when network structure or communi-
cation bandwidth introduce limitations on state observability.
In
general, the state observable by individual agents may be insuf-
ficient to faithfully reconstruct the overall state transition model,
iS, a1, a2, . . . , an). Motivated by the idea of interaction spar-
Pi(s(cid:48)
sity [29], we address this difficulty by observing that sparsity in Pi
may allow observations collected from a small number of neigh-
boring agents to be used in order to reasonably estimate Pi.
4. OVERVIEW OF THE METHOD
Before introducing the technical details of the method we pro-
pose, we start by presenting a high-level summary of the steps
involved in determining sharing opportunities by grouping agents
based on their local learning environments (or contexts):
1. each agent collects observations from its local environment
in the form of state, action, reward, and next state tuples.
Every K time steps (the reporting interval), agents report
such observations to their corresponding supervisors;
2. supervisors use the received information and their local
knowledge about the interactions between subordinate agents
to compute context summary vectors, one per agent. These
vectors correspond to dynamic local characterizations of the
environment under which agents operate, and are used to
identify possible sharing experiences;
3. supervisors measure the similarity between the context sum-
mary of each subordinate agent with respect to a covariance-
appropriate and scale-independent metric; similar agents are
organized into sharing groups;
4. supervisors relay experiences (state, action, reward, next
state tuples) between all members of each sharing group;
5. return to step (1) and adaptively regroup agents according to
updated context information.
Intuitively, a supervisor periodically collects information from a
small number of subordinate agents in its supervisory group and
computes context features. These are embedded in a summary
space in which similarity stochastically determines sharing oppor-
tunities -- not all agents within a same supervisory group need to
share experiences. Note that the method we propose here does not
aim at finding optimal subordinate-supervisor assignments, but on
efficiently identifying sharing opportunities within a given super-
visory structure. Sharing opportunities between agents are dynam-
ically re-evaluated by our method based on updated information
collected in a reporting interval. The overall context-creation and
data-transfer process is depicted in Figure 2.
Figure 2: Overview of the context-based transfer process.
5. CONTEXT FEATURES AS
ABSTRACTIONS
STATE
Context features are compact abstractions of the local learning
environment under which an agent operates. Carrol & Seppi [2]
discuss the difficulty of constructing such abstractions given the
difficulty of defining similarity metrics between learning environ-
ments. In the single-agent RL setting, many metrics are possible;
most compare local environments by comparing policies, Q-values,
or reward function differences [1, 7, 25]. In the multi-agent setting,
however, we wish to capture a measure of compatibility in the local
learning environment of agents.
Compatibility-based metrics are helpful to avoid issues arising
by using policy, Q-function or model similarity as proxies of en-
vironment similarity. Consider, for instance, the problem of expe-
rience imbalance: the policies or Q-values under comparison need
to be constructed with enough samples so that they are accurate
estimates of an optimal policy [22]. Metrics based on policy com-
parison are also difficult to define since optimal policies for Markov
games are not unique [14], thus making policy similarity a poor in-
dication of environment similarity. Metrics based on Q-function
comparison are also non-trivial to define since latent features often
cause agents to operate in different state spaces.
To avoid these problems, we propose reasoning over the under-
lying latent model of a stochastic game via contextual comparison.
In particular, if we can identify that agents are working under a
same local transition and reward model, we can infer (from the
homogeneity of the system) that they are facing the same learn-
ing problem. Experiences gathered by each compatible agent are
interchangeable and can be transferred. Since estimating system-
wide transition and reward models is impractical in large systems,
we rely on the use of context features to form broad-scope sum-
maries, or abstractions, of transition and reward models as experi-
enced by individual agents. Abstractions in RL have been stud-
ied extensively, and are not the focus of this paper; we simply
rely on any of the many existing methods (e.g. [4, 23, 15, 8, 13,
21, 10]) to construct features capable of abstracting the state of a
particular problem at hand.
If prior domain knowledge is avail-
able, context features may also be manually specified to abstract an
agent's state variables -- and possibly those of its observable neigh-
bors. The way in which sufficient statistics and features for such
abstractions are defined and computed depends on the structure of
the MAS and needs to be defined in light of the characteristics of
the application at hand.
In this paper we empirically show that
even simple (and possibly noise-corrupted) abstractions are often
sufficient to allow for experience sharing opportunities to be iden-
tified in large-scale non-stationary systems composed of hundreds
of concurrently-learning agents.
6. CONTEXT-BASED LEARNING
As mentioned above, context features are abstractions of the lo-
cal learning environment of an agent, and are constructed based
on data collected at a particular time t.
In order to characterize
the broader context of an agent's learning environment -- for in-
stance, the medium-term effects of its interactions with neighbor-
ing agents -- one needs to combine context observations across a
time window2. We refer to this aggregate context information as a
context summary vector. Context summaries are computed by hav-
ing neighborhoods of agents (called supervisory groups) send local
context information to their common supervisor. The supervisor
annotates these with information about the states and actions ex-
perienced by the agents it oversees. The contextual information re-
ceived by the supervisor is then used to compute a context summary
according to the method described in Section 6.2 and Algorithm 1.
When comparing context summaries to identify possible sharing
experiences, it is necessary to select an appropriate distance met-
ric in context space. This metric depends on the distribution from
which context features are drawn. Unless a designer has a priori in-
formation about this distribution, we assume that it can be approxi-
mated by a multivariate Gaussian. This is justified by two reasons:
1) by the central limit theorem (CLT), applied here in the limit of
K;3 and 2) because this is the distribution that imposes minimal
prior structural constraints (it is the maximum-entropy distribution
for these parameters) in the absence of prior knowledge. From this
assumption, it follows that context features can be summarized into
a context summary via a mean context vector and its covariance
matrix, and that a natural scale-invariant distance metric to com-
pare context summaries exists: the Mahalanobis distance [17].
Mahalanobis metrics generalize Euclidean distances in a way
that naturally takes correlations of the dataset (i.e., correlations be-
tween context features of different agents) into account. Further-
more, these distances are preserved under full-rank linear transfor-
mations of the space spanned by the data. This implies that con-
text distances are preserved even if the context features are further
abstracted or transformed under non-degenerate projections down
onto any other context space. Even though this metric is quite gen-
eral (it is naturally data-adaptive, scale-invariant, and preserved un-
der non-degenerate transformations) and imposes the least prior
structural constraints in the absence of expert knowledge, other
metrics may be used. In a designer chooses to do so, our method
changes trivially -- the Gram matrix used to stochastically identify
sharing opportunities (see Algorithm 1) is in that case computed
according to the alternative selected metric.
6.1 Agent Organization
Agents are dynamically organized by our method in sharing
groups whenever they are close (in context space) to other agents
within a supervisory group. This organization process involves a
trade-off between how many supervisors exist in the system and
the number of agents within each sharing group. As the num-
ber of supervisors grows, the sharing assignment problem becomes
increasingly distributed, reducing the computational requirements
2As will be discussed later, selecting a time window has impli-
cations on runtime, communication overhead, and reliability. See
Section 7 for more details.
3A variant of the CLT for weakly dependent processes can also be
applied assuming that sufficiently separated agents have approxi-
mately independent experiences [3].
Supervisora1a2a3Raw ObservationsContext FeaturesSubordinatesV1V2V3Context SummariesRaw Observatonsa1a2a3SubordinatesRaw Observations Context Features Context Summaries Raw Observations imposed on any individual supervisor. On the other hand, super-
visors overseeing larger groups of subordinates are capable of se-
lecting from a larger pool of experiences, increasing the likelihood
that similar agent groups can be identified. Each supervisor is gen-
erally responsible for a set of subordinates selected through self-
organization [32] or directly given the network structure. In this
work we do not focus on finding optimal subordinate-supervisor
assignments, but on efficiently identifying sharing opportunities
within a given supervisory structure. In the experiments presented
in Section 7 we evaluate our method in a network of hundreds
of agents cooperatively solving a large-scale distributed task allo-
cation problem; in this case, subordinate-supervisor assignments
are determined in a way that supervisors span physical regions of
the network consistent with agent interaction strength [32]. Such
an agent-organization criterion is justified in this task due to the
assumption of interaction sparsity [29], a general characterist of
nearly-decomposable systems. Many other real-world multi-agent
systems with similar local sparse network-like interactions exist
and could be organized similarly -- ranging from disaster planning
systems [9, 19] to sensing networks [31].
6.2 Assessing Context Similarity
We now present a method for computing contextual similarity
between agents and forming sharing groups. Suppose that an agent
communicates K observations to its supervisor every K time units.
Our solution easily extends to cases where agents do not make one
observation per time step. Let Oi be a time-indexed experience
vector of agent i:
(cid:48)
Oi = [(s, a, s
(cid:48)
, r)t1 , (s, a, s
(cid:48)
, r)t2 , . . . , (s, a, s
, r)tK ]
(cid:62)
.
1, . . . , Ot
A supervisor overseeing n agents computes contextual informa-
tion by mapping Ω = {O1, . . . , On} into an n-tuple of con-
text summary vectors V = (V1, . . . Vn). Assume we are given
a function f for computing context features for agent i at time
t, given the history of observations Ωt = {Ot
n}, where
i = {(s, a, s(cid:48), r)hh ≤ t}. That is, Ωt contains the observations
Ot
of all n agents in the neighborhood up to some time t. The con-
text features for agent i at some time t ∈ [t1, . . . , tK ] are given
by f (i, Ωt). Note that f may use information about neighboring
agents when constructing features that describe i's local learning
environment. This yields a total of nK context features vectors per
supervisor. Each context feature vector is a sample drawn at a par-
ticular time from the (latent, unknown) underlying context distribu-
tion of an agent. These samples can be combined by the supervisor
to compute an unbiased estimate of the true mean of the underlying
context distribution. Unbiased estimates of the true mean vector
of the context distribution are called context summary vectors, and
are compact descriptions of the learning environment of each agent
within a supervisory group. The supervisor stores the context sum-
mary vectors of its n subordinates in a tuple V :
(cid:33)
tK(cid:88)
t=t1
(cid:32)
tK(cid:88)
t=t1
1
K
V =
f (1, Ωt), . . . ,
1
K
f (n, Ωt)
Note that each element of V , as computed above, is an unbi-
ased estimate of the true mean context vector under the distribution
assumptions made in Section 6. If a different distance metric is se-
lected by a domain expert, the elements of V need to be defined
so that they correspond to unbiased estimators of the mean of the
corresponding underlying distribution posited by the designer.
Our method for identifying sharing opportunities is based on a
stochastic sampling process that probabilistically partitions agents
into sharing groups, based on their contextual similarity. In partic-
ular, membership of an agent to a sharing group is stochastically
determined based on the similarity of that agent's context sum-
mary and the context summaries of other agents in the same sharing
group. This stochastic process partitions subordinate agents within
a given supervisory group so that agents operating under similar
underlying local dynamics have a higher probability of undergoing
experience sharing.
The sampling process that we define consists in a two-stage se-
lection routine. First, agents are partitioned into potential sharing
groups C1, . . . , Ck, based on to their similarity. Here, similarity is
measured with respect to the selected context distance metric. Po-
tential sharing groups are sets of agents (within a same supervisory
group) that, given their contextual similarity, are deemed to be fea-
sible candidates to undergo experience sharing. Partitioning agents
into potential sharing groups is an unsupervised process that can
be implemented via any standard clustering algorithm. Its purpose
is to ensure that the computational costs of stochastically sampling
agents in order to construct sharing groups is approximately con-
stant, independently of the number of agents within a supervisory
group. In particular, it ensures that the Gram matrix used to define
the sampling distribution (see next paragraph and Algorithm 1) has
dimensions that scale linearly with k. We say that agent Ai ∈ Cj if
that agent's context summary Vi belongs to potential sharing group
Cj.
Next, the tuple V of context summaries collected by a supervisor
is used to define a probability distribution that stochastically deter-
mines whether particular pairs of subordinate agents should belong
to a same sharing group. First, pairwise distances are computed
over the context summaries of every pair of agents Ah and Aj in
a potential sharing group Ci. This distance is stored in the (h, j)-
th entry of a Gram matrix M. Each h-th row of M corresponds,
therefore, to distances in context space between an agent Ah and all
other agents in a potential sharing group Ci. We use M to define
a sampling distribution Pi that stochastically determines whether
pairs of subordinate agents in Ci should belong to a same sharing
group, given their contextual similarities.
In particular, Pi(h, j)
denotes the probability that any two agents h and j within Ci will
be assigned to a same sharing group, based on their distances in
context space. In this paper we define Pi as a Boltzmann distribu-
tion constructed based on the pairwise context summary distances
between agents in Ci:
(cid:1)
exp(cid:0)M i
(cid:16)
h,j
(cid:17)
a,b∈Ci
exp
M i
a,b
(cid:80)
Pi(h, j) ≡
Boltzmann distributions are widely used in machine learning when
one needs to define probability distributions that depend on the rel-
ative difference between numerical quantities associated with each
element in a given population. Here, they depend on the distance
between context summaries of any given pair of agents in Ci. Note
that Pi, as defined above, assigns a probability to every pair of
agents in Ci and reflects how likely it is that those agents will be
selected for membership in a same sharing group. Agents are se-
lected for membership in a sharing group by sampling from Ci
without replacement; this ensures that agents will belong to at most
a single sharing group -- see Algorithm 1 for details. This selection
process is repeated in order to construct a group-sharing function
Ψ, which maps agents to sharing groups. Once Ψ has been estab-
lished, supervisors relay all observations within the K-unit time
window from all agents in Ψ(i) to agent i. Agents incorporate
these experiences into their policies using any off-policy learning
algorithm. Note that an agent i within a potential sharing group
Ci is not necessarily associated to any sharing partners; if i is dis-
similar from all other (n − 1) subordinates, Ψ(i) = ∅ with high
probability.
groups (P(·) denotes powerset)
Input: Set of agents A = {1, 2, . . . , n}
Input: Tuple V = (V1, . . . , Vn) of context summaries
Output: Mapping Ψ : A → P(A) from agents to sharing
Let M be a selected context distance metric.
Partition V into k potential sharing groups C1, . . . , Ck w.r.t.
M
for i ← {1, 2, . . . , k} do
Mh,j ← M(Ah, Aj) (Gram matrix over Agents(Ci))
(cid:17) (sampling distribution)
(cid:16)
Pi(h, j) ←
exp(M i
h,j)
exp
for each agent a ∈ Ci do
for each agent b ∈ Agents(Ci) − {a} do
a,b∈Ci
With probability Pi(a, b)
Let Ψ(a) ← Ψ(a) ∪ {b}
Let Agents(Ci) ← Agents(Ci) − {b}
(cid:80)
M i
a,b
Algorithm 1: Selection of Sharing Partners
The process for selecting sharing partners, described in Algo-
rithm 1, is repeated (in parallel) by each supervisor once every K
steps. Let n be the number of agents in a supervisory group (which
can be defined to include only a small and bounded fraction of the
total number of agents in the system) and d be the dimensionality
of the context feature vector computed by f. Under mild assump-
tions4 it is possible to show that the complexity of Algorithm 1 is
O(dn3). In practical terms, the time-complexity is dominated by
the inversion of an (n× n) matrix, needed in order to compute dis-
tances according to the metric proposed in Section 6. If other dis-
tance metrics are used (e.g., Euclidean distances) the complexity of
the method becomes quadratic in n and linear in d. Importantly,
notice that because this process is executed separately and inde-
pendently by each supervisor, the overall complexity of the process
is independent of the number of supervisors in the system -- it de-
pends crucially only on the number of agents being overseen by
each supervisor. The communication complexity of the method is
linear in the number k of potential sharing groups, linear in the
number of agents in each potential sharing group (i.e., at most n)
and linear in the reporting interval K: O(kKn). Empirically, the
communication costs of Algorithm 1 seem to scale linearly with the
supervisor-to-subordinate ratio (see Section 7.2 for more details).
7. EXPERIMENTS
We evaluate our algorithm on large network-distributed task al-
location problems (Figure 1). An agent maintains two queues of
tasks: a processing queue, with tasks that it has committed to work
on; and a routing queue, with tasks that are not actively being
worked on and that can be forwarded to a neighbor or processed lo-
cally. Each task has a duration s. The reward function is defined as
the reciprocal of the average service time over a time window; ser-
vice time is the time incurred from task creation to completion. In
all experiments, task duration is an exponentially-distributed ran-
dom variable with mean 10. Tasks are generated by the environ-
ment according to patterns that are unknown to the agents, which
(along with the fact that agents cannot observe their neighbors'
states and policies) makes the problem non-stationary. When a task
4e.g., that the number of iterations executed by the clustering al-
gorithm is proportional to n and that the complexity of computing
f is proportional to the number of observations used to construct
such features.
is created, it is associated with some agent v and placed in its rout-
ing queue. Upon executing an action (to either process or forward
a task) agents receive a reward of 1
d , where d is the estimated ser-
vice time of the agent receiving the task. To estimate service time,
agents keep track of the time taken to complete past tasks. Agents
learn policies using an extension of Q-Learning to the multi-agent
case with stochastic policies, which is known to outperform related
methods in domains similar to ours [30].
In our experiments, context features for agent i are composed
of three quantities: i's load relative to the mean load of its neigh-
bors, and the rate at which each of its neighbors receives tasks from
the environment and from other agents. Since agents with differ-
ent neighborhood sizes have different actions spaces, their obser-
vations have different dimensionality; we therefore restrict context
comparisons to agents with the same number of neighbors. Expe-
rience sharing between agents with different action spaces is be-
yond the scope of this paper and would require learning inter-task
mappings (e.g. see [27]). Supervisory groups in these experiments
were defined according to the criterion discussed in Section 7.1.
When applying Algorithm 1, we used the K-means algorithm with
Mahalanobis distance. We automatically set the number k of clus-
ters based on the gap statistic [28]. The task-allocation networks
used in our experiments are lattices of up to 729 agents, where
each agent directly interacts with 4 neighbors. Different network
instances were obtained by varying two parameters regulating the
type of task distribution to be tackled by the agents in the system:
• Task Concentration/Pattern: This parameter regulates
whether tasks originate at the outer edges of the network or at
central nodes. Each pattern requires a qualitatively different
system behavior. A policy for the border concentration re-
quires boundary agents to forward tasks inward, and central
agents to accept tasks; a policy for the center concentration
requires the opposite arrangement;
• Task Frequency: Tasks are generated with frequency gov-
erned by a Poisson distribution. For agents that do not re-
ceive tasks from the environment, λ = 0. For all others,
a fixed λ > 0 is used. We consider a set of 11 λ values
selected uniformly along the range [0.25, 0.35]. We do not
consider λ < 0.25 since even random policies perform well
in this case, nor λ > 0.35, since this leads to queues that
grow indefinitely even under optimal policies.
7.1 Performance under Experience Sharing
We first examine the impact of the number of supervisors on sys-
tem performance. On one hand, a single-supervisor configuration
results in a nearly centralized system which benefits significantly
from sharing opportunities. This corresponds to the case where all
agents are placed in a same potential sharing group and may un-
dergo experience sharing. Alternatively, no supervisors could ex-
ist, in which case the system corresponds to a conventional MAS
with no information sharing. Intermediate sharing configurations
are possible, with different numbers of supervisors and correspond-
ing subordinate agents. Note that the single-supervisor configu-
ration is often infeasible in real environments, as it is burdened
with high communication costs (see Section 7.2). Four alternative
supervisory structures were considered in our experiments. First,
we evaluated two baseline configurations: one with no supervision,
corresponding to a conventional MAS with no information sharing;
and one with a single supervisor, corresponding to a system where
all agents may share information. Intermediate sharing configura-
tions, with 4 and 9 supervisors, were also investigated. The single-
supervisor configuration has a supervisor-subordinate ratio of 1:99,
the 4-supervisor configuration 1:24, and the 9-supervisor configu-
ration roughly 1:10. Subordinate agents were assigned to supervi-
sors in a way that minimizes the network distance between pairs of
subordinates. Results discussed in this section correspond to 440
runs of our algorithm; in particular, we executed five trials of each
combination of task concentration pattern, value of λ, and supervi-
sory structure were performed, for a total of 2 × 11 × 4 × 5 = 440
runs.
To appreciate the difference between the least challenging task
allocation setting (λ = 0.25) and the most challenging one (λ =
0.35), we analyze the average system-wide service time obtained
by the single-supervisor configuration throughout 10,000 steps,
with tasks concentrated on the border of a 100-agent lattice. Fig-
ure 3 shows the evolution of service time as time progresses. At
first, poor policies lead to a heavily saturated system, which de-
grades service times, with a peak of approximately 100 steps per
task occurring about 25% of the way into the simulation. As agents
learn appropriate policies, they more rapidly complete tasks, ulti-
mately converging to a service time of about 25 steps per task. This
level of performance is reached regardless of λ, though the amount
of time taken to reach it, and the performance of the system during
learning, are both of central importance.
Figure 3: Performance under different difficulty settings λ.
In all experiments that follow we define performance as the area
under the curve of service time as a function of time. When the
system converges quickly, this area is small. We treat the mini-
mum service time ever attained by any configuration as zero, so
that running the system at the optimal performance does not accu-
mulate area; i.e., performance of an optimally-performing system
is invariant with respect to simulation duration (see Figure 4 for
an example). Figure 5 shows that the single-supervisor configura-
information-sharing agents accumulating nearly half the area un-
der the curve compared to agents that do not share experiences. As
additional supervisors are introduced, this benefit diminishes, since
there are fewer experience sharing opportunities within each su-
pervisory group. Note, however, that even with a high supervisor-
subordinate ratio of 1:10 (which corresponds, in this experiment,
to having approximately as many supervisors as agents in each su-
pervisory group), experience sharing still allows us to reduce the
learning curve area by more than 25%.
Figure 5: Performance of different supervisory configurations
in a 100-agent network; smaller values correspond to faster
learning.
7.2 Scalability and Communication Over-
head
We intuitively expect that information sharing becomes more
beneficial as the size of the system grows: larger systems typically
have a more diverse pools of agents which may benefit from shar-
ing. To test this hypothesis, we constructed simulations sweeping
across a large number of settings for task concentration and fre-
quency, and varied the number of agents through {100, 324, 729}
(i.e., lattices of dimension 10, 18, and 27). Two supervisory config-
urations were considered: a 9-supervisor configuration and a base-
line (or no-sharing) arrangement. Our goal is to characterize how
a 9-supervisor setting fares compared to the baseline as the num-
ber of subordinates per supervisor increases. This was achieved by
varying the network size (see Figure 6). Performance in the 100-
agent network was roughly 30% higher than the baseline. As net-
work size increased to 729 agents, performance median improved
by 40% compare to the baseline.
Figure 4: Performance of the single-supervisor configuration
vs. no-sharing. Smaller areas under the curve indicate faster
convergence.
tion far outperforms the baseline approach with no transfer, with
Figure 6: Performance of the 9-supervisor configuration as a
function of network size.
These gains come at the cost of increased communication. Note,
however, that the communication overhead of Algorithm 1 scales
with the supervisor-to-subordinate ratio, not with the total num-
ber of agents (see Section 6.2 for a formal complexity analysis);
e.g., the 9-supervisor configuration undergoes 9 times less com-
munication than the single-supervisor configuration. In our experi-
ments we further observed that communication volume was invari-
ant with respect to K: on average 43 bytes per step per subordinate
using a loss-less compression scheme. This suggests 1) that com-
munication costs (i.e., the total amount of bytes exchanged between
agents sharing experiences) scales linearly with the supervisor-to-
subordinate ratio; and 2) that even when accounting for communi-
cation costs, more distributed configurations tend to perform better.
In fact, all evaluated information-sharing configurations surpassed
the baselines while incurring very low communication overhead --
as previously mentioned, on average 43 bytes per step per subor-
dinate using a loss-less compression scheme in a 100-agent net-
work. The fact that communication volume was empirically ob-
served to be invariant with respect to K is not a trivial statement --
in the worst case, communication volume could still increase lin-
early with the number of potential sharing groups within a supervi-
sory group, k, and linearly with the reporting window, K. The fact
that it does not suggests that Algorithm 1 is capable of effectively
identifying and exploiting useful sharing opportunities, instead of
always relaying all K observations to all n agents within a super-
visory group.
We also explored the use of lossy experience compression
schemes, which significantly reduced communication costs and
incurred negligible performance penalties. One lossy compres-
sion technique that we evaluated is a sparse polynomial spline
interpolator -- a method that approximately represents a set of ex-
periences with as few coefficients as possible. Supervisors may
use such a sparse interpolator in order to model how the observed
data (i.e., sequences of states, actions, and rewards within a report-
ing window) vary with time. Because states and rewards usually
change smoothly, the number of coefficients needed to represent the
corresponding set of observations Oi is typically much smaller than
the number of observations (K). Note that actions within the set of
observations Oi of an agent i are categorical features, and therefore
are not compressed. We constructed each compressed model of Oi
according to different compression degrees. Compression degree
refers to the frequency with which we subsample elements of Oi
in order to construct the training set for the interpolator. A com-
pression degree of R typically results in models requiring O( K
R )
coefficients in order to approximate a set of K observations. When
employing a lossy experience compression scheme such as this, su-
pervisors relay not complete sets of experiences to all agents within
a sharing group, but only the coefficients of the corresponding lossy
model. Figure 7 depicts how the use of a compression scheme im-
pacts the communication framework by which subordinate and su-
pervisor agents in a network share experiences.
Figure 7: Relaying compressed experiences through a supervi-
sor.
Figure 8 shows the system-wide communication volume (in
bytes) resulting from the use of different compression degrees. In
particular, this graph presents the average system-wide communi-
cation volume when evaluated over all supervisory structures dis-
cussed in Section 7.1 and tested in a network with 100 agents. The
reporting interval in this experiment was K = 100. When per-
forming these experiments observed an interesting trend: the use
of lossy models with compression degree up to 15 had negligi-
ble effect on the performance of the method (smaller than error
bars in Figure 5). This occurs because the set of observations of
an agent (states and rewards) is highly temporally correlated, and
can, therefore, be efficiently compressed via a sparse model and re-
constructed with very little information loss. Compression degrees
higher than 15, on the other hand, resulted in negligible positive
impact on the overall system-wide communication volume, since
the size of the (uncompressed) action time series begins to domi-
nate. These observations suggest that in systems where states and
rewards vary smoothly over time, it is possible to deploy effective
compression schemes for lowering the overall communication costs
of the method -- in this application, resulting in a 5-fold decrease
when compared to an architecture that uses loss-less compression.
Figure 8: System-wide communication volume resulting from
the use of different lossy compression degrees.
7.3 Robustness
In the previous experiments, unless noted otherwise, we used a
reporting interval of K = 115 steps, selected by cross-validation
to minimize a balance between performance and communication
overhead. Smaller values of K lead to more frequent communi-
cation, whereas larger values of K decrease the likelihood that an
agent's transition and reward models will remain static across the
K-timestep interval. The latter case results in reports containing
mixed observations arising from multiple underlying local learn-
ing contexts, which makes sharing less effective. We evaluate the
robustness of our algorithm by studying the effect of using subopti-
mal reporting intervals K. We ran 10 trials of the single-supervisor
and baseline configurations for each of eight reporting intervals,
using λ = 0.3 in a 100-agent network with boundary-based task
distribution (Figure 9). As larger reporting intervals are used, per-
formance degrades, as heterogeneity is induced in transition and
reward samples and the agent learns a policy that averages obser-
vations from different local learning contexts.
We also analyzed our method's robustness by studying the im-
pact of using corrupted or suboptimal context features. Context
SupervisorSubordinate ASubordinate BCompressed ExperiencesCompressed ExperiencesDecompressConstruct Context FeaturesAssessSimilarityCompressSubordinate CCompressed Experiences policy, Q-function or model similarity as proxies of environment
similarity. Although other transfer mechanisms are possible, to our
knowledge no other methods exist that address our particular set-
ting and scale. We believe this is the first algorithm that allows ex-
perience sharing in a concurrent and interacting MAS with ∼1000
agents while undergoing low communication and computational
overhead. Importantly, the time complexity of our method scales
with the number of agents within each local supervisory group, not
with the total number of agents in the network. Experiments further
suggest that the method provides significant improvements over
baseline settings with no experience sharing, and quantitative anal-
yses demonstrate that sharing becomes increasingly advantageous
as the system size grows. Finally, we have shown that our method is
robust to noise-corrupted or suboptimal context features, that com-
munication costs scale linearly with the supervisor-to-subordinate
ratio, and that sparse lossy compression schemes may be deployed
and provide a 5-fold improvement in communication costs while
inducing negligible negative impact on system-wide performance.
REFERENCES
[1] G. Boutsioukis, I. Partalas, and I. Vlahavas. Transfer learning
in multi-agent reinforcement learning domains. In Recent
Advances in Reinforcement Learning, pages 249 -- 260.
Springer, 2012.
[2] J. L. Carroll and K. Seppi. Task similarity measures for
transfer in reinforcement learning task libraries. In
Proceedings of the International Joint Conference on Neural
Networks, pages 803 -- 808. IEEE, 2005.
[3] R. M. De Jong and J. Davidson. The functional central limit
theorem and weak convergence to stochastic integrals i:
weakly dependent processes. Econometric Theory, pages
621 -- 642, 2000.
[4] T. G. Dietterich. State abstraction in maxq hierarchical
reinforcement learning. In Advances in Neural Information
Processing Systems 12, pages 994 -- 1000. MIT Press, 2000.
[5] D. Garant, B. C. da Silva, V. Lesser, and C. Zhang.
Accelerating multi-agent reinforcement learning with
dynamic co-learning. Technical Report UM-CS-2015-004,
School of Computer Science, University of Massachusetts
Amherst, 2015.
Figure 9: Effect of varying reporting intervals K on system
performance.
features that do not properly abstract the underlying local learning
environment make it difficult to identify appropriate sharing oppor-
tunities. To evaluate the sensitivity of our algorithm to this issue
we added different levels of normally-distributed noise to context
features. Noise degrades the quality of the signal encoded in the
features, up to a point where they are entirely uncorrelated with
the underlying system dynamics. The magnitude of the noise was
varied relative to the standard deviation of context features; when
noise level is 1, the standard deviation of the normally-distributed
noise term is greater than (or equal) to the standard deviation of
any context feature, effectively eliminating any signal that they en-
coded. Figure 10 shows that when noise dominates (approaches 1),
performance becomes increasingly volatile. The performance dis-
tribution, with mean approximately 1, suggests that as context fea-
tures become less meaningful, the sharing mechanism is equally
likely to achieve a 50% reduction in the area under the learning
curve as it is to increase this area by 100%. In other words, as the
information-sharing process tends to be guided by biased or incor-
rect features, there is no consistent positive or negative impact on
performance; the most prominent impact is on performance vari-
ability.
[6] C. Guestrin, D. Koller, and R. Parr. Multiagent planning with
factored mdps. In Proceedings of Advances in Neural
Information Processing Systems, pages 1523 -- 1530, 2001.
[7] Y. Hu, Y. Gao, and B. An. Learning in multi-agent systems
with sparse interactions by knowledge transfer and game
abstraction. In Proceedings of the International Conference
on Autonomous Agents and Multiagent Systems, pages
753 -- 761. IFAAMAS, 2015.
Figure 10: Robustness of experience sharing to the use of sub-
optimal/corrupted context features.
8. DISCUSSION
We have presented a solution for experience transfer among
RL agents in large multi-agent systems. Our method adaptively
identifies opportunities to transfer experiences between context-
compatible agents, where contexts provide abstract characteriza-
tions of local learning environments. By explicitly identifying
context-compatible groups, we avoid issues arising from the use of
[8] P. W. Keller, S. Mannor, and D. Precup. Automatic basis
function construction for approximate dynamic
programming and reinforcement learning. In Proceedings of
the 23rd International Conference on Machine Learning,
pages 449 -- 456, New York, NY, USA, 2006. ACM.
[9] H. Kitano, S. Tadokoro, I. Noda, H. Matsubara, T. Takahashi,
A. Shinjou, and S. Shimada. Robocup rescue: Search and
rescue in large-scale disasters as a domain for autonomous
agents research. In Systems, Man, and Cybernetics, 1999.
IEEE SMC'99 Conference Proceedings. 1999 IEEE
International Conference on, volume 6, pages 739 -- 743.
IEEE, 1999.
[10] G. Konidaris and A. G. Barto. Efficient skill learning using
abstraction selection. In Proceedings of the 21st
[27] M. E. Taylor, S. Whiteson, and P. Stone. Transfer via
inter-task mappings in policy search reinforcement learning.
In Proceedings of the 6th International Joint Conference on
Autonomous Agents and Multiagent Systems, May 2007.
[28] R. Tibshirani, G. Walther, and T. Hastie. Estimating the
number of clusters in a data set via the gap statistic. Journal
of the Royal Statistical Society: Series B (Statistical
Methodology), 63(2):411 -- 423, 2001.
[29] S. J. Witwicki and E. H. Durfee. Influence-based policy
abstraction for weakly-coupled dec-pomdps. In Proceedings
of the International Conference on Automated Planning and
Scheduling, pages 185 -- 192, 2010.
[30] C. Zhang and V. Lesser. Multi-Agent Learning with Policy
Prediction. In Proceedings of the 24th AAAI Conference on
Artificial Intelligence, pages 927 -- 934, Atlanta, 2010.
[31] C. Zhang and V. Lesser. Coordinating Multi-Agent
Reinforcement Learning with Limited Communication. In
Proceedings of the 12th International Conference on
Autonomous Agents and Multiagent Systems, pages
1101 -- 1108. IFAAMAS, 2013.
[32] C. Zhang, V. Lesser, and S. Abdallah. Self-Organization for
Coordinating Decentralized Reinforcement Learning. In
Proceedings of the 9th International Conference on
Autonomous Agents and Multiagent Systems, pages 739 -- 746,
2010.
International Joint Conference on Artificial Intelligence,
pages 1107 -- 1112, 2009.
[11] R. M. Kretchmar. Parallel reinforcement learning. In
Proceedings of the 6th World Conference on Systemics,
Cybernetics, and Informatics, 2002.
[12] A. Lazaric, M. Restelli, and A. Bonarini. Transfer of samples
in batch reinforcement learning. In Proceedings of the
International Conference on Machine Learning, pages
544 -- 551. ACM, 2008.
[13] L. Li, T. J. Walsh, and M. L. Littman. Towards a unified
theory of state abstraction for mdps. In Proceedings of the
Ninth International Symposium on Artificial Intelligence and
Mathematics, pages 531 -- 539, 2006.
[14] M. L. Littman. Value-function reinforcement learning in
markov games. Cognitive Systems Research, 2(1):55 -- 66,
2001.
[15] S. Mahadevan. Proto-value functions: Developmental
reinforcement learning. In Proceedings of the 22nd
International Conference on Machine Learning, pages
553 -- 560, New York, NY, USA, 2005. ACM.
[16] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. P. Lillicrap,
T. Harley, D. Silver, and K. Kavukcuoglu. Asynchronous
methods for deep reinforcement learning. CoRR,
abs/1602.01783, 2016.
[17] K. P. Murphy. Machine learning: a probabilistic perspective.
MIT press, 2012.
[18] R. Nair, P. Varakantham, M. Tambe, and M. Yokoo.
Networked distributed pomdps: A synthesis of distributed
constraint optimization and pomdps. In Proceedings of the
National Conference on Artificial Intelligence, volume 5,
pages 133 -- 139, 2005.
[19] F. A. Oliehoek, M. T. Spaan, S. Whiteson, and N. Vlassis.
Exploiting locality of interaction in factored dec-pomdps. In
Proceedings of the 7th international joint conference on
Autonomous agents and multiagent systems-Volume 1, pages
517 -- 524. International Foundation for Autonomous Agents
and Multiagent Systems, 2008.
[20] G. Owen. Game Theory. Academic Press, 2nd edition, 1982.
[21] R. Parr, C. Painter-Wakefield, L. Li, and M. Littman.
Analyzing feature generation for value-function
approximation. In Proceedings of the Twenty-Fourth
International Conference on Machine Learning, page
737â A¸S744, 2007.
[22] B. Price and C. Boutilier. Accelerating reinforcement
learning through implicit imitation. Journal of Artificial
Intelligence Research, 19:569 -- 629, 2003.
[23] B. Ravindran and A. G. Barto. SMDP homomorphisms: An
algebraic approach to abstraction in semi-markov decision
processes. In Proceedings of the Eighteenth International
Joint Conference on Artificial Intelligence, pages
1011 -- 1018, 2003.
[24] H. A. Simon. The sciences of the artificial, volume 136.
1996.
[25] A. Taylor, I. Duparic, E. Galván-López, S. Clarke, and
V. Cahill. Transfer learning in multi-agent systems through
parallel transfer. In Theoretically Grounded Transfer
Learning at the International Conference on Machine
Learning, 2013.
[26] M. E. Taylor and P. Stone. Transfer learning for
reinforcement learning domains: A survey. The Journal of
Machine Learning Research, 10:1633 -- 1685, 2009.
|
1506.05154 | 1 | 1506 | 2015-06-16T21:17:05 | SNA-based reasoning for multiagent team composition | [
"cs.MA",
"cs.AI",
"cs.SI"
] | The social network analysis (SNA), branch of complex systems can be used in the construction of multiagent systems. This paper proposes a study of how social network analysis can assist in modeling multiagent systems, while addressing similarities and differences between the two theories. We built a prototype of multi-agent systems for resolution of tasks through the formation of teams of agents that are formed on the basis of the social network established between agents. Agents make use of performance indicators to assess when should change their social network to maximize the participation in teams | cs.MA | cs | International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
SNA-BASED REASONING FOR MULTI-
AGENT TEAM COMPOSITION
André Filipe de Moraes Batista1 and Maria das Graças Bruno Marietto2
1Metropolitan-United Faculty (FMU), São Paulo, Brazil.
2Federal University of ABC (UFABC), São Paulo, Brazil.
ABSTRACT
The social network analysis (SNA), branch of complex systems can be used in the construction of multi-
agent systems. This paper proposes a study of how social network analysis can assist in modeling multi-
agent systems, while addressing similarities and differences between the two theories. We built a prototype
of multi-agent systems for resolution of tasks through the formation of teams of agents that are formed on
the basis of the social network established between agents. Agents make use of performance indicators to
assess when should change their social network to maximize the participation in teams.
KEYWORDS
Multi-Agent Systems, Social Network Analysis, Complex Systems, Reasoning, Task Resolvers
1. Introduction
One of the research topics in the Multi-Agent Systems (MAS) area is the definition of models that
represents social structures, such as organizations and alliances, in order to analyze more
objectively the emergent behavior of open systems. Individuals who are related to each other by
different types of relationships, such as dependencies on goals, conflicts over resources, similar
beliefs and so on compose organizations and alliances.
Since the 80s several studies were carried out using the theory of MAS to represent models of
social phenomena. Axelrod in [1] argues that the object of these studies is the use of the theory of
MAS to break up simplistic definitions on a particular subject, due to the need of the model be
mathematically tractable. Thus, social phenomena models normally used concepts such as
homogeneity, ignoring interactions.
With the use of agent-based models, this research area has been benefited by being able to
represent most of the behaviors of autonomous agents and the interactions between them. Agent-
based models are in the most cases suitable for decentralized processing and decision making,
particularly when individual interactions lead to the emergence of collective patterns, as well as
complex systems. Therefore, there is a close relationship between MAS and Complex Systems
Theory, and this relationship can be represented in the Social Network context.
An important question is how to represent these relationships, characterized by a high degree of
dynamism. Dealing with social issues, MAS area was inspired mainly by the economic
organizational theory and legal theory. Although, there is a lack of attention in the area of
research that describes the relationships between individuals within human organizations and
their dynamics. To this area, we give the name of Social Network Analysis (SNA). The social
network analysis has emerged as a key technique in modern sociology, anthropology, social
DOI : 10.5121/ijaia.2015.6305 51
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
psychology, communication studies, information science, organizational studies, and economics,
among others.
This work aims to integrate the techniques of multi-agent systems and social network analysis,
allowing an agent society make use of SNA metrics in the decision-making process considering
attributes such as the position in the network and the value and the importance of an agent in the
network. Since both techniques use the social element, the agent, the synergy between them will
allow easy use of models based on social network analysis, so that the multi-agent systems
become more reliable in the representation of social phenomena. For this, an application was
developed representing process of selecting teams for solving tasks in a multi-agent society, in
which agents have specific skills and their position in the social network will influence their
performance in the multi-agent society.
This work is structured as follows: in Section presents a review of the issues of multi-agent
systems and social network analysis area, while Section presents prospects of integration between
both techniques. Section presents and discuss about the prototype of multi-agent system for a
task-solving problem by agents teams making use of SNA technique. Section presents the final
considerations of this work.
2. Background
In this section, we present a review of the literature on multi-agent systems (MAS) and social
network analysis (SNA) research areas.
2.1. Multi-Agents Systems
Multi-agent systems consists in multiple agents that interact to each other in order to perform a
specific set of tasks. The metaphor of intelligence used by these systems is the intelligent
community in which social behavior is the basis for the system intelligence. From this, we can
distribute the agents in areas of specialization and with the interaction among them a complex
problem can be solved in a faster and more dynamically way[2].
The metaphor often used to structure the agents society is the human social groups, where expert
teams can solve problems cooperatively, and the complexity exceeds the individual capacities of
each of their members [3].
The autonomy of an agent allows it to take its own decisions to achieve its goals [4]. Thus, agents
can get in and get out of the society, change their rules, roles, inter-dependent relationships with
other agents, etc. This feature leads to a new generation of systems and distributed applications
intrinsically dynamic, open and complex.
The notion of agents and multi-agent systems have been adopted in the modeling of various
complex systems involving urban planning, biology, logistics and production, and many others [5,
6, 7, 8, 9, 10, 11, 12, 13]. Lynne and Nigel [14] propose a model based on agents for social
networks. At the same step that the proposed model is simplistic, it can represent a wide variety
of social networks. Teresa and Nina [15] implemented a Java tool that presents the dynamic
model of social behavior associated with the recruitment of terrorists based on prescriptive
models. They used concepts of MAS and SNA to represent this dynamic. Ronald et al. [16]
proposed a model based on agents to analysis the social influence in the travel activities decision
process. The agents have a travel schedule, and interact with each other in order to schedule social
activities, in particular trade based on the nature of the activity, who will participate, time and
52
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
location. The structure formed between the agents described as a complex social network is the
core of the decision process.
2.2. Social Network Analysis
Social Network Analysis (SNA) is a scientific research area derived from areas such as
Sociology, Social Psychology and Anthropology. This area studies the relational links
(a.k.arelational tie) between social actors. An actor in the SNA can be both individuals and
companies, analyzed as individual units, or collective social units such as, for example,
departments within an organization, public service agencies in a city, nation-states of the
continent, among others. The SNA fundamentally differs from other studies in that the emphasis
is not on the attributes (characteristics) of the actors, but the links between them.
Relations between actors pairs are made by relational ties or linkages. The most common types of
links are: individual assessment (eg, friendship or respect); the transaction and the transfer of
material resources (a purchase and sale transaction between two companies); the transfer of
nonmaterial resources (the exchange of electronic messages) or not; the association or affiliation
that occurs when actors participate in joint events (parties); interaction (sitting next to another
person); the movement and the physical and social connection; links between formal roles (boss-
subordinate authority loop in a company); biological relationships (father and son) and so on.
SNA is a method for enhancing the sharing of knowledge by analyzing the position and structure
between actors, i.e their relationships. According to [17], the network analysis process considers
basically two analytical perspectives that complement each other:
other nodes/actors in the network with others ego nodes that it maintains relations.
Therefore, the number, the magnitude and the diversity of the direct or indirectly
connections established with the ego define the other network nodes;
1. Egocentric - this type of analysis has attention in facing the particular node/actor (ego) and
2. Socio centric - in this type of analysis, information about the pattern of links between all
1. Individual positions - from the viewpoint of the actors;
2. Whole network - from the viewpoint of the assembly of links forming the network;
3. clusters and components - from the viewpoint of the groups formed due to some kind of
The interpretation of the results obtained with the SNA metric can be made from three points of
view, namely:
nodes/actors in the network is used broadly to identify reticulated subgroups with high
degree of internal cohesion and the nodes that have similar positions on the network.
relationship.
The social network analysis metrics provides mathematical mechanisms to analyze a given
society or group. Among the most significant metrics for understanding the role of an actor in the
network, we can cite:
• Centrality Degree: it indicates the prestige and power that the author has in that particular
• Betweenness Degree: it shows how much an actor is between others actors on the network.
community. In the SNA context, the power dimension is derived from the relational ties of
an actor. The more relationship an actor makes, the more power it has.
It helps to locate actors in positions of influence, those with the network information, etc.
An actor with a high degree of betweenness may be functionally operating as a broker in
the network;
53
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
• Closeness Centrality: shows the closeness of an actor to others on the network and thus its
access to resources circulating in the network, based on the evaluation of the shortest path
within the network.
Many of these metrics can be used in a multi-agent scenario, since it is composed of independent
individuals who interact with each other. However, there is still a gap on the theory and practice
on this issue, that is, how to implement and model a MAS with many different relationships
between agents in order to capture and analyze them.
3. Social Network Analysis and Multi-Agent Systems
One of the biggest questions about the social network analysis is“it is known that social networks
can be used to analyze human societies. But these can also be used for societies of agents?”. An
inconvenient and often questionable point, on social network analysis is that normally a
transcription of the human social structures in sociological meanings is made from a clear and
artificial manner, due to the complexities of individuals and their relationships. On the opposite
side the simplicity in social terms of multi-agents systems suggests that the analysis of social
networks can be applied to these, including may get even better results.
In the interconnection of two areas of study, one of the points of attention is the level of approach
to be carried out. A multi-agent system is built from the specification of the smallest entity: the
agent. The behavior and the structures are emerging consequences of their interactions. The same
is true for Social Networks; the interaction of the actors will result in the emergence of the
structure of the social network.
Many MAS consist of a complex network of autonomous and interdependent agents. In most of
these systems, agents must select a set of other agents with whom will interact, based on factors
such as limited resources, exchange of interests of knowledge, etc.
[18] Showed that the structure of existing artificial social network in multi-agent society, when
used to govern the actions of the agents, is extremely connected with the performance of the
multi-agent society.
Modeling scenarios in which agents are geared to social networks presents a set of challenges.
First, agents must make adaptations in the network by decisions made based on local system
information. This local information can give the agent a figure partially or completely wrong
about the current environment. In this case, it is said that an agent has a limited horizon of the
system. Since agents are organized as networks, and how various agents perform their local
decisions (adapting your social network), these simultaneous changes may neglect the benefits of
adaptation made by an agent.
Focusing on the social network adaptation strategy of an agent in a cooperative MAS (the agents
cooperate towards a single goal), the following questions should be considered when modeling
such agents, which are
• Local perception of overall performance: how an agent can estimate the collective
performance of the organization? These estimates may be unreliable, since they are each
based on partial views of the organization. A possible solution to this difficulty is to explore
the communication skills of the agents. With use of some communication protocols, an
agent can estimate whether its perception of overall performance is correct, or if it can use
the perception of neighboring agents to improve it;
54
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
• Triggers of adaptation: when an agent must decide to adapt its structure of local
• Network adaptation: how an agent decides which has to be removed connection and how it
connectivity? There are many possibilities of these triggers be fired. An agent, for example,
may decide to adapt its structure based on estimates of performances in relation to the
attendance of its goal;
selects a new agent to establish a new social connection. A good strategy advocated by [18]
is the adaptation based on reference where an agent asks for a reference about another agent
for one of your neighbors.
Social Network Analysis’ based Reasoning for Multi-Agent Team
4.
Composition
Based on [18], the proposed prototype on this work is a simple but intuitive multi-agent
cooperative system. It is a multi-agent society for dynamic team formation of task resolvers.
This prototype presents a model of dynamic formation of teams, where teams of agents are
formed spontaneously and in the decentralized form, as soon as the decision of an agent in
adhering to a team is done by base in its local social network.
In the proposed model, a set of tasks are generated periodically and are globally advertised to the
multi-agent society. The role of agents in this society is forming teams to solve these tasks. The
participating agents of this society are involved for a social network. For an agent to be on a team,
it must possess a social connection (that is, an edge in the network) with at least one member of
the team.
Once this prototype focuses on the team formation process, when these teams are complete the
task is considered done. Thus, in this model the multi-agent society consists of N agents, A =
{a1,a2,...,aN}, where each agent can be considered as a single node on the social network. The
network is modeled as an adjacency matrix E, where each adjacency matrix elements eij= 1if there
is an edge between the agents ai and aj, or eij= 0 otherwise. The degree (number of connections) of
an agent ai is defined by
In multi-agent society, each agent has a single ability, σi∈ [1,σ], where σ is the number of different
types of abilities present in the society. During the process of forming teams, each agent can be in
one of three states: uncommitted, committed, active. An agent is in the state uncommitted if it is
available to join teams and therefore is not in any team. When an agent selects a task in which it
is able to perform, this is in the committed state. When this agent effectively is accepted for the
team, the same is in the active state.
Agents that are in the active state can no longer get out of the team until it is complete or a
timeout is reached. In this case, the team was not formed in the required time to resolve the task.
The tasks are advertised in multi-agent society in a fixed interval of time µ. Each task is a vector
Tk of size k with required skills. Each task is announced by a finite number of time γ, ensuring
that resources (i.e., agents) linked to a task become available if the team is not formed.
The requirement of a team be an induced connected subgraph of the social network means that for
any agent on the team, ai∈ Mk - where Mk is the k team of this society, there
55
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
should be at least some other agent aj∈ Mk,i≠j such that eij= 1. This implies that an agent in the
Figure 1: Process of forming teams of agents to task resolution, making use of social network metrics.
uncommitted state is only eligible to enter in the committed state in two situations: 1) team
startup, when the team is empty, and 2) joining a team, when at least one of the neighbors of the
agent is in committed state related to this task.
Unlike the model presented by cite [18], in this model all agents in the uncommitted state can try
to belong to a team when they receive a task in which its ability is being requested. At this point,
a Scheduler agent should coordinate the multi-agent society. It is up to this agent accepts or not
the agents making use of SNA heuristics. Figure 1 presents an indicative flow chart of the multi-
agent process, highlighting where the techniques and metrics from social network analysis are
used.
In the flow shown in Figure 1 interactions, the role of the manager agent is generating tasks (skill
set) and send them to the scheduler agent. The scheduler agent has the following set of functions:
receive tasks sent by the manager agent and transmit them to all task agents; receive the proposals
from task agents; select the agents to composition of the teams and check the conclusion
(satisfactory or not) of the tasks.
Task agents receive tasks and if they have the necessary skills, send an application to the
scheduler agent. Right now, they are awaiting a confirmation or a rejection message. If it is
rejected then returns to receive tasks. If accepted, goes to the team and so it is waiting for the
complete formation of the team.
Social strategies are implemented as follows:
56
1. Scheduler Agent: it knows the entire network of agents. When an agent task sends a
message stating the interest in participating in the team, the scheduler agent checks whether
the team is empty (in this case the agent is starting the team), or if the agent knows other
agents that make up the team. This strategy that how the structure of social networks can be
used for decision-making. Such approach can be used to solve problems of allocation in a
general way. These problems have a high degree of complexity, and this occurs for two
main reasons. First, find the optimal scheduling is a NP-hard problem. In addition, each
scheduling problem has particular details, involving changes in the search algorithm of the
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
solutions. Thus, it becomes necessary to use some type of heuristic to reduce the search
space. SNA metrics can be a valid heuristic in this situation;
2. Task Agent: each task agent has a local measure of performance. This measure will be the
gear adaptation of the agent’s social network strategy. This strategy is based on
performance and reference [18]. The trigger of this adaptation is the measure of local
performance Y (ai), which each agent ai has. This measure is the ratio between the number
of times the agent participated effectively in a team (it was accepted for a team) by the
number of attempts to participate. The measure of performance is considered valid when
the agent tried to enter a team at least v times. With each iteration, the agent ai chooses
adapt their social network if this has a measure of performance valid, and if the measure of
performance is below the average of the sum of the performance of all its neighbors. That
is, an agent chooses adapt its social network if:
If an agent decides to adapt its network based on the performance value, this adaptation is
given both by performance and by reference. The agent will remove the connection to its
immediate neighbor aj which has the smallest measurement of performance:
(2)
The agent asks for a reference to the neighbor with the highest performance. Similarly, this
agent will refer to its neighbor with the highest performance. Is al the agent to whom the
agent ai calls for reference, the agent will establish a new connection with ak, the neighbor
of greater performance of al:
Such strategy represents how the knowledge of the network can be used in a MAS. In this
case, the task agent is making use of the centralization degree of the SNA theory. This is a
measure of power, influence, and which in this case is a measure of performance. Thus, an
agent is able to adapt its social network site so that this attempt to get more success in being
a member of a team.
4.3. Communication Protocols
All SNA-based reasoning may be mapped using a set of communication protocols in order to
establish a standard mechanism for establishing team for solving tasks. This protocol consists of a
set of messages with ACL (Agent Communication Language) perfomatives that indicate what
actions are being performed on the social network of agents.
The process begins when the manager agent sends a propose message to the scheduler agent. The
scheduler agent sends this message to all task agents. If a particular task agent has one of the
skills required by the task, this sends an accept-proposal message to the scheduler agent.
After receiving all the proposals, the scheduler agents reasons using SNA metrics and sends
accept-proposal messages for the agents that were accepted, and reject-proposal
57
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
Figure 2: Prototype developed in JADE platform: social network dynamically built by agents, monitoring of
messages exchanged and log of actions.
for those that were rejected. Agents who received the acceptance, are now committed to a team
and will no longer respond to the proposals made by the scheduler. The scheduler agent checks
after a time γ if the team was complete. If it is not complete, this sends a failure message to staff
agents of this team, stating the fact. When a team is complete, the scheduler agent considers that
the task was performed with success, sending a message confirm, releasing the agents in order to
be a member of new teams.
The mechanism of adaptation by performance and reference was implemented as follows: when
an agent reaches a valid value of performance (for example, after 10 attempts of formation of
teams), it sends a query-if message questioning to its neighbors about their values of performance.
The neighbors receive this message and respond to the agent with the value of its respective
performances, with a inform-if message. After receiving all the performances, the agent checks if
is advantageous to perform the change. If yes, this sends a proxy message to its neighbor with
higher performance by asking another agent by reference. After receiving this referenced agent
the agent adds it to its network, removing the neighbor with the lowest performance.
Figure 2 shows the control screen developed in JADE platform, allowing to observe in real-time
the changes in the structure of the social network, the messages exchanged between them, as well
as a detailed action log, which enables us to analyzing the SNA metrics used for decision-making
by each agent.
5. Final Considerations
Multi-agent systems and social networks are closely united. These two theories help to shape and
understand social phenomena in many different ways. This work was concerned to understand
how social networks could improve the multi-agent system modeling process.
58
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
The integration between the two areas can be in two ways: macro and micro modeling. The macro
modeling consists in using the existing metrics in SNA to analyze multi-agent behaviors. Given
the inherent distribution of multi-agents systems and the high degree of interaction in the MAS
society, it is important to have a tool that assists in validation of proposed model.
On the other hand, in the micro modeling agents are built making use of social networking
metrics in their behavior. With this, the interaction between the agents can generate new social
phenomena, as well as allow them to take a decision based on your social network. The prototype
used this approach and showed how an agent society can be built to solve a set of tasks where
teams are formed based on the social network structure.
As future work, we intend to improve the modeling of new social behaviors, such as anarchy, lack
of collaboration, and new strategies of relationship between agents that allow a multi-agent
society to act on non-cooperative scenarios.
References
[1] Axelrod, R.: Advancing the art of simulation in the social sciences. Working Papers 97-05-048, Santa
Fe Institute (May 1997)
[2] dasGraças Bruno Marietto, M.: DefiniçãoDinâmica de EstratégiasInstrucionaisemSistemas de
TutoriaInteligente: Uma AbordagemMultiagentesna WWW. PhD thesis, InstitutoTecnol´ogico de
Aeron´autica (2000)
[3] de Lira Lino, N.: Modelo de percep¸cao de agentesbaseados no sistema de emo¸coes (2006)
[4] Castelfranchi, C., Falcone, R.: From automaticity to autonomy: The frontier of artificial agents. In
Hexmoor, H., Castelfranchi, C., Falcone, R., eds.: Agent Autonomy. Volume 7 of Multiagent
Systems, Artificial Societies, and Simulated Organizations. Springer US (2003) 103–136
Jiang, Y., Jiang, J.: Understanding social networks from a multiagent coordination perspective. IEEE
Transactions on Parallel and Distributed Systems 99(PrePrints) (2013) 1
[6] Sless, L., Hazon, N., Kraus, S., Wooldridge, M.: Forming coalitions and facilitating relationships for
completing tasks in social networks. In: Proceedings of the 2014 International Conference on
Autonomous Agents and Multi-agent Systems. AAMAS ’14, Richland, SC, International Foundation
for Autonomous Agents and Multiagent Systems (2014) 261–268
[5]
[7] Pujol, J.M., Sanguesa, R., Delgado, J.: Extracting reputation in multi agent systems by means of
social network topology. In: Proceedings of the First International Joint Conference on Autonomous
Agents and Multiagent Systems: Part 1. AAMAS ’02, New York, NY, USA, ACM (2002) 467–474
[8] Pitt, J., Ramirez-Cano, D., Draief, M., Artikis, A.: Interleaving multi-agent systems and social
networks for organized adaptation. Comput. Math. Organ. Theory 17(4) (November 2011) 344–378
[9] Hu, X., Zhao, J., Zhou, D., Leung, V.C.: A semantics-based multi-agent framework for vehicular
social network development. In: Proceedings of the First ACM International Symposium on Design
and Analysis of Intelligent Vehicular Networks and Applications. DIVANet ’11, New York, NY,
USA, ACM (2011) 87–96
[10] Aziz, A.A., Ahmad, F.: A multi-agent model for supporting exchange dynamics in social support
networks during stress. In: Proceedings of the International Confer-
ence on Brain and Health Informatics - Volume 8211. BHI 2013, New York, NY, USA, Springer-
Verlag New York, Inc. (2013) 103–114
[11] Heppenstall, A.J., McFarland, O.E., Evans, A.J.: Application of multi-agent systems and social
network theory to petrol pricing on uk motorways. In: Proceedings of the 4th International Central
and Eastern European Conference on Multi-Agent Systems and Applications. CEEMAS’05, Berlin,
Heidelberg, Springer-Verlag (2005) 551–554
[12] Ma, L., Zhang, X.: Hierarchical social network analysis using a multi-agent system: A school system
case. Int. J. Agent Technol. Syst. 5(3) (July 2013) 14–32
[13] Kumokawa, S., Kryssanov, V.V., Ogawa, H.: Sona: A multi-agent system to support human
navigation in a community, based on social network analysis. In: Proceedings of the 8th International
Conference on Intelligent Virtual Agents. IVA ’08, Berlin, Heidelberg, Springer-Verlag (2008) 509–
510
59
International Journal of Artificial Intelligence & Applications (IJAIA) Vol. 6, No. 3, May 2015
[14] Hamill, L., Gilbert, N.: A simple but more realistic agent-based model of a social network. The Centre
for Research in Social Simulation (2008)
[15] Ko, T.H., Berry, N.M.: Agent-based modeling with social networks for terrorist recruitment. In
McGuinness, D.L., Ferguson, G., eds.: AAAI, AAAI Press / The MIT Press (2004) 1016–1017
[16] Ronald, N., Arentze1, T., Timmermans, H.: An agent-based framework for modelling social influence
on travel behaviour. In: MODSIM. (2009) 2955–2961
[17] Wellman, B.: The persistence and transformation of community: From neighbourhood groups to
social networks (2001)
[18] Gaston, M.E., Desjardins, M.: Agent-organized networks for dynamic team formation. In: In
Proceedings of clustering diameter the Fourth International Joint Conference on Autonomous Agents
and Multiagent Systems. (2005)
60
|
cs/0605032 | 1 | 0605 | 2006-05-08T12:27:59 | A framework of reusable structures for mobile agent development | [
"cs.MA",
"cs.AI",
"cs.SE"
] | Mobile agents research is clearly aiming towards imposing agent based development as the next generation of tools for writing software. This paper comes with its own contribution to this global goal by introducing a novel unifying framework meant to bring simplicity and interoperability to and among agent platforms as we know them today. In addition to this, we also introduce a set of agent behaviors which, although tailored for and from the area of virtual learning environments, are none the less generic enough to be used for rapid, simple, useful and reliable agent deployment. The paper also presents an illustrative case study brought forward to prove the feasibility of our design. | cs.MA | cs | A framework of reusable structures for mobile agent development
Bogdan Dumitriu
Tudor Marian
Computer Science Dept.
Computer Science Dept.
Tech. Univ. of Cluj-Napoca Tech. Univ. of Cluj-Napoca
26-28, G. Baritiu St.
26-28, G. Baritiu St.
400027 Cluj-Napoca
400027 Cluj-Napoca
Romania
Romania
[email protected]
[email protected]
Mihaela Dinsoreanu
Computer Science Dept.
Tech. Univ. of Cluj-Napoca
26-28, G. Baritiu St. 400027
Cluj-Napoca
Romania
[email protected]
Ioan Salomie
Computer Science Dept.
Tech. Univ. of Cluj-Napoca
26-28, G. Baritiu St.
400027 Cluj-Napoca
Romania
[email protected]
Abstract – Mobile agents research is clearly aiming towards
imposing agent based development as the next generation of
tools for writing software. This paper comes with its own
contribution to this global goal by introducing a novel
unifying framework meant to bring simplicity and
interoperability to and among agent platforms as we know
them today. In addition to this, we also introduce a set of
agent behaviors which, although tailored for and from the
area of virtual learning environments, are none the less
generic enough to be used for rapid, simple, useful and
reliable agent deployment. The paper also presents an
illustrative case study brought forward to prove the feasibility
of our design.
Keywords: mobile agents, agent behavior, unifying agent
platform framework
I. INTRODUCTION
Although agent platforms are becoming increasingly
widespread and powerful these days, although they are
evolving rapidly towards what some call a “second
generation” and although they are the centerfold of
extensive research all around the world, there still appears
to be something which prevents them from being
universally adopted as a natural evolution of the object
oriented world. In our opinion, there are actually two
things, not just one, which can be held responsible for the
current situation: lack of simplicity and lack of
interoperability.
Let us discuss lack of simplicity first. Most agent
platforms nowadays find themselves in one of the two
extremes: they either offer enormous flexibility at the cost
of usability (the user gets tangled in aspects that are
completely irrelevant to her application, thus losing
perspective) or they offer extended built-in functionality at
the price of interoperability and extensibility (the user has
plenty of predefined agent services to choose from and
plug into her application, but is confined to the specific
agent platform). The JADE [7] and the ADK [8] agent
platforms both fall into one of the two above categories.
Neither of these extremes can completely satisfy the user's
requirements. Ideally, an agent platform should offer the
means for the user to easily set up and run her agents (if so
required) and also the flexibility of changing virtually
everything about an agent if the need arises.
We also mentioned lack of interoperability in the
argument above. This is a bit more difficult to see,
especially for the untrained eye, due to the standardization
efforts such as those of FIPA [6] which allow a certain
degree of interoperability between agents built on top of
different agent platforms, or due to included semantic
support, such as that provided by ontologies. The catch,
however, is that compatibility among agents belonging to
different frameworks which are FIPA-compliant is virtually
restricted to what FIPA defines. This is a serious drawback,
if one cares to analyze it, because FIPA defines nothing in
terms of components which can be used for rapid agent
development and deployment. The implication of this fact
is that implementation efforts employed by various agent
platform developers are not reusable outside their own
platform (at least not without serious adaptations efforts).
The ultimate goal should be for the user to be able to
simultaneously benefit not only from a single, but from all
development work conducted by researchers and
implementors all around the globe.
These are the two main deficiencies which we have tried
to address with our work. Our idea was to overcome such
limitations by defining an abstract, platform independent,
framework which would allow the user to easily:
•
•
•
define her agents in terms of high level abstractions;
be able to “plug in” various common predefined
behaviors;
switch from one platform to another as the need arises.
Given that the area where we make use of agents falls
roughly into the category of computer based education, the
idea described before was primarily focused on the
development of a framework which would allow us to
benefit from all the aforementioned improvements in our
line of work. As a consequence, we narrowed our agent
directed research down to field of virtual learning
environments (VLE). This does not mean, however, that we
didn't design our framework to be as general as possible. It
merely implies the fact that the behaviors we have
identified and introduced into our implementation were
drawn mainly from the VLE field. Even so, by reading the
paper, it will become obvious that most of behaviors
described are equally applicable to a variety of other fields
as well.
In short, a virtual learning environment represents a
space available on-line where both students and teachers
are brought together to interact similarly to the way they do
in reality. There are a few things pertaining to a VLE which
allowed us to make certain simplifying assumptions in our
work. In the first place, a VLE is usually what one calls a
“closed environment”, meaning it has few (if any)
interactions with other systems. This has a direct
consequence in that our agents will usually be confined to
using “interior” (as opposed to “border”) protocols only.
Secondly, strong security requirements are normally absent
from a VLE, which makes concerns for agent
communication security a rather unimportant issue in our
case. Finally, the tasks our agents need to perform are
generally simple ones and can easily be achieved by the
composition of two or more basic behaviors, thus enabling
us to disregard some of the more advanced behaviors
(which might indeed be appropriate for other
environments) from our analysis.
We have based our agent work on that explained in [4].
What we have done was pick up where the authors of [4]
left off and change the agents employed in the assessment
service (AS) described there in order to make it benefit
from the advantages of our agent framework. This was
actually a means to prove the feasibility of our research.
In the first part of our paper we introduce and justify the
need for simple agent based reusable structures, especially
in an environment such as a VLE. We do this by describing
our vision of a unifying agent platform framework that puts
more (sample) real platforms all under the confines of the
same abstraction. We also bring forward a behavior
taxonomy which, in our opinion, can successfully be used
both in a VLE and in other environments to create
significantly useful agents “on the fly”. The detailed
description of such a behavior (namely, the Itinerary
behavior) is also addressed in last subsection of this part.
Since we have also made use of our platform in practice
as well, the second part of this article is dedicated to a case
study which covers the realization of a VLE assessment
service based on the agent technology. The study places the
service within the virtual university we have created,
defines the technical challenges that we have encountered
and discusses the solutions we have proposed. We close up
with some conclusions and ideas we still intend to pursue in
this area.
II. AGENT BASED REUSABLE STRUCTURES IN
VLE
Reusable structures [5] are indeed a very powerful tool
aiding the rapid and well-structured development of
applications, and it is our belief that agent-based software
should benefit of such mechanisms too. We have tried to
extract the behavioral patterns that agents are enacting
while deployed in their environment, namely the virtual
learning environment (VLE).
It is well known that complexity severely compromises
the usability of models, therefore reducing the chances of
adoption. Indeed, simple things have the reputation of
offering a better solution, if applicable, since they are both
easier for the user to understand and, therefore, use and for
developers to implement. The actual challenge was to find
the balance between expressive power and simplicity. This
required the analysis of the VLE domain we have
previously explored in [3] and [4] in order for us to:
•
•
determine the set of adequate abstractions for our
conceptual model
avoid artificial inclusion of complex abstractions,
especially of those that may easily be composed on
top of the basic ones, as well as ones that may be
needed on rare, possibly particular, occasions
To achieve this goal we strived to find the minimal set of
features and abstractions that would be both useful and
powerful enough to naturally describe the complex
behaviors of coexisting agents working together to perform
an assigned task.
A. Unifying Agent Platform Framework
Recurrently occurring patterns in the use of agent-based
development led us to pinpoint certain behaviors
encountered on numerous occasions, enforcing the idea of
reusable structures when handling this paradigm as well.
These structures were embodied into a comprehensive
agent behavioral model shaped on top of a unifying
framework. By means of such a framework we managed to
make the agent platform transparent to the user and, in the
same time, decouple the reusable patterns from the
underlying mobile agent platform. It thus becomes clear
that the model was structured to be highly independent,
encompassing a handful of abstract features that allow it to
be equally expressive regardless of the underlying agent
support.
Entities common to every agent platform (location,
agent, message, behavior, agent identifier along with other
relevant ones) provide the context within which we were
able to define the reusable patterns. These patterns produce
an environment that ultimately separates the behavioral
model from the actual skeleton upon which the patters are
enacted (i.e. the JADE agent platform) and, as such, once
they are created, rewriting them will not be necessary for
every new platform. Simply put, one has only to write new
adapters if needed, or use the available ones along with the
already existing framework items to integrate (coalesce)
the component she requires.
Adapters were employed to provide the bridge between
the framework and agent platforms. In this way, adding
support for a new system only requires the writing of the
appropriate wrappers. The following diagrams (Fig. 1 and
Fig. 2) illustrate the most relevant adapter hierarchies and
their position in the system.
Fig. 1. Agent framework class diagram – agent adapters
Fig. 2. Agent framework class diagram - location adapters
Other components are adapted similarly, following the
general line of the architectural pattern. We have refrained
from including more diagrams due to space issues.
The ultimate purpose of the framework was to provide a
fast and easy method of deploying agents by connecting
together the necessary pieces, in our case behavior
patterns. These patterns are to be presented in a more
detailed manner during the following section.
B. Agent Behavior Pattern Taxonomy
The basic patterns we have identified can be grouped under
headings such as creation-based (Role Factory, Clone),
task-oriented (Observer, Task), communicational (Client-
Server, Listener) and mobility-related patterns (Itinerary).
Extending this hierarchy should be a fairly easy task, and
one that we encourage as well, given that we only provide
those concepts related to the VLE context – but their
overall level of generalization should not be overlooked
when developing a particular solution. The following figure
(Fig. 3) shows the hierarchical structure of the behaviors.
Role Factory: It allows adding a behavior to an agent
dynamically, at runtime, the agent being unaware of the
concrete behavior type it will enact.
This particular
behavior can be used to provide additional behaviors to
various agents, behaviors that these agents should be
unaware of at the time of their creation. It can be used, for
example, in a scenario where a social group of agents is
created, performs some sort of initial task and later awaits
to be assigned a distinct behavior by this factory.
Itinerary: It provides the agent with the ability to
perform given tasks at several locations along its route.
Basically, it enables the agent to travel from one
destination to another, attempting to best meet a
preestablished schedule at the same time. For example, this
kind of behavior can be used in route discovery scenarios.
Although some could argue that it is rather VLE domain
specific, we believe that this pattern is useful throughout
various scenarios, and our assessment service provides a
pertinent example in this direction. It is why it shall be
detailed in the later sections.
Observer: This pattern allows the given agent to check
periodically for the occurrence of various events. Should
any such event occur before the next time quanta has
elapsed, an event handler will be triggered, performing the
required action. It can be used when a number of tasks have
to be performed preconditioned by particular events. For
example, consider a VLE assessment service and the agent
that is given the task of delivering the mandatory tests to
recipients at given dates (the midterm date). Such an agent
can use this behavior in the following manner: if the time
has come (the trigger in this case would be the reaching of
a certain date) the action to be taken is the delivery of tests
to students (one could use the Itinerary behavior for such a
task).
Listener: It is a communication pattern - it brings the
agent it belongs to in a state of waiting for a message. The
message is set up with a type, thus enabling the agent to be
awoken only when a message with the expected type
arrives (naturally, one can use a generic type to grab every
message received). Only then will all the registered event
callbacks be fired. Such a behavior can be used in a large
number of scenarios, for instance when a message
addressed to an agent is supposed to originate a certain
action (or sequence of actions). Multiple event listeners can
be registered in order to be fired on the arrival of a
message; this behavior can be used either cyclically or in a
“one shot” manner.
Fig. 3. Agent framework behavior hierarchy
Client – Server: The two roles are connected, since both
require their counterpart to properly perform its task within
the communication protocol. Since the messages are
considered to have a high degree of reliability, the protocol
requires only one acknowledgment from the server
(actually a server worker thread) towards the client. Both
the sever and the client take turns in communicating with
each other, and the semantic is as follows:
• The Client, knowing the location of the server, initiates
the communication protocol, by sending a request. This
request encapsulates a task the server is required to
perform (and a container for the result of the operation).
The client blocks awaiting an acknowledgment
message, and, if received, it awaits the result of the task
performed by the server worker. Timeout is used to
detect a possible failure and the agent terminates the
behavior (to emphasize the flexibility of our framework
let us mention that one could, for example, use this
behavior together with the Observer to retry connecting
after a certain period of time has elapsed). The
acknowledgment message is required in this phase
because it it the most vulnerable period within the
sequence, when the server could not be available due to
different adversities.
• The Server behavior, when started, awaits requests from
its clients. Upon the arrival of a valid message a new
server worker thread is spawned (enabling multi
threaded server response policy) which takes over all
further communication with that particular calling
client. Each server worker sends an acknowledge
message back to the originator of the message, namely
the corresponding client, starts processing the task he
was given, again by the client along with the initial
message, and sends back the result. This result message
is the one that brings the client behavior to its end, thus
terminating the current communication session with the
server (the thread).
This pattern is susceptible to variations and further
adaptations; we tried to reach a fair level of functionality
without imposing too many bounds to the model, as that
would most likely discourage the user from its
employment.
Task: This is the standard “one shot” behavior expiring
after the assigned job has completed. It is a very general,
simple pattern one can use in every situation, and should be
considered when extending simple behaviors.
While we have tried to identify the main abstractions one
could use, surely there are others we have failed to
recognize, or that may become relevant in the future. In
addition, since these abstractions are extracted solely from
the VLE context, horizontal as well as vertical domains
may recognize the need for adding several other relevant
components, whether they are more general or more
domain-specific. Consequently, the model was devised to
be extensible in that it allows the adding of any newly
identified property (or behavior for that matter) to the
framework with very little effort and no modifications.
Clearly, one can infer the need for composition when
handling such basic behaviors in order to increase the
complexity and also the power of the model. As a response
to this, three such types of behaviors have been introduced:
sequential, parallel and Finite State Machine (FSM), and
more can be added as the circumstances require it.
• Sequential behavior deploys its elements one at a time,
following a predefined, dynamically changeable, order;
• Parallel behavior deploys all its elements on separate
“threads”. The components are enacted in a parallel
manner;
• FSM conforms to a somewhat sophisticated description
of state machines (the composite branch can be
observed in Fig. 4 as well). States and transitions are
the main concepts used when deploying this behavior,
while the Task pattern is used to model the activity per
state.
C. Itinerary Pattern Description
To better comprehend the inner workings of the
behaviors, in the following we shall provide a more
thorough description of one such entity, namely the
Itinerary Behavior, that was deliberately under-described in
the above section to avoid redundancy. It is meant to
provide agents with the ability to travel along a path
following a time schedule and perform assigned tasks at
each location, the locations being part of its initial route
description.
Basically, to deploy this pattern one has to initially
Route,
configure the behavior with a
several
ObjectiveReachedListener
items,
and
an
ObjectiveMissedBehavior. None of these elements can be
changed at any time after the behavior has been enacted.
The sequence diagram and the state-chart diagram (Fig. 4
and Fig. 5) offer a graphical representation of the entities,
relations, flow of messages and main states defining this
behavior. Hopefully they will provide enough insightful
ancillary information for one to better understand our
design approach.
The Route is a list of objectives; each objective contains
a location that must be reached and a pair of time
indicatives, the earliest and the latest possible time of
arrival at that given location. The time units are increments
to a base value established when the behavior is initially
created. This solution was adopted due to the nature of the
agent environment's possible distribution (since one agent
could be required to travel at different locations, possibly
on different time zones to perform its assigned task) and
also due to the need to consider the variable travel delay
and the delay raised by the deactivation and reactivation of
the agent before and after the migration process. Should the
agent reach one location on schedule, i.e. at some time
between the two above mentioned time indicatives, the
ObjectiveReachedListener registered listeners are fired.
Upon completion the agent migrates to the next location,
chosen to best meet the deadline of arrival. If the agent
reaches one location early it enters a wait state until enough
time elapses, awakes and performs the assignments, then
continues along its path. If, on the contrary, one location
has not been reached on time (the delays are considered,
although ways of properly computing them are still under
development) the agent is free to enact the
ObjectiveMissedBehavior if one exists (otherwise the agent
falls into a complete stop).
The ObjectiveMissedBehavior is what the agent
performs if it misses a location (arrives late). The outcome
of the overall action depends on this behavior, for example
one could code it to abruptly terminate the agent's
endeavor, or to simply skip to the next location and
possibly register the failure using some logging facilities.
When the last location has been reached, the agent stops
after fulfilling all assignments.
Fig. 4. Itinerary behavior sequence diagram
III. CASE STUDY: VLE ASSESSMENT SERVICE
Nowadays e-learning virtual environments emerge as
useful domains available on the Internet. In exploring the
bounds and expressive power of the behavioral patterns
presented in the previous section we chose to approach one
specific aspect related to VLE, namely the student
assessment service (AS). Its design and implementation
were employed within the context of an already existing
VLE system we had previously created - a web-based
virtual university [3].
one particular test to assess all students' knowledge
level. The test results shall be stored by the system.
From a multi-tier architectural perspective the agents and
their behavioral patterns are part of the business process
management layer, while tests and all other persistent
objects are stored into a repository and accessed via a
database management layer. Presentation components
include the GUI and user interaction capabilities agents are
capable of.
Embracing a concise tone, in the following few
paragraphs we shall confine ourselves to presenting the
student oriented part of the AS, leaving aside the teaching
authority's interaction with the system. Consequently, the
agents' social order and their bearing behaviors for the self-
assessment and for the compulsory exam scenarios shall be
portrayed next. We believe that these two examples are
sufficient for the reader to grasp the features offered by the
reusable agent structures.
Throughout the explanations (regarding agents and
behaviors) to follow, one should keep in mind that we are
dealing with the case of an unbalanced distribution of
computational resources due to the locality of the persistent
data stores. Hence, a server side is defined as the host
where the data is, while a client side is any host where an
agent platform has been deployed. Agents are found on
both sides, and there are some that migrate from one to
another. On both sides we have tried to minimize the
number of all-time active agents, managing to reach the
point where there are no such objects on the client side.
This is a fairly good thing since dormant, all-time active
agents on client machines are unacceptable (due to
resources overhead and inadequate schedule). Thus, clients
activate solely the agent platform and even that only when
needed (taking a self-assessment) or required (awaiting for
an exam).
For the mandatory tests (i.e. midterms), agents bearing
the Observer behavior exist on the server side for each
exam planned by the teaching authorities (sure enough, one
professor can schedule several tests that will ultimately be
handled by the same agent, but we shall not approach such
details). The pattern allows one such agent to check
periodically if the time for the evaluation has come. If so,
an Itinerary behavior is enacted (by the same agent or by
the newly created one) and the agent starts its journey to
deliver the tests. This solution is used to release the client
platforms from the overhead of dormant agents (see above).
Once the agent has reached a location in good time it
spawns a new user agent, sets it up with a Task behavior
that takes over the delivery of the test to the user, detaches
and starts its migration towards the next objective. After
the last location is visited, the agent returns at the starting
point where it awaits test solutions. Data concerning
missed/unreachable objectives is gathered as well. The test
related material (questions, choices and answers) is packed
and carried along, ultimately being copied to each user
agent that stays behind to trigger and conduct the
assessment. Upon completion of the last action the user
agent's Task behavior stipulates that it is to send the
student's answers back to the server agent that spawned it,
and then terminate itself. Once the test solutions arrive
from all valid clients and are safely stored in the data
repository, the server agent ceases its existence. Note that
the client agent is the one that interacts with the user
Fig. 5. Itinerary behavior state chart
One of the major problems of creating a virtual
environment involves traditional domain modeling and
implementing such abstractions using the most suitable
technologies [4].
Our approach models a closed
organization (no alien/rogue agents are allowed),
containing benevolent, cooperative agents that are
deployed in a particular configuration and working
together to enact the AS.
In what follows we try to prove that employing
behavioral patterns is a facile, elegant and fast
development agent based solution. Some adjacent design
considerations are included as well (i.e. the data source),
though we shall not insist on presenting them thoroughly
since that would be beyond the scope of our case study and
would impair conciseness. Moreover, we want to
emphasize the fact that the solution is bound to no specific
platform, thus making its porting to another environment a
fairly quick and easy task, which implies modifications
only at the adapter level (or even just adapter usage, if such
adapters already exist for the desired platform). In our case
the JADE [7] agent platform has been chosen as supporting
mobile agent platform for the assessment service.
The AS provides a means for evaluating a learning
entity's (i.e. a student) acquired knowledge and, at the same
time, for providing feed-back on her progress. The
teaching authority is the one responsible for providing the
tests, while the learning entity is the one the tests are
addressed to. A test comprises of a number of questions,
each having a weight within the total score. To enable
automatic evaluation, we use only questions with exact
answers, whether they are single choice, multiple choice,
true/false, or filling in the blank space (numeric and text
are treated separately). A key issue is that the AS is
independent from the counterpart system that provides the
student with the learning material, thus offering a great deal
of flexibility in deploying integrated VLE solutions.
Although tests can have various types, yet they all fall
into one of the two approaches considered below:
•
•
a pull scenario, the self-assessment approach, when the
learning actor is initiating the assessment to poise her
current level of knowledge. The test results are not to
be stored within the system, yet for individual progress
monitoring, some data is stored.
a push scenario, i.e. the compulsory exam, initiated by
the teaching authority (directly or indirectly) enforcing
through the GUI and also the one containing the evaluation
engine, thus performing the assessment on the spot.
Since the self-assessment scenario requires a user to
commence the interaction, on the client side there is no
issue regarding permanent dormant agents. It is only the
server side that has to start a permanently active agent (with
a Server behavior deployed on top) that delegates the initial
requests from clients (to server workers). It does so by
creating a new agent for every requesting entity and
handing the responsibility of continuing to serve the client
over to them. All subsequent communication between the
client and the server is taken care of by the newly spawned
agent from this moment on. By contrast, the learning actor
initially creates an agent with a Client behavior that
initiates the communication with the server, revealing the
self-assessment intention. Once the communication is
established, two virtual channels are created between the
two counterparts, one for commands and one for data. To
support this, a pair of Server and Listener behaviors are
composed within a Parallel behavior on server side, while
a Client and a Task behavior are deployed in the same
concurrent manner on the client side. The client-server pair
deals with data communication, while the listener-task pair
handles the commands (the task is responsible for
triggering the listener with a message containing the
command). Amongst the valid commands we mention the
retrieval of the list of possible valid self-assessment tests,
the retrieval of one test, the sending back of the results, and
the termination of the session. Actions are performed
accordingly employing the necessary data structures. The
evaluation engine is integrated within the client and the
results are computed on the spot.
In what was previously presented no details concerning
how the behaviors are actually deployed were included due
to space limitations. One must imagine that the event
listeners, triggers and other collaborating parts are coded
together throughout each scheme very naturally once the
necessary behaviors to endow the agent with are chosen.
Hence, the problem faced by the developer has shifted
towards the proper orchestration of behavioral patterns,
since it is known that computation is orthogonal to
coordination [13].
IV. CONCLUSION AND FUTURE WORK
In this paper we have tried to present how we envision
ways of unifying mobile agent platforms through the use of
reusable patterns. A new vocabulary of abstractions had to
be enforced, one broad enough to encompass the major
concepts required by all possible underlying environments.
Since one of the initial requirements was that one should be
able to change almost everything about agents
instantaneously, behaviors are employed to enact agent
activity. Our reusable structures are thus behavioral
patterns that can be the building blocks of even more
complex composite objects. We have identified a number
of basic patterns, provided composition mechanisms for
these patterns, and created a model that is extensible
enough for enabling the developer to add new structures
with a minimal amount effort. The framework ensures low
coupling between its high abstractions (the patterns) and
the sustaining agent platforms through the use of adapters.
Since the starting point of our analysis was a VLE, it is
only natural that the strengths and weaknesses of the
framework were eventually tested within this context (by
means of the described assessment service, to be exact). So
far, the model has proved itself to be both fairly flexible
and usable, thus adding a significant building block to the
path towards different design approaches in the world of
mobile agents (though we are aware that using it can be
somewhat cumbersome, at least until some “hands on”
experience is gained).
We are aware that many improvements are possible and,
in the future, we intend to enhance the behavioral patterns
and make them easier to use. Also we shall try to provide
support for more agent platforms as well. Another
important step would be a quality-oriented enhancement of
each pattern, with the most stringent aspect being that of
better delay computation at the time of agent migration (an
issue that appears particularly in the case of the Itinerary
behavior).
V. REFERENCES
[1] M. N. Huhns and M. P. Singh, Agents and multiagent
systems: Themes, approaches, and challenges,
chapter 1 in [2], pp. 1-23, 1998.
[2] M. N. Huhns and M. P. Singh, Readings in Agents,
Morgan Kaufmann, San Francisco, 1998.
[3] T. Marian and B. Dumitriu, “Web based virtual
university,” Technical University of Cluj-Napoca,
2003. Technical Report.
[4] M. Dinsoreanu, C. Godja, C. Anghel, I. Salomie and T.
Coffey, “Mobile Agent Based Solutions for
Knowledge Assessment in eLearning Environments”,
in Proceedings of the 2003 Euromedia Conference,
[5] E. Gamma, R. Helm, R. Johnson and J. Vlissides,
Design Patterns, Addison Wesley Longman
Publishing, 1994.
[6] The Foundation of Intelligent Physical Agents (FIPA),
http://www.fipa.org/
[7] The Java Agent Development Framework (JADE),
http://sharon.cselt.it/projects/jade/
[8] The Agent Development Kit (ADK),
http://www.tryllian.com/technology/product1.html
[9] DeLoach S.A. 2000, “Specifying agent Behavior as
Concurrent Tasks: Defining the Behavior of Social
Agents”, AFIT/EN-TR-00-03. Technical Report.
[10] F . Zambonelli, N.R. Jennings, A. Omicini, M.
Wooldridge, “Agent-Oriented Software Engineering
for Internet Applications”, in Coordination of Internet
Agents: Models, Technologies and Applications,
2000, Springer-Verlag.
[11] F . Zambonelli, N.R. Jennings, M. Wooldridge,
„Organisational abstractions for the Analysis and
Design of Multi-Agent Systems”, in P. Ciancarini and
M. Wooldridge, editors, Agent-Oriented Software
Engineering, Springer-Verlag Lecture Notes in AI
Volume 1957, January 2001.
[12]Weiss G, Multi-Agent Systems, A Modern Approach to
DAI, MIT Press, 1999.
[13] D. Gelernter, and N . Carriero, “Coordination
Languages and their Significance”, Communication of
the ACM, 35(2), 1992, pp. 97-107.
|
1607.05540 | 2 | 1607 | 2016-09-20T14:26:12 | Exploiting Vagueness for Multi-Agent Consensus | [
"cs.MA",
"cs.AI"
] | A framework for consensus modelling is introduced using Kleene's three valued logic as a means to express vagueness in agents' beliefs. Explicitly borderline cases are inherent to propositions involving vague concepts where sentences of a propositional language may be absolutely true, absolutely false or borderline. By exploiting these intermediate truth values, we can allow agents to adopt a more vague interpretation of underlying concepts in order to weaken their beliefs and reduce the levels of inconsistency, so as to achieve consensus. We consider a consensus combination operation which results in agents adopting the borderline truth value as a shared viewpoint if they are in direct conflict. Simulation experiments are presented which show that applying this operator to agents chosen at random (subject to a consistency threshold) from a population, with initially diverse opinions, results in convergence to a smaller set of more precise shared beliefs. Furthermore, if the choice of agents for combination is dependent on the payoff of their beliefs, this acting as a proxy for performance or usefulness, then the system converges to beliefs which, on average, have higher payoff. | cs.MA | cs | Exploiting Vagueness for Multi-Agent Consensus
Michael Crosscombe and Jonathan Lawry
Department of Engineering Mathematics,
University of Bristol,
[email protected] · [email protected]
BS8 1UB, United Kingdom
6
1
0
2
p
e
S
0
2
]
A
M
.
s
c
[
2
v
0
4
5
5
0
.
7
0
6
1
:
v
i
X
r
a
Abstract. A framework for consensus modelling is intro-
duced using Kleene's three valued logic as a means to express
vagueness in agents' beliefs. Explicitly borderline cases are
inherent to propositions involving vague concepts where sen-
tences of a propositional language may be absolutely true,
absolutely false or borderline. By exploiting these intermedi-
ate truth values, we can allow agents to adopt a more vague
interpretation of underlying concepts in order to weaken their
beliefs and reduce the levels of inconsistency, so as to achieve
consensus. We consider a consensus combination operation
which results in agents adopting the borderline truth value as
a shared viewpoint if they are in direct conflict. Simulation
experiments are presented which show that applying this op-
erator to agents chosen at random (subject to a consistency
threshold) from a population, with initially diverse opinions,
results in convergence to a smaller set of more precise shared
beliefs. Furthermore, if the choice of agents for combination is
dependent on the payoff of their beliefs, this acting as a proxy
for performance or usefulness, then the system converges to
beliefs which, on average, have higher payoff.
Keywords: Agent-Based Modelling · Many-Valued Logics ·
Belief Aggregation · Consensus
1 Introduction
Reaching a consensus by agreeing a shared viewpoint or posi-
tion is a fundamental part of many multi-agent decision mak-
ing and negotiation scenarios. In this paper we argue that
by exploiting vagueness in the form of explicitly borderline
cases we can define an operator for belief combination which
not only allows a population of agents to reach consensus but
also results in them adopting, on average, a more useful set
of beliefs. The basic intuition underlying this operator is that
conflicting agents can agree to allocate borderline truth values
to propositions about which they hold inconsistent beliefs. For
example, two individuals, one of which believes that 'Cameron
is an effective prime minister' whilst the other believes that
'Cameron is ineffective', may agree, in some circumstances,
to adopt the shared view that 'Cameron is borderline effec-
tive/ineffective'.
Of course, beliefs about the world do not exist in isolation
but inform and influence our decisions and actions. From this
perspective, some sets of beliefs are more positive or useful
than others, resulting in better long term performance, per-
haps by making the individuals concerned richer, happier or
just better able to survive. More generally, in a multi-agent
context, different beliefs result in different actions, collecting
different payoffs. In this paper we present simulation studies
which show that implementing our proposed operator across
a population of agents, initially holding diverse beliefs, re-
sults in convergence to a smaller subset of more precise shared
opinions. Furthermore, under the assumption that better per-
forming agents, i.e. those with higher payoff, are more likely
to interact to reach consensus, we show that the range of be-
liefs obtained at steady state are on average better, i.e have
higher payoff, than the agents' initial beliefs. The formalism
adopted here is that of Kleene's three valued logic and the op-
erator investigated has been proposed for single propositions
in [10] and extended to multi-propositional languages in [6].
An outline of the paper is as follows: Section 2 gives a brief
overview of consensus modelling. Section 3 introduces Kleene
logic and the three valued consensus combination operator.
Section 4 describes simulation experiments in which agents
are selected at random to form a consensus provided that they
are sufficiently consistent with one another. In section 5 we
introduce a payoff function for beliefs, so that the payoff of a
particular set of beliefs acts as a proxy for the performance of
an agent holding those beliefs. We then adapt the experiments
described in section 4 so that the probability of an agent being
selected for consensus is proportional to their payoff. Finally,
in section 6 we give some discussions and conclusions.
2 Background and Related Work
A number of models for consensus have been proposed in
the literature which have influenced the development of the
framework described in this paper. [3] introduced a model for
reaching a consensus involving a weighted, global updating of
beliefs, iterating until an agreement is reached. In DeGroot's
model, agents assign a weight distribution to the population
before forming a new opinion. By applying their assigned
weights to the other agents' beliefs, an agent can control the
influence that others have on their own beliefs.
As an alternative to DeGroot's model, the Bounded Con-
fidence (BC) model introduced in [5] provides agents with a
confidence measure. An agent quantifies their level of confi-
dence in their own opinions and are then able to limit their
interactions to those agents who possess similar beliefs if
they are highly confident (small bounds), or extend the range
of possible interactions if the agents possess low confidence
(large bounds). In this model agents do not a priori assign
weights to the beliefs of others, but instead determine such
weightings based on similarity and on their own confidence
levels. This is similar in essence to the inconsistency threshold
that we introduce in section 3, but applied on an individual
basis.
The Relative Agreement (RA) model [2] then extends the
Bounded Confidence model to allow agents to assign weights
to the beliefs of others by quantifying the extent of the over-
lap of their respective confidence bounds. By having agents
declare a confidence interval for their beliefs, the model then
restricts interactions to those pairs of agents with overlap-
ping intervals. Consequently, agents are only required to as-
sess their own beliefs and are not required to make explicit
judgements about those of other agents. [2] also moved to a
model of pair-wise interactions to better capture social in-
teractions of individuals, the latter being a setting in which
group-wide updates to beliefs are unintuitive in that they do
not reflect typical social behaviour.
A fundamental difference between our approach and the
above models is that we use Kleene's three valued logic to
represent beliefs in a propositional logic setting, rather than
identify opinions with real values or intervals. [10] have shown
that through use of a three-state model for networked consen-
sus of complete graphs, nodes converge to a consensus much
faster and with greater accuracy when compared to a restric-
tive binary model. In the sequel we extend this approach to a
more general setting involving larger languages and incorpo-
rating a measure of payoff for beliefs.
3 A Three Valued Consensus Model
In this section we introduce Kleene's three valued logic [4]
as a model of explicitly borderline cases resulting from the
inherent vagueness of propositions. We adopt a propositional
logic setting as follows: Let L be a finite language of propo-
sitional logic with connectives ∧, ∨ and ¬, and propositional
variables P = {p1, . . . , pn}. Also, let SL denote the sentences
of L generated by recursively applying the connectives to the
propositional variables in the usual manner. A Kleene valua-
tion then allocates truth values 0 (false), 1
2 (borderline) and
1 (true) to the sentences of L as follows:
Definition 1. Kleene Valuations
such that ∀θ, ϕ ∈ SL the following hold:
• v(¬θ) = 1 − v(θ)
• v(θ ∧ ϕ) = min(v(θ), v(ϕ))
• v(θ ∨ ϕ) = max(v(θ), v(ϕ))
A Kleene valuation v on L is a function v : SL → {0, 1
2 , 1}
The truth table for Kleene valuations are shown in table 1.
¬
1
1
2
0
0
1
2
1
∧
1
1
2
0
1
1
1
2
0
1
2
1
2
1
2
0
0
0
0
0
∨
1
1
2
0
1
1
1
1
1
2
1
1
2
1
2
0
1
1
2
0
Table 1. Kleene truth tables.
v(pi) = 1} and N = {pi ∈ P : v(pi) = 0}. Notice that P ∩
N = ∅ and that (P ∪ N )c corresponds to the set of borderline
propositional variables.
Kleene valuations have been proposed as a suitable formal-
ism in which to capture explicitly borderline cases as resulting
from inherent flexibility in the definition of vague concepts in
natural language [8, 7]. For example, consider the proposition
'Ethel is short'. For the concept short, we might identify a
lower height threshold h below which any height is classed as
being absolutely short, and similarly there may be an upper
threshold h above which any height is absolutely not short. If
Ethel's height lay between h and h then this would result in
a borderline truth value for the statement 'Ethel is short'.
It is important to note that the middle truth value 1
2 is not
intended to represent epistemic uncertainty, but rather explic-
itly borderline cases resulting from the inherent vagueness of
natural language propositions. Hence, if we say that the state-
ment 'Ethel is short' is borderline true/false we are not saying
that the truth or falsity of this proposition is unknown. In-
stead we are indicating that Ethel's height is a borderline case
of the predicate short. In order to emphasise the difference be-
tween the epistemic and the borderline interpretation of three
valued logic it is helpful to think in terms of conditioning. For
instance, if we learn that it is unknown whether or not Ethel
is short, then this provides us with no new information about
her height. In contrast, learning that Ethel is borderline short
does provide us with new information about Ethel's height,
namely that it lies on the borderline between short and not
short. A more comprehensive discussion of these issues can
be found in [1]. A consequence of using this interpretation of
the middle truth value is that in the current paper we only
model consensus for sets of propositions which admit bor-
derline cases. In other words, our approach can be used for
propositions such as 'Ethel is short' but not, for example, for
the proposition 'Ethel is strictly less that 1.4 metres tall'.
The following three valued consensus operator was de-
scribed in detail in [6]:
Definition 2. Consensus Operator
Let v1 and v2 be Kleene valuations on L with associated
orthopairs (P1, N1) and (P2, N2). Then the consensus v1 (cid:12) v2
is the Kleene valuation with the orthopair
((P1−N2) ∪ (P2−N1), (N1−P2) ∪ (N2−P1))
The corresponding truth table for this operator is shown
in table 2. From this we can see that the operator preserves
the non-borderline truth values 0 or 1 except in the case of
a direct conflict i.e. when one agent has truth value 1 and
the other 0. In this case both agents adopt the middle truth
value 1
as an operator which initially weakens both opinions so as to
remove direct inconsistencies, before then combining them.
2 . Alternatively, from definition 2 we can think of (cid:12)
(cid:12)
1
1
2
0
1
1
1
1
2
1
2
1
1
2
0
0
1
2
0
0
I
1
1
2
0
1
0
0
1
1
2
0
0
0
0
1
0
0
It is sometimes convenient to represent a Kleene valuation
v by its associated orthopair (P, N ) [6], where P = {pi ∈ P :
Table 2. Truth table for the
consensus operator.
Table
truth table.
3.
Inconsistency
We now introduce two measures that will be used through-
out the subsequent simulation experiments.
Definition 3. A Measure of Vagueness
Let v be a Kleene valuation on L with orthopair (P, N ) and
n propositional variables. Then we measure the vagueness of v
by the proportion of propositional variables which it classifies
as being borderline. That is:
V (v) =
(P ∪ N )c
n
Definition 4. Inconsistency Measure
Let v1 and v2 be Kleene valuations on L with corresponding
orthopairs (P1, N1) and (P2, N2). Then we define the incon-
sistency measure of v1 and v2 to be the proportion of propo-
sitional variables which are in direct conflict between the two
valuations i.e. v1(pi) (cid:54)= 1
2 and v1(pi) = 1 − v2(pi).
That is:
2 , v2(pi) (cid:54)= 1
I(v1, v2) =
(P1 ∩ N2) + (P2 ∩ N1)
n
Table 3 shows the inconsistency truth table of two valua-
tions for a propositional variable, highlighting the cases where
two valuations are inconsistent, and consistent otherwise. We
can see that there is a probability of 2
9 that two valuations
will be inconsistent for each propositional variable in the lan-
guage. In the sequel we will propose a threshold γ ∈ [0, 1] on
inconsistency so that valuations v1 and v2 can be combined
only if I(v1, v2) ≤ γ.
4 Simulation Experiments based on
Random Selection of Agents
We introduce simulation experiments in order to investigate
the convergence properties of the three valued logic operator
when implemented across a multi-agent system. The experi-
mental set up is loosely based on those proposed in [2] and
[9], although our representation of opinions is quite different
with beliefs taking the form of Kleene valuations on L, rather
than vectors of bounded real numbers.
We will consider two distinct initialisations of the beliefs of
a population of agents. The random three valued initialisation
allocates the truth values 0, 1
2 and 1 to each agent and each
propositional variable at random i.e. with probability 1
3 for
each truth value. In contrast, the random Boolean initialisa-
tion only allocates the binary truth values 0 and 1, each with
a probability of 1
2 . This latter initialisation will be required
in section 5 in order to directly compare the proposed three
valued combination operator with a similar two valued op-
erator. In this section we will use the random three valued
initialisation in order to investigate the extent to which the
three valued operator results in convergence to a shared set
of opinions across the population of agents.
We set a fixed maximum number of 50, 000 iterations1. At
each time step a pair of agents are selected at random from
the population. An inconsistency threshold value γ ∈ [0, 1] is
set, so that for any pair of agents with respective valuations v1
and v2, if I(v1, v2) ≤ γ then both agents replace their beliefs
1 In preliminary experiments we found that 50, 000 was an upper
bound on the number of iterations required for the system to
reach steady state across a range of parameter settings.
with the consensus valuation v1 (cid:12) v2, while if I(v1, v2) > γ
then no combination is performed and both agents retain their
original beliefs. For γ = 1 we obtain what is equivalent to the
totally connected graph model described in [10], in which any
pair of agents can combine their beliefs, whilst taking γ = 0
corresponds to the most conservative scenario in which only
absolutely consistent beliefs can be combined. The parameters
for the simulation experiments are then as follows:
• Population size: 100
• Language size i.e P = n: 5, 10, 50, 100
• Initial beliefs: Random three valued.
• Inconsistency threshold: γ ∈ [0, 1].
Figures 1 and 2 show the results for the experiments after
50, 000 iterations. In each case the plots show mean values
with error bars representing standard deviation across 100
independent runs of the simulation. Figure 1 shows the aver-
age vagueness determined by taking the mean value of V (v)
(definition 3) across the population. Note that for a random
three valued initialisation of beliefs we expect a mean vague-
ness value of 1
3 at the start of the simulation. As the threshold
γ increases then the average vagueness decreases to zero, so
that for γ ≥ 0.3 we are left with almost entirely crisp (i.e.
Boolean) opinions. In general the more conservative the com-
bination rules (i.e. requiring higher levels of consistency) then
the more it is that vague beliefs are maintained in the popu-
lation. Figure 2 shows the number of distinct valuations (i.e.
different opinions) remaining in the population after 50, 000
iterations. Again this decreases with γ and for γ > 0.4 agents
have on average converged to a single shared belief. This is
consistent with the analytical results presented in [10] for the
single propositional, γ = 1 case.
5 Simulation Experiments Incorporating a
Payoff Model
In this section we extend the simulation framework described
in section 4 to allow for different payoffs for different beliefs.
As outlined in section 1, payoff is introduced as a proxy for
performance, and is motivated by the intuition that different
beliefs result in different actions which then, over time, lead
to different levels of performance. Here we adopt an abstract
simplification of this process in which each Kleene valuation is
allocated a real valued payoff. Then, instead of being selected
at random for combination, an agent is picked from the pop-
ulation according to a probability which is proportionate to
the payoff value of their beliefs. The idea, then, is that agents
with better or more useful opinions will be more successful
and furthermore, it will be these successful agents who will
be most likely to need to reach a consensus between them.
Here the underlying intuition is that, in real systems it is the
most successful agents, with the highest payoff values, who are
most likely to find themselves in conflict with one another, and
who will most benefit from reaching an agreement. We adopt
a simple summative payoff model in which each propositional
variable pi is allocated a value in the range [−1, 1], denoted
f (pi), and the payoff for a valuation with orthopair (P, N ) is
then calculated as follows:
(cid:88)
pi∈P
f (pi) − (cid:88)
pi∈N
f (P, N ) =
f (pi)
Figure 1. Average vagueness after 50, 000 for varying inconsistency thresholds γ and language sizes.
Figure 2. Number of distinct valuations after 50, 000 iterations for varying inconsistency thresholds γ and language sizes.
Another perspective on this type of payoff function is as fol-
lows: For each propositional variable pi, a truth value of 1
results in a payoff f (pi) (which can be either positive or neg-
ative), a truth value of 0 results in the opposite signed payoff
−f (pi), and a borderline truth value 1
2 results in a neutral
payoff of 0. The payoff value for a Kleene valuation v is then
simply taken to be the sum of the payoffs for each proposi-
tional variable under the truth values allocated by v.
Based on payoff values we define a probability distribution
over the agents in the population according to which the prob-
ability that an agent with beliefs (P, N ) is selected for possible
consensus combination is proportional to f (P, N )+n. At each
iteration a pair of agents are selected at random according to
this distribution. For each such pair the inconsistency mea-
sure (definition 4) is evaluated and either both the valuations
are replaced with the consensus valuation, or both are left
unchanged, depending on the threshold γ as in section 4. The
parameters for the simulation experiments are as follows:
• Population size: 100
• Language size: 5
• Initial beliefs: Random Boolean.
• Inconsistency threshold: γ ∈ [0, 1]
binary operator
0
0
1
0
2 , 1 : 1
2
0 : 1
0 : 1
1
2 , 1 : 1
1
2
Table 4. The truth table for the stochastic Boolean consensus
operator
Notice that here we are initialising the beliefs as random
Boolean valuations (see section 4)2. This allows us to make a
direct comparison between the performance of the three val-
ued combination operator and a similar two valued operator.
For the latter we assume that only binary truth values are
available to represent an agent's beliefs. In this context, in
order for two agents with conflicting truth values for pi (i.e.
one 0 and the other 1) to reach consensus, we propose that
they simply agree to pick one of the truth values at random
e.g. by tossing a fair coin. Table 4 gives the truth table for
the operator in which directly conflicting truth values leads
to a stochastic outcome.
The focus on simulations with 5 propositional variables is
intended to increase the number of opinions relative to the
size of the population, in order to achieve a good distribu-
2 As a result of this Boolean initialisation, a language size of 5 now
produces a total of 25 (32) possible valuations, as opposed to 35
(243) possible valuations.
0.00.10.20.30.40.5Threshold γ0.050.000.050.100.150.200.250.300.350.40Average Vagueness510501000.00.10.20.30.40.5Threshold γ020406080100120Distinct Valuations51050100 Figure 3. Average payoff after 50, 000 iterations for varying inconsistency thresholds γ, shown as a percentage of the maximal payoff.
Figure 4. Number of distinct valuations after 50, 000 iterations for varying inconsistency thresholds γ.
tion of valuations. For example, a language size of 5 allows
for 32 possible Boolean valuations. With a population of 100
agents, it is therefore very likely that each opinion will occur
at least once. In comparison, a language size of 10 produces
1, 024 possible Boolean valuations which severely decreases
the probability of an opinion being present in a population of
the same size.
Figures 3, 4 and 5 show the results for simulation experi-
ments with agent selection based on payoff. The results shown
are mean values with error bars taken over 100 independent
runs of the simulation. Figure 3 shows the average popula-
tion payoff after 50, 000 iterations given as a percentage of the
maximal possible payoff value i.e. the payoff for the valuation
(P, N ) where P = {pi : f (pi) > 0} and N = {pi : f (pi) < 0}.
For both the binary and the three valued operators we show
results for simulations in which agents are selected according
to payoff (three-valued, Boolean) and at random as in section
4 (three-valued random, Boolean random). We see that for all
values of γ, the three valued operator with payoff based selec-
tion outperforms all of the other approaches. For the former
we can also see that average payoff increases with γ. In con-
trast, for the other approaches, including the payoff operator
with payoff based selection, the mean of the average popula-
tion payoff remains close to 0 after 50, 000 iterations. Figure
4 shows the mean number of distinct valuations across the
population of agents after 50, 000 iterations. All four versions
of the operators converge on a small set of shared beliefs for
sufficiently large γ. For γ ≥ 0.4 the mean number of distinct
valuations is less than 5 while for γ ≥ 0.8 it is 1. Figure 5
shows a trajectory of how the number of distinct valuations
varies with each iteration when γ = 0.7. We can see that
both the three-valued models converge quickly (in just over
2000) iterations while the Boolean models require consider-
ably longer to converge (over 20, 000 iterations).
6 Conclusions
In this paper we have explored the use of Kleene's three valued
logic as a framework in which to model multi-agent consensus
formation. We have proposed a three valued combination op-
erator, the intuition behind which is that conflicting binary
truth values are replaced with a borderline (middle) truth
value. A number of simulation experiments have been pre-
sented employing this operator. These can be divided into
two main categories. For the first type of experiments, agents
are selected at random from the population and form a con-
sensus valuation providing that the level of inconsistency of
their respective opinions is below a threshold parameter γ.
0.00.20.40.60.81.0Threshold γ604020020406080100Average PayoffThree-ValuedThree-Valued randomBooleanBoolean random0.00.20.40.60.81.0Threshold γ05101520253035Distinct ValuationsThree-ValuedThree-Valued randomBooleanBoolean random Figure 5. Trajectory showing the number of distinct valuations plotted against iterations for γ = 0.7.
Otherwise they do not form a consensus and instead retain
their current opinions. For these experiments we found that
there is convergence to a smaller subset of shared opinions
across the population. For higher γ values there is conver-
gence on average to a single shared opinion and furthermore
this opinion is crisp i.e. it admits no borderlines. For inter-
mediate values of γ the system convergences to a small set of
opinions which to some extent remain vague.
In the second type of experiments a payoff function over
beliefs is introduced, and agents are selected for possible com-
bination with probability proportional to the payoff value of
their current beliefs. Here we compare the three value opera-
tor with a similar stochastic Boolean operator. We find that
the three valued operator with payoff based agent selection
results in convergence to a smaller shared set of beliefs with
significantly higher average payoff than that of the initial pop-
ulation. The Boolean operator does not perform well in this
context and does not result in a significant increase in av-
erage payoff, which instead remains close to 0 after 50, 000
iterations.
The results of the payoff based experiments show how a
three valued model for consensus provides a number of im-
provements over a traditional Boolean model. Firstly, we have
shown that the introduction of Kleene valuations to capture
the inherent vagueness of propositions does not, in the long
run, lead to the mass adoption of borderline truth values as
a result of conflict occurring in the population. Instead, we
have seen how vagueness is reduced at lower γ values, and at
higher γ values the population converges towards completely
crisp opinions on average, admitting no borderline cases. In
addition to this, we can see that the introduction of a payoff
based model drives consensus towards those valuations which
result in higher payoff on average. By selecting pairs of agents
based on their perceived success, we can achieve an increase
to overall payoff in a small number of iterations, compared
to no significant increase in payoff for the Boolean model.
Therefore, we have shown that the three valued approach in-
corporating a payoff model can drive convergence across the
population towards more successful opinions.
We suggest that the experiments presented in this paper
show the potential of using three valued logic in consen-
sus modelling. There is also significant scope to extend the
research presented in several new directions. For example,
the above studies concern consensus defined at the level of
propositional variables. However, in many cases agents will
be most concerned to reach agreement about a relevant set of
compound statements. For example, they may need to reach
agreement about a particular set of conditional statements,
or equivalences. Hence, an important question is that of how
best to extend our proposed consensus model so as to be ap-
plicable to compound logical expressions. Another significant
question concerns uncertainty. Suppose that in addition to
vagueness agents also quantify their uncertainty about be-
liefs. [6] propose an extension of the three valued framework
in which agents' beliefs are represented by a probability distri-
bution over Kleene valuations. Ongoing research concerns the
design of simulation studies in which to evaluate the conver-
gence and payoff based performance of this extended model.
Finally, it would be interesting to consider extensions to the
operator which allows for consensus between groups rather
than just pairs of agents.
Acknowledgements
This research is partially funded by an EPSRC PhD stu-
dentship as part of a doctoral training partnership (grant
number EP/L504919/1).
All underlying data is included in full within this paper.
REFERENCES
[1] D. Ciucci, D. Dubois, and J. Lawry. Borderline vs. unknown:
comparing three-valued representations of imperfect informa-
tion. In International Journal of Approximate Reasoning. 55.
pp 1866-1889. Elsevier, 2014.
[2] G. Deffuant, F. Amblard, G. Weisbuch, and T. Faure. How
can extremism prevail? a study based on the relative agree-
ment interaction model. In Journal of Artificial Societies and
Social Simulation. 5(4). JASSS, 2002.
[3] M. H. DeGroot. Reaching a consensus.
In Journal of the
American Statistical Association. 69(345). pp 118-121. JS-
TOR, 1974.
[4] S. C. Kleene.
Introduction to Metamathematics. North-
Holland, The Netherlands, 1st edition, 1952.
[5] U. Krause. A discrete nonlinear and non-autonomous model
In Communications in Difference
of consensus formation.
01000020000300004000050000Iterations05101520253035Distinct Valuations3-Valued3-Valued randomBooleanBoolean random Equations: Proceedings of the Fourth International Con-
ference on Difference Equations, pp 227-237. Gordon and
Breach, 1998.
[6] J. Lawry and D. Dubois. A bipolar framework for combining
beliefs about vague propositions. In Proceedings of the Thir-
teenth International Conference on the Principles of Knowl-
edge Representation and Reasoning. pp 530-540. AAAI, 2012.
[7] J. Lawry and I. Gonzlez-Rodrguez. A bipolar model of as-
sertability and belief. In International Journal of Approxi-
mate Reasoning. 52(1). pp 76-91. Elsevier, 2011.
[8] J. Lawry and Y. Tang. On truth-gaps, bipolar belief and the
assertability of vague propositions. In Artificial Intelligence.
191. pp 20-41. Elsevier, 2012.
[9] M. Meadows and D. Cliff. Reexamining the relative agreement
model of opinion dynamics. In Journal of Artificial Societies
and Social Simulation. 15(4). JASSS, 2012.
[10] E. Perron, D. Vasudevan, and M. Vojnovic. Using three states
for binary consensus on complete graphs. In Proceedings of
IEEE Infocom. pp 2527-2535. IEEE Communications Society,
2009.
|
1312.5941 | 1 | 1312 | 2013-12-20T13:51:11 | Developing a model of evacuation after an earthquake in Lebanon | [
"cs.MA"
] | This article describes the development of an agent-based model (AMEL, Agent-based Model for Earthquake evacuation in Lebanon) that aims at simulating the movement of pedestrians shortly after an earthquake. The GAMA platform was chosen to implement the model. AMEL is applied to a real case study, a district of the city of Beirut, Lebanon, which potentially could be stricken by a M7 earthquake. The objective of the model is to reproduce real life mobility behaviours that have been gathered through a survey in Beirut and to test different future scenarios, which may help the local authorities to target information campaigns. | cs.MA | cs | Developing a model of evacuation after an earthquake in Lebanon
Hong Van Truong, Elise Beck, Julie Dugdale, and Carole Adam
Abstract— This article describes the development of an
agent-based model (AMEL, Agent-based Model for
Earthquake evacuation
in Lebanon) that aims at
simulating the movement of pedestrians shortly after an
earthquake. The GAMA platform was chosen to
implement the model. AMEL is applied to a real case
study, a district of the city of Beirut, Lebanon, which
potentially could be stricken by a M7 earthquake. The
objective of the model is to reproduce real life mobility
behaviours that have been gathered through a survey in
Beirut and to test different future scenarios, which may
help
the
local authorities
to
target
information
campaigns.
I. INTRODUCTION
Every year earthquakes cause many victims. Although the
survival of people is largely related to the resilience of
buildings, the way that people behave after an earthquake
also influences the number of fatalities (Rojo, Beck, Lutoff
and Schoeneich, 2011), especially for low magnitude
earthquakes. The rareness of strong earthquakes and the
impossibility of forecasting them prevent us from assessing
the effect that behaviours may have on survival rates. In this
context, computer simulation is an appropriate and powerful
tool that helps us to assess different scenarios, thus helping
local authorities to develop their risk management policies
and information campaigns.
The objective of the AMEL model (Agent-based Model
for Earthquake evacuation in Lebanon) is to simulate
pedestrians’ mobility shortly after an earthquake. Although
many victims in buildings are generally killed during the
main shock, the mobility of any survivors can greatly affect
their chances of surviving any aftershocks. We want to
understand whether a better knowledge of the safety
procedures and the location of safe places in a city could
decrease people’s vulnerability. Firstly of all, the model aims
at reproducing the behaviours adopted by individuals.
Secondly, the simulation tests some optimistic or pessimistic
fictive scenarios, in order to evaluate the impact of individual
behaviours on the number of fatalities or on the number of
people in exposed and dangerous situations. This work is
applied to the real case of a district of Beirut, Lebanon. This
article presents a work in progress and focuses on explaining
the model and how it was developed. Section 2 presents the
H. V. Truong is with the Institut de la Francophonie pour l’Informatique,
Hanoi, Vietnam. (e-mail: [email protected]).
E. Beck is with the University of Grenoble-Alpes. France (e-mail:
[email protected])
J. Dugdale is with the University of Grenoble-Alpes, France and
affiliated
to
the University
of Agder, Norway.
(e-mail:
[email protected])
C. Adam is with the University of Grenoble-Alpes, France (e-mail:
[email protected])
overall approach and gives the state of the art, whilst section
3 describes the methodology used for developing AMEL.
The model is described in section 4 and our first attempts at
validation are detailed in section 5. Section 6 concludes the
paper and presents future work.
II. APPROACH AND STATE OF THE ART
Agent-based social simulation (ABSS) is a branch of
distributed artificial intelligence and multi-agent systems
whose focus is on understanding, modelling and simulating
social behaviours. Rather than purely focusing on cognition,
this approach recognises the social complexity of a society
and looks at how social phenomena, such as cooperation,
emerge through human behaviours.
The use of agents in modelling human systems has several
advantages over other approaches (Bonabeau, 2002). Firstly,
agent based systems are able
to capture emergent
phenomena that are so representative of complex adaptive
systems. Secondly, they provide a natural description of a
system, which as Bonabeau notes makes the agent based
approach much closer to reality. Finally, they are flexible,
allowing us to study social systems at different levels of
abstraction by varying the complexity of our agents or by
aggregating agents into subgroups.
Hundreds of agent-based social simulators have been
developed. These have been used for predicting future
situations, as training tools, for developing and formalizing
theories, or for testing new technological designs or new
ways of organizing work (Gilbert and Troitzsch, 1999). Here
we focus on the use of a simulator for understanding and
predicting human behaviour when faced with an earthquake,
depending on their level of vulnerability.
Using agent based simulation as a way to assess
emergency situations has become increasingly popular in
recent years. The RoboCup Rescue Agents Simulation
project has attracted many researchers who are interested in
using the specially developed platform to develop intelligent
agents that undertake the role of Police Forces, Fire Brigades,
and Ambulance Teams in a disaster response scenario
(Skinner and Ramchurn, 2010). Likewise, other works, such
as the REScUE research project (Hawe, Coates, Wilson and
Crouch, 2011) and EQ-Rescue (Fiedrich, 2006) have been
developed to evaluate different rescue plans, or to optimize
resource allocation during an emergency. These works, like
many of this kind of simulation, focus specifically on the
response and rescue aspects, with the main accent being on
modelling rescue agents, on coordination between different
teams, or on the allocation of resources. Our work differs in
this respect since we do not look at response activities, but
just the behavior of the potential victims.
Pelechano and Badler present a model of building
evacuation
to
study
the
influence of
inter-agent
communication and the effect of training (Pelechano and
Badler, 2006). The movements of agents in rooms are
modelled by acceleration equations. The authors show that
the evacuation may be made more efficient by introducing
communication between agents to share their knowledge of
blocked routes. Similarly the evacuation is also more
efficient if there are a small number of leaders (trained staff)
in the crowd that the other dependent agents group around
and follow. Although we have designed AMEL to take into
account leader/follower behaviours, our work is still in
progress and therefore we have not yet experimented with
this aspect. Concerning inter-agent communication, we hope
to build on
this work by considering neighbourhood
evacuation.
Nguyen et al. present a hybrid model of the pedestrian
flow on road networks, applied to the evacuation of Nhatrang
(Vietnam) in case of a tsunami (Nguyen, Zucker, Nguyen,
2011). This model combines micro (agent-based, fine-grained
but slow) and macro (equation-based, fast) models of
pedestrians’ movement, in order to improve the efficiency
(speed) of simulations involving a large number of agents.
This hybrid model was shown to be more efficient than a
micro-model, and of better quality than a macro-model.
III. METHODOLOGY
A controversy in ABSS concerns the fundamental
question of how to develop useful models of real-life social
situations. Broadly there are two schools of thought. One
follows the KISS (Keep It Simple, Stupid) philosophy where
the aim is to develop simplistic models and where much of
the real world detail has been abstracted away. Although
there are obvious benefits, e.g. in terms of ease of
constructing the models, the approach has been widely
criticised. The arguments can be reduced to the idea that
models that are too simple only address simplistic problems
that are not representative of the real world. The other
extreme is a KIDS (Keep It Descriptive Stupid) approach
(Edmonds and Moss, 2004) where the model is constructed
by taking into account the widest possible range of evidence,
including anecdotal accounts and expert opinion. Although
we obtain a much truer representation of reality it may be
very difficult to obtain the data to build the model,
implementation is more complicated, and validation of the
model and simulator are problematic. The approach adopted
in this work falls in between these two extremes and follows
that proposed by Rosaria Conte: “Keep it Simple as Suitable”
(Conte, 2000). Here models are abstract enough to achieve an
adequate level of generality, but no less complex than what is
required by the purpose of the simulation. Whilst the original
KISS, KIDS, and reformulated KISS approaches provide
advice on designing models it is very general and somewhat
vague, and the approaches lack a complete modelling
method. In response, several methods and modelling
techniques have been proposed,
for example GAIA
(Wooldridge, Jennings and Kinny, 2000), VOWELS
(Demazeau, 1995), CoMoMAS (Glaser, 1996), MMTS
(Kinny, Geogeff and Rao, 1996), and Unified Approach
(Sabas, Delisle and Badr, 1996). These all provide the
standard framework for modelling the agent dimension, some
taking into account the deliberative behaviour of agents.
However, they are largely intended for developing general
multi-agent systems and are not specifically focused on
modelling the social elements that are required in ABSS.
Furthermore, they fail to provide a structure for analysing
human agents in the design phase and for validating the
model with respect to the observed human behaviour. Some
methods have been developed that focus particular on agent
based simulation. Notably the ODD protocol by Grimm and
his colleagues aimed to standardising descriptions of models
to aid understanding and ensure reproducibility (Grimm,
Berger, Bastiansen, Eliassen, Ginot, Giske, Goss-Custard,
Grand, Heinz, Huse, Huth, Jepsen, Jørgensen, Mooij, Müller,
Pe’er, Piou, Railsback, Robbins, Robbins, Rossmanith,
Rüger, Strand, Souissi, Stillman, Vabø, Visser and DeAngelis
DL. 2006; Grimm, Berger, DeAngelis, Polhill, Giske and
Railsback, 2010). Although very useful, the protocol only
concerns model description and does not suggest how the
model itself may be developed.
The methodology adopted in this work is shown in figure
1 and is adapted from that of Edmonds (Edmonds, 2000).
However it has several important differences. Firstly, it
focuses on analysing human behaviour in the real world
situation through the use of extensive field studies. This
provides a solid corpus of empirical data through video
recordings and observations, etc. Secondly, it puts validation
at the heart of the process ensuring that the results of the
simulator can be more readily trusted (Dugdale, Bellamine-
Ben Saoud, Pavard, Pallamin, 2001). Finally, it reinforces
iteration; this allows us to revisit previous steps such as
undertaking additional targeted field studies and refining the
model and code as necessary.
Field
Model
Studies
design
and
data
collection
Validation
Simulation
(experimen
tation
and
analysis)
Develop-‐
ment
Figure 1. Methodology adopted for AMEL
Following figure 1, the first step covers performing
detailed field studies of the real situation in order to assess
the human behaviours and their underlying motivations. The
second step, Model design, involves developing the formal
model, for example by using UML, of what has been
obtained through the analysis of field studies data. Validation
has been put at the heart of the methodology and relies
heavily on the data obtained via field studies in the analysis
phase. The final step, Simulation, covers experimentation.
The bi-directional arrows ensure that iteration plays a major
role. The above methodology was first described in 1999
(Dugdale, Pavard and Soubie, 1999; Dugdale, Pavard and
Soubie 2000) and has been used for designing and
developing agent-based simulators in several works over the
years, for example in (Bellamine-Ben Saoud, Ben Mena,
Dugdale, Pavard and Ahmed, 2006; Kashif, Binh, Dugdale
and Ploix, 2001; Kasif, Ploix, Dugdale and Binh, 2013).
IV. FIELD STUDIES AND DATA COLLECTION
The studied district is characterized by a high density of
buildings that have a differing number of storeys, date and
construction material. Several types of data were collected
using a multidisciplinary approach, involving the disciplines
of geography, psychology and geotechnics. Concerning the
geographic aspect, the buildings, streets, and green spaces
correspond to spatial data represented by polygons and
polylines. These objects correspond to geographic layers and
were processed through geographic information systems.
The data first comes from a national reference database
(such as the French Geographic Institute database) and was
corrected and updated using high spatial resolution satellite
imaging. Concerning demographic data, the last Lebanon
population census was carried out in 1932. Therefore, the
population was estimated at 3.8 persons per apartment.
Because of the high heterogeneity of the building types,
which leads to high variations of physical vulnerability to
earthquakes, each of the 357 buildings was documented
through a field survey that aimed at characterizing its
vulnerability. For example,
the survey form
included
information such as number of storeys, year of construction,
construction material. Some geotechnical data was collected
with material samples for different types of construction.
This data was then computed with the FEMA’s method
called HAZUS (Kircher, Whitman and Holmes, 2006) in
order to estimate the damage rate for each building in case of
an M7 earthquake on the Yammouneh fault.
Finally, the model requires social data concerning the
behaviours of individuals. This data was collected through a
field sociological survey that interviewed 88 persons of the
studied district (Beck, Colbeau-Justin, Cartier and Saikali,
2011). The questionnaire focused on several subjects
including knowledge about earthquakes, risk perception,
earthquake experience and associated behaviours, etc. The
statistical analysis of the survey allowed us to define several
categories of behaviours. Some extra behaviours, that were
not observed in the survey but have been reported in similar
cases, were also considered and implemented into the model
(leader-follower behaviour, for example).
V. MODEL DESIGN
As a first step, the purpose of the simulation is to
reproduce, based on the survey information, the mobility of
people and their behaviours following an earthquake in
Lebanon. In an earthquake, natural and artificial obstacles
(e.g. escarpments and stairways), and obstacles induced by
the earthquake itself, can injure people and constrain their
movement towards safe zones. The simulator will provide
information concerning the physical and human damage
incurred (the number of damaged buildings, blocked streets,
number of victims, number of people in dangerous zones,
evacuation time, etc.).
Once we have reproduced mobility behaviours that align to
the survey information, as a second step, we experiment with
difference scenarios by changing the ways that people behave
and analysing the results. We want to see what happens
(damages, number of fatalities and exposure to danger) if
people act in different ways. These results can serve to
inform people about how to improve their behaviours, for
example by forming the basis of an information campaign.
A. Entities and their attributes
The model consists of six entities: Human, Street,
Building, Green space, Quake, and Obstacle (figure 2). Their
attributes are divided
into
three
types: position and
visualization (location, colour, etc.), entity characteristics
(e.g. Human entity has attributes about age, sex, profession,
etc., and the Building entity has attributes about height,
capacity, etc.). Finally there are special attributes concerning
the
entity’s
behaviours
(e.g.
target
to
reach,
street_knowledge, etc. of the Human entity).
Figure 2. Class diagram
B. Human agent
Since the Human agent is the most important entity we
focus on describing this agent in detail. Human agents
represent people in the district. After an earthquake, agents
can decide to move, on roads that are not blocked, or stay at
their current location. They can become informal leaders
who know the safe areas and can guide other agents to their
targets. Agents can also become followers by searching for a
leader in their perception zones and then following that
agent. If a follower cannot find a leader, it wanders around.
Human agents can observe obstacles and other agents
inside its perception radius (attribute radius_perception).
After meeting a big obstacle, the Human agent can act in
different ways. If it is a leader, it will choose another way to
reach target. If it is a wanderer, it will choose another target
to go to. Each Agent update their street knowledge,
remembering which streets are blocked so that they may be
avoided in future.
C. Obstacle agent
Obstacle agents can affect the behaviour of Human
agents. In this work, we focus on the obstacles induced by an
earthquake itself, particularly by the buildings damaged by
the earthquake. Based on the damage level of buildings, we
can group obstacles into three levels: big, medium and small.
The big obstacles destroy all the streets and people in those
zones, so no agent can pass these obstacles. In the zone of
medium and small obstacles, people are exposed to the
danger (attribute exposed); this means that the agent is in the
danger zone and there is a high risk of being injured. Medium
obstacles can constrain the movement of agents in their zone,
making them move more slowly.
D. Other agents
The Quake agent has attributes for the intensity and time
of the earthquake. Building agents have attributes for
damage level of the building. Street agents are considered as
a weighted graph where each street is an edge and each
intersection is a vertex. Each unaffected street has a small
weight (e.g. the length of street), but blocked streets have a
much higher weight than normal streets (e.g. a billion). By
using this weighted graph approach each agent has their own
representation (beliefs) about the state of the streets.
The spatial scale in our model is in meters and each
simulation cycle lasts one second. The simulation begins
after all human agents have decided upon their target, and
ends when all the leaders have reached their targets.
E. Agent behaviours
Human behaviours:
How people behave in reality is very complex. In our
model, we just focus on the mobility of people and do not
consider more complex behaviours. From the survey, we
synthesize 6 behaviours:
1. Move without changing activity. This covers
people that had intended to move and do not
change
their activity as a result of
the
earthquake (e.g. people who are on their way to
work)
2. Stay in place without changing activity. This
covers people who remain where they are and
continue doing the same activity (e.g. those
who are at home and do not have employment).
3. Change activity in order to go to a safe place.
This covers people who intentionally move to a
safe place just after an earthquake.
4. Change activity in order to go to an unsafe
place. This covers peoples who intentionally
move to an unsafe place (e.g. someone who
goes to a relative’s home even though it may
be unsafe).
5. Change activity and stay in place. This covers
people who had planned to go somewhere, but
because of the earthquake they remain where
they are.
Imitate others. This addresses leader-follower
behaviours.
6.
The simulation begins when all of the agents have
determined where to go. Thus, from a mobility point of
view, behaviours 1, 3 and 4 are equivalent (the agent
moves), and behaviours 2 and 5 are also equivalent (the
agent does not move).
Human behaviours are synthesized into three groups:
movement, obstacle perception and
imitation. While
movement behaviour concerns the mobility of people,
obstacle perception and imitation relate to determining and
updating target locations and knowledge of the blocked
streets.
• Movement behaviour
The streets are considered as a weighted graph on which
the Human agent moves and determines the shortest path
to reach its target. Moreover, the street knowledge of agent
is considered as list of weights applied for the graph. Each
agent has a different knowledge of the streets, so the
weights applied for the graph are different according to
each agent. Thus the shortest path to the target depends on
two factors: the agent’s target and the agent’s street
knowledge (figure 3).
Figure 3. The movement process
• Obstacle perception behaviour
Agents within the zone of medium and small
obstacles are considered to be exposed to danger. In
addition, medium obstacles constrain the movement of
agents, making them move along their route more
slowly.
Agents can observe obstacles in their zone of
perception and when they encounter a big obstacle they
cannot proceed. However, they remember blocked
streets so that they may be avoided in the future when
finding another path to the target. The difference
between the behaviour of leaders and followers is that
if they meet an obstacle, leaders will find other ways to
reach their target, while followers will find another
target to go to. In our model, we consider 5 blocked
streets configurations (Fig. 4).
repeatedly doesn’t have success it will eventually give up.
Figure 6 shows the obstacle perception behaviour.
Figure 4. Blocked street configurations (with damages)
In figure 4 the red circle is an obstacle and a blue square is
the Human agent. In (1) the Agent is in the area with no exit.
In (2) there is a long street section before the agent moves
towards the obstacle. In (3) there is a short street section
before the agent moves towards the obstacle. In (4) the
Agent is in the street with two obstacles on both sides. In (5)
the Agent sees an obstacle on another street.
The simplest method to implement this behaviour is to
update an agent’s knowledge of blocked streets and target
path when the agent encounters an obstacle. However, this
leads to several problems.
First, with configuration (2), the agent is still in the
blocked street, it must go back to choose another street, but
the street section to go back to is longer than the rest of the
street, so the agent often chooses to enter into the obstacle.
To overcome this problem we make all agents update their
street knowledge and go back to the previous intersection
after meeting an obstacle (Figure 5). After that, the weights
of blocked streets are updated, leader agents build other paths
to reach their targets, and wandering agents choose other
targets and build paths to go. If an agent encounters an
obstacle as in the configuration (5), figure 4, then it will
normally update the weights.
•
Figure 6. Obstacle perception process
Imitation behaviour
For the followers and wanderers, at the beginning, they
search randomly around them for a leader. If they find a
leader, then they become followers. They update their
targets to the position of leader and update paths to follow
it. Figure 7 shows the process of imitation.
Figure 7. Imitation process for followers and wanderers
For the leaders, every 10 cycles of the simulation, they
perceive their followers and adjust their speeds to the
slowest follower (figure 8).
Figure 5. Agent moves to the previous intersection to update its street
knowledge
Secondly, with configurations (1) and (4), the agent has no
choice but to pass over obstacles to get out. In order to fix
this problem we add a parameter concerning the maximum
number of times that the agent encounters an obstacle (2-3
times). After this, the agent stops and stays where it is. This
feature relates to the psychological aspect of people in the
evacuation. If a person tries to find a way to its target but
Figure 8. Speed adjustment process for leader
• Obstacle behaviours
The Obstacle agent has two behaviours: street damage
and human damage. With street damage, each agent finds
all the streets around it that it overlaps and updates the
“blocked” attribute of that Street agent to “true”. With
human damage, each Obstacle agent finds all of the
Human agents with which it overlaps and destroys them.
VI. DEVELOPMENT
The AMEL model has been implemented on the GAMA
platform
(Grignard, Taillandier, Gaudou, Huynh and
Drogoul, 2013); GAMA is an open-source generic agent-
based modelling and simulation platform. It provides an
intuitive modelling language with high-level primitives to
define agents and their environment. GAMA includes a
powerful Integrated Development Environment to help non-
computer scientists
to develop complex models with
powerful features in terms of Geographical Information
Systems (GIS)
integration and high-level
tools (e.g.
decision-making or clustering algorithms). In addition, both
the language and the software have been designed to allow
the development of big models with a huge number of
agents (with various architectures from reflex-based to
belief-desire-intention architectures). GAMA also allows
modellers to manage various kinds of complex environments
such as square, hexagonal or irregular grids, networks, or
continuous environments linked to GIS data.
GAMA enables a hierarchical and dynamic organization
of agents. The platform is also easily extensible to add new
features to models (e.g. to give the possibility of integrating
equation-based (ODE) models into agents) or new agent
architectures or features. GAMA has been successfully used
to develop various large-scale applications that share the
need for a tight integration and management of huge GIS
data, and for strong
interactions between a complex
environment and the agents. It was used for instance in the
MAELIA platform for simulating water management
problems (Gaudou, Sibertin-Blanc, Thérond, Amblard,
Arcangeli, Balestrat, Charron-Moirez, Gondet, Hong,
Mayor, Panzoli, Sauvage, Sanchez-Perez, Taillandier,
Nguyen, Vavasseur and Mazzega, 2013).
Figure 9 shows a screenshot of the simulation. The main
display represents the district with the streets in black, green
space in green, and the buildings in yellow. Human agents
are represented by small circles of different colours: hot
colours (red, pink) for agents who stay where they are; cold
colours (blue, purple) for agents who move; and the colour
cyan for imitating agents. We also created other displays:
charts of the different behaviours; charts about the number of
victims, people exposed, leaders and followers, etc; and a
chart about the total exposition time. On the right, we can see
the parameters of the simulator; these concern the location of
the human agents and their behaviours, as well as some
global simulator parameters. By varying the value of these
parameters we can create different scenarios which may help
to see the effect of the different information campaigns. For
example, if we want to see the effect on the number of
victims if an earthquake occurs during the night, we can set
the value for the “people in buildings” parameter (probability
of people who are in the buildings) equal to 100 (percent),
indicating that at night everyone is inside a building.
Figure 9.
Screenshot of the simulation
VII. VALIDATION
Once the model has been implemented, it must be
validated. This is current work in progress. Every newly
implemented “brick” of the model is tested individually and
all the associated scenarios are evaluated in order to
understand the impact of the new brick. As we developed the
simulator incrementally it eventually becomes increasingly
complex. The challenge is to understand the impact of any
new changes on the global model. For example, once human
exposure to hazards is implemented, we have to verify that
the fatalities figures are realistic. For these kinds of
validations we ask the experts who collected the data and
those that have real life experience of the consequences of
earthquakes. This process of validation involves all other
steps of the methodology and is central to the development
process. In practice we have found this to be a highly
iterative step where we frequently request validation from
experts when changing our model and after implementation.
This validation constitutes a major step that should not be
underestimated.
VIII. FIRST RESULTS
Although we are currently in the validation stage, we
provide some first results so that the reader may see the type
of results that may be achieved with the simulator. The
results concern two scenarios: 1. The survey scenario, this is
the real situation where most people stay where they are and
do not go to the safe place; 2. An optimistic scenario where
everyone goes to a safe place (e.g. green space, schools,
outside of the city, or even just into the streets).
We can see in the graphs below that the number of
exposed people in the optimistic scenario is less than that in
the survey scenario. Also the number of ‘others’ (i.e. those
who are not exposed and injured, in other words they are in
the safe places) and those who go to outside of the city
(considered as a safe place) in the optimistic scenario is more
than in the survey scenario.
to many small earthquakes, a significant number of them do
not change their behaviours during an earthquake and
continue to do their originally planned activities. Whilst in
small earthquakes this is not problematic, the consequences
of being unprepared in the case of a large one could be
devastating.
We have developed a model and first prototype of a
simulator that shows pedestrians’ movements following
different human behaviours after an earthquake. Using
GAMA we were able to incorporate the actual map of the
neigbourhood in Beirut, including current streets and actual
buildings. This allowed us to develop a realistic simulator
with agents moving from known buildings along current
streets. One particular feature of how we have designed the
model is that each agent has its own personal view of the
state of the road infrastructure. This replicates the real
situation where people are only aware of the routes open to
them after an earthquake as
they move around
the
neigbourhood towards their target. As agents explore their
neigbourhood,
their knowledge of
the
streets
is
incrementally updated.
We are currently in the process of validating the model
and simulator. In particular we are addressing the problem of
an agent’s exposure to danger and how fatalities may be
quantified given the different severities of building damage.
Validation is a long process since we need to interact with
local experts from different domains and the data required is
not always available. Nevertheless, despite some difficulties
with validation, the results of the simulator are encouraging.
ACKNOWLEDGEMENTS
This study is funded by the French National Research
Agency (ANR RISKNAT 2009 LIBRIS). The authors would
like to thank the different persons involved in the collection
of data: Jocelyne Adjizian-Gérard, Rita Zaarour, Nada
Saliba, Pierre-Charles Gérard and the students involved in
the
survey
for
the building
inventory
(Geography
Department of
the Saint Joseph University, Beirut,
Lebanon); Jacques Harb for the damages modelling (Notre
Dame University, Beirut, Lebanon); Maud Saikali and
Carine Azzam (Saint Joseph University, Beirut, Lebanon),
Ludvina Colbeau-Justin and Stéphane Cartier (CNRS-
PACTE Laboratory, Grenoble, France)
for
the
risk
perception survey. The authors also thank Patrick Taillandier
(Rouen University, France), Nicolas Marilleau (Institute for
Research and Development, France) and Benoit Gaudou
(Toulouse University, France) for their precious help on
GAMA. Julie Dugdale would like to acknowledge the
support of the University of Agder, to which she is affiliated.
REFERENCES
[1] N. Bellamine-Ben Saoud, T. Ben Mena, J. Dugdale, B. Pavard and B.
Ahmed, “Assessing large scale emergency rescue plans: an agent
based Approach. Special Issue on Emergency Management Systems.
International Journal of Intelligent Control and Systems. Vol. 11, No.
4, Dec. 2006. 260-271.
[2] M. Bertran Rojo, E. Beck, C. Lutoff and Ph. Schoeneich, "Exposition
sociale face aux séismes : la mobilité en question. Le cas de Lorca
(Espagne) – May 2011”. Géorisques.
Figure 10. Fatality and exposure graphs. On the left, the survey
scenario. On the right an optimistic scenario
Concerning the total exposure time, we can see that in the
current situation (survey scenario) there are more people who
are exposed for a long time, whereas in the optimistic
scenario there are more people who are exposed for the short
time.
Figure 11. Number of agents exposed. On top, the survey scenario.
On the bottom the optimistic scenario. The horizontal axis shows time slot
periods and the vertical axis shows the number of exposed people in a
corresponding time period.
IX. SUMMARY
Lebanon is an earthquake prone area, lying on the Mount
Lebanon Thrust Fold and the Yamouneh faults, which have
the capacity to generate a 7.5 magnitude earthquake.
Although major earthquakes are rare, minor ones occur
frequently and there is a deep concern that a large
earthquake could destroy the city, as it did in 551 AD.
Despite this threat the study conducted as part of this project
showed that, possibly because people are constantly exposed
Agent World (MAAMAW). LNAI, 1996, Volume 1038, 56-71.
Springer-Verlag.
[20] T.N.A. Nguyen, J.-D. ZuckerD., H.D. Nguyen., A. Drogoul. and D.-
A. Vo, “A Hybrid Macro-Micro Pedestrians Evacuation Model to
Speed Up Simulation in Road Networks”, In Advanced Agent
Technology workshop at AAMAS 2011, Taipei, Taiwan, May 2-6,
2011. Revised selected papers. Pp 371-383. Lecture Notes in
Computer Science, volume 7068. Springer Berlin Heidelberg. 2012.
[21] N. Pelechano and N. I. Badler. “Modeling Crowd and Trained Leader
Behavior during Building Evacuation”, IEEE Computer Graphics and
Applications, Volume 26, Issue 6, November-December 2006, pages
80-86.
[22] A. Sabas, S. Delisle and M. Badri, “A Comparative Analysis of
Multiagent System Development Methodologies: Towards a Unified
Approach”, In Proceedings of the Third International Symposium
From Agent Theory to Agent Implementation, 16th European Meeting
on Cybernetics and Systems Research, 1996.
[23] Skinner, C. and Ramchurn, S. (2010) The RoboCup Rescue
simulation platform, Proceedings of the 9th International Conference
on Autonomous Agents and Multiagent Systems, pp. 1647-1648,
Toronto, Canada.
[24] M. Wooldridge, N.R. Jennings. and D. Kinny, “The GAIA
Methodology for Agent-Oriented Analysis and Design”, Journal of
Autonomous Agents and Multi-Agent Systems, 2000, 3(3) 285-312.
[6]
[8]
[7]
[3] Bonabeau E. “Agent-based modeling: Methods and techniques for
simulating human systems”. In: Adaptive Agents, Intelligence, and
Emergent Human Organization: Capturing Complexity
through
Agent-Based Modeling. Proceedings of the National Academy of
Sciences (PNAS) U S A. 2002 May 14; 99(Suppl 3): pp. 7280–7287.
[4] R. Conte, “The necessity of intelligent agents in social simulation”. In
Ballot, G. and Weisbuch, G. (eds.), Applications of Simulation to
Social Sciences. Paris: Hermes Science, 2000.
[5] Y. Demazeau, “From interactions to collective behaviour in agent
based-systems”. In Proceedings of the First Conference on Cognitive
Science. Saint-Malo, France, 1995.
J. Dugdale, N. Bellamine-Ben Saoud, B. Pavard and N. Pallamin,
“Simulation and Emergency Management”.In Van de Walle, B.,
Turoff, M. and Hiltz, R.H. (eds) Information Systems for Emergency
Management. Series: Advances in Management Information Systems.
Sharp, 2001.
J. Dugdale, B. Pavard and J.L. Soubie, “Design Issues in the
Simulation of an Emergency Call Centre”. In Proceedings of the 13th
European Simulation Multiconference (ESM 99). June 1-4, 1999.
Warsaw, Poland.
J. Dugdale, B. Pavard and J.L. Soubie, “A Pragmatic Development of
a Computer Simulation of an Emergency Call Centre”. In Dieng, R.
(ed) Designing Cooperative Systems. Frontiers
in Artificial
Intelligence and Applications. IOS Press, 2000.
[9] B.Edmonds, “The Use of Models - making MABS actually work”. In.
Moss, S. and Davidsson, P. (eds.), Multi Agent Based Simulation,
2000. Lecture Notes in Artificial Intelligence, 1979:15-32.
[10] B. Edmonds and S. Moss, “From KISS to KIDS: An “anti-simplistic”
In Proceedings of Multi-Agent Based
modelling approach”.
Simulations (MABS) Conference, 2004, pp. 130-144.
[11] F. Fiedrich, An HLA-based multiagent system for optimized resource
allocation after strong earthquakes. Winter Simulation Conference
2006, 486-492.
[12] B. Gaudou, C. Sibertin-Blanc, O. Thérond, F. Amblard, J.-P.
Arcangeli, M. Balestrat, M.-H. Charron-Moirez, E. Gondet, Y. Hong,
Th. Louail, E. Mayor, D. Panzoli, S. Sauvage, J.-M. Sanchez-Perez, P.
Taillandier, V.B. Nguyen, M. Vavasseur and P. Mazzega,
"The MAELIA multi-agent platform for integrated assessment of low-
water management issues” (regular paper). In International Workshop
on Multi-Agent-Based Simulation (MABS 2013), Saint-Paul, MN,
USA, 06/05/2013-07/05/2013, 2013 (to appear).
[13] N. Gilbert and K.G. Troitzsch, Simulation for the Social Scientist.
Open University Press, London, 1999.
[14] N. Glaser, Contribution to Knowledge Acquisition and Modelling in a
Multi-Agent Framework — The CoMoMAS Approach, PhD Thesis
Universite Henry Poincare, Nancy, F, December 1996.
[15] A. Grignard., P. Taillandier, B. Gaudou, N.Q. Huynh, D.-A. Vo and
A. Drogoul A. GAMA v. 1.6: Advancing the art of complex agent-
based modeling and simulation. PRIMA 2013.
[16] Grimm V, Berger U, Bastiansen F, Eliassen S, Ginot V, Giske J,
Goss-Custard J, Grand T, Heinz S, Huse G, Huth A, Jepsen JU,
Jørgensen C, Mooij WM, Müller B, Pe’er G, Piou C, Railsback SF,
Robbins AM, Robbins MM, Rossmanith E, Rüger N, Strand E,
Souissi S, Stillman RA, Vabø R, Visser U, DeAngelis DL. 2006. A
standard protocol for describing individual-based and agent-based
models. Ecological Modelling 198:115-126.
[17] Grimm V, Berger U, DeAngelis DL, Polhill G, Giske J, Railsback SF.
2010. The ODD protocol: a review and first update. Ecological
Modelling 221: 2760-2768
[18] Hawe, G.I., Coates, G., Wilson, D.T. & Crouch, R.S. (2011). Design
Decisions in the Development of an Agent-Based Simulation for
Large-Scale Emergency Response. 8th International Conference on
Information Systems for Crisis Response and Management, Lisbon,
Portugal.
[19] D. Kinny, M. Georgeff and A. Rao, “A Methodology and Modeling
Technique for Systems of BDI agents”. In Proceedings of the Seventh
European Workshop on Modelling Autonomous Agents in a Multi-
|
cs/0510089 | 1 | 0510 | 2005-10-30T16:06:56 | Automata-based adaptive behavior for economic modeling using game theory | [
"cs.MA",
"cs.DM"
] | In this paper, we deal with some specific domains of applications to game theory. This is one of the major class of models in the new approaches of modelling in the economic domain. For that, we use genetic automata which allow to buid adaptive strategies for the players. We explain how the automata-based formalism proposed - matrix representation of automata with multiplicities - allows to define a semi-distance between the strategy behaviors. With that tools, we are able to generate an automatic processus to compute emergent systems of entities whose behaviors are represented by these genetic automata. | cs.MA | cs |
AUTOMATA-BASED ADAPTIVE BEHAVIOR FOR ECONOMIC
MODELING USING GAME THEORY
R. Ghnemat, K. Khatatneh, S. Oqeili
C. Bertelle(1), G.H.E. Duchamp(2)
Al-Balqa' Applied University,
Al-Salt, 19117
Jordan
(1) LIH - University of Le Havre,
(2) LIPN - University of Paris XIII,
France
ABSTRACT
this way a good tool to modelize such adaptive strategy.
In this paper, we deal with some specific domains of ap-
plications to game theory. This is one of the major class
of models in the new approaches of modelling in the eco-
nomic domain. For that, we use genetic automata which
allow to buid adaptive strategies for the players. We ex-
plain how the automata-based formalism proposed - ma-
trix representation of automata with multiplicities - allows
to define a semi-distance between the strategy behaviors.
With that tools, we are able to generate an automatic pro-
cessus to compute emergent systems of entities whose be-
haviors are represented by these genetic automata.
1. Introduction: Adaptive Behaviour Modeling for Ga-
me Theory
Since the five last decades, game theory has become a ma-
jor aspect in economic sciences modelling and in a great
number of domains where strategical aspects has to be in-
volved. Game theory is usually defined as a mathematical
tool allowing to analyse strategical interactions between
individuals.
Initially funded by mathematical researchers, J. von
Neumann, E. Borel or E. Zermelo in 1920s, game theory
increased in importance in the 1940s with a major work
by J. von Neumann and O. Morgenstern and then with
the works of John Nash in the 1950s [9]. John Nash has
proposed an original equilibrium ruled by an adaptive cri-
terium. In game theory, the Nash equilibrium is a kind of
optimal strategy for games involving two or more players,
whereby the players reach an outcome to mutual advan-
tage. If there is a set of strategies for a game with the prop-
erty that no player can benefit by changing his strategy
while the other players keep their strategies unchanged,
then this set of strategies and the corresponding payoffs
constitute a Nash equilibrium.
We can understand easily that the modelization of a
player behavior needs some adaptive properties. The com-
putable model corresponding to genetic automata are in
The plan of this paper is the following. In the next sec-
tion, we present some efficient algebraic structures, the au-
tomata with multiplicities, which allow to implement pow-
erful operators. We present in section 3, some topological
considerations about the definition of distances between
automata which induces a theorem of convergence on the
automata behaviors. Genetic operators are proposed for
these automata in section 4. For that purpose, we show
that the relevant "calculus" is done by matrix represen-
tions unravelling then the powerful capabilities of such al-
gebraic structures. In section 5, we focus our attention on
the "iterated prisonner dilemma" and we buid an original
evolutive probabilistic automaton for strategy modeling,
showing that genetic automata are well-adapted to model
adaptive strategies. Section 6 shows how we can use the
genetic automata developed previously to represent agent
evolving in complex systems description. An agent behav-
ior semi-distance is then defined and allows to propose an
automatic computation of emergent systems as a kind of
self-organization detection.
2. Automata from boolean to multiplicies theory (Au-
tomata with scalars)
Automata are initially considered as theoretical tools. They
are created in the 1950's following the works of A. Tur-
ing who previously deals with the definition of an abstract
"machine". The aim of the Turing machines is to define
the boundaries for what a computing machine could do
and what it could not do.
The first class of automata, called finite state automata
corresponds to simple kinds of machines [21]. They are
studied by a great number of researchers as abstract con-
cepts for computable building. In this aspect, we can re-
call the works of some linguist researchers, for example N.
Chomsky who defined the study of formal grammars.
In many works, finite automata are associated to a rec-
ognizing operator which allows to describe a language [2,
10]. In such works, the condition of a transition is simply
a symbol taken from an alphabet. From a specific state S,
the reading of a symbol a allows to make the transitions
which are labeled by a and come f romS (in case of a
deterministic automaton - a DFA - there is only one tran-
sition - see below). A whole automaton is, in this way,
associated to a language, the recognized language, which
is a set of words. These recognized words are composed
of the sequences of letters of the alphabet which allows to
go from a specific state called initial state, to another spe-
cific state, called final state.
A first classification is based on the geometric aspect :
DFA (Deterministic Finite Automata) and NFA (Nonde-
terministic Finite Automata).
• I : Q 7→ K is a function over the set of states, which
associates to each initial state a value of K, called
entry cost, and to non- initial state a zero value ;
• F : Q 7→ K is a function over the set states, which
associates to each final state a value of K, called
final cost, and to non-final state a zero value;
• T is the transition function, that is T : Q×A×Q 7→
K which to a state Si, a letter a and a state Sj as-
sociates a value z of K (the cost of the transition) if
it exist a transition labelled with a from the state Si
to the state Sj and and zero otherwise.
• In Deterministic Finite Automata, for each state there
is at most one transition for each possible input and
only one initial state.
• In Nondeterministic Finite Automata, there can be
none or more than one transition from a given state
for a given possible input.
Remark 1 Automata with multiplicities are a generalisa-
tion of finite automata.
In fact, finite automata can be
considered as automata with multiplicities in the semiring
K, the boolan set B = {0, 1} (endowed with the logical
"or/and"). To each transition we affect 1 if it exists and 0
if not.
Besides the classical aspect of automata as machines
allowing to recognize languages, another approach con-
sists in associating to the automata a functional goal. In
addition of accepted letter from an alphabet as the condi-
tion of a transition, we add for each transition an infor-
mation which can be considered as an output data of the
transition, the read letter is now called input data. We de-
fine in such a way an automaton with outputs or weighted
automaton.
Such automata with outputs give a new classification
of machines. Transducers are such a kind of machines,
they generate outputs based on a given input and/or a state
using actions. They are currently used for control appli-
cations. Moore machines are also such machines where
output depends only on a state, i.e.
the automaton uses
only entry actions. The advantage of the Moore model is
a simplification of the behaviour.
Finally, we focus our attention on a special kind of au-
tomata with outputs which are efficient in an operational
way. This automata with output are called automata with
multiplicities. An automaton with multiplicities is based
on the fact that the output data of the automata with out-
put belong to a specific algebraic structure, a semiring
[13, 22]. In that way, we will be able to build effective
operations on such automata, using the power of the alge-
braic structures of the output data and we are also able to
describe this automaton by means of a matrix representa-
tion with all the power of the new (i.e. with semirings)
linear algebra.
Definition 1 (Automaton with multiplicities)
An automaton with multiplicities over an alphabet A and
a semiring K is the 5-uple (A, Q, I, T, F ) where
• Q = {S1, S2 · · · Sn} is the finite set of state;
Remark 2 We have not yet, on purpose, defined what a
semiring is. Roughly it is the least structure which allows
the matrix "calculus" with unit (one can think of a ring
without the "minus" operation). The previous automata
with multiplicities can be, equivalently, expressed by a ma-
trix representation which is a triplet
• λ ∈ K 1×Q which is a row-vector which coefficients
are λi = I(Si),
• γ ∈ K Q×1 is a column-vector which coefficients are
γi = F (Si),
• µ : A∗ 7→ K Q×Q is a morphism of monoids (in-
deed K Q×Q is endowed with the product of matri-
ces) such that the coefficient on the qith row and qjth
column of µ(a) is T (qi, a, qj)
3. Topological considerations
If K is a field, one sees that the space A(n) of automata
of dimension n (with multiplicities in K) is a K-vector
space of dimension k.n2 + 2n (k is here the number of
letters). So, in case the ground field is the field of real
or complex numbers [3], one can take any vector norm
(usually one takes one of the Holder norms (xi)i∈I α :=
α for α ≥ 1, but any norm will do) and the
1
(cid:0)Pi∈I xiα(cid:1)
distance is derived, in the classical way, by
d(A1, A2) = norm(V (A1) − V (A2))
(1)
where V (A) stands for the vector of all coefficients of
A = (λ, µ, γ) arranged in some order one has then the
result of Theorem 1. Assuming that K is the field of real or
complex numbers, we endow the space of series/behaviours
with the topology of pointwise convergence (Topology of
F. Treves [23]).
Theorem 1 Let (An) be a sequence of automata with limit
L (L is an automaton), then one has
5. Applications to competition-cooperation modeling
using prisoner dilemma
Behaviour(L) = lim
n→∞
Behaviour(An)
(2)
where the limit is computed in the topology of Treves.
4. Genetic automata as efficient operators
We define the chromosome for each automata with multi-
plicities as the sequence of all the matrices associated to
each letter from the (linearly ordered) alphabet. The chro-
mosomes are composed with alleles which are here the
lines of the matrix [6].
In the following, genetic algorithms are going to gener-
ate new automata containing possibly new transitions from
the ones included in the initial automata.
The genetic algorithm over the population of automata
with multiplicities follows a reproduction iteration broken
up in three steps [14, 18, 17]:
We develop in this section how we can modelize competition-
cooperation processes in a same automata-based represen-
tation. The genetic computation allows to make automatic
transition from competition to cooperation or from coo-
peartion to competition. The basic problem used for this
purpose is the well-known prisoner dilemma [1].
5.1. From adaptive strategies to probabilistic automata
The prisoner dilemma is a two-players game where each
player has two possible actions: cooperate (C) with its ad-
versary or betray him (C). So, four outputs are possible
for the global actions of the two players. A relative payoff
is defined relatively to these possible outputs, as described
in the following table where the rows correspond to one
player behaviour and the columns to the other player one.
C
C (3,3)
C (5,0)
C
(0,5)
(1,1)
• Duplication: where each automaton generates a clone
of itself;
Table 1. Prisoner dilemma payoff
• Crossing-over: concerns a couple of automata. Over
this couple, we consider a sequence of lines of each
matrix for all. For each of these matrices, a permu-
tation on the lines of the chosen sequence is made
between the analogue matrices of this couple of au-
tomata;
• Mutation: where a line of each matrix is randomly
chosen and a sequence of new values is given for
this line.
Finally the whole genetic algorithm scheduling for a
full process of reproduction over all the population of au-
tomata is the evolutionary algorithm:
1. For all couple of automata, two children are cre-
ated by duplication, crossover and mutation mech-
anisms;
2. The fitness for each automaton is computed;
3. For all 4-uple composed of parents and children, the
performless automata, in term of fitness computed
in previous step, are suppressed. The two automata,
still living, result from the evolution of the two ini-
tial parents.
Remark 3 The fitness is not defined at this level of ab-
stract formulation, but it is defined corresponding to the
context for which the automaton is a model, as we will do
in the next section.
In the iterative version of the prisoner's dilemma, suc-
cessive steps can be defined. Each player do not know
the action of its adversary during the current step but he
knows it for the preceding step. So, different strategies
can be defined for a player behaviour, the goal of each one
is to obtain maximal payoff for himself.
In Figures 1 and 2, we describe two strategies with
transducers. Each transition is labeled by the input cor-
responding to the player perception which is the prece-
dent adversary action and the output corresponding to the
present player action. The only inital state is the state 1,
recognizable by the incoming arrow labeled only by the
output. The final states are the states 1 and 2, recognizable
with the double circles.
In the strategy of Figure 1, the player has systemati-
cally the same behaviour as its adversary at the previous
step. In the strategy of Figure 2, the player chooses defini-
tively to betray as soon as his adversary does it. The previ-
ous automaton represents static strategies and so they are
not well adapted for the modelization of evolutive strate-
gies. For this purpose, we propose a model based on a
probabilistic automaton described by Figure 3 [5].
This automaton represents all the two-states strategies
for cooperation and competitive behaviour of one agent
against another in prisoner's dilemma.
The transitions are labeled in output by the probabili-
ties pi of their realization. The first state is the state reached
C:C
C:C
2
C:C
C
1
C:C
Figure 1. Tit-for-tat strategy automaton
C
1
C:C
C:C
2
C:C
C:C
Figure 2. Vindictive strategy automaton
C:1−p4
C:p5
C:p4
C:1−p2
p1
1
2
1−p1
C:p3
C:1−p5
C:p2
C:1−p3
Figure 3. Probabilistic multi-strategies two-states au-
tomaton
after cooperation action and the second state is reached af-
ter betrayal.
For this automaton, the associated matrix representa-
tion, as described previously, is:
I = ( p1
1 − p1 ) ;
1 − p6(cid:19) ;
F = (cid:18) p6
T (C) = (cid:18) p2
T (C) = (cid:18) p4
p5
p3
1 − p2
1 − p3(cid:19) ;
1 − p5(cid:19)
1 − p4
(3)
(4)
(5)
(6)
5.2. From probabilistic automata to genetic automata
With the matrix representation of the automata, we can
compute genetic automata as described in previous sec-
tions. Here the chromosomes are the sequences of all the
matrices associated to each letter. We have to define the
fitness in the context of the use of these automata. The
fitness here is the value of the payoff.
5.3. General Genetic Algorithm Process for Genetic
Automata
A population of automata is initially generated. These au-
tomata are playing against a predefined strategy, named
S0.
Each automaton makes a set of plays. At each play, we
run the probabilistic automaton which gives one of the two
outputs: (C) or (C). With this output and the S0's output,
we compute the payoff of the automaton, according with
the payoff table.
At the end of the set of plays, the automaton payoff is
the sum of all the payoffs of each play. This sum is the
fitness of the automaton. At the end of this set of plays,
each automaton has its own fitness and so the selection
process can select the best automata. At the end of these
selection process, we obtain a new generation of automata.
This new generation of automata is the basis of a new
computation of the 3 genetics operators.
This processus allows to make evolve the player's be-
havior which is modelized by the probabilistic multi-stra-
tegies two-states automaton from cooperation to compe-
tition or from competition to cooperation. The evolution
of the strategy is the expression of an adaptive computa-
tion. This leads us to use this formalism to implement
some self-organisation processes which occurs in complex
systems.
6. Extension to Emergent Systems Modeling
In this section, we study how evolutive automata-based
modeling can be used to compute automatic emergent sys-
tems. The emergent systems have to be understood in the
meaning of complex system paradigm that we recall in the
next section. We have previously defined some way to
compute the distance between automata and we use these
principles to define distance between agents behaviours
that are modeled with automata. Finally, we defined a spe-
cific fitness that allows to use genetic algorithms as a kind
of reinforcement method which leads to emergent system
computation [15].
6.1. Complex System Description Using Automata-Ba-
sed Agent Model
According to General System Theory [4, 19], a complex
system is composed of entities in mutual interaction and
interacting with the outside environment. A system has
some characteristic properties which confer its structural
aspects, as schematically described in part (a) of Figure 4:
• The set elements or entities are in interactive depen-
dance. The alteration of only one entity or one in-
teraction reverberates on the whole system.
• A global organization emerges from interacting con-
stitutive elements. This organization can be identi-
fied and carries its own autonomous behavior while
it is in relation and dependance with its environ-
ment. The emergent organization possesses new prop-
erties that its own constitutive entities don't have.
"The whole is more than the sum of its parts".
• The global organization retro-acts over its constitu-
tive components. "The whole is less than the sum of
its parts" after E. Morin.
The interacting entities network as described in part
(b) of Figure 4 leads each entity to perceive informations
or actions from other entities or from the whole system
and to act itself.
A well-adapted modeling consists of using an agent-
based representation which is composed of the entity called
agent as an entity which perceives and acts on an environ-
ment, using an autonomous behaviour as described in part
(c) of Figure 4.
To compute a simulation composed of such entities,
we need to describe the behaviour of each agent. This one
can be schematically described using internal states and
transition processes between these states, as described in
part (d) of Figure 4.
There are several definitions of "agents" or "intelligent
agents" according to their behaviour specificities [11, 24].
Their autonomy means that the agents try to satisfy a goal
and execute actions, optimizing a satisfaction function to
reach it.
For agents with high level autonomy, specific actions
are realized even when no perception are detected from
the environment. To represent the process of this delib-
eration, different formalisms can be used and a behaviour
decomposed in internal states is an effective approach. Fi-
nally, when many agents operate, the social aspects must
also be taken into account. These aspects are expressed
as communications through agent organisation with mes-
sage passing processes. Sending a message is an agent ac-
tion and receiving a message is an agent perception. The
previous description based on the couple: perception and
action, is well adapted to this.
6.2. Agent Behavior Semi-Distance
We describe in this section the bases of the genetic algo-
rithm used on the probabilistic automata allowing to man-
age emergent self-organizations in the multi-agent simula-
tion.
For each agent, we define e an evaluation function of
its own behaviour returning the matrix M of values such
that Mi,j is the output series from all possible successive
perceptions when starting from the initial state i and end-
ing at the final state j, without cycle. It will clearly be 0
if either i is not an initial state or j is not a final one and
the matrix Mi,j is indeed a matrix of evaluations [2] of
subseries of
M ∗ := (Xa∈A
µ(a)a)∗
(7)
Notice that the coefficients of this matrix, as defined,
are computed whatever the value of the perception in the
alphabet A on each transition on the successful path1. That
means that the contribution of the agent behaviour for col-
lective organization formation is only based, here, on prob-
abilities to reach a final state from an initial one. This al-
lows to preserve individual characteristics in each agent
behaviour even if the agent belongs to an organization.
Let x and y two agents and e(x) and e(y) their respec-
tive evaluations as described above. We define d(x, y) a
semi-distance (or pseudometrics, see [3] ch IX) between
the two agents x and y as e(x) − e(y), a matrix norm
of the difference of their evaluations. Let Vx a neighbour-
hood of the agent x, relatively to a specific criterium, for
example a spatial distance or linkage network. We define
f (x) the agent fitness of the agent x as :
f (x) =
card(Vx)
d(x, yi)2
Xyi∈Vx
∞
d(x, yi)2 6= 0
if Pyi∈Vx
otherwise
1A succesful path is a path from an initial state to a final state
Global Dynamics
and structures
Global structures
Agent
Interacting Entities
Agent
Agent
(a) Global complex system
(b) Interacting entities network
Agent
behaviour
Agent
behaviour
perceptions
actions
Environment
perceptions
actions
Environment
(c) Agent−based model for entity
(d) automata−based model for agent behaviour
Figure 4. Multi-scale complex system description: from global to individual models
6.3. Evolutive Automata for Automatic Emergence of
Self-Organized Agent- Based Systems
In the previous computation, we defined a semi-distance
between two agents. This semi-distance is computed using
the matrix representation of the automaton with multiplic-
ities associated to the agent behaviour. This semi-distance
is based on successful paths computation which needs to
define initial and final states on the behaviour automata.
For specific purposes, we can choose to define in some
specific way, the initial and final states. This means that
we try to compute some specific action sequences which
are chararacterized by the way of going from some spe-
cific states (defined here as initial ones) to some specific
states (defined here as final ones).
Based on this specific purpose which leads to define
some initial and final states, we compute a behaviour semi-
distance and then the fitness function defined previously.
This fitness function is an indicator which returns high
value when the evaluated agent is near, in the sense of
the behaviour semi-distance defined previously, to all the
other agents belonging to a predefined neighbouring.
Genetic algorithms will compute in such a way to make
evolve an agent population in a selective process. So dur-
ing the computation, the genetic algorithm will make evolve
the population towards a newer one with agents more and
more adapted to the fitness. The new population will con-
tain agents with better fitness, so the agents of a population
will become nearer each others in order to improve their
fitness. In that way, the genetic algorithm reinforces the
creation of a system which aggregates agents with similar
behaviors, in the specific way of the definition of initial
and final states defined on the automata.
The genetic algorithm proposed here can be consid-
ered as a modelization of the feed-back of emergent sys-
tems which leads to gather agents of similar behaviour,
but these formations are dynamical and we cannot predict
what will be the set of these aggregations which depends
of the reaction of agents during the simulation. Moreover
the genetic process has the effect of generating a feed-
back of the emergent systems on their own contitutive ele-
ments in the way that the fitness improvement lead to bring
closer the agents which are picked up inside the emergent
aggregations.
[11] J. Ferber, Multi-agent system, Addison-Wesley,
1999.
[12] L.J. Fogel, A.J. Owens, M.J. Welsh, Artificial intelli-
gence through simulated evolution, John Wiley, 1966.
[13] J.S. Golan, Power algebras over semirings, Kluwer
Academic Publishers, 1999.
[14] D.E. Goldberg, Genetic Algorithms, Addison-
Wesley, 1989.
[15] J. H. Holland, Hidden Order - How adaptation builds
complexity, Persus books ed., 1995.
[16] J.E. Hopcroft, R. Motwani, J.D. Ullman, Introduc-
tion to automata theory, Languages and Computation,
Addison-Wesley, 2001.
[17] J. Koza, Genetic programming, Encyclopedia of
Computer Sciences and Technology, 1997.
[18] M. Mitchell, An introduction to Genetic Algorithms,
The MIT Press, 1996.
[19] J.-L. Le Moigne, La mod´elisation des syst`emes com-
plexes, Dunod, 1999.
[20] I. Rechenberg, Evolution strategies, Fromman-
Holzboog, 1973.
[21] M.P. Schutzenberger, "On the definition of a fam-
ily of automata", Information and Control 4, 245-270
(1961).
[22] R.P. Stanley, Enumerative combinatorics, Cam-
bridge University Press, 1999.
[23] F. Treves, Topological Vector Spaces, Distributions
and Kernels, Acad. Press 1967.
[24] G. Weiss, ed., Multiagent Systems, MIT Press, 1999.
For specific problem solving, we can consider that the
previous fitness function can be composed with another
specific one which is able to measure the capability of the
agent to solve one problem. This composition of fitness
functions leads to create emergent systems only for the
ones of interest, that is, these systems are able to be de-
veloped only if the aggregated agents are able to satisfy
some problem solving evaluation.
7. Conclusion
The aim of this study is to develop a powerful algebraic
structure to represent behaviors concerning cooperation-
competition processes and on which we can add genetic
operators. We have explained how we can use these struc-
tures for modeling adaptive behaviors needed in game the-
ory. More than for this application, we have described
how we can use such adaptive computations to automat-
ically detect emergent systems inside interacting networks
of entities represented by agents in a simulation.
8. References
[1] R. Axelrod, The complexity of cooperation, Princeton
University Press, 1997.
[2] J. Berstel and G. Reutenauer, Rational series and their
language, EATCS, 1988.
[3] Bourbaki N., Elements of Mathematics: General
Topology, Chapters 5-10, Springer-Verlag Telos (dec.
1998).
[4] L. von Bertalanffy, General System Theory, Georges
Braziller Ed., 1968.
[5] C. Bertelle, M. Flouret, V. Jay, D. Olivier, and J.-
L. Ponty "Adaptive behaviour for prisoner dilemma
strategies based on automata with multiplicities." In
ESS 2002 Conf., Dresden (Germany), October 2002.
[6] C. Bertelle, M. Flouret, V. Jay, D. Olivier, and J.-L.
Ponty "Genetic algorithms on automata with multi-
plicities for adaptive agent behaviour in emergent or-
ganizations" In SCI'2001, Orlando, Florida, USA, 22-
25th July 2001.
[7] G. Duchamp, Hatem Hadj-Kacem and Eric
Laugerotte, "Algebraic elimination of ǫ-transitions",
DMTCS, Volume 7 n. 1 (2005), pp. 51-70.
[8] G. Duchamp, J-M Champarnaud, Derivatives of ra-
tional expressions and related theorems, Theoretical
Computer Science 313 (2004).
[9] N. Eber, Th´eorie des jeux, Dunod, 2004.
[10] S. Eilenberg, Automata, languages and machines,
Vol. A and B, Academic press, 1976.
|
1206.6866 | 1 | 1206 | 2012-06-27T16:28:53 | Stochastic Optimal Control in Continuous Space-Time Multi-Agent Systems | [
"cs.MA",
"eess.SY",
"math.OC"
] | Recently, a theory for stochastic optimal control in non-linear dynamical systems in continuous space-time has been developed (Kappen, 2005). We apply this theory to collaborative multi-agent systems. The agents evolve according to a given non-linear dynamics with additive Wiener noise. Each agent can control its own dynamics. The goal is to minimize the accumulated joint cost, which consists of a state dependent term and a term that is quadratic in the control. We focus on systems of non-interacting agents that have to distribute themselves optimally over a number of targets, given a set of end-costs for the different possible agent-target combinations. We show that optimal control is the combinatorial sum of independent single-agent single-target optimal controls weighted by a factor proportional to the end-costs of the different combinations. Thus, multi-agent control is related to a standard graphical model inference problem. The additional computational cost compared to single-agent control is exponential in the tree-width of the graph specifying the combinatorial sum times the number of targets. We illustrate the result by simulations of systems with up to 42 agents. | cs.MA | cs | Stochastic Optimal Control in Continuous Space-Time Multi-Agent
Systems
Bart van den Broek
Wim Wiegerinck
SNN, Radboud University Nijmegen
6525 EZ Nijmegen, The Netherlands
{w.wiegerinck,b.vandenbroek,b.kappen}@science.ru.nl
Bert Kappen
Abstract
ronment and limitations of its resources.
Recently, a theory for stochastic optimal con-
trol in non-linear dynamical systems in con-
tinuous space-time has been developed (Kap-
pen, 2005). We apply this theory to col-
laborative multi-agent systems. The agents
evolve according to a given non-linear dy-
namics with additive Wiener noise. Each
agent can control its own dynamics. The goal
is to minimize the accumulated joint cost,
which consists of a state dependent term and
a term that is quadratic in the control. We fo-
cus on systems of non-interacting agents that
have to distribute themselves optimally over
a number of targets, given a set of end-costs
for the di(cid:11)erent possible agent-target com-
binations. We show that optimal control is
the combinatorial sum of independent single-
agent single-target optimal controls weighted
by a factor proportional to the end-costs
of the di(cid:11)erent combinations. Thus, multi-
agent control is related to a standard graphi-
cal model inference problem. The additional
computational cost compared to single-agent
control is exponential in the tree-width of the
graph specifying the combinatorial sum times
the number of targets. We illustrate the re-
sult by simulations of systems with up to 42
agents.
1
INTRODUCTION
A collaborative multi-agent system (MAS) is a collec-
tion of agents that autonomously control their behav-
ior to achieve a common goal or to maximize the per-
formance of the group. Examples are teams of soccer-
robots and teams of unmanned rescue vehicles in a
hazardous disaster area.
In practical applications, a
MAS often has to deal with uncertainty in the envi-
In this paper, we are interested in optimal control in
such systems. We focus on systems in which agents
in a stochastic environment have to distribute them-
selves e(cid:14)ciently over a number of targets. For exam-
ple, consider a system of n (cid:12)re(cid:12)ghter-agents and (cid:12)res.
The agents are at some initial positions and should
reach the (cid:12)res positions in the most e(cid:14)cient way, such
that each (cid:12)re is reached by an agent (see (cid:12)gure 1). In
this problem, the (cid:12)nal con(cid:12)guration, i.e., which agent
has reached exactly which (cid:12)re is not of importance
for the end performance. The MAS should continu-
ously control itself such that in the end one of these n!
con(cid:12)gurations is realized at minimal expected e(cid:11)ort.
The additional complexity is that due to the noise in
the dynamics, a con(cid:12)guration that seems optimal from
the initial positions may become suboptimal in a later
stage.
A common approach is to model such a system as a
Markov Decision Process (MDP) in discrete space and
time: the optimal actions in an MDP optimization
problem are in principle solved by backward dynamic
programming. Since both the joint action space and
the joint state space of the agents are assumed to be
large in the discretization, and increase exponentially
in the number of agents, simply taking a basic dy-
namic programming approach to solve the MDP will
generally be infeasible [1].
Typically one can describe the system more compactly
as a factored MDP. In such systems both the transition
probabilities and reward functions have some struc-
ture. Unfortunately, this structure is not conserved
in the value functions and exact computation remains
exponential in the system size. Recently, a number
of advanced and powerful approximate methods have
been proposed. The common denominator of these ap-
proaches is that they basically assume some prede(cid:12)ned
approximate structure of the value functions [2, 3].
In this paper, we take a di(cid:11)erent starting point.
Rather than discretizing, we will consider the stochas-
tic optimal control problem in continuous space and
time. As in discrete MDPs, this optimization prob-
lem is in principle solved by backward dynamic pro-
gramming. Usually this optimization is intractable.
However, if (1) both the noise and the control are ad-
ditive to the (nonlinear) dynamics, (2) the increment
in cost is quadratic in the control, and (3) the noise
satis(cid:12)es certain additional conditions, then it can be
shown that the stochastic optimization problem can be
transformed into a linear partial di(cid:11)erential equation,
which can be solved by forward stochastic integration
of a di(cid:11)usion process [4, 5]. This formalism contains
linear-quadratic control as a special case [6].
An interesting observation in [4, 5] is the phenomenon
of symmetry breaking in multi-modal systems, i.e, in
problems where several local minima of the cost co-
exist. This symmetry breaking manifests itself as a
delayed choice, keeping options open and using the
fact that the noise may help to come closer to one of
the options at no additional cost.
Formally the extension of this formalism to coopera-
tive MAS is straightforward. The question that we ask
ourselves, is how the formalism scales from the single-
agent single-target situation (e.g. one (cid:12)reman has to
go to a given (cid:12)re) to a collaborative system of n agents
and m targets. Although the dynamics of the agents is
assumed to be independent, with optimal control the
behavior of the agents will be coupled in a non-trivial
way in order to reach an end-con(cid:12)guration at minimal
cost.
The paper is organized as follows. In section 2 we pro-
vide a review of the general (single-agent) framework.
As an example, we will rederive linear quadratic con-
trol from control theory. In this system a single agent
has to control itself to a single target. Next, in sec-
tion 3 we show how the framework easily allows the
modeling of an agent that has to control itself to one
of m possible targets. It turns out that optimal con-
trol for this case can be written as a weighted sum of
m optimal controls in the presence of the single tar-
gets. This result will form the basis of the multi-agent
analysis later in the paper.
In section 4 we will consider the framework in the
multi-agent setting.
In general, the solution of this
type of problem consists of a sum of mn terms due
to a contribution from each agent-target combination.
For small problems, this summation can be performed
explicitely. For large n and m, the solution is generally
intractable.
Next, we consider models in which the end-costs of
agents-targets con(cid:12)gurations are factored as a sparse
graph. We show that the structure of the graph is con-
Figure 1: The (cid:12)remen problem. A number of (cid:12)remen
go to a number of (cid:12)res, each to a di(cid:11)erent one. How
should the agents coordinate when it is not yet de-
cided to which (cid:12)re each agent should go, and when
the actions of the agents are sub ject to noise?
served in the mn terms. Therefore this summation can
be performed more e(cid:14)ciently using e.g. the junction
tree algorithm [7]. The computation time is exponen-
tial in the induced tree-width of the graph times the
number of targets m, and is linear in the number of
agents n. This is in contrast to discrete MDPs, where,
as remarked earlier, sparsity in the reward function is
not retained in the value function and does not help
to reduce the computation costs.
In section 5, we illustrate the framework by simula-
tion results of stochastic optimal control in the two
toy problems considered in this paper.
2 STOCHASTIC OPTIMAL
CONTROL OF A SINGLE AGENT
In this section, we review the framework developed
in [4, 5].
We consider an agent moving in IRk .
obeys the stochastic dynamics
Its position x
dx = (b(x; t) + u)dt + d(cid:24) ;
(1)
with d(cid:24) a Wiener process with hd(cid:24)id(cid:24)j i = (cid:23)ij dt, b(x; t)
an arbitrary function of x and t, modeling the dynam-
ics due to the environment. The agent can in(cid:13)uence
the dynamics by the control u.
Given x at initial time ti , the problem is to (cid:12)nd a
control tra jectory u(ti ! T ) such that the expected
cost-to-go
C (x; ti ; u(ti ! T )) =
(cid:28)(cid:30)(x(T )) + Z T
u(t)>Ru(t) + V (x(t); t)(cid:17)(cid:29) (2)
dt(cid:16) 1
2
ti
is minimal. The expectation is taken over all noise re-
alizations, resulting in di(cid:11)erent tra jectories in state
space x(ti ! T ) that start in x. (cid:30)(x(T )) is the
end cost, depending only on the end state x(T ).
V (x(t); t)dt is the cost of being at position x(t) during
the time interval [t; t + dt], dt u(t)>Ru(t) is the cost
of the control during the same time interval. R is a
constant k (cid:2) k matrix.
The expected cost-to-go at time t needs to be min-
imized over all strategies u(t ! T ), this yields the
optimal (expected) cost-to-go
the cost to go fol lows from a Gaussian convolution with
e(cid:0)(cid:30)=(cid:21) , resulting in
jx (cid:0) (cid:22)j2
2(cid:23) (T (cid:0) t + R=(cid:11)) (cid:21) :
Z (x; t) / exp (cid:20)(cid:0)
The optimal control fol lows from (6) and (5) resulting
in
u(x; t) =
(cid:22) (cid:0) x
T (cid:0) t + R=(cid:11)
:
J (x; t) = min
u(t!T )
C (x; t; u(t ! T )):
(3)
This result is wel l known from control theory [6].
In the appendix, it is brie(cid:13)y explained that due to the
linear-quadratic form of the optimization problem
the dynamics (1) is linear in the action u, the cost
(2) is quadratic in the actionthe minimization can
be performed explicitly, yielding a non-linear partial
di(cid:11)erential equation in J . If, in addition, the matrices
(cid:23) and R can be linked via a scalar (cid:21) such that (cid:23) =
(cid:21)R(cid:0)1 , the optimal cost-to-go is re-expressed as the log
of an ordinary integral (equation (15) in the appendix),
J (x; t) = (cid:0)(cid:21) log Z (x; t)
(4)
(5)
with \partition function"
Z (x; t) = Z dy(cid:26)(y ; T jx; t)e(cid:0)(cid:30)(y)=(cid:21)
in which (cid:26)(y ; T jx; t) is the probability of arriving in
state y at time T , when starting in x at time t, under
the dynamics (14) in the appendix. This dynamics(14)
equals the stochastic system dynamics without control,
i.e., u = 0, with in addition the probability V (x;t)
(cid:21) dt of
being removed from the system between t and t + dt
and thus not arriving in state y . (cid:30)(y) is the end cost
in the state y .
The optimal control of the agent is directly obtained
from the optimal cost-to-go, by taking its gradient
(equation (10) in the appendix), which implies the fol-
lowing result
u(x; t) = (cid:23) @x log Z (x; t) :
(6)
Example 1 The running example is a system with
linear dynamics, b = 0, and zero potential V (x; t) = 0.
R and (cid:23) are proportional to the identity and are con-
sidered as scalars. Regard less the end costs, the di(cid:11)u-
sion process results in a Gaussian distribution
jy (cid:0) xj2
(cid:26)(y ; T jx; t) / exp (cid:20)(cid:0)
2(cid:23) (T (cid:0) t) (cid:21) :
If we take a quadratic end cost around a target (cid:22),
(cid:30)(x) =
(cid:11)
2
jx (cid:0) (cid:22)j2
Example 2 Now b and V are arbitrary, such that the
di(cid:11)usion process results in a distribution (cid:26)(y ; T jx; t).
To enforce an end state close to target (cid:22), a quadratic
end cost with a large (cid:11) can be chosen. The e(cid:11)ect
is that the factor e(cid:0)(cid:30)(y)=(cid:21) becomes proportional to a
delta-function centered around (cid:22), and Z fol lows di-
rectly from the value of (cid:26) at the target,
Z (x; t) / Z dy(cid:26)(y ; T jx; t)(cid:14)(y (cid:0) (cid:22)) / (cid:26)((cid:22); T jx; t)
from which J and u fol low directly.
3 MULTIPLE TARGETS
In this section, we apply the framework of the previous
section to the case where an agent has to reach one of
a number of possible end states. We show that the
optimal control can be constructed from a weighted
combination of single-target optimal controls.
When the agent has to reach one of a number of states
(cid:22)1 ; : : : ; (cid:22)m at the end time, this can be modeled by
letting e(cid:0)(cid:30)(y)=(cid:21) be a linear combination of functions
which each are peaked around a single target (cid:22)s (with
s = 1; : : : ; m), like e.g. a delta function (cid:14)(y (cid:0) (cid:22)s ) or
a Gaussian centered around (cid:22)s with small width. We
denote these functions as (cid:8)(y ; (cid:22)s ) (cid:17) (cid:8)(y ; s). If we put
an additional cost E (s) when target (cid:22)s is reached, this
combination becomes
e(cid:0)(cid:30)(y)=(cid:21) =
m
m
Xs=1
Xs=1
Substitution into (5) gives the partition function
e(cid:0)E (s)=(cid:21)(cid:8)(y ; s) (cid:17)
w(s)(cid:8)(y ; s);
Z (x; t) =
w(s)Z (x; t; s)
m
Xs=1
which is a weighted combination of single-target par-
tition function
Z (x; t; s) = Z dy(cid:26)(y ; T jx; t)(cid:8)(y ; s):
The optimal action is obtained from (6) and reads
4.1
INDEPENDENT DYNAMICS, JOINT
TASK
u(x; t) =
m
Xs=1
with single-target optimal controls
p(sjx; t)u(x; t; s);
u(x; t; s) = (cid:23) @x log Z (x; t; s)
and p(sjx; t) the probability
p(sjx; t) =
w(s)Z (x; t; s)
Pm
s0=1 w(s0 )Z (x; t; s0 )
Example 3 In the running example, optimal control
with multiple targets is
:
u(x; t) =
(cid:22)(cid:22) (cid:0) x
T (cid:0) t + R=(cid:11)
with (cid:22)(cid:22) the ‘expected target’
(cid:22)(cid:22) =
m
Xs=1
which is the expected value of the target according to
the probability
p(sjx; t)(cid:22)s
p(sjx; t) =
2(cid:23) (T (cid:0)t+R=(cid:11)) i
w(s) exp h(cid:0) jx(cid:0)(cid:22)s j2
2(cid:23) (T (cid:0)t+R=(cid:11)) i
s=1 w(s) exp h(cid:0) jx(cid:0)(cid:22)s j2
Pm
4 STOCHASTIC OPTIMAL
CONTROL OF A MAS
We now turn to the issue of optimally coordinating a
multi-agent system of n agents. In principle, a multi-
agent system can be considered as a system with a joint
state x = (x1 ; : : : ; xn ), where xa is the state of agent
a, a joint dynamics (1), and a joint cost (2) which is to
be minimized by a joint action u = (u1 ; : : : ; un ), where
ua is the control of agent a. The optimal control by
agent a follows from the appropriate components of
the gradient
ua (x1 ; : : : ; xn ; t) = (cid:23) @xa log Z (x1 ; : : : ; xn ; t):
(7)
We remark that in continuous space-time, the opti-
mal controls can be determined independently for each
agent, and coordination does not have to be imposed
explicitly. This is in contrast to discrete multi-agent
MDP models, in which coordination may be needed
since more than one optimal joint action can exist [1].
The reason is that in continuous time control results
in actions that are in(cid:12)nitesimal within an in(cid:12)nitesi-
mal time increment. This allow agents to adapt their
control immediately to each other.
In the remainder of the paper, we consider agents with
independent dynamics ba (x; t) = ba (xa ; t) and inde-
pendent noise (cid:23)ab = (cid:23)a (cid:14)ab with (cid:23)a a noise matrix re-
stricted to the domain of agent a. We also assume
individual contributions to the costs during the pro-
cess: Rab = Ra (cid:14)ab with Ra a matrix restricted to a,
and V (x; t) = Pa Va (xa ; t). We (cid:12)nally assume that
(cid:23) = (cid:21)R(cid:0)1 holds globally. Under these assumptions,
the agents behave like ‘non-interacting particles’, e.g.,
they can freely move through each other without costs
for collisions. Under these assumptions, the joint solu-
tion of the di(cid:11)usion process factorizes into a product of
solutions of independent (single agent) di(cid:11)usion pro-
cesses.
(cid:26)a (ya ; T jxa ; t) :
(cid:26)(y ; T jx; t) = Ya
The agents optimal control and the resulting dynamics
will be coupled by their joint task, expressed in the
end-costs (cid:30)(y). We consider the problem where the
agents have to distribute themselves over a number of
targets (cid:22) = (cid:22)1 ; : : : ; (cid:22)m .
The trivial case where each agent a has to go to a single
target (cid:22)sa is equivalent with a single control problem,
with joint target (cid:22)s = ((cid:22)s1 ; : : : ; (cid:22)sn ). Now s is the
vector of labels, s = (s1 ; : : : ; sn ). Of course, control by
the agents is independent of each other. The partition
function factorizes in single agent partition functions
Z (x; t; s) = Ya Z dya(cid:26)a (ya ; T jxa ; t)(cid:8)a (ya ; sa )
(cid:17) Ya
Za (xa ; t; sa ) :
More interesting is the case where the system has more
choices in how to distribute itself. Like in the single-
agent case, this is described by de(cid:12)ning e(cid:0)(cid:30)(y)=(cid:21) to be
a positive linear combination of peaked (multi-agent)
single-target functions, (cid:8)(y ; s) = Qa (cid:8)a (ya ; sa ), as in
section (3), with the di(cid:11)erence that in this sum s runs
over mn states (for all the possible distributions of
agents over targets). The partition function of this
system then reads
Z (x; t) = Xs
w(s) Ya
The optimal control of an individual agent a is ob-
tained using (7), and leads again to an average of
single-target optimal controls,
m
Xsa=1
p(sa jx; t)ua (xa ; t; sa );
Za (xa ; t; sa ) :
ua (x; t) =
where ua (xa ; t; sa ) = (cid:23)a@xa Za (xa ; t; sa ) and p(sa jx; t)
is the probability
p(sa jx; t) = Psnsa w(s) Qb Zb (xb ; t; sb )
Ps w(s) Qb Za (xb ; t; sb )
which can be interpreted as the probability that agent
a has to go to target sa given the joint state x of the
MAS and time t.
;
Example 4 The (cid:12)remen problem. We consider n
identical agents, and m targets modeled as in exam-
ples 1 and 3. The aim of the agents is to distribute
themselves with minimal action such that each target
is reached at time T by about n
m agents. We model the
system by an additional cost of
Xf =1 (cid:18) n
m
= c(cid:18) n
m (cid:19)2
n2
m (cid:19)
n
Xa=1
Xa;b=1
in which c > 0 is a constant indicating the costs of
suboptimal distributions. Optimal control of agent a is
given by
E (s) = c
(cid:14)sa ;sb (cid:0)
(cid:14)f ;sa (cid:0)
ua (x; t) =
(cid:22)(cid:22)a (cid:0) xa
T (cid:0) t + R=(cid:11)
(8)
with (cid:22)(cid:22) the expected target for agent a,
(cid:22)(cid:22)a = Xsa
p(sa jx; t)(cid:22)sa ;
where
p(sa jx; t) / Xsnsa
exp (cid:20)(cid:0)
E (s)
(cid:21)
(cid:0)
jx (cid:0) (cid:22)s j2
2(cid:23) (T (cid:0) t + R=(cid:11)) (cid:21) :
(9)
The additional computation e(cid:11)ort in multi-agent con-
trol compared to single agent control is the computa-
tion of p(sa jx; t), which involves a sum over mn states.
For small systems this is feasible. For large systems,
this is only feasible if the summation can be performed
e(cid:14)ciently.
4.2 FACTORED END-COSTS
in which (cid:11) are subsets of agents, we can apply the junc-
tion tree algorithm from probabilistic graphical models
to make inference more e(cid:14)cient [7]. The complexity is
then exponential in the induced tree-width of the un-
derlying graph. In the case of the (cid:12)remen problem, this
approach does not really help. The clusters (cid:11) contain
only two agents. However all pairs of agents appear in
a factor, which makes the graph fully connected, sim-
ilar to the fully connected Boltzmann machine. The
tree-width of the graph is n. Non-trivial tree width
(smaller than n) can be expected in systems where the
contribution of an agent to the end cost only depends
on the states of a limited number of other agents.
Example 5 Holiday resort problem. We consider a
set of n agents with identical dynamics, and m targets
(a few holiday resorts) modeled as in examples 1 and 3.
Each agent has relations with only a few other agents,
and only cares for related agents whether or not to have
holiday in the same resort (depending on the sign of the
relation). If two agents are unrelated they are indi(cid:11)er-
ent whether they end up in the same resort. Relations
are assumed to be symmetric. The aim of the agents
is that they have optimal ly distributed themselves with
minimal e(cid:11)ort over the resorts at time T . A way to
model such a system is to de(cid:12)ne for each related pair
of agents a and b a cost
Eab (sa ; sb ) = (cid:0)cab (cid:14)sa ;sb
with cab to weight the relevance of the relation. The
sign of cab is equal to the sign of the relation. The op-
timal control is again as in (8), with the current E (s)
substituted into (9). Note that the (cid:12)remen problem is a
special case, with a ful ly connected network of negative
relations of strength c.
Inference of graphical models is in general linear in
the number of variables n and exponential in the num-
ber of states in the cliques of the underlying junction
tree [7]. The number of states in the largest clique is
equal to the treewidth of the graph times the num-
ber of states per node. This implies that models with
sparse graphs and a limited number of targets are
tractable.
The issue is the computational complexity of the prob-
ability distribution
5 NUMERICAL EXAMPLES
1
Z (x; t)
p(sjx; t) =
Za (xa ; t; sa ) :
w(s) Ya
The complexity comes from the weights w(s) of the
end costs, which couples the agents. In the case that
the weights w(s) allow a factored representation
#
w(cid:11) (s(cid:11) ) = exp "X(cid:11)
w(s) = Y(cid:11)
E(cid:11) (s(cid:11) )
(cid:21)
(cid:0)
In this section we illustrate the optimally controlled
stochastic MASs by numeric simulations results.
In all simulations, we modeled agents in 1-d (for ex-
posure purposes). The models were as in the running
examples, with b = 0 and V = 0. In numerical simu-
lations, time is to be discretized. This has to be done
with a bit of care. In continuous time, udt should be
in(cid:12)nitesimal, regardless the size of u. In the discrete
1
0
s
n
o
i
t
i
s
o
p
−1
0
1
0
s
t
e
g
r
a
t
d
e
t
c
e
p
x
e
−1
0
time
time
1
1
s
n
o
i
t
i
s
o
p
1
0
−1
0
s
t
e
g
r
a
t
d
e
t
c
e
p
x
e
1
0
−1
0
time
time
1
1
Figure 2: Simulation of (cid:12)remen problem. Two agents
start in x = 0 at t = 0 and should reach at t = 1 the
two targets located at x = (cid:0)1 and x = 1 in a noisy
environment with minimal cost of control.
(a) The
positions xa of the agents as a function of time. (b)
The expected targets (cid:22)a of the agents as a function of
time.
Figure 3: Simulation of (cid:12)remen problem with 6 agents
start in x = 0 at t = 0 and should reach at t = 1 the
three targets located at x = (cid:0)1; 0; 1, preferably two
agents per target. (a) The positions xa of the agents
as a function of time. (b) The expected targets (cid:22)a of
the agents as a function of time.
approximation this implies that u(cid:1)t should be small
compared to typical scales in the system. For the run-
ning examples, in particular (8), this means that
u(cid:1)t = (cid:15)( (cid:22)(cid:22) (cid:0) x)
with a small (cid:15) (cid:28) 1. From (8) we solve
(cid:1)t = (cid:15)(T (cid:0) t + R=(cid:11))
which yields a (cid:12)nite discretization for (cid:12)nite (cid:11). In the
simulations, we took (cid:11) = 103 and (cid:15) = 0:01. Further-
more we took noise parameter (cid:23) = 1, and R = 1. We
started all agents at t = 0 at x = 0. End time is T = 1.
In the (cid:12)rst simulation, we have the (cid:12)remen problem
with two agents and two (cid:12)res located at (cid:0)1 and 1.
We model a preference of one agent per (cid:12)re in the end
situation. This is achieved by using the weight rep-
resentation w(s1 ; s2 ) = exp((cid:0)E (s1 ; s2 )=(cid:21)) and setting
E (1; 2) = E (2; 1) = 0 and E (1; 1) = E (2; 2) = 2. In
(cid:12)gure 2, the positions of the 2 agents x1 and x2 are
plotted, as well as the expected target locations (cid:22)(cid:22)1 and
(cid:22)(cid:22)2 . We see that the MAS reached a preferred goal: at
the end time each target is reached by exactly one
agent. During the whole tra jectory, (cid:22)(cid:22)1 (cid:25) (cid:0) (cid:22)(cid:22)2 since
the MAS mostly aim at an end-con(cid:12)guration with one
agent per (cid:12)re. Furthermore, note that in the (cid:12)rst part
of the tra jectory, the expected targets are close to zero,
while only after about t = 0:6 the agents seem to make
a clear choice for their targets. This delayed choice is
due to a symmetry breaking in the cost-to-go as time
increases. Before the symmetry breaking, it is bet-
ter to keep options open, and see what the e(cid:11)ect of
the noise is. After the symmetry breaking, time is too
short to wait longer and a choice has to be made. This
phenomenon is typical for multi-modal problems. For
more details we refer to [4, 5].
In the second simulation, we have the (cid:12)remen prob-
lem with six agents and three (cid:12)res located at (cid:0)1, 0
and 1. We modeled the end cost as in example 4. In
(cid:12)gure 3, the positions of the 3 agents are plotted, as
well as the expected target locations. From the (cid:12)gure
it can be concluded that the MAS has successfully dis-
s
n
o
i
t
i
s
o
p
2
1
0
−1
−2
0
1
0
s
t
e
g
r
a
t
d
e
t
c
e
p
x
e
−1
0
time
time
1
1
Figure 4: Simulation of holiday resort problem. 42
agents start in x = 0 at t = 0 and should reach at t = 1
the three targets located at x = (cid:0)1; 0; 1, preferably
together with positively related agents and not with
negatively related agents (relations are not shown in
the (cid:12)gure). (a) The positions xa of the agents as a
function of time. (b) The expected targets (cid:22)a of the
agents as a function of time.
tributed itself with two agents at each target. In this
simulation, two (local) symmetry breakings are clearly
visible. At about t=0.5, 2 agents seem to choose for
target (cid:22) = (cid:0)1 and the other four agents for an ex-
pected target of (cid:22)(cid:22) = 0:5. Then at about t = 0:8
there is a second symmetry breaking, where these four
agents make their (cid:12)nal choice.
In the last simulation, we have the holiday resort prob-
lem with 42 agents and three resorts located at (cid:0)1, 0
and 1. We modeled the end cost as in example 4. To
model the relations between the agents we generated
a random graph with 42 nodes, in which each node is
coupled to exactly three randomly chosen neighbors.
The relation strengths were randomly chosen cab = (cid:6)1
with equal probability.
In (cid:12)gure 4, the positions of the 42 agents are plotted,
as well as the expected target locations. From the
results it can be seen that each agent reached a target.
Actually, target (cid:0)1 is reached by 10 agents, target 0
by 23 agents, and target 1 is reached by 9 agents. In
this simulation, the coordination in terms of cluster
formation in (cid:22)(cid:22) is profound, despite the fact that the
positions of the agents seem to be quite chaotic.
In the graph in this simulation, there were 34 positive
and 29 negative relations. The treewidth is 7. Among
the agents that ended at target (cid:0)1, there were 5 pos-
itive relations and 0 negative ones. At target 0, there
were 18 positive relations and 1 negative one. At tar-
get 1, there were 6 positive relations and 0 negative
ones. So within the targets, there were a total of 29
positive relations and 1 negative one. Between agents
at di(cid:11)erent targets, there were 5 positive relations and
28 negative ones.
6 DISCUSSION
We studied optimal control
in collaborative multi-
agent systems in continuous space-time. A straightfor-
ward approach to discretize the system in space and
time would make the n agent MAS intractable due
to the exponential blow-up of the state-space. In this
paper, we took the approach developed in [4, 5]. We
showed that under given model assumptions, optimal
distributed control can be solved analytically and that
this solution is tractable in large sparsely connected
systems.
In dense multi-agent systems, however, the exact infer-
ence is intractable. A natural approach would be the
use of message passing algorithms for approximate in-
ference. This is currently studied and will be reported
in near future.
There are many possible model extensions that need
to be explored in future research. Obvious extensions,
such as a non-(cid:12)xed end-time, or systems with more
realistic environments, such as allowing for obstacles
are already of interest to study in the single agent sit-
uation. Others apply typically to the multi-agent sit-
uation, such as penalties for collisions between agents.
Typically, these types of model extensions will prohibit
an analytical solution of the control, and approximate
numerical methods will be required. Some proposals
can be found in [4, 5].
Finally we would like to stress, that although the
model class is quite speci(cid:12)c and maybe not generally
applicable, we think that the study of this class is in-
teresting because it is one of the few \exactly solvable"
multi-agent systems, allowing the study of non-trivial
collective optimal behaviour in large distributed sys-
tems, both analytically as well as in simulations, and
possibly providing insights that might help to develop
approximating methods for more general systems.
Acknowledgments
This research is part of the Interactive Collaborative
Information Systems (ICIS) pro ject, supported by the
Dutch Ministry of Economic A(cid:11)airs, grant BSIK03024.
A STOCHASTIC OPTIMAL
CONTROL
In this appendix we give a brief derivation of (4), (5)
and (6), starting from (3). Details can be found in
[4, 5].
The optimal cost-to-go J in a state x at time t is found
by minimizing C (x; t; u(t ! T )) over all sequences of
controls over the time interval [t; T ],
J (x; t) = min
u(t!T )
C (x; t; u(t ! T )):
It satis(cid:12)es the stochastic Hamilton-Jacobi-Bellman
(HJB) equation
xJ (cid:1)(cid:19);
u (cid:18) 1
u>Ru+V +(cid:0)b+u(cid:1)>
Tr(cid:0)(cid:23) @ 2
(cid:0)@tJ = min
2
with boundary condition J (x; T ) = (cid:30)(x). The mini-
mization with respect to u yields
@xJ +
1
2
u = (cid:0)R(cid:0)1@xJ;
(10)
which de(cid:12)nes the optimal control. Substituting this
control in the HJB equation gives a non-linear equa-
tion for J . We can remove the non-linearity by using
a log transformation: de(cid:12)ne (x; t) through J (x; t) =
(cid:0)(cid:21) log (x; t), with (cid:21) a constant to be de(cid:12)ned, then
1
2
1
2
(cid:21)2 (cid:0)2 (@x )>R(cid:0)1@x ;
u>Ru + u>@xJ = (cid:0)
Tr(cid:0)(cid:23) @ 2
xJ (cid:1) =
1
2
(cid:21) (cid:0)2 (@x )> (cid:23) @x
(cid:21) (cid:0)1Tr(cid:0)(cid:23) @ 2
x (cid:1):
The terms quadratic in vanish if there exists a scalar
(cid:21) such that
1
2
1
2
(cid:0)
(cid:23) = (cid:21)R(cid:0)1 :
(11)
In the one dimensional case, such a (cid:21) can always be
found.
In the higher dimensional case, this restricts
the matrices R(cid:0)1 / (cid:23) . When (11) is satis(cid:12)ed, the
HJB equation becomes
@t = (cid:18) V
(cid:21)
= (cid:0)H ;
x (cid:1)(cid:19)
Tr(cid:0)(cid:23) @ 2
where H a linear operator acting on the function .
Equation (12) must be solved backwards in time with
(cid:0) b>@x (cid:0)
(12)
1
2
(x; T ) = e(cid:0)(cid:30)(x)=(cid:21) . However, the linearity allows us
to reverse the direction of computation, replacing it by
a di(cid:11)usion process, as we will now explain.
The solution to equation (12) is given by
(x; t) = Z dy(cid:26)(y ; T jx; t)e(cid:0)(cid:30)(y)=(cid:21) ;
the density (cid:26)(y ; #jx; t) (t < # (cid:20) T ) satisfying a forward
Fokker-Planck equation
(13)
@#(cid:26)(y ; #jx; t) = H y(cid:26)(y ; #jx; t);
(14)
1
2
V
(cid:21)
H y(cid:26) = (cid:0)
(cid:26) (cid:0) @>
y b(cid:26) +
where H y the Hermitian adjoint of H ,
Tr(cid:0)(cid:23) @ 2
y (cid:26)(cid:1):
The potential V in H y implies an annihilation, the
di(cid:11)usion process is \killed" with a rate V
(cid:21) dt.
Finally we (cid:12)nd the optimal cost-to-go from equation
(13),
J (x; t) = (cid:0)(cid:21) log Z dy(cid:26)(y ; T jx; t)e(cid:0)(cid:30)(y)=(cid:21) :
References
(15)
[1] C. Boutilier. Planning, learning and coordination in
multiagent decision processes.
In TARK, volume 6,
pages 195{210, 1996.
[2] C. Guestrin, D. Koller, and R. Parr. Multiagent plan-
ning with factored MDPs. In NIPS, volume 14, pages
1523{1530, 2002.
[3] C. Guestrin, S. Venkataraman, and D. Koller. Context-
speci(cid:12)c multiagent coordination and planning with fac-
tored MDPs.
In AAAI, volume 18, pages 253{259,
2002.
[4] H. J. Kappen.
Linear theory for control of non-
linear stochastic systems. Physical Review Letters,
95(20):200201, November 2005.
[5] H. J. Kappen. Path integrals and symmetry breaking
for optimal control theory. Journal of statistical me-
chanics: theory and experiment, page P11011, Novem-
ber 2005.
[6] R. Stengel. Optimal Control and Estimation. Dover,
New York, 1993.
[7] S.L. Lauritzen and D.J. Spiegelhalter. Local compu-
tations with probabilties on graphical structures and
their application to expert systems (with discussion).
J. Royal Statistical Society Series B, 50:157{224, 1988.
|
1911.10055 | 1 | 1911 | 2019-11-22T14:13:36 | Towards a Goal-oriented Agent-based Simulation framework for High-Performance Computing | [
"cs.MA"
] | Currently, agent-based simulation frameworks force the user to choose between simulations involving a large number of agents (at the expense of limited agent reasoning capability) or simulations including agents with increased reasoning capabilities (at the expense of a limited number of agents per simulation). This paper describes a first attempt at putting goal-oriented agents into large agent-based (micro-)simulations. We discuss a model for goal-oriented agents in High-Performance Computing (HPC) and then briefly discuss its implementation in PyCOMPSs (a library that eases the parallelisation of tasks) to build such a platform that benefits from a large number of agents with the capacity to execute complex cognitive agents. | cs.MA | cs |
Towards a Goal-oriented Agent-based
Simulation framework for
High-Performance Computing
Dmitry GNATYSHAK a, Luis OLIVA-FELIPE a, Sergio ´ALVAREZ-NAPAGAO b,
Julian PADGET c, Javier V ´AZQUEZ-SALCEDA a, Dario GARCIA-GASULLA b and
Ulises CORT ´ES a,b
a Universitat Polit`ecnica de Catalunya - BarcelonaTECH (UPC)
C/Jordi Girona 1-3, E-08034 , Barcelona, Spain
b Barcelona Supercomputer Centre (BSC)
C/Jordi Girona 1-3, E-08034 , Barcelona, Spain
c Department of Computer Science
University of Bath, BATH BA2 7AY , United Kingdom
Abstract. Currently, agent-based simulation frameworks force the user to choose
between simulations involving a large number of agents (at the expense of lim-
ited agent reasoning capability) or simulations including agents with increased rea-
soning capabilities (at the expense of a limited number of agents per simulation).
This paper describes a first attempt at putting goal-oriented agents into large agent-
based (micro-)simulations. We discuss a model for goal-oriented agents in High-
Performance Computing (HPC) and then briefly discuss its implementation in Py-
COMPSs (a library that eases the parallelisation of tasks) to build such a platform
that benefits from a large number of agents with the capacity to execute complex
cognitive agents.
Keywords. agent-based simulation, high-performance computing, agent platform,
multi-agent system, goal-oriented agent
1. Introduction
Agent-based simulation (ABS) faces the conflicting demands of providing larger simu-
lations (in terms of the number of agents) and providing agents with better cognitive and
decision-making capabilities (which tend not to scale well in large simulations). This
paper is a first attempt at addressing this friction, through the use of High-Performance
Computing (HPC), with the aim of enabling scenarios in which large populations of in-
dividuals (e.g. traffic simulation, industrial and urban areas along geographical areas)
have the ability to perform normative reasoning, planning or even BDI-like behaviour,
in order to explore the analysis of phenomena resulting from more detailed behavioural
modelling. On the one hand, current HPC-based approaches[12] seem to focus on large
simulations with limited interaction or perception or with reactive-like agents. In some
cases, planning is even hard-coded. On the other hand, [2] lists a wide range of ABS
1
(a) PyCOMPSs runtime structure
(b) PyCOMPSs object handling scheme
Figure 1. Aspects of PYCOMPSs
which seem to hardly offer scalability and agents with high cognitive capabilities. This is
an indication of an existing trade-off between scalability and agents with complex delib-
eration processes. In Section 2, we discuss a model for goal-oriented agents in HPC and
briefly outline its implementation in PyCOMPSs (a library that eases task parallelisation
through annotation) to build such a platform. In Section 3, we present the test scenarios
used to evaluate the performance of the platform and demonstrate its capabilities and
potential. In Section 4, we summarise the main results, the contributions and we identify
the next steps.
2. Proposed Model
This section provides background about the COMPSs HPC framework and the corre-
sponding Python package, and then presents the proposed agent-based simulation model.
2.1. COMPSs and PyCOMPSs
COMPSs [15] is a framework based on the Grid Component Model (GCM)[1] and Ser-
viceSs model[17] developed by the Barcelona Supercomputing Center. Its main purpose
is to allow the development of distributed cloud- or grid-based applications without the
need to deal with the specifics of underlying execution systems. Thus, it provides an
abstraction layer, allowing the development of hardware-configuration-agnostic applica-
tions. These applications can be distributed automatically, saving effort in accounting
for low-level aspects of the target hardware. COMPSs analyses the data dependencies
among the user-specified functions (called tasks as in the GCM model) of a sequential
program and runs as many of those tasks in parallel as it is safe to do. A set of additional
commands and parameters can be used to fine-tune the runtime execution and paralleli-
sation. The core of this runtime is written in Java, but C/C++ and Python interfaces are
also available. As all of these commands are language-specific, for the purpose of this
work we choose to focus on the Python version: PyCOMPSs[16].
From a coding perspective, the COMPSs interface takes the form of annotations
(decorators) in the Python code and several API functions and procedures. The main
annotation is @Task(...), where the user specifies the return type (a type or class name)
and the data direction of inputs (in, out) for mutable types and classes.The following
code shows an example of a PyCOMPSs task annotation:
2
1 @Task ( par1 = INOUT , par3 = OUT , returns = list )
2 def my_func ( par1 , par2 , par3 ) :
By means of such annotations, we may mark global functions, class methods, and in-
stance methods as tasks to be automatically parallelised.
2.2. Conceptual Model: Multi-agent system elements
We define a Multi-agent System M as the tuple M = {E, A +, C} where:
1. E is an environment, in which the agents are situated, that they can perceive, and
on which they can act
2. A + is a non-empty set of agents
3. C is a controller that maintains the environment and handles communications be-
tween agents
An Agent Ai is defined as Ai = {ID,msgQs,Bh,B,G, P,outAcs} where:
2. msgQs = {I , O} is a set of the Ai's message queues
1. ID = {AgID,AgDesc} is Ai's identity data
(a) AgID is the unique identifier for Ai
(b) AgDesc is an arbitrary description of Ai
(a) I = {. . . ,msgi, . . .} is the inbox, a set of messages sent to Ai
(b) O = {. . . ,msgi, . . .} is the outbox, a set of messages sent by Ai
(c) msgi = {AgIDs,AgIDr, per f ormative,content, priority} is a message from
agent Ags to agent Agr with the given performative, content, and priority.
per f ormatives are FIPA-compliant.
3. Bh = {MendR,RG} is Ai's role behaviour
(a) MendR is a means-ends reasoner used to generate plans (see Section 2.5)
(b) RG is a set of goals associated with this role behaviour
4. B is the set of Ai's beliefs
5. G is the set of Ai's goals
6. P = {. . . ,abi, . . .} is the current plan of Ai
7. abi =(cid:8). . . ,ai j, . . .(cid:9) is an action block, an ordered set of agent actions, for which
we distinguish three types: internal actions executed by the agents and intended
to change their beliefs external actions sent to the controller to be executed by it
on the environment, and message actions to generate messages to other agents
(a) outAcsi = {senderID,ae} is a tuple of sender ID and an external action
8. outAcs is a set of external actions to be executed on the environment
The Controller C is defined by the tuple C = {I ,inAcs} where I is an inbox for
the reception of messages from all the agents and inAcs is the set of all the actions to be
applied on the environment.
2.3. Operational semantics of the Agents' deliberation cycle: agent transition rules
In the previous section we have defined the different elements that compose our frame-
work. Using these definitions, we can introduce the operational semantics of our model
by means of a set of transition rules. The model assumes that M evolves as a set of
simulation steps. These transition rules describe how the agents' internal states are trans-
3
B perceive(E)
−−−−−−−→ B(cid:48)
Ai{·,·,·,B,·,·,·} ,E →
Ai{·,·,·,B(cid:48),·,·,·} ,E
G goal check(B,G)
−−−−−−−−−−→ G(cid:48)
Ai{·,·,·,·,G,·,·} ,· →
Ai{·,·,·,·,G(cid:48),·,·} ,·
P (h,t)
−−−−−−→ P(cid:48) (t)
execute(h)
Ai{·,{I , O} ,·,B,·, P,outAcs} ,· →
Ai{·, (I , O(cid:48)) ,·,B(cid:48),·, P(cid:48),outAcs(cid:48)} ,·
(1)
(3)
(5)
−−−−−−→ B(cid:48),G(cid:48), O(cid:48)
B,G, O process(h)
Ai{·,{{h,t} , O} ,·,B,G,·,·} ,· →
Ai{·,{{t} , O(cid:48)} ,·,B(cid:48),G(cid:48),·,·} ,·
(2)
P = ∅∨ P = Fail MendR(B,G,RG)
−−−−−−−−−−→ P(cid:48)
Ai{·,·,·,·,·, P,·} ,· →
Ai{·,·,·,·,·, P(cid:48),·} ,·
(4)
(6)
Bh role check(B,G,RG)
−−−−−−−−−−−−→ Bh(cid:48)
Ai{·,·,Bh,·,·,·,·} ,· →
Ai{·,·,Bh(cid:48),·,·,·,·} ,·
Figure 2. Definitions for the agent transition (rules 1 to 6). We use · to denote elements in the agent tuple that
are not modified in the transition.
formed over each single step of the simulation. They show the part of the state that is
changed, the function that changes it, and a summary of the transformation.
Agents in our model are goal-oriented deliberative agents inspired by the Beliefs-
Desires-Intentions (BDI) Model[11]. Therefore, the agents' reasoning cycle is split into
a perception phase (where beliefs are updated), a deliberation phase (where goals are
updated and prioritised), a means-ends reasoning phase (where plans are selected/built)
and an actuation phase. As in many modern BDI-inspired agent languages (e.g., JADEX,
2APL, GOAL), in our model plans are selected directly from beliefs and goals without
the need for intermediate intention selection (which forces, in the theoretical BDI model,
to only pursue one goal at a time).
The first sub-step in the agent deliberation cycle happens when the agent perceives
the current state of the environment. This perception only modifies the agent's beliefs
which are then used in the following sub-steps. Formally, this sub-step is formalised in
Rule 1 (formulae are provided in Figure 2):
Rule 1: Perceiving the environment, where perceive (. . . ) is a user-defined function that
transforms the agent A 's beliefs based on the environment state E.
During the second sub-step, the agent processes the incoming messages. This is done
sequentially, in an arbitrary order. Message processing may modify the agent's beliefs,
as well as its goals. This is formalised in Rule 2:
Rule 2 Message processing, where process (. . . ) is a user-defined message processing
function that can modify the agent A 's beliefs, goals, and outbox.
In the third sub-step of the deliberation cycle, the agent gets an opportunity to reeval-
uate and change its own goals. The decision is based on its beliefs and current goals, and
only the latter is changed during the process. This transition is defined in Rule 3:
4
Rule 3 Goal check, where goal check (. . . ) is a user-defined function that may change
the agent A 's goals based on its beliefs.
One of the most important sub-steps is the actual reasoning. In this step, the agent
relies on a means-ends reasoner to generate a plan that will be followed by itself after-
wards. This step is only performed if the current plan has failed or finished.1 Formally,
this step is defined in Rule 4:
Rule 4 Means-ends reasoner, where MendR (. . . ) is a means-ends reasoner that gener-
ates plans based on the agent A 's beliefs and goals.
After the plan has been updated, the agent proceeds to execute it. The next action
block is extracted from the plan and the set of its actions are run one after another.
Rule 5 Action execution, where execute (. . . ) is the action execution function for action
h. If h is an internal action, execute runs h to modify the set of agent A 's beliefs; if h
is an external action, execute appends h to the set of outgoing external actions; if h is a
message action, execute runs h to generate messages to send.
Finally, the agent has an optional choice to reevaluate its role behaviour (i.e. Bh).
This sub-step is similar to the goal check and its result can be either a new role, or no
changes.
Rule 6 Role check (optional), where role check (. . . ) is a user-defined function that
may modify the agent A 's role behaviour and role goals, based on its beliefs and goals.
2.4. Operational semantics of Controller and Environment: MAS transition rules
After formally defining the internal deliberation process of agents in our framework, we
now describe how the Controller changes the environment on each simulation step.
First of all, there is an optional sub-step during which the controller can modify the
environment before agents' actions are applied to it. These actions are considered to be
commutative. Formally, we can express it in the following way:
Rule 7 Pre-action execution step (optional), where pre step () is a user-defined function
that modifies the environment before the agents' actions are collectively applied to it.
After that, the controller acts as a message dispatcher, processing all the outgoing
messages and passing them to the corresponding recipients in a sequential manner.
Rule 8 Message forwarding, where f wdmsh (. . . ) is a function that moves the specified
message from the sender's inbox, to the recipient's outbox.
The next sub-step executes, sequentially, all the actions sent by agents.
Rule 9 Action execution, where execute ac (. . . ) is a function that applies the action h
to the environment.
Finally, there is another optional sub-step to modify the environment. It is similar to
the first sub-step, differing only in the function used.
Rule 10 Post-action execution step (optional), where post step () is a user-defined
function that modifies the environment after the agents' actions are applied to it.
1See § 2.5 for more details.
5
E pre step()
−−−−−→ E(cid:48)
M {E,·,·} → M {E(cid:48),·,·}
−−−−−−−→ E(cid:48)
E execute ac(a)
M {E,·,{·,{a,t}}} →
M {E(cid:48),·,{·,{t}}}
(7)
(9)
Ar {·,{{x} , O} ,·,·,·,·,·} f wd msg(h,Ar)
−−−−−−−−→
r {·,{{x,h} , O} ,·,·,·,·,·}
A (cid:48)
M {·,{. . . , Ar, . . .} ,{{h,t} ,·}} →
r , . . .} ,{{t} ,·}}
M {·,{. . . , A (cid:48)
(8)
E post step()
−−−−−−→ E(cid:48)
M {E,·,·} → M {E(cid:48),·,·}
(10)
Figure 3. Definitions for the MAS transition rules (rules 7 to 10)
2.5. Means-ends reasoning: Hierarchical Task Networks planner
In our base model, we choose not to specify any fixed means-to-ends reasoning model
(MendR (B,G,RG) in agent Rule 4), leaving it to the user to select one (and to pro-
vide the required specification). However, for our implementation of the base model,
we choose a planner based on Hierarchical Task Networks (HTN)[8] to instantiate the
MendR function, on the grounds that HTN planners are a good fit to the means-end stage
of the BDI reasoning process[14].There are several HTN-based models, with different
representation and planning procedures. As most of these models deal only with high-
level planning, at a distance from the practical planning and following task execution,
we considered HTN planner models used for actual computer games. These models have
been adapted not only for effective abstract-to-concrete facts mapping, but also for re-
planning during execution.
We slightly modified [9] to suit both our needs and the COMPSs requirements.
These modifications cover the difference in the perception models (the original algorithm
used active sensors), and allow the partial sets of method sub-tasks to be accepted as
parts of the plan to gain flexibility. In this model, abstract tasks are divided into com-
pound tasks and methods. A compound task consists only of an ordered set of methods
while each method consists of an ordered set of compound or primitive tasks, conditions
and some additional information. Primitive tasks contain action blocks to be executed,
consisting of action of various types, and also some preconditions.
To connect the output of the deliberation sub-step with the HTN planner, we consider
the goals provided by the goal check function (rule 3) as tasks to be searched in the HTN.
By following the model planning rules associated with the (abstract) task [9], we can
easily obtain a sequence of actions to be executed based on the current beliefs and goals.
After the current plan has finished, we generate a new one, as per rule 4. Thus, the only
explicit goal type supported currently is the maintain goal[7], but other types of goals
(for example, achieve goals or perform goals) can be simulated by the appropriate use
of the goal check function. In the case of HTN as a means-ends reasoner, goal check ()
can easily switch between different compound tasks in HTN (or select a new HTN) as
6
a method of goal revision. This can be done either by traversing the graph of the HTN,
or by generating a new HTN and replacing the old one, although one should always be
careful about the belief sets and ontologies used by each planner.
3. Implementation and Results
In the previous section, we introduced the agent-based simulation model and described
its design features. This section describes an instantiation of the framework in a concrete
implementation and provides some preliminary test results.
3.1. Implementation
The whole system is implemented in Python. It is organised as a package with several
modules and a list of main classes that are easily accessible by the user:
Controller: defines the Controller class, the main utility entity in the system, respon-
sible for containing the MAS, running PyCOMPSs tasks, forwarding messages,
and executing all the utility methods.
Agent: defines the Agent class that is a container for the agent description.
Behaviour: defines the Behaviour class, that implicitly provides the agent's workflow;
the user may subclass this for complex behaviours.
Directory: defines the Directory and DirectoryEntry classes that are used by the
built-in directory facilitator.
HTN: defines all the HTN-related classes, as well as some related functionality;
the full list is: CompoundTask, Method, PrimitiveTask, Effect, Action,
ActionBlock, HTNPlanner, BeliefSet and Conditions.
Messaging: defines the Message and Messagebox classes used for agent messaging.
State: defines the State structure, used to transfer and work with persistent states.
The Behaviour class follows the model presented in Section 2.3 with its instance
functions corresponding to the functions in the rules. In complex scenarios, it is expected
that the user redefines the following functions:
1 state . beliefs = self . p e r c e i v e ( environment , state . beliefs )
2 for message in inbox :
3
state . beliefs , state . planner , reply = self . process ( message , state
. beliefs , state . planner )
4 state . planner = self . g o a l _ c h e c k ( state . beliefs , state . planner )
5 role = self . r o l e _ c h e c k ( state . beliefs )
A simulation is run by initialising a Controller object, registering agents in it, and
calling the run method. An example of the latter two:
1 c o n t r o l l e r . g e n e r a t e _ a g e n t ( " Test agent " , b e h a v i o r = M y B e h a v i o r () ,
beliefs = myBeliefs , i n i t _ b l o c k = None , planner = myPlanner ,
d e f a u l t _ b l o c k = None , s e r v i c e s =[ " testing " ] , r e g i s t e r = True )
2 c o n t r o l l e r . run ( n u m _ i t e r =10 , p e r f o r m a n c e = False )
An HTN planner can be generated following the structure defined in Section 2.5.
The user needs to create a structure of PrimitiveTasks, Methods, and CompoundTasks
and use the root CompoundTask as a parameter for the planner's initialisation.
7
Number
of agents
10
50
100
200
300
500
1000
2000
Time, s
Total
29.5460
75.2455
132.4631
308.3213
539.4989
972.8603
1775.0650
2872.2980
Tasks
28.5788
72.3717
128.8426
302.9179
530.3990
959.3281
1747.6440
2834.4890
Table 1. Results for the agent count test. All the results are averages of 5 replications.
3.2. Performance experiments
In order to test the system's performance, we have designed a number of experiment sets
to test different aspects of the system. The overall time performance of the system was
evaluated in accordance to 4 parameters: number of agents (100), number of requested
processing units (246), number of messages sent (10), and size of messages sent (0Kb).
For each series of tests one of the parameters was modified while the others were fixed
at default values (the ones inside the brackets above). Due to space limitations we will
only focus on the first two.
The Agents' behaviours are guided by a trivial HTN with a single compound task,
single method, and single primitive task with 2 actions: send messages and increment
step counter belief. On each turn, each agent sent a specified number of messages con-
taining lists of zeroes of specified length to random agents and incremented its "step"
counter belief. Upon the reception of a message, each agent incremented its "counter"
belief.
All of these experiments, with two exceptions, were performed on the BSC NordIII
cluster using the default setting of 256 processors cores per test.
In the case of the test of number of agents (Table 1), the behaviour of the chart fits
the expectations. The chart increases linearly w.r.t the increase of the total number of
agents (and, correspondingly, the number of PyCOMPSs tasks). Also, even if it is not
clear enough from the plot, the controller time noticeably increases due to the fact that
the controller has to process additional messages and handle these new agents.
But the most interesting results come from the experiments on the number of pro-
cesses (Table 2). For them we have two sets of columns: the left ones show the results
for the default tests with a specific agent behaviour, the right ones are for the slightly
modified version of it, where on each step we have added a one-second delay. We can see
that, in the case without delays, there is only a slight dependency between the number
of processes and the execution time. But as soon as we imitate the harder tasks that take
more time to execute, even just adding one second, the difference becomes massive as
we get a clear exponential dependency.
Also please notice that after the number of processes exceeds the number of tasks,
the changes in the computation time are almost non-existent, as COMPSs has enough
resources to get all the tasks distributed with a minimal delay.
Besides the results on the tested metrics, we have obtained insights about both possi-
ble weaknesses of the system and directions for future research. The main issue we faced
8
Number
of processes
Number
of nodes
2
4
6
8
10
12
14
32
64
96
128
160
192
224
Time, s
(w/o delay)
Time, s
(with delay)
Tasks
223.4435
159.7769
161.5091
179.6115
164.2693
125.6789
128.8578
Total
220.9405
157.7344
159.1821
176.8729
159.5374
122.7972
125.1992
Tasks
829.7442
373.3008
252.5065
194.2150
195.0013
195.0007
196.3293
Total
826.4027
370.3567
250.0795
192.0736
191.9677
192.0937
193.1141
Table 2. Results for the test for the number of processes. 1 second delays were introduced to each task in the
second case.
was disk space usage. As COMPSs transforms custom objects into files for transfer, this
puts a strain on the data transfer infrastructure. This may result in exceeding cluster disk
space quotas. This limits the scale of the results we are able to achieve (although they
still exceed the standard capabilities of the non-HPC BDI platforms).
4. Conclusions
This paper has presented a model for agent-based simulation in HPC and its implemen-
tation in PyCOMPSs. Our simulation model is driven by the controller, which can be
effectively considered as the synchronisation point for the simulation, following the Bulk
Synchronous Parallel model[18]. The main advantage of our model when compared to
the closest works in literature[4,6,5,13] is that our model would benefit from simulation
domains that require goal-driven agents being able to perform more complex reasoning
or planning. Another limitation of these frameworks (except for [4]) is that the agent
communication is very limited, based on direct method calls. This kind of communica-
tion only works properly within agents executed on the same processor or in the over-
lapping zones, but not with agents in other processors. We are starting to test our model
with a real scenario base on wastewater management of a full river basin with hundreds
of pollutant producers and dozens of pollutant processors coordinating their effort to en-
sure pollution levels are law-compliant. Future research lines include other simulation
problems and implementing normative reasoning by incorporating a norm monitor [3] or
deontic sensors [10], which would also facilitate the introduction of normative contexts.
Acknowledgements
This work is partially supported by the BSC-IBM Deep Learning Center agreement, the Spanish
Government through Programa Severo Ochoa (SEV-2015-0493), the 11th call on the Severo Ochoa
Mobility Program in BSC, the Spanish Ministry of Science and Technology through TIN2015-
65316-P project and the Generalitat de Catalunya (contract 2017-SGR-1414).
References
[1] Basic features of the grid component model (assessed). CoreGRID Deliverable D.PM.04, 2007.
9
[2] Sameera Abar, Georgios K. Theodoropoulos, Pierre Lemarinier, and Gregory M.P. O'Hare. Agent based
modelling and simulation tools: A review of the state-of-art software. Computer Science Review, 24:13
-- 33, 2017.
[3] Sergio ´Alvarez Napagao. Bringing social reality to multiagent and service architectures: practical
reductions for monitoring of deontic-logic and constitutive norms. PhD thesis, Universitat Polit`ecnica
de Catalunya, 2016.
[4] S. Coakley, M. Gheorghe, M. Holcombe, S. Chin, D. Worth, and C. Greenough. Exploitation of high
performance computing in the flame agent-based simulation framework. In 2012 IEEE 14th Interna-
tional Conference on High Performance Computing and Communication 2012 IEEE 9th International
Conference on Embedded Software and Systems, pages 538 -- 545, June 2012.
[5] Nicholson Collier and Michael North. Repast hpc: A platform for large-scale agent-based modeling.
Large-Scale Computing, pages 81 -- 109, 04 2012.
[6] Gennaro Cordasco et al. A framework for distributing agent-based simulations. In Michael Alexander et
al., editor, Euro-Par 2011: Parallel Processing Workshops, pages 460 -- 470, Berlin, Heidelberg, 2012.
Springer Berlin Heidelberg.
[7] Mehdi Dastani, M. Birna van Riemsdijk, and John-Jules Ch. Meyer. Goal types in agent programming.
In Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Sys-
tems, AAMAS '06, pages 1285 -- 1287, New York, NY, USA, 2006. ACM.
[8] Kutluhan Erol, James Hendler, and Dana Nau. Htn planning: Complexity and expressivity. Proceedings
of the National Conference on Artificial Intelligence, 2, 05 1994.
[9] Troy Humphreys. Exploring HTN Planners through Example, chapter Architecture, pages 149 -- 167.
[10]
CRC Press, September 2013.
In Proceedings of the 27th
Julian Padget, Marina De Vos, and Charlie Ann Page. Deontic sensors.
International Joint Conference on Artificial Intelligence, IJCAI'18, pages 475 -- 481. AAAI Press, 2018.
[11] Anand S. Rao and Michael P. Georgeff. Modeling rational agents within a BDI-architecture. In James
Allen, Richard Fikes, and Erik Sandewall, editors, Proceedings of the 2nd International Conference on
Principles of Knowledge Representation and Reasoning, pages 473 -- 484. Morgan Kaufmann publishers
Inc.: San Mateo, CA, USA, 1991.
[12] Alban Rousset, B´en´edicte Herrmann, Christophe Lang, and Laurent Philippe. A survey on parallel
and distributed multi-agent systems for high performance computing simulations. Computer Science
Review, 22:27 -- 46, 2016.
[13] Xavier Rubio-Campillo. Pandora: A versatile agent-based modelling platform for social simulation.
Proceedings of SIMUL, pages 29 -- 34, 2014.
[14] Sebastian Sardina and Lin Padgham. A bdi agent programming language with failure handling, declar-
ative goals, and planning. Autonomous Agents and Multi-Agent Systems, 23:18 -- 70, 07 2011.
[15] Enric Tejedor and Rosa M. Badia. Comp superscalar: bringing grid superscalar and gcm together. In
2008 Eighth IEEE International Symposium on Cluster Computing and the Grid (CCGRID), pages 185 --
193. IEEE, 06 2008.
[16] Enric Tejedor, Yolanda Becerra, Guillem Alomar, Anna Queralt, Rosa M. Badia, Jordi Torres, Toni
Cortes, and Jes´us Labarta. Pycompss: Parallel computational workflows in python. The International
Journal of High Performance Computing Applications, 31(1):66 -- 82, 2017.
[17] Enric Tejedor, Jorge Ejarque, Francesc Lordan, Roger Rafanell, Javier ´Alvarez Cid-Fuentes, Daniele
Lezzi, Raul Sirvent, and Rosa M. Badia. A cloud-unaware programming model for easy development
of composite services. In 2011 IEEE Third International Conference on Cloud Computing Technology
and Science, pages 375 -- 382. IEEE, 11 2011.
[18] Leslie G. Valiant. A bridging model for parallel computation. Commun. ACM, 33(8):103 -- 111, August
1990.
10
|
1910.02591 | 1 | 1910 | 2019-10-07T03:32:41 | Multi-Agent Reinforcement Learning for Order-dispatching via Order-Vehicle Distribution Matching | [
"cs.MA",
"cs.LG"
] | Improving the efficiency of dispatching orders to vehicles is a research hotspot in online ride-hailing systems. Most of the existing solutions for order-dispatching are centralized controlling, which require to consider all possible matches between available orders and vehicles. For large-scale ride-sharing platforms, there are thousands of vehicles and orders to be matched at every second which is of very high computational cost. In this paper, we propose a decentralized execution order-dispatching method based on multi-agent reinforcement learning to address the large-scale order-dispatching problem. Different from the previous cooperative multi-agent reinforcement learning algorithms, in our method, all agents work independently with the guidance from an evaluation of the joint policy since there is no need for communication or explicit cooperation between agents. Furthermore, we use KL-divergence optimization at each time step to speed up the learning process and to balance the vehicles (supply) and orders (demand). Experiments on both the explanatory environment and real-world simulator show that the proposed method outperforms the baselines in terms of accumulated driver income (ADI) and Order Response Rate (ORR) in various traffic environments. Besides, with the support of the online platform of Didi Chuxing, we designed a hybrid system to deploy our model. | cs.MA | cs | Multi-Agent Reinforcement Learning for Order-dispatching via
Order-Vehicle Distribution Matching
Ming Zhou1, Jarui Jin1, Weinan Zhang1, Zhiwei Qin2, Yan Jiao2,
Chenxi Wang2, Guobin Wu3, Yong Yu1, Jieping Ye2
1Shanghai Jiao Tong University, 2DiDi AI Labs, 3DiDi Research
{zhouming,yyu}@apex.sjtu.edu.cn,[email protected],[email protected],{qinzhiwei,yanjiao,wangchenxi,
wuguobin,yejieping}@didiglobal.com
9
1
0
2
t
c
O
7
]
A
M
.
s
c
[
1
v
1
9
5
2
0
.
0
1
9
1
:
v
i
X
r
a
ABSTRACT
Improving the e(cid:129)ciency of dispatching orders to vehicles is a re-
search hotspot in online ride-hailing systems. Most of the existing
solutions for order-dispatching are centralized controlling, which
require to consider all possible matches between available orders
and vehicles. For large-scale ride-sharing platforms, there are thou-
sands of vehicles and orders to be matched at every second which is
of very high computational cost. In this paper, we propose a decen-
tralized execution order-dispatching method based on multi-agent
reinforcement learning to address the large-scale order-dispatching
problem. Di(cid:130)erent from the previous cooperative multi-agent re-
inforcement learning algorithms, in our method, all agents work
independently with the guidance from an evaluation of the joint
policy since there is no need for communication or explicit co-
operation between agents. Furthermore, we use KL-divergence
optimization at each time step to speed up the learning process and
to balance the vehicles (supply) and orders (demand). Experiments
on both the explanatory environment and real-world simulator
show that the proposed method outperforms the baselines in terms
of accumulated driver income (ADI) and Order Response Rate (ORR)
in various tra(cid:129)c environments. Besides, with the support of the
online platform of Didi Chuxing, we designed a hybrid system to
deploy our model.
CCS CONCEPTS
•Computing methodologies→ Multi-agent reinforcement learn-
ing; •(cid:135)eory of computation → Multi-agent reinforcement learn-
ing; •Applied computing → Transportation;
KEYWORDS
Deep Reinforcement Learning; Multi-agent Reinforcement Learn-
ing; Ride-Hailing; Order-Dispatching
ACM Reference format:
Ming Zhou1, Jarui Jin1, Weinan Zhang1, Zhiwei Qin2, Yan Jiao2,
Chenxi Wang2, Guobin Wu3, Yong Yu1, Jieping Ye2. 2019. Multi-Agent
Reinforcement Learning for Order-dispatching via
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for pro(cid:128)t or commercial advantage and that copies bear this notice and the full citation
on the (cid:128)rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permi(cid:138)ed. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior speci(cid:128)c permission and/or a
fee. Request permissions from [email protected].
CIKM '19, Beijing, China
© 2019 ACM. 978-1-4503-6976-3/19/11...$15.00
DOI: 10.1145/3357384.3357799
Order-Vehicle Distribution Matching. In Proceedings of (cid:138)e 28th ACM In-
ternational Conference on Information and Knowledge Management, Beijing,
China, November 3 -- 7, 2019 (CIKM '19), 9 pages.
DOI: 10.1145/3357384.3357799
1 INTRODUCTION
With the booming of mobile internet, it becomes feasible and
promising to establish the modern large-scale ride-hailing systems
such as Uber, Didi Chuxing and Ly(cid:137) which allow passengers book
routes with smartphones and match available vehicles to them
based on intelligent algorithms. To some extent, these ride-hailing
systems improve the e(cid:129)ciency of the transportation system.
In ride-hailing systems, a key point is how to dispatch orders
to vehicles to make the system work more e(cid:129)ciently and generate
more impact. We illustrate the order-dispatching in Figure. 1, where
one can see that the algorithm used by the decision maker is critical
for (cid:128)nding suitable matches because the result of order-dispatching
has direct in(cid:131)uences on the platform e(cid:129)ciency and income.
(cid:140)e general strategies of automatically order-dispatching sys-
tems are to minimize the waiting time and taxi cruising time through
route planning or matching the nearest orders and vehicles [4, 5,
11, 23]. In recent research, another approach to solve the order-
dispatching problem is to leverage combinatorial optimization [25]
to improve the success rate of order-dispatching [39]. It makes
a signi(cid:128)cant improvement in the online test, but it su(cid:130)ers from
high computational cost, and strongly relies on appropriate feature
engineering. More importantly, the above strategies are myopic:
they may (cid:128)nd suitable matches in the current stage, but ignore the
potential future impact.
In this paper, we focus on developing a method to maximize the
accumulated driver income (ADI), i.e., the impact of orders served
in one day, and the order response rate (ORR), i.e., the proportion
of served orders to the total orders in one day. Intuitively, match-
ing vehicles with high-price orders can receive high impact at a
single order-dispatching stage. However, if the served orders re-
sult in the mismatch between the orders and vehicles in the near
future, it would harm the overall service quality in terms of ORR
and the long-term ADI. Hence, in order to (cid:128)nd a balance between
the long-term ADI and ORR, it is necessary to develop an order-
dispatching algorithm which takes the future supply and demand
into consideration.
Xu et al. [37] proposed a planning and learning method based on
decentralized multi-agent deep reinforcement learning (MARL) and
centralized combinatorial optimization to optimize the long-term
ADI and ORR. (cid:140)e method formulates the order-dispatching task
into a sequential decision-making problem and treats a vehicle as
an agent. However, for centralized approaches, a critical issue is
the potential "single point of failure" [18], i.e., the failure of the cen-
tralized authority control will fail the whole system [16]. Another
two related work using multi-agent to learn order-dispatching is
based on mean-(cid:128)eld MARL [13] and knowledge transferring [35].
(cid:140)ere are some challenges to be solved when we apply the MARL
to the real-time order-dispatching scenario. First, handling the non-
stationary environment in MARL is a major problem, which means
that all agents learn policies concurrently, while each individual
agent does not know the policies of other agents [10]. (cid:140)e state
transition in a multi-agent environment is driven by all agents
together, so it is important for agents to have knowledge about
other agents' policies. In the order-dispatching scenario, we only
care about the idle status of a vehicle since they are available for
order-dispatching. However, as the duration of each order is non-
deterministic, compared to the traditional multi-agent scenarios
which have deterministic time interval, it is di(cid:129)cult to learn the
interactions between agents in successive idle states, which makes
many MARL methods including opponent modeling [2, 26] and
communication mechanism [7] hard to work well. Second, the
number of idle vehicles keeps changing during the whole episode,
i.e., there will always be some vehicles ge(cid:138)ing o(cid:132)ine or online,
thus the general MARL methods which require (cid:128)xed agent number
cannot be directly applied in such a case [6, 40].
In addition, we believe that a higher ORR usually means a higher
ADI, and if we can maintain a higher long-term ORR, we will get
a higher long-term ADI. With regard to the correctness of this
point, we also conducted a corresponding experimental analysis in
Section 4.2.2.
To the best of our knowledge, this is the (cid:128)rst work that utilizes
this character to improve both ORR and ADI. In detail, we propose a
centralized learning and decentralized execution MARL method to
solve the above challenges with an extension of Double Q-learning
Network [20] with Kullback-Leibler (KL) divergence optimization.
Besides, the KL-based backward learning optimization method also
speeds up the agents learning process with the help of others'.
Considering the large scale of agents, and they are homogeneous,
we learn only one network using parameter sharing, and share
learning experiences among all agents at the training stage, as that
in [30, 40]. To address the non-stationary action space problem, in
our implementation, the input of deep Q-learning network consists
of the state and selected action.
Extensive experiments with di(cid:130)erent tra(cid:129)c and order condi-
tions and real-world simulation experiments are conducted. (cid:140)e
experimental results demonstrate that our method yields a large im-
provement on both ADI and ORR compared to the baseline methods
in various tra(cid:129)c environments. We also claim the proposed method
is highly feasible to be deployed on the existing order-dispatching
platform.
2 RELATED WORK
Taxi-order Dispatching. (cid:140)ere have been several GPS-based
order-dispatching systems to enhance the accuracy, communica-
tions, and productivity of taxi dispatching [14, 15, 23]. (cid:140)ese sys-
tems do not o(cid:130)er detailed dispatching algorithms, which means
Figure 1: Ride-hailing order-dispatching process
these platforms are more like information sharing platforms, help-
ing vehicles choose orders to serve by o(cid:130)ering orders information.
Other automatic order-dispatching methods [12, 19] focus on re-
ducing the pick-up distance or waiting time by (cid:128)nding the nearest
orders. While these methods usually fail to reach a high success
rate on order-dispatching and ignore many potential orders in the
waiting list which may be more suitable for vehicles. Zhang et
al. [39] proposed a centralized control dispatching system based
on combinatorial optimization. Although it is a simple method,
the requirement of computing all available order-vehicle matches
can be of much high computational cost in a large-scale taxi-order-
dispatching situation. Moreover, it requires appropriate feature
engineering. (cid:140)us it greatly increases the system implementation
di(cid:129)culty and human e(cid:130)orts of applying the method in a practical
situation.
Multi-agent Reinforcement Learning. Multi-agent reinforce-
ment learning has been applied in domains like collaborative deci-
sion support systems. Di(cid:130)erent from the single agent reinforcement
learning (RL), multi-agent RL needs the agents to learn to cooper-
ate with others. It is generally impossible to know other policies
since the learning process of all agents is simultaneous. (cid:140)us for
each agent, the environment is non-stationary [3]. It is problem-
atical that directly apply the independent reinforcement learning
methods into the multi-agent environment. (cid:140)ere are several ap-
proaches proposed to relieve or address this problem, including
sharing the policy parameters [8], training the Q-function with
other agents' policy parameters [31], centralized training [17] and
opponent modeling [2, 26]. Besides, there are also some methods
which use explicit communication to o(cid:130)er a relatively stationary
environment for peer agents [7, 9, 30]. In the large-scale multi-
agent systems, the non-stationary problem will be ampli(cid:128)ed. To
address this problem, Yang et al. [38] proposed a novel method
which converts multi-agent learning into a two-player stochastic
game [28] by applying mean (cid:128)eld theory in multi-agent reinforce-
ment learning to make it possible in large-scale scenarios. Since the
mean-(cid:128)eld MARL method only takes a mean (cid:128)eld on states/actions
input into consideration, it ignores the agent interactions. Our
proposed method provides another way to enable large-scale multi-
agent learning and retain the interactions between agents, which
makes agents receive global feedback from the next moments and
adjust their strategies in time. Furthermore, our proposed method
matchidle vehiclespassengersvehicle listorder listpack updecision-making platform provides a backward stationary learning method and has a rapid
reaction to the feedback from the environment.
Multi-agent Taxi Dispatching. A lot of previous work models
the taxi dispatching into multi-agent learning, like [1], it divides
the city into many dispatching areas, and regards an area as an
agent, then uses self-organization techniques to decrease the total
waiting time and increase the taxi utilization. NTuCab [27] is a
collaborative multi-agent taxi dispatching system which a(cid:138)empts
to increase custom satisfaction more globally, and it can dispatch
multiple orders to taxis in the same geographical regions. NTuCab
thinks that it is not feasible to compute the shortest-time path for
each of a possibly large number of available taxis nearby a customer
location since it is computationally costly. We follow these se(cid:138)ings
in our proposed model and divide the city into many dispatching
regions. Each dispatching region is controlled in a given distance,
which indirectly limits the maximum waiting time. (cid:140)e NTuCab
achieves a signi(cid:128)cant improvement in reducing the wait time and
taxi cruising time, but it is also a computational cost method. Xu et
al. [37] proposed a learning and planning method based on MARL
and combinatorial optimization recently, and some other methods
[16, 24, 36] focus on (cid:131)eet management to improve the ADI or de-
crease the waiting time. But considering the current operational
ride-sharing scenarios, it is hard to perform (cid:131)eet management for
it is impossible to force drivers to designated regions. (cid:140)e men-
tioned MARL method[37] is an independent MARL method, which
ignores the interactions between agents. However, it is a consensus
to consider that the agent interactions have a positive impact on
making optimal decisions. Our proposed method considers the
interaction between agents by applying constraints on the joint
policies using KL-divergence optimization, and the experiments
demonstrate that the proposed method outperforms baselines on
all metrics in di(cid:130)erent tra(cid:129)c environments.
3 METHODOLOGY
In this section, we (cid:128)rst give a de(cid:128)nition of order-dispatching from
a perspective of multi-agent reinforcement learning process, and
then discuss the main challenges when applying the MARL method
for order-dispatching, and give our methods.
3.1 Order-dispatching as a Markov Game
We regard the order-dispatching task as a sequential decision task,
where the goal is to maximize the long-term ADI and ORR per
day. According to the characters of the practical environment, each
vehicle can only serve the surrounding orders, thus we model the
order-dispatching task using Partially Observable Markov Decision
Process (POMDP) [29] in multi-agent se(cid:138)ings. With the multi-agent
se(cid:138)ings, we can decompose the original global order-dispatching
task into many local order-dispatching tasks, and transform a high-
dimensional problem into multiple low-dimensional problems.
(cid:140)e POMDP framework to the multi-agent order-dispatching
problem can be formulated as a tuple (cid:104)S, P, A, R, G,N, γ(cid:105), where
S, P, A, R, G, N, γ represent the sets of states, state transition
probability function, sets of action spaces, reward functions, set of
grids, the number of agents and the future reward discount factor
respectively.
For each agent i, Si ∈ S, Ai ∈ A, Ri ∈ R represent the state
space, action space and reward function respectively, and Gi ∈ G
represents the grid which the agent in. (cid:140)e state transition occurs
a(cid:137)er the decision making, i.e. agents executed their actions, then
the state St of environment at time t transform to St +1 at time
t + 1, and agents will get rewards given by the environment. Based
on the above de(cid:128)nitions, the main purpose of each agent is to learn
to maximize the cumulative reward Gt:T from t to T
T
t =0
max Gt:T = max
γ t rt(st , at) , where at ∼ πθ(st) .
sents the policy with respect to the state at time t.
In reinforcement learning, the πθ(·) parameterized with θ repre-
It is common to divide the city into regional dispatch areas
[16, 27]. In our se(cid:138)ings, we use a grid-world to represent the real
world and divide the real world into several order-dispatching re-
gions. Each grid represents an individual order-dispatching region
which contains some orders and vehicles, and we regard vehicles as
agents here. Based on the above MARL se(cid:138)ings, we specify the de(cid:128)-
nitions of the order-dispatching task as follows from a mathematical
perspective.
• State: (cid:140)e state input used in our method is expressed as
a four elements tuple, namely, S = (cid:104)G, N , M, Ddest(cid:105). Ele-
ments in the tuple represent the grid index, the number of
idle vehicles, the number of valid orders and the distribu-
tion of orders' destinations respectively. (cid:140)e distribution of
order's destination is a mean over the destination vectors
of orders in grid G, which roughly re(cid:131)ects the overall the
orders information. In our se(cid:138)ings, agents in the same grid
share the same state.
• Action: (cid:140)e action input used in our method is expressed
as A = (cid:104)Gsource, Gdest,T , C(cid:105). Elements in the tuple repre-
sent the source grid index, target grid index, order duration,
and price respectively. We regard the set of orders in the
grid j at time t as the candidate actions of the agent i. Since
agents are homogeneous, so agents in grid j share the same
action space. In practice, sometimes there is no order in
some regions. Under the se(cid:138)ing of MARL, agents need to
select orders at each timestep, but some grids may not have
orders, so in order to ensure the feasibility and sustainabil-
ity of the MDP, we arti(cid:128)cially add some virtual orders
whose Gsource = Gdest, and set the price C to 0. When idle
vehicles select these virtual orders, it means they will stay
where they are.
• State Transition: (cid:140)e agent which serves one order will
migrate to the destination grid given by the taken order
a(cid:137)er T time step, where T is de(cid:128)ned with the served order
duration, then the state of agent will be updated to the
newest state, namely, the stage of destination grid.
• Reward: (cid:140)e reward function is very important for rein-
forcement learning to a great extent which determines the
direction of optimization. Because of the goal of learning
is to (cid:128)nd a solution which maximizes the ADI with high
ORR, so we design a reward function which is proportional
to the price of each order.
3.2 Non-stationary Action Space
Traditional deep Q-learning network accepts a state input and out-
puts a vector of Q values whose dimension is equal to the dimension
of action space, i.e.,
dim(cid:16)
Q(s, A)(cid:17)
= dim(cid:16)A(cid:17)
.
(1)
It is correct when the action space is (cid:128)xed, while it is problematic
in our se(cid:138)ings. (cid:140)ere is a fact that for the grid j, the orders produced
at time t are always di(cid:130)erent from the orders produced at other
moments. It cannot ensure that the action space is consistent along
with the whole episode, so it is problematical to regard the orders
as an action while ignoring the distribution of the variant action
space. In our proposed method, we use the tuple (cid:104)S, a(cid:105) to represent
the input of Q-learning, then evaluate all available state-order pairs.
3.3 Action Selection Q-learning
For convenience, we name the Q-learning network with a state-
action input as action selection Q-learning shown in Figure 2.
For agent i, supposing there are M available orders, which re-
quires M state-action evaluation. In the case of N agents, the com-
putational complexity will be O(N · M). To decrease the original
complexity to O(M), we use parameter sharing and state sharing
mentioned in previous sections to achieve it.
From the perspective of agent i, we suppose that st denotes
the state at time t, at denotes the set of orders, then the Bellman
Equation in our se(cid:138)ings can be expressed as
Q(st , at) = αQ(st , at)+
(1 − α)(cid:104)
rt + γ · Eat +1∼π(st +1)[Q(st +1, at +1)](cid:105)
,
(2)
where γ ∈ [0, 1] is the discount factor, α is the step size. (cid:140)e value
of the next timestep is a expectation of all available state-order
pairs. When the policy π(st +1) is greedy, then Eq. (2) represents
the traditional Q-learning algorithm.
To balance the exploitation and exploration, the Q values related
to the same orders set are converted into a biased strategy Boltzman
exploration
π(a
j
t st) =
eQ(st ,a j
t ∈Ai
a j
t)/τ
eQ(st ,a j
t)/τ
,
(3)
where τ is the temperature to balance the exploitation and explo-
ration.
3.4 KL Divergence Optimization
In the multi-agent system, the main method to relieve or overcome
the non-stationary problem is learning multi-agent communication
[6, 7, 9, 30], while most of them require a (cid:128)xed agent number or
observations from other agents before making decisions. In the
order-dispatching case, explicit communication occurs between
agents is o(cid:137)en time-consuming and di(cid:129)cult to adapt. As illustrated
in Figure 3, supposing that triangles in each grid represent orders,
and dots represent vehicles. It shows that the order-dispatching pro-
cess of each grid at time t, and di(cid:130)erent order has di(cid:130)erent duration
of d, so the vehicles will arrive at the destination grids at di(cid:130)erent
time, and vehicles serve di(cid:130)erent orders will be assigned to di(cid:130)erent
Figure 2: Action selection Q-learning. Di(cid:130)erent from the tra-
ditional Q-learning network, our model accepts both state
input and action feature vector. A(cid:133)er embedding them re-
spectively, there is a concatenation followed by 2 dense lay-
ers which follows the embedding layers. (cid:135)en the network
outputs a scalar value Q(s, a).
Figure 3: Grid-based order-dispatching. Supposing that tri-
angles in each grid represent orders, and dots represent ve-
hicles. It shows that the order-dispatching process of each
grid at time t, and di(cid:130)erent order has di(cid:130)erent duration of d,
so the vehicles will arrive at the destination grids at di(cid:130)er-
ent time, and vehicles serve di(cid:130)erent orders will be assigned
to di(cid:130)erent grids, then it is hard to form continuous interac-
tions and communication between vehicles. For these two
reasons, applying the communication mechanism or learn-
ing others' policies is not a good choice.
grids, then it is hard to form continuous interactions and communi-
cation between vehicles. Also, it o(cid:137)en su(cid:130)ers from computational
cost, especially in large-scale se(cid:138)ings. Taking the aforementioned
reasons, we introduce a centralized training method using KL diver-
gence optimization, which aims to optimize the agents' joint policy
and try to match the distribution of vehicles with the distribution
of orders.
Notice that we have two goals need to achieve in our proposed
method: (1) maximize the long horizontal ADI; (2) optimize the
order response rate.
If there are always enough vehicles in the
ReLU%(s,a)State sAction0ReLUReLUReLUGrid iGrid jGrid pGrid qt+d1t+d2t+d3 dispatching grid, it is easy to decrease the rate of idle vehicles and
improve the order response rate, also the long horizontal ADI, while
there is a fact that we cannot control the distribution of orders. So
we want to make the order and vehicle distribution as similar as
possible through (cid:128)nding feasible order-vehicle matches. We do not
require explicit cooperation or communication between agents, but
an independent learning process with centralized KL divergence
optimization.
Supposing at time t, the agents (cid:128)nd a feasible order set Ot by
executing their policies, namely,
Ot ∼ {πθ, j(st) j = 1, . . . , N} .
(4)
Our purpose is to (cid:128)nd an optimal order set O∗
t . Focusing on
a certain grid j, it supposes that the policy πθj
at time t is param-
eterized by θj. A(cid:137)er all policies have been executed, we get the
newest distribution of vehicles Dv
t +1, and the newest distribution
of orders is Do
t +1 shows
the margin between the joint policy Π at time t and Do
t +1, so the KL
optimization is actually (cid:128)nding an optimal joint policy Π∗ which
has a minimal margin:
t +1. (cid:140)e KL divergence from Dv
t +1 to Do
∗ = arg
Π
Π
min DKL(Do
t +1 (cid:107) Dv
t +1(Π)) ,
(5)
where Π = {πθ, j
j = 1, ..., N}. For the convenience, we replace
t +1(Π)) with DK L. We want to decrease the KL
t +1 (cid:107) Dv
DKL(Do
divergence from the distribution of vehicles to the distribution of
orders to balance the demand and supply at each order-dispatching
grid. Formally, our KL policy optimization can be wri(cid:138)en as:
L =(cid:107) Qθ(s, a) − Q
min
θ
s.t . DK L ≤ β ,
∗ (cid:107)2
where β ∈ R. (cid:140)en the objective function can be expressed as
L =(cid:107) Qθ(s, a) − Q
∗ (cid:107)2 +λDK L ,
min
θ
(6)
(7)
(8)
t +1 =N
where Q∗ is the target Q-value, λ ∈ R parameterizes the contribu-
tion of KL item. To formulate the relationship between min L and
θj, we make some de(cid:128)nitions of notations in advance. Considering
that there is N grids in total, ni
t +1 represents the number of idle
vehicles in grid i at time step t + 1, which can be formulated as
t · πj→i, where c
represents the idle driver number
j
ni
t
at last time step t, πj→i represents the probability of dispatching
orders which from grid j to grid i to idle vehicles at time t, and these
vehicles will arrive at grid i at time t +1. q
t +1 is the rate of idle vehi-
j
cles in grid j which can be formulated into q
t +1.
t +1 represents the rate of orders in grid i at time t + 1 here. Using
pi
t +1/N
j
t +1 = n
k =1 nk
j=1 c
j
j
i =1
i =1
pi
t +1
pi
i =1
=
=
=
log
k =1
log
πj→k c j
t
1
qi
t +1
t +1∇πj
pi
i =1
t +1∇πj
pi
(cid:17) · ∇θi πj
chain rule, we can decompose the gradient of DKL to θ as
∇θj DKL = ∇πj DK L · ∇θj πj
log qi
t +1
pi
t +1
· ∇θj πj
= −(cid:16) N
N
(cid:104)∇πj
N
N
N
N
(cid:104) ∇πj
N
N
k =1N
(cid:104)
N
N
k =1N
(cid:104)
N
where Nvehicle =N
(cid:105) · ∇θj πj
N
N
(cid:105) · ∇θj πj
N
(cid:105) · ∇θj πj
t − ∇πj
t)
l (cid:44)j cl
l =1 πl→k cl
t
1
l =1 πl→k cl
t
− 1
(9)
ni
t +1
t +1. (cid:140)e gradient of πj to θj is ∇Qj(s,a)πj(a
j
s) · ∇θ Q(s, a). We use the δ =(cid:107) Q − Q∗ (cid:107)2, then the (cid:128)nal gradient
of Lθ(s, a) to θ is calculated as
N
(cid:105) · ∇θj πj ,
l =1
k =1 πj→k(c j
+N
k =1 πj→k c j
t
k =1 πj→k c j
t
k =1 πj→k c j
t
k =1
∇πj
(10)
For convenience, we give a summary for some important nota-
∇θj L = ∇θj δ + λ∇θj DKL .
i =1
i =1
= c j
t
= c j
t
pi
t +1
pi
t +1
1
Nvehicle
−
1
j=1 n
πl→k cl
t
−
log
tions in Table 1.
Table 1: Important notations
N
ni
t +1
j
c
t
πj→i
j
q
t +1
pi
t +1
the number of grids
the number of idle vehicles in grid i at time t + 1
the number of idle vehicles in grid j at time t
the probability of dispatching orders which from grid
j to grid i to idle vehicle at time t
the rate of idle vehicles in grid j
the rate of orders in grid i at time t + 1
4 EXPERIMENTS
We examine the correctness of our model in a toy grid-based order-
dispatching environment and the practicality of our model using
real-world data from three cities. Considering the constraint of a
grid-based environment, we did not compare with order-dispatching
algorithms based on coordinate systems. To compare with existing
methods, and investigate the e(cid:130)ectiveness of our method on the
metrics of ORR and ADI, we select three typical algorithms as base-
lines, namely, Independent Deep Q-learning Network (IL), Nearest
order-dispatching (NOD) and MDP respectively, and we will give a
brief description at the (cid:128)rst.
• IL: A variant of Double DQN [33] which takes a tuple of
state and action as an input. Compared with our method,
the only di(cid:130)erence in IL is that KL optimization is not used.
• NOD: Nearest-distance Order Dispatching (NOD) algo-
rithm, which dispatches orders to idle vehicles with consid-
ering the shortest distance. (cid:140)e reason why we use NOD
as one of the baselines is that it is a fairly representative
algorithm which is used frequently and easy to implement
in practice. However, in our environment se(cid:138)ing, because
there are reasonable regional division strategies (for ex-
ample, in our later real-world data experiments, the size
of the division area guarantees the maximum order wait-
ing time is 10 minute), we have no need to distinguish
the speci(cid:128)c position of vehicles in the same dispatching
region. (cid:140)at is to say, the principle of matching orders
based on distance is equivalent to random matching in our
environment se(cid:138)ing.
• MDP: Proposed by Xu et al. [37], a planning and learning
method based on decentralized multi-agent deep reinforce-
ment learning and centralized combinatorial optimization.
Considering the fairness of experiments, we use the same reward
function for the reinforcement learning methods.
4.1 Model Settings
Our model is an extension of Double DQN with so(cid:137) update. All
neural-based models used in our experiments are implemented by
the MLP with 2 hidden layers, and the active function used here
for all neural-based algorithm is recti(cid:128)ed linear unit (ReLU). (cid:140)e
replay bu(cid:130)er stores experience tuples, which can be formulated into
(cid:104)st , at , At , st +1, At +1, ∇π DKL(cid:105). Elements of the tuple represent
the state of time t, action selected at time t, the action set at time
t, state at time t + 1, action set at time t + 1 and the (cid:128)rst gradient
item of ∇θ DKL, respectively. (cid:140)e temperature is 1.0, discount factor
γ = 0.95 and learning rate is α = 10−4.
4.2 Particle Order-dispatching Experiment
(cid:140)e grid-based order-dispatching environment showed in Figure 4
is implemented based on the multi-agent particle environment sup-
ported by Mordatch et al. [21]. (cid:140)is toy environment abstracts
the real-world order-dispatching, where one grid represents one
dispatching region, and orders have the same duration. (cid:140)e blue
particles and the red particles represent vehicles and orders respec-
tively. All of the blue particles and red particles sca(cid:138)er in a 10 × 10
grid world. Each of the red particles owns a direction vector v and
a reward r, the direction vector v = (cid:104)#source, #tar❕et(cid:105) denotes a di-
rection from the source grid to the target grid. In this environment,
the order price is simpli(cid:128)ed with Euler distance between grids, so
reward r is proportional to the Euler distance between #source and
#tar❕et, i.e.,
r = 0.1× (cid:107) #source − #tar❕et (cid:107)2 .
(11)
(cid:140)e blue particles will get reward by picking red particles in
the same grid, one blue particle can only pick one red particle at
each time-step. Blue particles will migrate to the target grids given
by the red particles at the next timestep. In our se(cid:138)ings, the time
horizon is T = 144. At t = 0, we produce the reds and blues using
speci(cid:128)c distributions respectively. At the next timesteps from t = 1
to t = T − 1, there are some new red particles born with a speci(cid:128)c
distribution with respect to the grid position and time, while there
is no new blue particles born expect t = 0, that means all blue
particle movements fully dependent on picking red particles. In
order to match the real world situation, the amount of blues is less
than the reds in our se(cid:138)ings.
(a)
(b)
Figure 4: (cid:135)e toy order-dispatching environment.
(a) (cid:135)e
10x10 toy order-dispatching environment. Red particles and
blue particles represent orders and vehicles. (b) (cid:135)e migrate
process of blue particles, where the dotted arrow indicates
that the red one is dispatched to the blue one, the solid ar-
row indicates that the blue one will migrate to another grid
at the next timestep. For the red particles in each grid, the
surrounding 8 grids are feasible destinations.
4.2.1
Influence of KL Divergence. In order to verify the feasibil-
ity of KL-divergence optimization, we adopt three cases correspond
to di(cid:130)erent degree of order distribution changes. (cid:140)e degree of
order distribution changes means the margin between two adjacent
order distributions. In the particle order-dispatching environment,
we generate orders with a given 2-dimensional Gaussian distribu-
tion N(µt , σt) at each timestep. To quantify and explicitly compare
the margin between order distributions at di(cid:130)erent timesteps, we
can change µt at each timestep. (cid:140)e degree of order distribution
changes is equivalent to the distance between adjacent order distri-
butions. Farther distance means a higher degree, that is, the greater
the degree of changes. In our experiment se(cid:138)ings, the degree of
changes from low to high orders correspond to a distance of 1, 2, 4
grids respectively, Figure 5 shows an example of order distribution
changes, the distance between µt and µt +1 is 8 grid. Since the desti-
nations of orders are random, if we want to let vehicles serve more
orders at the next timestep, we need to let the algorithm learns to
pick suitable orders at current decision stage to assign the vehicles
to suitable grids, so that we can ensure that there is a be(cid:138)er ORR at
next timestep. Intuitively, a long-term higher ORR corresponds to
a higher long-term ADI. Table 2 shows the performance at metrics
of ORR and ADI at di(cid:130)erent degree of order distribution changes.
Influence of λ. λ plays an important role in our method,
so it is necessary to investigate how it a(cid:130)ects the performance at
di(cid:130)erent degree of order distribution changes. In our experiments,
the value of λ ranges from 0.0 to 0.6 with stepping 0.05. When
λ = 0.0, it means our method is equivalent to IL. Figure 6 shows
curves at di(cid:130)erent degree of order distribution changes.
4.2.2
4.2.3 Result Analysis. We train 300 episodes for all algorithms
in the three cases which are related to di(cid:130)erent degrees of KL-
divergence. We compare the three baselines from metrics of ADI
and ORR. As shown in Table 2, it shows the average experimen-
tal results of 5 groups with di(cid:130)erent random seeds. (cid:140)e particle
Table 2: Performance comparison in terms of ADI and ORR with respect to NOD. We compare against with baselines in
three di(cid:130)erent order distribution changes degree, namely, low, medium and high. KL-Based is our proposed method, which
outperforms all baselines on all metrics.
Order Distribution Divergence
Metrics
IL
MDP
KL-Based
Low
Medium
High
ADI
ORR
ADI
ADI
+11.5%
+6.68% +2.32%
+12.5%
+14.5%
+13.3% +6.69% +7.28% +3.42%
+25.12% +13.40% +20.94% +7.89% +13.47% +4.61%
ORR
+6.94%
+8.94%
ORR
+6.3%
Table 3: Performance comparison in terms of ADI and ORR
with respect to NOD. We compare against baselines using dif-
ferent datasets from three cities. KL-Based is our proposed
method, which outperforms all baselines on all metrics.
City
Metrics
City A
City B
City C
ADI
ORR
ORR
ADI
+4.69% +1.68% +2.96% +1.11% +4.72% +2.05%
+5.80% +1.89% +3.69% +2.63% +5.98% +2.14%
KL-Based +6.46% +3.07% +4.94% +3.30% +6.12% +3.01%
IL
MDP
ORR
ADI
(a)
(c)
(b)
(d)
Figure 5: Order distribution changes. From (a) to (d), the or-
der distribution of four consecutive timesteps in the 10 × 10
grid world is shown. (cid:135)e distance of µt and µt +1 is 8 grids.
order-dispatching environment generates orders with random desti-
nations, that is, the probabilities of long and short orders appear are
equivalent in a grid. Although choosing long orders means higher
ADI, the degree of order distribution changes we set requires the
algorithm to choose more non-longest orders, thus ensuring both
higher ORR and long-term ADI. In the three di(cid:130)erent degrees of
order distribution changes, our KL-based outperforms all baselines
on all metrics. It means our method can be(cid:138)er counterweight the
margin of order distribution.
Figure 6 shows the learning curve of λ at di(cid:130)erent degree of
order distribution changes. In the three cases of order dispatching
changes, our method achieves highest ORR at λ = 0.05, λ = 0.45,
λ = 0.5 respectively. (cid:140)e results in Figure 6 also show that higher
ORR o(cid:137)en corresponds to higher ADI. In practice, if there is a low
degree of order distribution changes, in order to achieve a higher
ORR, the algorithm needs to pick shorter orders, so that in the
future, agents in the closer regions can still serve more orders,
namely, higher ORR. When it comes to a high degree of order
distribution changes, the dispatching algorithm needs to perform
more greedy to achieve a higher ORR, that is, it prefers to pick
longer orders. Also, the result of Figure 6c shows our algorithm
achieves be(cid:138)er performance on ORR and ADI than IL in the case of
a high degree of order distribution changes. (cid:140)erefore, combined
with the results of Table. 2 and Figure 6, our method can (cid:131)exibly
choose long or short orders.
4.3 Real World Data Experiments
4.3.1 Dispatching Simulator. Since our model is implemented
on the se(cid:138)ing of dividing the city into many order dispatching
regions, so we conduct experiments on an open source grid-based
environment simulator provided by Didi Chuxing [16]. (cid:140)e simula-
tor divides the city into N hexagonal grids which depends on the
size of the city. At each time t, the simulator provides a set of idle
vehicles and a set of available orders. Each order is featured with
its origin, destination, and duration, and vehicles in the same re-
gion share the same state. (cid:140)e travel distance between neighboring
regions is approximately 2.2km and the time interval is 10min.
4.3.2 Data Description. (cid:140)e real-world datasets provided by
Didi Chuxing include order and trajectories of vehicles informa-
tion of three cities in one month. (cid:140)e order information includes
price, origin, destination, and duration. (cid:140)e trajectories contain
the positions (latitude and longitude) and status (on-line, o(cid:130)-line,
on-service). We divide the three cities into 182, 126, 112 hexagonal
grids respectively.
Result Analysis. We compare our model with three baselines
a(cid:137)er 300 episodes training. As shown in Table 3, it lists the average
results of 5 groups experiments with di(cid:130)erent random seeds. (cid:140)e
real datasets contain more changes in the order distribution. From
the results, our method can still be(cid:138)er discover the changes of
order distribution and improve the ORR and ADI via order-vehicle
distribution matching.
5 DEPLOYMENT
Taking both model se(cid:138)ing in this paper and online platforms of
Didi Chuxing, we design a hybrid system and incorporate with
other components including routing planning technique [32] and
estimating time of arrival (ETA) [34] as illustrated in Figure 7.
02468100246810Heatmap(µ=(1,0))51015202502468100246810Heatmap(µ=(9,0))51015202502468100246810Heatmap(µ=(9,8))51015202502468100246810Heatmap(µ=(1,8))510152025 (a) (cid:107) µt − µt +1 (cid:107)2= 1
(b) (cid:107) µt − µt +1 (cid:107)2= 2
(c) (cid:107) µt − µt +1 (cid:107)2= 4
Figure 6: ORR and ADI performance under di(cid:130)erent λ settings. (cid:135)e horizontal axis represents di(cid:130)erent λ, the le(cid:133) and right
vertical axis represent ORR and ADI respectively.
As aforementioned mentioned in Section 3, there are several
assumptions prevent this model from deploying in real-world set-
tings: (i) vehicles in the same grid share the same se(cid:138)ing, and
this isomorphic se(cid:138)ing ignores the intra-grid information; (ii) this
paper adopts the grid-world map to simplify the real-world envi-
ronment which replace coordinate position information with grid
information. To address these issues, we adapt estimate travel time
techniques proposed in and incorporate with our action selection Q-
learning mentioned in Section 3.3. For example, the duration time
of each order in our model is regarded as one of the already known
order features. However, in the real-world scenario, each order's
travel time obtained via the ETA model is dynamic and depends
on current tra(cid:129)c and route conditions. Since coordinate position
information is taken into consideration in the ETA model, this hy-
brid system is able to deal with the assumption (ii) and feasible to
be deployed in real-world.
We extend the Matching System and the Routing System a(cid:137)er
obtaining Q-value via the hybrid system as illustrated in Figure. 7.
Speci(cid:128)cally, in each time slot, the goal of the real-time order dis-
patch algorithm is to determine the best matching between vehicles
and orders (see Figure 1) in the matching system and plan a routine
for drivers to serve the orders. Formally, the principle of Matching
System can be formulated as:
where
(cid:40)1,
0,
aij =
if order j is assigned to driver i
if order i is not assigned to driver i
where i ∈ [1, 2, ..., m] and j ∈ [1, 2, ..., n] present all idle drivers and
available orders at each time step respectively. Q(i, j) is the output
from hybrid system and represents the action-value function driver
i performing an action of serving order j. Note that constraints in
Eq. (13) guarantee that each driver will select one available real
m
n
i =0
j=0
argmaxai j
Q(i, j)aij ,
s.t .
m
n
i =0
j=0
aij = 1, j = 1, 2, 3..., n
aij = 1, i = 1, 2, 3..., m
(12)
(13)
orders or doing nothing while each order will be assigned to one
driver or stay unserved at each time step.
(cid:140)is Matching System used in Xu et al. [37] and Wang et al.
[35] is implemented using Kuhn-Munkres (KM) algorithm [22].
In detail, they formulated Eq. (12) as a bipartite graph matching
problem where drivers and orders are presented as two set of nodes.
(cid:140)en, each edge between order i and driver j is valued with Q(i, j),
and the best matches will be (cid:128)ned using KM algorithm. Di(cid:130)erent
from them, since we implemented our method based on assumption
(i), that is, there is no di(cid:130)erence in the drivers in a same grid. So the
KM algorithm will degenerate into a sorting algorithm here. We
just need to select the top m orders with the highest Q(i, j) values.
Figure 7: Illustration of deployment. (cid:135)e hybrid system con-
sists of two modules, namely, Action Selection Q-learning
(ASQ) and Estimate Travel time modules. (cid:135)e ASQ will inter-
act with simulator periodically, and it will be trained o(cid:130)line
in the simulator. Matching System accepts value estimation
and outputs (cid:104)vehicle, order(cid:105) matches to Routing System.
Once the matching pairs of orders and vehicles has been selected
from the matching system, we then deliver these pairs with co-
ordinate information to the routing system. (cid:140)e routing system
equipped with route planning techniques [32] allows drivers to
serve the order. (cid:140)is process will give feedback, i.e. reward to the
hybrid system and help the whole system training to achieve be(cid:138)er
performance.
6 CONCLUSIONS
In this paper, we proposed a multi-agent reinforcement learning
method for order-dispatching via matching the distribution of or-
ders and vehicles. Results on the three cases in the simulated
0.00.10.20.30.40.50.60.2060.2080.2100.2120.2140.2160.2180.220ORRADI4.354.404.454.504.554.600.00.10.20.30.40.50.60.2480.2500.2520.2540.2560.258ORRADI5.605.655.705.755.805.850.00.10.20.30.40.50.60.2700.2750.2800.2850.290ORRADI6.006.056.106.156.206.256.306.356.40Data: (2,4)Action Selection Q-learningEstimate Travel timeHybrid SystemMatchingSystemRoutingSystemSimulatorOfflineData: (2,4)Q-value order-dispatching environment have demonstrated that our pro-
posed method achieves both higher ADI and ORR than the three
baselines, including one independent MARL method, one planning
algorithm, and one rule-based algorithm, in various tra(cid:129)c environ-
ments. (cid:140)e experiments on real-world datasets also show that our
model can obtain higher ADI and ORR. Furthermore, our proposed
method is a centralized training method and can be executed de-
centralized. In addition, we designed the deployment system of the
model with the support of the existing platform of Didi Chuxing. In
future work, we plan to deploy the model to do online tests through
the designed deployment system.
ACKNOWLEDGMENTS
We thank the support of National Natural Science Foundation of
China (61702327, 61772333, 61632017).
[8]
[6]
[7]
REFERENCES
[1] Aamena Alshamsi, Sherief Abdallah, and Iyad Rahwan. 2009. Multiagent self-
organization for a taxi dispatch system. In 8th international conference on au-
tonomous agents and multiagent systems. 21 -- 28.
[2] Darse Billings, Denis Papp, Jonathan Schae(cid:130)er, and Duane Szafron. 1998. Oppo-
nent modeling in poker. Aaai/iaai 493 (1998), 499.
[3] Lucian Busoniu, Robert Babuska, and Bart De Schu(cid:138)er. 2006. Multi-agent rein-
forcement learning: A survey. In Control, Automation, Robotics and Vision, 2006.
ICARCV'06. 9th International Conference on. IEEE, 1 -- 6.
[4] Stephen C Chadwick and Charles Baron. 2015. Context-aware distributive taxi
cab dispatching. (March 19 2015). US Patent App. 14/125,549.
[5] Lee Chean Chung. 2005. GPS taxi dispatch system based on A* shortest path
algorithm. Ph.D. Dissertation. Master's thesis, Submi(cid:138)ed to the Department of
Transportation and Logistics at Malausia University of Science and Technology
(MUST) in partial ful(cid:128)llment of the requirements for the degree of Master of
Science in Transportation and Logistics.
Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon White-
son. 2016. Learning to communicate with deep multi-agent reinforcement learn-
ing. In Advances in Neural Information Processing Systems. 2137 -- 2145.
Jakob N Foerster, Yannis M Assael, Nando de Freitas, and Shimon Whiteson.
2016. Learning to communicate to solve riddles with deep distributed recurrent
q-networks. arXiv preprint arXiv:1602.02672 (2016).
Jayesh K Gupta, Maxim Egorov, and Mykel Kochenderfer. 2017. Cooperative
multi-agent control using deep reinforcement learning. In International Confer-
ence on Autonomous Agents and Multiagent Systems. Springer, 66 -- 83.
[9] Ma(cid:138)hew John Hausknecht. 2016. Cooperation and communication in multiagent
deep reinforcement learning. Ph.D. Dissertation.
Junling Hu, Michael P Wellman, et al. 1998. Multiagent reinforcement learning:
[10]
theoretical framework and an algorithm.. In ICML, Vol. 98. Citeseer, 242 -- 250.
[11] Der-Horng Lee, Hao Wang, Ruey Cheu, and Siew Teo. 2004. Taxi dispatch
system based on current demands and real-time tra(cid:129)c conditions. Transportation
Research Record: Journal of the Transportation Research Board 1882 (2004), 193 --
200.
[12] Bin Li, Daqing Zhang, Lin Sun, Chao Chen, Shijian Li, Guande Qi, and Qiang
Yang. 2011. Hunting or waiting? Discovering passenger-(cid:128)nding strategies from a
large-scale real-world taxi dataset. In Pervasive Computing and Communications
Workshops (PERCOM Workshops), 2011 IEEE International Conference on. IEEE,
63 -- 68.
[13] Minne Li, Yan Jiao, Yaodong Yang, Zhichen Gong, Jun Wang, Chenxi Wang,
Guobin Wu, Jieping Ye, et al. 2019. E(cid:129)cient Ridesharing Order Dispatching with
Mean Field Multi-Agent Reinforcement Learning. arXiv preprint arXiv:1901.11454
(2019).
[14] Ziqi Liao. 2001. Taxi dispatching via global positioning systems. IEEE Transactions
[15] Ziqi Liao. 2003. Real-time taxi dispatching using global positioning systems.
on Engineering Management 48, 3 (2001), 342 -- 347.
Commun. ACM 46, 5 (2003), 81 -- 83.
[16] Kaixiang Lin, Renyu Zhao, Zhe Xu, and Jiayu Zhou. 2018. E(cid:129)cient Large-Scale
Fleet Management via Multi-Agent Deep Reinforcement Learning. arXiv preprint
arXiv:1802.06444 (2018).
[17] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor
Mordatch. 2017. Multi-agent actor-critic for mixed cooperative-competitive
environments. In Advances in Neural Information Processing Systems. 6379 -- 6390.
[18] Gary S Lynch. 2009. Single point of failure: (cid:138)e 10 essential laws of supply chain
risk management. John Wiley & Sons.
[19] Fei Miao, Shuo Han, Shan Lin, John A Stankovic, Desheng Zhang, Sirajum Munir,
Hua Huang, Tian He, and George J Pappas. 2016. Taxi dispatch with real-time
sensing data in metropolitan areas: A receding horizon control approach. IEEE
Transactions on Automation Science and Engineering 13, 2 (2016), 463 -- 478.
[20] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness,
Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg
Ostrovski, et al. 2015. Human-level control through deep reinforcement learning.
Nature 518, 7540 (2015), 529.
Igor Mordatch and Pieter Abbeel. 2017. Emergence of Grounded Compositional
Language in Multi-Agent Populations. arXiv preprint arXiv:1703.04908 (2017).
James Munkres. 1957. Algorithms for the assignment and transportation prob-
lems. Journal of the society for industrial and applied mathematics 5, 1 (1957),
32 -- 38.
[23] David Myr. 2013. Automatic optimal taxicab mobile location based dispatching
[22]
[21]
system. (May 14 2013). US Patent 8,442,848.
[24] Takuma Oda and Yulia Tachibana. 2018. Distributed Fleet Control with Maximum
Entropy Deep Reinforcement Learning. (2018).
[25] Christos H Papadimitriou and Kenneth Steiglitz. 1998. Combinatorial optimiza-
[26] Frederik Schadd, Sander Bakkes, and Pieter Spronck. 2007. Opponent Modeling
tion: algorithms and complexity. Courier Corporation.
in Real-Time Strategy Games.. In GAMEON. 61 -- 70.
[27] Kiam Tian Seow, Nam Hai Dang, and Der-Horng Lee. 2010. A collaborative
multiagent taxi-dispatch system. IEEE Transactions on Automation Science and
Engineering 7, 3 (2010), 607 -- 616.
[28] Lloyd S Shapley. 1953. Stochastic games. Proceedings of the national academy of
sciences 39, 10 (1953), 1095 -- 1100.
[29] Ma(cid:138)hijs TJ Spaan. 2012. Partially observable Markov decision processes. In
Reinforcement Learning. Springer, 387 -- 414.
[30] Sainbayar Sukhbaatar, Rob Fergus, et al. 2016. Learning multiagent communica-
tion with backpropagation. In Advances in Neural Information Processing Systems.
2244 -- 2252.
[31] Gerald Tesauro. 2004. Extending Q-learning to general adaptive multi-agent
systems. In Advances in neural information processing systems. 871 -- 878.
[32] Yongxin Tong, Yuxiang Zeng, Zimu Zhou, Lei Chen, Jieping Ye, and Ke Xu. 2018.
A uni(cid:128)ed approach to route planning for shared mobility. Proceedings of the
VLDB Endowment 11, 11 (2018), 1633 -- 1646.
[33] Hado Van Hasselt, Arthur Guez, and David Silver. 2016. Deep Reinforcement
Learning with Double Q-Learning.. In AAAI, Vol. 2. Phoenix, AZ, 5.
[34] Zheng Wang, Kun Fu, and Jieping Ye. 2018. Learning to estimate the travel time.
In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining. ACM, 858 -- 866.
[35] Zhaodong Wang, Zhiwei Qin, Xiaocheng Tang, Jieping Ye, and Hongtu Zhu.
2018. Deep Reinforcement Learning with Knowledge Transfer for Online Rides
Order Dispatching. In 2018 IEEE International Conference on Data Mining (ICDM).
IEEE, 617 -- 626.
[36] Chong Wei, Yinhu Wang, Xuedong Yan, and Chunfu Shao. 2018. Look-Ahead
Insertion Policy for a Shared-Taxi System Based on Reinforcement Learning.
IEEE Access 6 (2018), 5716 -- 5726.
[37] Zhe Xu, Zhixin Li, Qingwen Guan, Dingshui Zhang, Qiang Li, Junxiao Nan,
Chunyang Liu, Wei Bian, and Jieping Ye. 2018. Large-Scale Order Dispatch in
On-Demand Ride-Hailing Platforms: A Learning and Planning Approach. In
Proceedings of the 24th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining. ACM, 905 -- 913.
[38] Yaodong Yang, Rui Luo, Minne Li, Ming Zhou, Weinan Zhang, and Jun
Wang. 2018. Mean Field Multi-Agent Reinforcement Learning. arXiv preprint
arXiv:1802.05438 (2018).
[39] Lingyu Zhang, Tao Hu, Yue Min, Guobin Wu, Junying Zhang, Pengcheng Feng,
Pinghua Gong, and Jieping Ye. 2017. A taxi order dispatch model based on
combinatorial optimization. In Proceedings of the 23rd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining. ACM, 2151 -- 2159.
[40] Lianmin Zheng, Jiacheng Yang, Han Cai, Weinan Zhang, Jun Wang, and Yong Yu.
2017. MAgent: A Many-Agent Reinforcement Learning Platform for Arti(cid:128)cial
Collective Intelligence. NIPS Demo (2017).
|
1603.05208 | 3 | 1603 | 2017-06-08T16:47:14 | Safe Sequential Path Planning Under Disturbances and Imperfect Information | [
"cs.MA",
"eess.SY"
] | Multi-UAV systems are safety-critical, and guarantees must be made to ensure no unsafe configurations occur. Hamilton-Jacobi (HJ) reachability is ideal for analyzing such safety-critical systems; however, its direct application is limited to small-scale systems of no more than two vehicles due to an exponentially-scaling computational complexity. Previously, the sequential path planning (SPP) method, which assigns strict priorities to vehicles, was proposed; SPP allows multi-vehicle path planning to be done with a linearly-scaling computational complexity. However, the previous formulation assumed that there are no disturbances, and that every vehicle has perfect knowledge of higher-priority vehicles' positions. In this paper, we make SPP more practical by providing three different methods to account for disturbances in dynamics and imperfect knowledge of higher-priority vehicles' states. Each method has different assumptions about information sharing. We demonstrate our proposed methods in simulations. | cs.MA | cs | Safe Sequential Path Planning Under Disturbances and Imperfect Information
Somil Bansal*, Mo Chen*, Jaime F. Fisac, and Claire J. Tomlin
7
1
0
2
n
u
J
8
]
A
M
.
s
c
[
3
v
8
0
2
5
0
.
3
0
6
1
:
v
i
X
r
a
Abstract- Multi-UAV systems are safety-critical, and guar-
antees must be made to ensure no unsafe configurations occur.
Hamilton-Jacobi (HJ) reachability is ideal for analyzing such
safety-critical systems; however, its direct application is limited
to small-scale systems of no more than two vehicles due to an
exponentially-scaling computational complexity. Previously, the
sequential path planning (SPP) method, which assigns strict
priorities to vehicles, was proposed; SPP allows multi-vehicle
path planning to be done with a linearly-scaling computational
complexity. However, the previous formulation assumed that
there are no disturbances, and that every vehicle has perfect
knowledge of higher-priority vehicles' positions. In this paper,
we make SPP more practical by providing three different
methods to account for disturbances in dynamics and imperfect
knowledge of higher-priority vehicles' states. Each method has
different assumptions about information sharing. We demon-
strate our proposed methods in simulations.
I. INTRODUCTION
Recently, there has been an immense surge of interest in
using unmanned aerial systems (UASs) for civil purposes
[1]–[4]. Many of these applications will involve unmanned
aerial vehicles (UAVs) flying in urban environments. As a
result, government agencies such as the Federal Aviation
Administration (FAA) and National Aeronautics and Space
Administration (NASA) of the United States are trying to
develop new scalable ways to organize an air space in which
potentially thousands of UAVs can fly together [5], [6].
One essential problem that needs to be addressed is how
a group of vehicles in the same vicinity can reach their
destinations while avoiding collision with each other. In some
previous studies that address this problem, specific control
strategies for the vehicles are assumed, and approaches such
as induced velocity obstacles have been used [7]–[9]. Other
researchers have used ideas involving virtual potential fields
to maintain collision avoidance while maintaining a specific
formation [10], [11]. Although interesting results emerge
from these studies, simultaneous trajectory planning and
collision avoidance were not considered.
Trajectory planning and collision avoidance problems in
safety-critical systems have been studied using Hamilton-
Jacobi (HJ) reachability analysis, which provides guarantees
on the success and safety of optimal system trajectories [12]–
[16]. In this context, one computes the reachable set, defined
as the set of states from which the system can be driven
to a target set. HJ reachability has been successfully used
This work has been supported in part by NSF under CPS:ActionWebs
(CNS-931843), by ONR under the HUNT (N0014-08-0696) and SMARTS
(N00014-09-1-1051) MURIs and by grant N00014-12-1-0609, by AFOSR
under the CHASE MURI (FA9550-10-1-0567). The research of J. F. Fisac
has received funding from the "la Caixa" Foundation.
* Both authors contributed equally to this work. All authors are with the
Department of Electrical Engineering and Computer Sciences, University of
California, Berkeley. {somil, mochen72, jfisac, tomlin}@eecs.berkeley.edu
in applications involving systems with no more than two
vehicles [13], [17]–[19]. However, HJ reachability cannot be
directly applied to systems involving multiple vehicles due
to its exponentially scaling computational complexity.
To overcome this problem, [20] presents sequential path
planning (SPP),
in which vehicles are assigned a strict
priority ordering. In SPP, higher-priority vehicles ignore
the lower-priority vehicles, which must take into account
the presence of higher-priority vehicles by treating them
as induced time-varying obstacles. Under this structure,
computation complexity scales just linearly with the number
of vehicles. In addition, a structure like this has the potential
to flexibly divide up the airspace for the use of many UAVs;
this is an important task in NASA's concept of operations
for UAS traffic management [6].
The formulation in [20], however, ignores disturbances and
assumes perfect information about other vehicles' trajecto-
ries. In presence of disturbances, a vehicle's state trajectory
evolution cannot be precisely known a priori; thus, it is
impossible to commit to exact trajectories as required in
[20]. In such a scenario, a lower-priority vehicle needs to
account for all possible states that the higher-priority vehicles
could be in. To do this, the lower-priority vehicle needs to
have some knowledge about the control policy used by each
higher-priority vehicle. The main contribution of this paper
is to take advantage of the computation benefits of the SPP
scheme while resolving some of its practical challenges. In
particular, we achieve the following:
• incorporate disturbances into the vehicle models,
• analyze three different assumptions on information to
which lower-priority vehicles may have access,
• for each information pattern, propose a reachability-
based method to compute the induced obstacles and the
reachable sets that guarantee collision avoidance as well
as successful transit to the destination.
II. PROBLEM FORMULATION
Consider N vehicles, denoted Qi, i = 1, . . . , n, whose
dynamics are described by the ordinary differential equation
t ≤ tSTA
i = 1, . . . , N
i
xi = fi(t, xi, ui, di),
ui ∈ Ui, di ∈ Di,
(1)
where xi ∈ Rni, ui denote the state and control of ith vehicle
Qi respectively, and di denotes the disturbance experienced
by Qi. In general,
the physical meaning of xi and the
dynamics fi depend on the specific dynamic model of Qi,
and need not be the same across the different vehicles. tSTA
in (1) denotes the scheduled time of arrival of Qi.
For convenience, we will use the sets Ui, Di to denote
the set of functions from which the control and disturbance
i
functions ui(·), di(·) can be drawn. Let pi ∈ Rp denote the
position of Qi. Denote the rest of the states hi, so that xi =
(pi, hi). The initial state of Qi is given by xi0. Under the
worst case disturbance, each vehicle aims to get to some set
of target states, denoted Ti ⊂ Rni, by some scheduled time
. On its way to Ti, each vehicle must avoid
of arrival tSTA
the danger zones Aij(t) of all other vehicles j (cid:54)= i for all
time. In general, the danger zone can be defined to capture
any undesirable configurations between Qi and Qj. In this
paper, we define Aij(t) as
i
Aij(t) = {xi ∈ Rni : (cid:107)pi − pj(t)(cid:107)2 ≤ Rc}
(2)
ferential game of dimension (cid:80)
the interpretation of which is that a vehicle is in another
vehicle's danger zone if the two vehicles are within a
Euclidean distance of Rc apart.
The problem of driving each of the vehicles in (1) into
their respective target sets Ti would be in general a dif-
i ni. However, due to the
exponential scaling of the complexity with the problem
dimension, an optimal solution is computationally intractable
even for N > 2, with ni as small as 3.
In this paper, we assume that vehicles have assigned
priorities as in the SPP method [20]. Since the analysis in
[20] did not take into account the presence of disturbances di
and limited information available to each vehicle, we extend
the work in [20] to answer the following:
1) How can each vehicle guarantee that it will reach its
target set without getting into any danger zones, despite
the disturbances it and other vehicles experience?
2) How should each vehicle robustly handle situations
with limited information about the state, control policy,
and intention of other vehicles?
III. BACKGROUND
This section provides a brief summary of [20], in which
the SPP scheme is proposed under perfect information and
absence of disturbances. Here, the dynamics of Qi becomes
where the difference compared to (1) is that the disturbance
di is no longer a part of the dynamics.
In order to make the N-vehicle path planning problem
safe and tractable, a reasonable structure is imposed to the
problem: the vehicles are assigned a strict priority ordering.
When planning its trajectory to its target, a higher-priority
vehicle can disregard the presence of a lower-priority vehicle.
In contrast, a lower-priority vehicle must take into account
the presence of all higher-priority vehicles, and plan its
trajectory in a way that avoids the higher-priority vehicles'
danger zones. For convenience and without lost of generality,
let Qi be the vehicle with the ith highest priority.
Under the above convention, each vehicle Qi must take
time-varying obstacles induced by vehicles
into account
Qj, j < i, denoted Oj
i (t) and represent the set of states
that could possibly be in the danger zone of Qj. Optimal
xi = fi(t, xi, ui),
ui ∈ Ui,
i = 1, . . . , N
t ≤ tSTA
i
max
(3)
safe path planning of each lower-priority vehicle Qi then
consists of determining the optimal path that allows Qi to
reach its target Ti while avoiding the time-varying obstacles
Gi(t), defined by
i−1(cid:91)
j=1
Gi(t) =
Oj
i (t)
(4)
Such an optimal path planning problem can be solved by
computing a backward reachable set (BRS) Vi(t) from a
target set Ti using formulations of HJ variational inequalities
(VI) such as [12], [14], [16], [21]. For example, to compute
BRSs under the presence of time-varying obstacles,
the
authors in [21] augment system with the time variable, and
then applied reachability theory for time-invariant systems.
To avoid increasing the problem dimension and save compu-
tation time, for the simulations of this paper we utilize the
formulation in [16], which does not require augmentation of
the state space with the time variable.
Starting from the highest-priority vehicle Q1, one com-
putes the BRS V1(t), from which the optimal control and
trajectory x1(·) to T1 can be obtained. Under the absence
of disturbances and perfect information, obstacles induced
by a higher-priority vehicle Qj, starting with j = 1, for a
lower-priority vehicle Qi is simply the danger zone centered
around the position pj(·) of each point on the trajectory:
i (t) = {xi : (cid:107)pi − pj(t)(cid:107) ≤ Rc}
Oj
(5)
Given Oj
i (t), j < i, and continuing with i = 2, the optimal
safe trajectories for each vehicle Qi can be computed. All
of the trajectories are optimal in the sense that given the
requirement that Qi must arrive at Ti by time tSTA
, the
i (·) that
and the optimal control u∗
latest departure time tLDT
guarantees arrival by tSTA
can be obtained.
To compute Vi(t) using the method in [16], we solve the
following HJ VI for t ≤ tSTA
(cid:110)
min(cid:8)DtVi(t, xi) + Hi (t, xi, DxiVi) ,
(cid:111)
li(xi) − Vi(t, xi)(cid:9),−gi(t, xi) − Vi(t, xi)
, xi) = max(cid:8)li(xi),−gi(0, xi)(cid:9)
= 0
(6)
:
Vi(tSTA
i
i
i
i
i
Hi (t, xi, λ) = min
ui∈Ui
λ · fi(t, xi, ui)
(7)
where λ is the gradient of the value function, DxiVi, and
li(xi), gi(t, xi), Vi(t, xi) are implicit surface functions rep-
resenting the target Ti, the time-varying obstacles Gi(t), and
the backward reachable set Vi(t), respectively:
xi ∈ Ti ⇔ li(xi) ≤ 0
xi(t) ∈ Gi(t) ⇔ gi(t, xi) ≤ 0
xi(t) ∈ Vi(t) ⇔ Vi(t, xi) ≤ 0
The optimal control is given by
u∗
i (t, xi) = arg min
ui∈Ui
λ · fi(t, xi, ui)
(8)
(9)
IV. DISTURBANCES AND INCOMPLETE INFORMATION
Disturbances and incomplete information significantly
complicate the SPP scheme. The main difference is that the
vehicle dynamics satisfy (1) as opposed to (3). Committing
to exact trajectories is therefore no longer possible, since
the disturbance di(·) is a priori unknown. Thus, the induced
obstacles Oj
i (t) are no longer just the danger zones centered
around positions. We present three methods to address the
above issues. The methods differ in terms of control policy
information that is known to a lower-priority vehicle about a
higher-priority vehicle, and have their relative advantages and
disadvantages depending on the situation. The three methods
are as follows:
• Centralized control: A specific control strategy is en-
forced upon a vehicle; this can be achieved, for example,
by some central agent such as an air traffic controller.
• Least restrictive control: A vehicle is required to arrive
at its targets on time, but has no other restrictions.
• Robust trajectory tracking: A vehicle declares a nom-
inal trajectory which can be robustly tracked.
In general, the above methods can be used in combination
in a single path planning problem, with each vehicle inde-
pendently having different control policies. Lower-priority
vehicles would then plan their paths while taking into ac-
count the control policy information known for each higher-
priority vehicle. For clarity, we will present each method as
if all vehicles are using the same method of path planning.
For simplicity of explanation, we assume that no static
obstacles exist. If static obstacles do exist, the time-varying
obstacles Gi(t) simply become the union of the induced
obstacles Oj
A. Method 1: Centralized Control
i (t) in (4) and the static obstacles.
The highest-priority vehicle Q1 first plans its path by
computing the BRS (with i = 1)
Vi(t) = {xi : ∃ui(·) ∈ Ui,∀di(·) ∈ Di, xi(·) satisfies (1),
∀s ∈ [t, tSTA
], xi(s) /∈ Gi(s),∃s ∈ [t, tSTA
], xi(s) ∈ Ti}
(10)
Since we have assumed no static obstacles exist, we have
that for Q1,G1(s) = ∅ ∀s ≤ tSTA
, and thus the above BRS
is well-defined. This BRS can be computed by solving the
HJ VI (6) with the following Hamiltonian:
i
i
i
Hi (t, xi, λ) = min
ui∈Ui
max
di∈Di
λ · fi(t, xi, ui, di)
(11)
where li(xi), gi(t, xi), Vi(t, xi) are implicit surface functions
representing the target Ti,Gi(t),Vi(t), respectively. From the
BRS, we can obtain the optimal control
u∗
i (t, xi) = arg min
ui∈Ui
max
di∈Di
λ · fi(t, xi, ui, di)
(12)
i
departure time tLDT
Here, as well as in the other two methods, the latest
is then given by arg supt xi0 ∈ Vi(t).
If there is a centralized controller directly controlling each
of the N vehicles, then the control law of each vehicle
can be enforced. In this case, lower-priority vehicles can
safely assume that higher-priority vehicles are applying the
enforced control law. In particular, the optimal controller for
i (t, xi) can be enforced. In this case,
getting to the target, u∗
the dynamics of each vehicle becomes
i (t, xi, di) = fi(t, xi, u∗
xi = f∗
di ∈ Di,
i = 1, . . . , N,
i (t, xi), di)
, tSTA
t ∈ [tLDT
i
i
(13)
]
where ui no longer appears explicitly in the dynamics.
From the perspective of a lower-priority vehicle Qi, a
higher-priority vehicle Qj, j < i induces a time-varying
obstacle that represents the positions that could possibly
be within the capture radius Rc of Qj under the dynamics
f∗
j (t, xj, dj). Determining this obstacle involves computing a
forward reachable set (FRS) of Qj starting from xj(tLDT
) =
xj0. The FRS Wj(t) is defined as follows:
j
Wj(t) = {y ∈ Rnj : ∃dj(·) ∈ Dj,
xj(·) satisfies (13), xj(tLDT
The FRS can be computed using the following HJ VI:
, tSTA
DtWj(t, xj) + Hj
]
j
(cid:0)t, xj, Dxj Wj
) = xj0, xj(t) = y}
(cid:1) = 0, t ∈ [tLDT
j
j
(14)
j
, xj) = ¯lj(xj)
Wj(tLDT
Hj (t, xj, λ) = max
dj∈Dj
λ · f∗
j (t, xj, dj)
j
(15)
where ¯l is chosen to be1 such that ¯l(y) = 0 ⇔ y = xj(tLDT
).
The FRS Wj(t) represents the set of possible states at
time t of a higher-priority vehicle Qj given all possible
disturbances dj(·) and given that Qj uses the feedback
controller u∗
j (t, xj). In order for a lower-priority vehicle Qi
to guarantee that it does not go within a distance of Rc to
Qj, Qi must stay a distance of at least Rc away from the set
Wj(t) for all possible values of the non-position states hj.
This gives the obstacle induced by a higher-priority vehicle
Qj for a lower-priority vehicle Qi as follows:
Oj
i (t) = {xi : dist(pi,Pj(t)) ≤ Rc}
(16)
where the dist(·,·) function represents the minimum distance
from a point to a set, and the set Pj(t) is the set of states in
the FRS Wj(t) projected onto the states representing position
pj, and disregarding the non-position dimensions hj:
Pj(t) = {pj : ∃hj, (pj, hj) ∈ Wj(t)}.
(17)
Finally, taking the union of the induced obstacles Oj
i (t)
as in (4) gives us the time-varying obstacles Gi(t) needed to
define and determine the BRS Vi(t) in (10). Repeating this
process, all vehicles will be able to plan paths that guarantee
the vehicles' timely and safe arrival.
B. Method 2: Least Restrictive Control
Here, we again begin with the highest-priority vehicle
Q1 planning its path by computing the BRS Vi(t) in (10).
However, if there is no centralized controller to enforce the
control policy for higher-priority vehicles, weaker assump-
tions must be made by the lower-priority vehicles to ensure
collision avoidance. One reasonable assumption that a lower-
priority vehicle can make is that all higher-priority vehicles
1In practice, we define the target set to be a small region around the
vehicle's initial state for computational reasons.
follow the least restrictive control that would take them to
their targets. This control would be given by
j (t, xj) given by (12)} if xj(t) ∈ ∂Vj(t),
uj(t, xj) ∈
(cid:40){u∗
Uj otherwise
(18)
Such a controller allows each vehicle to use any controller,
except when it is on the boundary of the BRS, ∂Vj(t), in
which case u∗
j (t, xj) given by (12) must be used to get to
the target safely and on time. This assumption is the weakest
one that could be made by lower-priority vehicles given that
the higher-priority vehicles will get to their targets on time.
Suppose a lower-priority vehicle Qi assumes that higher-
priority vehicles Qj, j < i use the least restrictive control
strategy in (18). From the perspective of Qi, a higher-priority
vehicle Qj could be in any state that is reachable from Qj's
) = xj0 and from which the target Tj can
initial state xj(tLDT
be reached. Mathematically, this is defined by the intersection
of a FRS from the initial state xj(tLDT) = xj0 and the BRS
defined in (10) from the target set Tj, Vj(t) ∩ Wj(t). In
this situation, since Qj cannot be assumed to be using any
particular feedback control, Wj(t) is defined as
j
Wj(t) = {y ∈ Rnj : ∃uj(·) ∈ Uj,∃dj(·) ∈ Dj,
) = xj0, xj(t) = y}
xj(·) satisfies (1), xj(tLDT
j
(19)
This FRS can be computed by solving (15) without
λ · fj(t, xj, uj, dj)
(20)
obstacles, and with
Hj (t, xj, λ) = max
uj∈Uj
max
dj∈Dj
In turn, the obstacle induced by a higher-priority Qj for
a lower-priority vehicle Qi is as follows:
Oj
i (t) = {xi : dist(pi,Pj(t)) ≤ Rc}, with
Pj(t) = {pj : ∃hj, (pj, hj) ∈ Vj(t) ∩ Wj(t)}
(21)
C. Method 3: Robust Trajectory Tracking
Although it is impossible to commit to and track an exact
trajectory in the presence of disturbances, it may still be
possible to robustly track a nominal trajectory with a bounded
error at all times. If this can be done, then the tracking error
bound can be used to determine the induced obstacles. Here,
computation is done in two phases: the planning phase and
the disturbance rejection phase. In the planning phase, we
compute a nominal trajectory xr,j(·) that is feasible in the
absence of disturbances. In the disturbance rejection phase,
we compute a bound on the tracking error.
In the planning phase, planning is done for a reduced
control set U p ⊂ U, as some margin is needed to reject
unexpected disturbances while tracking the nominal trajec-
tory. In the disturbance rejection phase, we determine the
error bound independently of the nominal trajectory. Let
xj and xr,j denote the states of the actual vehicle Qj and
an arbitrary nominal trajectory, respectively, and define the
tracking error ej = xj − xr,j. When the error dynamics are
independent of the absolute state as in (22) (and also (7) in
[13]), we can obtain error dynamics of the form
To obtain bounds on the tracking error, we first conser-
vatively estimate the error bound around any reference state
xr,j, denoted Ej = {ej : (cid:107)pej(cid:107)2 ≤ REB}, where pej denotes
the position coordinates of ej and REB is a design parameter.
We next solve a reachability problem with its complement
E c
j , the set of tracking errors violating the error bound, as
the target in the space of the error dynamics. From E c
j , we
compute the following BRS:
VEB
j (t, 0) = {y : ∀uj(·) ∈ Uj,∃ur,j(·) ∈ Up
ej(·) satisfies (22), ej(t) = y,∃s ∈ [t, 0], ej(s) ∈ E c
j},
where the Hamiltonian to compute the BRS is given by:
λ · fej (ej, uj, ur,j, dj).
HEB
j ,∃dj(·) ∈ Di,
(23)
j (t, 0)(cid:1)c.
:= limt→−∞(cid:0)VEB
(24)
Letting t → −∞, we obtain the infinite-horizon control-
invariant set Ωj
is
nonempty, then the tracking error ej at flight time is guaran-
teed to remain within Ωj ⊆ Ej provided that the vehicle starts
inside Ωj and subsequently applies the feedback control law
λ · fej (ej, uj, ur,j, dj).
(25)
The induced obstacles by each higher-priority vehicle Qj
κj(ej) = arg max
uj∈Uj
min
j ,dj∈Dj
ur,j∈U p
If Ωj
j (ej, λ) = max
uj∈Uj
min
j ,dj∈Dj
ur,j∈U p
can thus be obtained by:
Oj
i (t) = {xi : ∃y ∈ Pj(t),(cid:107)pi − y(cid:107)2 ≤ Rc}
Pj(t) = {pj : ∃hj, (pj, hj) ∈ Ωj + xr,j(t)},
where the "+" in (26) denotes the Minkowski sum.
(26)
Since each vehicle Qj, j < i, can only be guaranteed to
stay within Ωj, we must make sure during the path planning
of Qi that at any given time, the error bounds of Qi and Qj,
Ωi and Ωj, do not intersect. This can be done by augmenting
the total obstacle set by Ωi:
Gi(t) = Gi(t) + Ωi.
(27)
Finally, given Ωi, we can guarantee that Qi will reach its
target Ti if Ωi ⊆ Ti; thus, in the path planning phase, we
modify Ti to be Ti := {xi : Ωi + xi ⊆ Ti}, and compute a
BRS, with the control authority U p
i , that contains the initial
state of the vehicle. Mathematically,
i , xi(·) satisfies (3),
) ={y : ∃ui(·) ∈ Up
Vrtt
i (t, tSTA
i
∀s ∈ [t, tSTA
∃s ∈ [t, tSTA
i
], xi(s) /∈ Gi(s),
], xi(s) ∈ Ti, xi(t) = y}
i
(28)
i
i (t, tSTA
The Hamiltonian to compute Vrtt
) and the optimal
control for reaching Ti are given by (7) and (9) respectively.
The nominal trajectory xr,i(·) can thus be obtained by using
i (·). From
vehicle dynamics (3), with the optimal control urtt
the resulting nominal trajectory xr,i(·), the overall control
policy to reach Ti can be obtained via (25).
V. NUMERICAL SIMULATIONS
ej = fej (ej, uj, ur,j, dj),
uj ∈ Uj, ur,j ∈ U p
j , dj ∈ Dj,
t ≤ 0
(22)
We demonstrate our proposed methods using a four-
vehicle example. Each vehicle has the following model:
px,i = vi cos θi + dx,i
py,i = vi sin θi + dy,i
θi = ωi + dθ,i
v ≤ vi ≤ ¯v,ωi ≤ ¯ω
(cid:107)(dx,i, dy,i)(cid:107)2 ≤ dr,dθ,i ≤ ¯dθ
where pi = (px,i, py,i), θi, d = (dx,i, dy,i, dθ,i) respectively
represent Qi's position, heading, and disturbances in the
three states. The control of Qi is ui = (vi, ωi), where vi is
the speed of Qi and ωi is the turn rate; both controls have a
lower and upper bound. For illustration purposes, we choose
v = 0.5, ¯v = 1, ¯ω = 1; however, our method can easily
handle the case in which these inputs differ across vehicles
and cases in which each vehicle has a different dynamic
model. The disturbance bounds are chosen as dr = 0.1, ¯dθ =
0.2, which correspond to a 10% uncertainty in the dynamics.
The initial states of the vehicles are given as follows:
1 = (−0.5, 0, 0),
x0
3 = (−0.6, 0.6, 7π/4) ,
x0
(29)
Each of the vehicles has a target set Ti that is circular in
their position pi centered at ci = (cx,i, cy,i) with radius r:
(30)
x0
2 = (0.5, 0, π),
x0
4 = (0.6, 0.6, 5π/4) .
Ti = {xi ∈ R3 : (cid:107)pi − ci(cid:107) ≤ r}
(a)
(b)
(c)
(d)
Fig. 1: Initial configuration and simulated trajectories of the
vehicles for the three proposed methods.
i
i
i
For the example shown, we chose c1 = (0.7, 0.2), c2 =
(−0.7, 0.2), c3 = (0.7,−0.7), c4 = (−0.7,−0.7) and r =
0.1. The setup of the example is shown in Fig. 1a.
Using the SPP algorithms presented, we obtain tLDT
, i =
= 0. Note that even though tSTA
1, 2, 3, 4 assuming tSTA
is assumed to be same for all vehicles in this example for
simplicity, our method can easily handle the case in which
tSTA
i
is different for each vehicle.
For each proposed method of computing induced obsta-
cles, we show the vehicles' entire trajectories (colored dotted
lines), and overlay their positions (colored asterisks) and
headings (arrows) at a point in time in which they are in
relatively dense configuration. In all cases, the vehicles are
able to avoid each other's danger zones (colored dashed
circles) while getting to their target sets in minimum time.
In addition, we show the evolution of the BRS over time for
Q3 (green boundaries) as well as the obstacles induced by
the higher-priority vehicles (black boundaries).
Fig. 2: Evolution of the BRS and the obstacles induced by
Q1 and Q2 for Q3 in the centralized control method.
Fig. 1b shows simulated trajectories in the situation where
each vehicle uses u∗
i (t, xi) in (12). In this case, vehicles ap-
pear to deviate slightly from a straight line trajectory towards
their targets, just enough to avoid higher-priority vehicles.
The deviation is small since the centralized controller is quite
restrictive, making the possible positions of higher-priority
vehicles cover a small area. In the dense configuration at
t = −1.0, the vehicles are close to each other but still outside
each other's danger zones.
Fig. 2 shows the evolution of the BRS for Q3 (green
boundary), as well as the obstacles (black boundary) induced
by the higher-priority vehicles. The size of the obstacles re-
mains relatively small. tLDT
numbers for the four vehicles (in
order) in this case are −1.35,−1.37,−1.94 and −2.04. They
are relatively close for the vehicles, because the obstacles
generated by higher-priority vehicles are small and hence do
not affect tLDT of the lower-priority vehicles significantly.
A. Least Restrictive Control
i
Fig. 1c shows the simulated trajectories in the situation
where each vehicle assumes that higher-priority vehicles use
the least restrictive control to reach their targets, as described
in IV-B. Fig. 3 shows the BRS and induced obstacles for Q3.
Q1 (red) takes a relatively straight path to reach its target.
From the perspective of all other vehicles, large obstacles
are induced, since lower-priority vehicles make the weak
assumption that higher-priority vehicles are using the least
restrictive control. Because the obstacles induced are so
large, it is optimal for lower-priority vehicles to wait until
higher-priority vehicles pass. As a result, a dense configura-
tion is never formed, and trajectories are relatively straight.
values for vehicles are −1.35,−1.97,−2.66 and
The tLDT
−3.39. Compared to the centralized control method, tLDT
's
decrease significantly except for Q1, which need not account
for any moving obstacles.
From Q3's (green) perspective, the large obstacles induced
by Q1 and Q2 are shown in Fig. 3 as the black boundaries. As
the BRS (green boundary) evolves over time, its growth gets
inhibited by the large obstacles for a long time, as evident
i
i
-0.500.5-0.500.5t = -0.34-0.500.5-0.500.5t = -1.04Initial pos. and headingObstacleBRSTargets-0.500.5-0.500.5t = -1.94-0.500.5-0.500.5Initial setupVehicle 1Vehicle 2Vehicle 3Vehicle 4-0.500.5-0.500.5Centralized control-0.500.5-0.500.5Least restrictive control-0.500.5-0.500.5Robust trajectory trackingTargetsPositions, HeadingsDanger ZonesTrajectories-0.500.5-0.500.5Initial setupVehicle 1Vehicle 2Vehicle 3Vehicle 4-0.500.5-0.500.5Centralized control-0.500.5-0.500.5Least restrictive control-0.500.5-0.500.5Robust trajectory trackingTargetsPositions, HeadingsDanger ZonesTrajectories-0.500.5-0.500.5Initial setupVehicle 1Vehicle 2Vehicle 3Vehicle 4-0.500.5-0.500.5Centralized control-0.500.5-0.500.5Least restrictive control-0.500.5-0.500.5Robust trajectory trackingTargetsPositions, HeadingsDanger ZonesTrajectories-0.500.5-0.500.5Initial setupVehicle 1Vehicle 2Vehicle 3Vehicle 4-0.500.5-0.500.5Centralized control-0.500.5-0.500.5Least restrictive control-0.500.5-0.500.5Robust trajectory trackingTargetsPositions, HeadingsDanger ZonesTrajectories-0.500.5-0.500.5Initial setupVehicle 1Vehicle 2Vehicle 3Vehicle 4-0.500.5-0.500.5Centralized control-0.500.5-0.500.5Least restrictive control-0.500.5-0.500.5Robust trajectory trackingPositions, HeadingsDanger ZonesTargetsTrajectories at t = −0.89. Eventually, the boundary of the BRS reaches
the initial state of Q3 at t = tLDT
3 = −2.66.
made. In all of the methods, all vehicles are guaranteed
to successfully reach their respective destinations without
entering each other's danger zones despite the worst-case
disturbance the vehicles could experience.
REFERENCES
Fig. 3: Evolution of the BRS for Q3 in the least restrictive
control method. tLDT
is significantly lower than that in the
centralized control method (−1.94 vs. −2.66).
B. Robust Trajectory Tracking
3
In the planning phase, we reduced the maximum turn rate
of the vehicles from 1 to 0.6, and the speed range from
[0.5, 1] to exactly 0.75 (constant speed). With these reduced
control authorities, we determined from the disturbance
rejection phase that any nominal trajectory from the planning
phase can be robustly tracked within a distance of 0.075.
Fig. 1d shows vehicle trajectories in the situation where
each vehicle robustly tracks a nominal trajectory. Fig. 4
shows the BRS evolution and induced obstacles for Q3.
Fig. 4: Evolution of the BRS for Q3 in the robust trajectory
tracking method. Note that a smaller target set is used to
ensure target reaching for any allowed tracking error.
i
In this case, the tLDT
values for the four vehicles are
−1.61,−3.16,−3.57 and −2.47 respectively. In this method,
vehicles use reduced control authority for path planning
towards a reduced-size effective target set. As a result,
higher-priority vehicles tend to have lower tLDT compared
to the other two methods, as evident from tLDT
. Because of
this "sacrifice" made by the higher-priority vehicles during
the path planning phase, the tLDT's of lower-priority vehicles
may increase compared to those in the other methods, as
evident from tLDT
will
change for a vehicle compared to the other methods, as the
conservative path planning increases tLDT
for higher-priority
vehicles and decreases tLDT
for lower-priority vehicles.
is unclear how tLDT
. Overall,
it
1
4
i
i
i
VI. CONCLUSIONS
We have proposed three different methods to account
for disturbances and imperfect control policy information
in sequential path planning;
these three methods can be
used independently across the different vehicles in the path
planning problem. In each method, different assumptions
about
the control strategy of higher-priority vehicles are
[1] W. M. Debusk, "Unmanned aerial vehicle systems for disaster relief:
Tornado alley," in Infotech@Aerospace Conferences, 2010.
[2] Amazon.com, Inc. (2016) Amazon prime air. [Online]. Available:
http://www.amazon.com/b?node=8037720011
[3] AUVSI News. (2016) UAS aid in south carolina tornado investigation.
[Online]. Available: http://www.auvsi.org/blogs/auvsi-news/2016/01/
29/tornado
[4] BBC
Technology.
service for 2017.
technology-34704868
(2016) Google
delivery
[Online]. Available: http://www.bbc.com/news/
drone
plans
[5] Jointed Planning and Development Office (JPDO), "Unmanned aircraft
systems (UAS) comprehensive plan – a report on the nation's UAS
path forward," Federal Aviation Administration, Tech. Rep., 2013.
[6] P. Kopardekar, J. Rios, T. Prevot, M. Johnson, J. Jung, and J. E. R.
III, "UAS traffic management (UTM) concept of operations to safely
enable low altitude flight operations," in AIAA Aviation Technology,
Integration, and Operations Conference, 2016.
[7] P. Fiorini and Z. Shillert, "Motion planning in dynamic environments
using velocity obstacles," International Journal of Robotics Research,
vol. 17, pp. 760–772, 1998.
[8] G. C. Chasparis and J. S. Shamma, "Linear-programming-based multi-
vehicle path planning with adversaries," in Proceedings of American
Control Conference, June 2005.
[9] J. van den Berg, M. C. Lin, and D. Manocha, "Reciprocal velocity
obstacles for real-time multi-agent navigation," in IEEE International
Conference on Robotics and Automation, May 2008, pp. 1928–1935.
[10] R. Olfati-Saber and R. M. Murray, "Distributed cooperative control
of multiple vehicle formations using structural potential functions," in
IFAC World Congress, 2002.
[11] Y.-L. Chuang, Y. Huang, M. R. D'Orsogna, and A. L. Bertozzi, "Multi-
vehicle flocking: Scalability of cooperative control algorithms using
pairwise potentials," in IEEE International Conference onRobotics and
Automation, April 2007, pp. 2292–2299.
[12] E. N. Barron, "Differential Games with Maximum Cost," Nonlinear
analysis: Theory, methods & applications, pp. 971–989, 1990.
[13] I. Mitchell, A. Bayen, and C. Tomlin, "A time-dependent Hamilton-
Jacobi formulation of reachable sets for continuous dynamic games,"
IEEE Transactions on Automatic Control, vol. 50, no. 7, pp. 947–957,
July 2005.
[14] O. Bokanowski, N. Forcadel, and H. Zidani, "Reachability and mini-
mal times for state constrained nonlinear problems without any con-
trollability assumption," SIAM Journal on Control and Optimization,
pp. 1–24, 2010.
[15] K. Margellos and J. Lygeros, "Hamilton-Jacobi Formulation for
Reach-Avoid Differential Games," IEEE Transactions on Automatic
Control, vol. 56, no. 8, Aug 2011.
[16] J. F. Fisac, M. Chen, C. J. Tomlin, and S. S. Shankar, "Reach-avoid
problems with time-varying dynamics, targets and constraints," in
18th International Conference on Hybrid Systems: Computation and
Controls, 2015.
[17] J. Ding, J. Sprinkle, S. S. Sastry, and C. J. Tomlin, "Reachability
calculations for automated aerial refueling," in IEEE Conference on
Decision and Control, Cancun, Mexico, 2008.
[18] H. Huang, J. Ding, W. Zhang, and C. Tomlin, "A differential game
approach to planning in adversarial scenarios: A case study on
capture-the-flag," in Robotics and Automation (ICRA), 2011 IEEE
International Conference on, 2011, pp. 1451–1456.
[19] A. M. Bayen, I. M. Mitchell, M. Oishi, and C. J. Tomlin, "Aircraft
autolander safety analysis through optimal control-based reach set
computation," Journal of Guidance, Control, and Dynamics, vol. 30,
no. 1, 2007.
[20] M. Chen, J. Fisac, C. J. Tomlin, and S. Sastry, "Safe sequential path
planning of multi-vehicle systems via double-obstacle hamilton-jacobi-
isaacs variational inequality," in European Control Conference, 2015.
[21] O. Bokanowski and H. Zidani, "Minimal time problems with moving
targets and obstacles," {IFAC} Proceedings Volumes, vol. 44, no. 1,
pp. 2589 – 2593, 2011.
-0.500.5-0.500.5t = -0.39-0.500.5-0.500.5t = -0.89Initial pos. and headingObstacleBRSTargets-0.500.5-0.500.5t = -2.66-0.500.5-0.500.5t = -0.50-0.500.5-0.500.5t = -2.59Initial pos. and headingObstacleBRSTargets-0.500.5-0.500.5t = -3.57 |
1406.7487 | 3 | 1406 | 2015-03-22T21:37:41 | Coalition Formation and Combinatorial Auctions; Applications to Self-organization and Self-management in Utility Computing | [
"cs.MA",
"cs.GT"
] | In this paper we propose a two-stage protocol for resource management in a hierarchically organized cloud. The first stage exploits spatial locality for the formation of coalitions of supply agents; the second stage, a combinatorial auction, is based on a modified proxy-based clock algorithm and has two phases, a clock phase and a proxy phase. The clock phase supports price discovery; in the second phase a proxy conducts multiple rounds of a combinatorial auction for the package of services requested by each client. The protocol strikes a balance between low-cost services for cloud clients and a decent profit for the service providers. We also report the results of an empirical investigation of the combinatorial auction stage of the protocol. | cs.MA | cs |
Abstract
In this paper we propose a two-stage protocol for re-
source management in a hierarchically organized cloud.
The first stage exploits spatial locality for the formation
of coalitions of supply agents; the second stage, a com-
binatorial auction, is based on a modified proxy-based
clock algorithm and has two phases, a clock phase and a
proxy phase. The clock phase supports price discovery;
in the second phase a proxy conducts multiple rounds of
a combinatorial auction for the package of services re-
quested by each client. The protocol strikes a balance
between low-cost services for cloud clients and a decent
profit for the service providers. We also report the re-
sults of an empirical investigation of the combinatorial
auction stage of the protocol.
Coalition Formation and Combinatorial Auctions; Applications to
Self-organization and Self-management in Utility Computing
Dan C. Marinescu and Ashkan Paya
Computer Science Division
Department of Electrical Engineering and Computer Science
University of Central Florida, Orlando, FL 32816, USA
Email: [dcm, apaya]@cs.ucf.edu
John P. Morrison
Computer Science Department
University College Cork. Cork, Ireland
Email: [email protected]
May 22, 2021
Control theory tells us that accurate state informa-
tion and a tight feedback loop are the critical elements
for effective control of a system.
In a hierarchical or-
ganization the quality of state information degrades as
we move from the bottom to the top; only local infor-
mation about the state of a server is by definition accu-
rate. Moreover, this information is volatile, it must be
acted upon promptly because the state changes rapidly.
Our recent results [22] confirm that hierarchical control
has considerably larger overhead than a simple economic
model for cloud resource management. The communi-
cation complexity of hierarchical control based on moni-
toring is more than two orders of magnitude higher and
consumes a significant fraction of the available band-
width at all levels of the interconnection network.
Existing solutions for cloud resource management are
neither effective nor scalable and they require detailed
models of the system and accurate information about
the state of individual servers. The polices for cloud re-
source management must support: (i) admission control;
(ii) capacity allocation; (iii) load balancing; (iv) energy
optimization, and (v) quality of service (QoS) [21]. Some
of the existing and future challenges for cloud resource
management which affect these policies are:
-1. The cloud infrastructure is increasingly more hetero-
geneous; servers with different configurations of multi-
core processors, GPUs, FPGAs, and data flow engines
are expected to become elements of the cloud computing
landscape.
-2. The spectrum of cloud services and cloud applica-
tions widens. For example, AWS added new services,
including Elastic Cache, and Dynamo DB, offers several
types of EC2 (Elastic Cloud Computing) profiles includ-
ing C3 - compute optimized, R3 - memory optimized;
1 Introduction and Motivation
Nowadays large farms of computing and storage servers
are assembled to support several cloud delivery models
including Software as a Service (SaaS), Platform as a
Service (PaaS), and Infrastructure as a Service (IaaS).
In such systems users pay only for computing resources
they use, similarly to other utilities such as electricity
and water.
Computer clouds raise the question of how far we can
push the limits of composability of computing and com-
munication systems, while still being able to support
effective policies for resource management and their im-
plementation mechanisms. The software, the glue al-
lowing us to build increasingly more complex systems,
consists of more and more layers thus, the challenge of
controlling large-scale systems is amplified.
1
each instance type provides different sets of computer
resources measured by vCPUs (vCPU is a hyper-thread
of an Intel Xeon core for M3, C3, R3, HS1, G2, and I2).
This process is expected to continue at a faster pace to
accommodate the so-called Big Data applications.
-3. Cloud over-provisioning demands high initial costs
and leads to a low system utilization; this strategy is
not economically sustainable [10]. Elasticity allows cloud
users to increase or decrease their resource consumption
based on their needs; elasticity is now based on over-
provisioning, assembling pools of resources far larger
than required to satisfy the average needs. As a re-
sult the average cloud server utilization is in the 18% to
30% range [4, 5]. Power consumption of clouds based on
over-provisioning is excessive and has a negative ecolog-
ical impact [4, 26]. A 2010 survey [6] reports that idle
or under utilized servers contribute 11 million tonnes of
unnecessary CO2 emissions each year and that the total
yearly cost for the idle servers is $19 billion.
-4. The cloud computing landscape is fragmented. CSPs
support different cloud delivery models. and this leads
to the vendor lock-in; once becoming familiar and stor-
ing her data on one cloud it is very costly for the user
to migrate to another CSP.
In a large scale-system tensions between local and
global objectives exist. These tensions manifest them-
selves in questions such as: How to balance the individ-
ual cost of autonomous servers with global goals e.g.,
maximizing the CSP payoff? How to adapt the price
for services to the actual demand? How to find an
equilibrium between system reconfiguration and contin-
uous system availability? Moreover, cooperation must
reflect the particular characteristics of the physical or-
ganization. Locality is important; indeed, communica-
tion across multiple layers of the networking infrastruc-
ture is less desirable as the latency increases and the
bandwidth decreases. The hypothesis of our research
is that self-organization and self-management could ad-
dress these challenges and provide effective means for
cloud resource management.
Informally, self-organization means synergetic activ-
ities of elements when no single element acts as a co-
ordinator and the global patterns of behavior are dis-
tributed. Self-management means that individuals can
effectively set their own goals, make decisions on how
to achieve those goals, plan and schedule their activities
independently, and evaluate the progress towards these
goals. Self-management can lead to faster and more ac-
curate resource management decisions.
Self-management as a result of auctions eliminates the
need for a system model and requires only local thus,
more accurate information about the state of individ-
ual components. This approach has the potential of
optimizing the use of resources and allow Cloud Ser-
vice Providers (CSPs) to offer services at a lower cost
for the consumers [13, 19]. Though the virtues of self-
management have long been recognized, there is, to our
knowledge, no cloud computing infrastructure, based on
self-organizing principles and self-management. This is
in itself proof of the difficulties to apply these concepts
in practice.
Self-management has to be coupled with some mech-
anisms for coalition formation allowing autonomous
agents, the servers, to act in concert. Autonomous sys-
tems have to cooperate to guarantee QoS by distribut-
ing and balancing the workload, replicate services to
increase reliability, and implement other global system
policies. Cooperation means that individual systems
have to partially surrender their autonomy.
Self-organization cannot occur instantaneously in an
adaptive system. It is critical to give the autonomous
cloud platforms interconnected by a hierarchy of net-
works the time to form coalitions in response to ser-
vices requests thus, self-management requires an effec-
tive reservation system. Reservations are ubiquitous for
systems offering services to a large customer population,
e.g., airline ticketing, chains of hotels, and so on. Ex-
isting clouds, e.g., the Amazon Web Services, offer both
reservations and spot access, with spot access rates lower
than those for reservations.
The solution discussed in this paper involves concepts,
policies, and algorithms from several well-established ar-
eas of economics and computer science: self-organization
and self-management of complex systems; coalition for-
mation and virtual organizations; auction theory and
practice; and system organization and computer archi-
tecture. We discuss related work and our contributions
in Section 2 and in Section 3 we describe the system
model. Algorithms for the formation of sub-coalitions
and for clock-proxy auction are the subjects of Sections
4 and 5, respectively. The results of a simulation exper-
iment and the conclusions of our work are presented in
Sections 6 and 7.
2 Related Work
The present and future challenges outlined in Section
1 motivate the search for effective and scalable policies
and mechanisms for cloud resource management [7, 9,
18, 23, 25, 29, 36]. In this section we survey some of the
research in this area focused on market mechanisms.
Coalition formation.
Informally, a coalition is a
group of entities which have agreed to cooperate for
achieving a common goal. A virtual organization in-
volves entities that require a communication infrastruc-
ture and dedicated software to support their activities.
Coalition formation is a widely used method for increas-
ing the efficiency of resource utilization and for providing
convenient means to access these resources [24]. In re-
cent years, the emergence of large-scale electronic mar-
kets, grid and cloud computing, sensor networks, and
robotics have amplified the interest in coalition forma-
tion and virtual organizations [16, 17, 31]. For example,
2
self-organization of sensor networks through bottom-up
coalition formation is discussed in [20, 32].
Different aspects of resource management in compu-
tational grids including load balancing, job-allocation,
and scheduling, as well as revenue sharing when agents
form coalitions or virtual organizations are discussed in
[8, 14, 15, 27, 33, 35]. Grid resource allocation is mod-
eled as cooperative games [15] or non-cooperative games
[27]. Resource co-allocation is presented in [35].
There is little surprise that the interest in coalition
formation migrated in recent years from computational
grids to cloud resource management. The vast majority
of on-going research in this area is focused on game-
theoretic aspects of coalition formation for cloud federa-
tions. A cloud federation is of a set of CSPs collaborating
to provide services to a cloud user community.
A stochastic linear programing game model for coali-
tion formation is presented in [25]; the authors analyze
the stability of the coalition formation among cloud ser-
vice providers and show that resource and revenue shar-
ing are deeply intertwined. An optimal VM provisioning
algorithm ensuring profit maximization for CSPs is in-
troduced in [9].
A cloud federation formation described as a hedonic
game and focused on the stability and the fairness of the
game is discussed in [23]. The profit maximization for
each federation is formulated as an integer programming
problem (IP) and the game is augmented with a pref-
erence relation over the set of federations. The paper
assumes that the Virtual Machines (VMs) contributed
by each CSP to a federation are characterized by sev-
eral attributes, a ∈ A including the number of cores,
the amount of memory and of secondary storage. The
IP problem for CSP Ci in federation F is formulated as
j=1 ni,j(pj − ci,j) subject to the set of con-
j ni,j ≤ Ai, ∀a ∈ A and (cid:80)Ci∈F ni,j = rj
with: ni,j - the number of VMs of type j; pj - the price
for a VM running an instance of type j; ci,j - the cost of
an instance of type j provided by Ci; qa
j - the quantity
of resource of type a in a VM of type j; Ai - the total
amount of resource of type a offered by Ci; and rj - the
number of V M s of type j requested. The paper adopts
a payoff division based on the Banzhaf value [23].
max(cid:80)Ci∈F(cid:80)n
ditions (cid:80)n
j=1 qa
A combinatorial coalition formation problem is de-
scribed in [17]. The paper assumes that a seller has a
price schedule for each item. The larger the quantity re-
quested, the lower is the price a buyer has to pay for each
item; thus, buyers can take advantage of price discounts
by forming coalitions. A similar assumption is adopted
by the authors of [16] who investigate systems where
the negotiations among deliberate agents are not feasi-
ble due to the scale of the system. The paper proposes a
macroscopic model and derives a set of differential equa-
tions describing the evolution in time of coalitions with a
different number of participants. The results show that
even a low rate of leaving away participants allows a
coalition to achieve a steady state.
3
An algorithm to find optimal coalition structures in
cooperative games by searching through a lattice like
the one in Figure 2, was introduced by [30]. A more
refined algorithm is described in [28]; in this algorithm
the coalition structures are grouped according to the so-
called configurations reflecting the size of the coalitions.
Auctions. Auctions are a widely used mechanism for
resource allocation [11, 12]. Among the numerous appli-
cations of auctions are: the auctioning of airport take-
off and landing slots, spectrum licensing by the Federal
Communication Commission (FCC), and industrial pro-
curement. An online auction mechanism for resource
allocation in computer clouds is presented in [35].
A combinatorial auction is one where a buyer requires
simultaneous access to a package of goods. An auction
allows the seller to obtain the maximum feasible profit
for the auctioned goods; it is organized by an auctioneer
for every request of a consumer. A proxy is an inter-
mediary who collects individual bids from the buyers
participating at an auction, computes the total cost of
the package from the bids, and communicates this price
to the auctioneer. A vast literature including [1, 2, 3, 34]
covers multiple aspects of combinatorial auctions includ-
ing bidding incentives, stability, equilibrium, algorithm
testing, and algorithm optimality.
Package bidding assumes that a seller offers N differ-
ent types of items. A buyer bids for packages of items.
A package is a vector of integers Z = {z1, z2, . . . , zN}
which indicates the quantity of each item in the package;
the price of items is given by M = {m1, m2, . . . , mN}.
Package bidding can be traced back to generalized
Vickerey auctions based on the Vickerey-Clarke-Groves
mechanisms [11, 12]. In Vickerey auctions a bidder re-
ports its entire demand schedule. The auctioneer then
selects the allocation which maximizes the total value
of the package and requires a bidder to pay the lowest
bid it would have made to win its portion of the final
allocation, considering all other bids.
In an ascending package auction (APA) there are K
participants identified by an index, k = 0 is the seller
and k = 1, 2, . . . ,K are the buyers [1]. Each buyer has
a valuation vector vi = (νi(z), z ∈ [0,M]); νk(z) repre-
sents the value of package z to the bidder k. Some of
the rules for this type of auction are: all bids are firm,
a bid cannot be reduced or withdrawn; the auctioneer
identifies after each round the set of the bids that max-
imize the total price, the so-called provisional winning
bids. The auction ends when a new round fails to elicit
new bids; then the provisional winning bids become the
winers of the auction.
In an ascending package auction a bidder can be de-
terred from bidding for the package she really desires by
the threat that competitors could drive prices up; this
would threaten the equilibrium. This problem does not
exist in ascending proxy auctions when each bidder in-
structs a proxy agent to bid on her behalf [1]. The proxy
accepts as input the bidder's valuation profile and bids
following a "sincere strategy." Nash equilibrium can be
reached when the bid increments are negligibly small [1].
In a clock auction the auctioneer announces prices and
the bidders indicate the quantities they wish to buy at
the current price. When the demand for an item in-
creases, so does its price until the there is no excess de-
mand. On the other hand, when the offering exceeds the
demand, the price decreases [1]. In a clock auction the
bidding agents see only aggregate information, the price
at a given time, and this eliminates collusive strategies
and interactions among bidding agents. The auction is
monotonic, the amounts auctioned decrease continually
and this guarantees that the auction eventually termi-
nates. When the price of a package can be computed as
the sum of products of prices and quantities it is said
that auction benefits from linear pricing.
The clock-proxy-auction is a hybrid auction based on
an iterative process with two phases [3]. A clock phase
is followed by a proxy round. During the proxy round
the bidders report the values they have submitted to
the proxy which in turn submit bids for the package
to the auctioneer. A bidder has a single opportunity
to report the quantity and the price to the proxy, bid
withdrawals are not allowed, and the bids are mutually
exclusive. The auctioneer then selects the winning bids
that maximize the seller's profit.
The contribution of this paper. The reservation
system we propose has two stages; coalitions of servers
are formed periodically during the first and in the sec-
ond the coalitions participate in combinatorial auctions
organized in each allocation slot. To our knowledge this
is the first attempt to address cloud self-organization
and resource management based on coalition forma-
tion and combinatorial auctions when individual servers
learn from past behavior, see Figure 1.
Figure 1: A protocol with two stages; feedback about
past values of individual coalitions is used to determine
the value of individual coalition structures as shown in
Section 4.
We discuss coalition formation for a realistic model of
the cloud infrastructure, hierarchical organization, while
most of the research reported in the literature is focused
on coalition formation for cloud federations. The coali-
tion formation problem has different formulations and
different constraints in the two cases.
At this time individual CSPs believe that they have a
competitive advantage due to the unique value of their
services and are not motivated to disclose relevant infor-
4
mation about the inner working of their systems as we
have re-discovered when investigating the energy con-
sumption of AWS instances [26]. Thus, the practical
realization of cloud federations seems a rather remote
possibility [21].
A rare glimpse at the architecture of a cloud is pro-
vided in [5] and we are taking advantage of it to base
our research on a realistic model of the cloud infrastruc-
ture. We investigate coalition formation subject to the
physical constraints of the hierarchical cloud organiza-
tion model. As more diverse applications, including Big
Data applications, are likely to use computer clouds, the
demand for computing resources allocated to a single ap-
plication will increase and could be considerably larger
than any server can provide; only coalitions of servers
will be capable to offer such resources. It is critical for
the members of a coalition to communicate effectively;
this requires coalition member to be in close proximity
of each other in a system consisting of a hierarchy of net-
works with different bandwidth and latency. This adds
additional constrains to the coalition formation protocol.
To respond to the needs of increasingly more com-
plex applications consisting of multiple phases and re-
quiring workflow management, CSPs are already offer-
ing workflow management services such as SWS (Simple
Workflow Management) and EBS (Elastic Bean Stock)
at AWS. Different phases of an application may require
coalitions of servers with different types of resources and
this is the reason why we decided to investigate combina-
torial auctions where packages of items are auctioned.
3 System Model
System architecture. We assume a hierarchical orga-
nization of the cloud infrastructure similar to the one
described in [5]. A data center consists of multiple
warehouse-scale computers (WSCs), each WSC has mul-
tiple cells, each cell has multiple racks and each rack
houses multiple servers. A WSC connects 50, 000 to
100, 000 servers and uses a hierarchy of networks. The
servers are housed in racks; typically, the 48 servers in a
rack are connected by a 48 port Gigabit Ethernet switch.
The switch has two to eight up-links which go to higher
level switches in the network hierarchy [5]. The band-
width to communicate outside the rack is much smaller
than the one within the rack; this has important impli-
cations for resource management policies and becomes
increasingly difficult to address in systems with a large
number of servers.
Model assumptions. For simplicity we assume that
the racks are homogeneous, they have identical proces-
sors with the same number of cores and an identical con-
figuration of GPUs, FPGAs, workflow engines, or other
hardware, the same amount of main storage, cache, and
secondary storage. We also assume that all servers in
a rack are identically configured and support the same
type of services. The same service may be offered by
Coalition formationCombinatorial auctionsRequests for serviceCoalition values from previous auctionsSuccessful bids multiple racks; for example, multiple racks could offer
configurations with GPUs.
The system we envision supports a reservation system
and spot resource allocation. The reservation system has
two stages: (A) coalition formation, and (B) combina-
torial auctions. The spot allocation is done through a
bidding process for each type of service. The time is
quantified, reservations are made as a result of auctions
carried out at the beginning of each allocation slot of
duration τ ; for example, a allocation slot could be one
hour.
Coalition formation. The rational for coalition for-
mation is that applications may need resources beyond
those provided by an individual server. For example, a
Map-Reduce application may require a set of 20 servers
during the Map phase to process a data set of several
PB (Petabytes).
If the algorithm requires the servers
to communicate during this phase then the application
should start at the same time on all servers and run at
the same pace, a condition known as co-scheduling. Co-
scheduling is only feasible if the set of 20 servers form
a coalition dedicated to the application; moreover, the
hardware configuration of the coalition members should
be optimal for the algorithms used by the application,
e.g., have attached GPUs.
Combinatorial auctions. Combinatorial auctions
allow cloud users making the reservations to acquire
packages consisting of coalitions of servers with different
types and amounts of resources. Combinatorial auctions
are necessary because different phases of an application
may require systems with different configurations or sys-
tems supporting different functions. In our previous ex-
ample the Reduce phase of the Map-Reduce application
may require several servers with a very large amount of
secondary storage.
4 Coalition Formation
First, we discuss the formulation of the coalition for-
mation problem as a cooperative game. Then we intro-
duce the algorithms for determining the optimal coali-
tion structure and for coalition formation in the context
of our model.
Coalition formation as a cooperative game. The
coalition formation is modeled as a cooperative game
where the goal of all agents is to maximize the reward
due to the entire set of agents. We consider a set of N
servers {s1, s2, . . . , sN}, located in the same rack.
A coalition Ci is a non-empty subset of N . A coalition
structure is set of m coalitions S = {C1, C2, . . . , Cm}
satisfying the following conditions
Ci = N and i (cid:54)= j ⇒ Ci
(cid:92) Cj = ∅.
m(cid:91)
i=1
(1)
5
Figure 2 shows a lattice representation of the coali-
tion structures for a set of four servers s1, s2, s3 and s4.
This lattice has four levels, L1, L2, L3 and L4 contain-
ing the coalition structures with 1, 2, 3 and 4 coalitions,
respectively. In general, the level k of a lattice contains
all coalition structures with k coalitions; the number of
of coalitions structures at level k for a population of N
agents is given by the Sterling Number of Second Kind:
k(cid:88)
i=0
(cid:18)k
(cid:19)
i
S(N, k) =
1
k!
(−1)i
(k − i)N .
(2)
In the case illustrated in Figure 2 N = 4 and the number
of coalition structures at levels L1 − L4 are 1, 7, 6, 1,
respectively.1 The total number of coalition structures
with N agents is called the Bell number
N(cid:88)
k=0
N(cid:88)
k(cid:88)
1
k!
k=0
i=0
(cid:18)k
(cid:19)
i
B(N ) =
S(N, k) =
(−1)i
(k − i)N .
(3)
The number of coalitions structures increases exponen-
tially with the number of agents. For example, for N =
40, a typical number of servers in a rack, the logarithm
of the number of coalition structures is close to 1035 and
S(40, 14) = 3.5859872255621803491428554E + 34. The
logarithm of number of coalitions is close to E + 10.
Searching for the optimal coalition structure C is com-
putationally challenging due to the size of the search
space. The first step for determining the optimal coali-
tion structure is to assign a value v reflecting the utility
of each coalition. The second step is the actual coalition
formation.
Rack-level coalition formation. Recall from Sec-
tion 3 that in our model a rack is homogeneous, all
servers have an identical configuration. This realistic as-
sumption simplifies considerably the complexity of the
search for an optimal coalition structure as the servers
are indistinguishable from one another.
The second important observation is that we have
a system with two stages and feedback, see Figure 1.
In the second stage the coalitions created during the
first stage are included in successfully auctioned pack-
ages thus, we can determine precisely the value of all
coalitions structures. The third important observation
is that only available servers, servers with no commit-
ments for the current slot, can participate to coalition
formations and then to the auction organized in that
slot; call Na ≤ N the number of available servers.
An elected rack leader collects information about all
successful coalitions - coalitions that have been included
in packages auctioned successfully during a window of w
successive past allocation slots. The current rack-leader
records an entry for the corresponding partial coalition
1For N = 5 and N = 6 the Stirling Numbers of the Second
Kind are respectively 1, 15, 25, 10, 1 and 1, 31, 90, 65, 15, 1.
Figure 2: A lattice with four levels L1, L2, L3 and L4 shows the coalition structures for a set of 4 servers, s1, s2, s3
and s4. The number of coalitions in a coalition structure at level Lk is equal to k.
structure (PCS) including nk - the coalition size, mk -
the multiplicity of occurrence, the value ¯vk calculated as
the average price over all auctions when a PCS including
a coalition of size nk was part of a package successfully
auctioned during the past w allocation slots.
Call L the PCL-list. For a window of size w the list
L is the list of all triplets Lk = [nk, mk, ¯vk] ordered first
by 1 ≤ nk ≤ Na then by mk. The list includes only
entries Lk with ¯vk > 0. Given Na a coalition structure
(CS) Sk among the entries Lk1,Lk2, ... . . . ,Lkn is feasi-
j nk × mk = Na. Then the value of the coalition
j ¯vj. Note that we force the for-
mation of coalitions involving all available servers. An
example of a PCS list L follows
ble if(cid:80)
structure Sk is vk =(cid:80)
----------------------------------------------
a
b
\* 4 PCS of 1-server {s}
\ *15 PCS of 1-server {s}
[1,4,35]
[1,15,682]
.........
[2,3,78]
\* 3 PCS of 2-servers {s,s}
........
[3,2,502]
[3,4,812]
\* 2 PCS of 3-servers {s,s,s}
\* 4 PCS of 3-servers {s,s,s}
.........
[16,1,751] \* 1 PCS of 16-servers {s,...s}
[16,2,740] \* 2 PCS of 16-servers {s,...s}
c
d
e
f
g
.........
-----------------------------------------------
In this example some of the feasible coalitions struc-
tures when Na = 16 are: Sg with vg = 751; Sa,b with
va,b = 35 + 682 = 712; Sa,e with va,e = 35 + 812 = 837;
Sa,c,d with va,c,d = 35 + 78 + 502 = 615, and so on. Note
that the value of a coalition reflects also the length of
time the coalition was active in response to successful
auction. We see that a PCS of 15 coalitions of 1 server
6
have been active for larger number of slots than a PCS
of 4 coalitions of 1 server. The value attributed to a
coalition of k servers is a distributed equally among the
servers; the value of a package of several coalitions auc-
tioned successfully is divided among the coalitions based
on the resource supplied by each one of them.
Coalition formation. The protocol for coalition for-
mation proceeds as follows:
1. Server si sends to the current rack leader:
i , β1
i ], [ν2
i , β2
i , βN
i ], . . . ....[νN
(a) A vector ([ν1
i ]) with
i , 1 ≤ k ≤ N the total value due to the par-
νk
ticipation of si in successful coalitions, of k
servers and βk
i a bit vector with w components
with βk,j
i = 1 if si was included in a successful
coalition of k servers in slot j of window w.
(b) Availability, ai = 1 if available, 0 otherwise.
2. After receiving the information from all servers the
current rack leader:
(a) Determines Na =(cid:80)N
(b) Computes mk =(cid:80)Na
(c) Computes ¯vk =(cid:80) νk
i=1
i
(cid:80)w
i=1 ai.
j=1 βk,j
i
, 1 ≤ k ≤ N .
(d) Computes the optimal coalition structure.
(e) Assigns a server to coalition of size k a based
(f) Chooses the best performer as the next coali-
tion leader. The best performer is the one with
on the values νk
i .
the largest value(cid:80)
j νj
i .
Finding the optimal CS requires at most L operations
with L the size of the PCL-list. The system starts with a
predetermined coalition structure and coalition values.
{s1}{s2}{s3}{s4}{s1}{s2}{s3,s4}{s1}{s3}{s2,s4}{s1}{s4}{s2,s3}{s2}{s3}{s1,s4}{s2}{s4}{s1,s3}{s3}{s4}{s1,s2}{s1}{s2,s3,s4}{s2}{s1,s3,s4}{s3}{s1,s2,s4}{s4}{s1,s2,s3}{s1,s4}{s2,s3}{s1,s2}{s3,s4}{s1,s3}{s2,s4}{s1,s2,s3,s4}L4L3L2L1 5 A Reservation System Based on a
Combinatorial Auction Protocol
The protocol introduced in this section targets primar-
ily the IaaS cloud delivery model represented by Amazon
Web Services (AWS). Reservation systems are currently
used by CSPs. For example, AWS supports reserva-
tions as well as spot allocation and offers a limited num-
ber of instance families, including M3 (general purpose),
C3 (compute optimized), R3 (memory optimized), I2
(storage optimized), G2 (GPU) and so on. An in-
stance is a package of system resources; for example,
the c3.8xlarge instance provides 32 vCPU, 60 GiB of
memory, and 2× 320 GB of SSD storage. The resources
auctioned are supplied by coalitions of servers in differ-
ent racks and the cloud users request packages of re-
sources.
The combinatorial auction protocol is inspired by the
clock-proxy auction [3]. The clock-proxy auction has a
clock phase, where the price discovery takes place, and a
proxy phase, when bids for packages are entertained. In
the original clock-proxy auction there is one seller and
multiple buyers who bid for packages of goods.
For example, the airways spectrum in the US is auc-
tioned by the FCC and communication companies bid
for licenses. A package consist of multiple licenses; the
quantities in these auctions are the bandwidth allocated
times the population covered by the license.
Individ-
ual bidders choose to bid for packages during the proxy
phase and pay the prices they committed to during the
clock phase.
Our protocol supports auctioning service packages; a
packages consist of combinations of services in one or
more time slots. The items sold are services advertised
by coalitions of autonomous servers and the bidders are
the cloud users. Each service is characterized by
1. A type describing the resources offered and the con-
ditions for service,
2. The time slots when the service is available.
Protocol specification. The terms used to describe
the protocol are discussed next. An allocation slot (AS)
is a period of fixed duration, e.g., one hour, that can be
auctioned. An auction, At, is organized at time t if there
are pending reservation requests which require immedi-
ate attention. Figure 3 shows two consecutive auctions
at times t and s; during the first slot of auction At new
reservation requests are received and the allocation slot
1 for As.
2 is not fully covered; this slot becomes ASs
ASt
A service A is described by a relatively small number
of attributes, {a1, a2, . . . ,}. Each attribute ai can take a
number of distinct values, vi = {vi,1, vi,2, . . .}. The first
attribute is the coalition size or equivalently the number
of vCPS provided; other attributes could be the type of
service and server architecture with two values "32-bit"
and "64-bit;" another attribute could be "organization"
with values "vN" (von Neumann), "DF" (data-flow), or
"vN-GPU" (vN with graphics co-processor).
Call S t the set of services the clients want to reserve
during auction At
S t = {St
1, St
2, . . . , St
νt} with St
A reservation bundle, αt
requested by client i in slot j of auction At
i = [sId, (aj, vj,k)]
(4)
i,j ⊂ S t, is the set of services
i,j = {(St
αt
i,j,1, rt
i,j,1), (St
i,j,2, rt
i,j,2), . . .}
(5)
with rt
i,j,l a measure of the quantity; for example, if the
attribute is "service intensity" the quantity is the num-
ber of vCPUs.
k,j ⊂ S t, is the set of services
An advertised bundle, βt
advertised by coalition k in slot j of auction At
k,j = {(St
βt
k,j,1, qt
k,j,1, pt
k,1), (St
k,j,2, qt
k,j,2, pk,2) . . .}
(6)
with qt
k,j,l a measure of the quantity of service l and pk,l
the price per ECU of service St
l determined by coalition
k. A package, P t
is a set of reservations for services
requested by client i for slots j1, j2, . . . during auction
At.
i
P t
i = {αt
i,j1, αt
i,j2, . . .}
(7)
The clock phase. Figure 4 illustrates the basic idea
of a clock phase: the auctioneer announces prices and
the bidders indicate the quantities they wish to buy at
the current price. When the demand for an item in-
creases, so does its price until there is no excess demand;
on the other hand, when the offering exceeds the de-
mand, the price decreases.
During the clock phase of auction At the price discov-
ery is done for each time slot and for each type of service;
a clock runs for each one of the κt slots and for each one
of the νt services. Next we describe the clock phase for
service St
l in slot j. Assume that there are n coalitions
C = {C1, C2, . . . , Cn} offering the service and m requests
for reservations from clients D = {D1, D2, . . . , Dm}.
A clock auction starts at clock time t = 0 and at price
per unit of service for Sl
(8)
Call C0 the available capacity at this price and D0 the
demand for service St
p0
l = minCk
l in slot j
{pk,l}
n(cid:88)
k=1
C0 =
l offered at price p0
k,j,l and D0 =
qt
m(cid:88)
rt
i,j,l.
i=1
(9)
If C0 < D0 the clock c advances and the next price per
unit of service is
p1
l = p0
l + I
(10)
7
Figure 3: Auctions At and As conducted at times t and s, respectively. τ t
slots, ASt
0 are the start of the first allocation
1 of the two auctions. The number of slots auctioned in each case are κt and κs, respectively.
1 and ASs
0 and τ s
with I the price increment decided at the beginning of
auction. There is an ample discussion in the literature
regarding the size of the price increment; if too small,
the duration of the clock phase increases, if too large, it
introduces incentives for gaming [3].
The process is repeated at the next clock value start-
ing with the new price. The clock phase for service St
l
and slot j terminates when there is no more demand.
Figure 4: The clock phase for service St
l and slot j. The
starting price is p0
l given by Equation 8. The clock ad-
vances and the price increases from pc to pc + I when
the available capacity at that price given by Equation 9
is exhausted; the demand is given by Equation 9.
The proxy phase. In a traditional clock-proxy auc-
tion the bidders do not bid directly, they report to a
proxy the price and the quantity of each item in the
8
package they desire. The proxy then bids in an ascend-
ing package auction.
In our application, the proxy phase of the auction con-
sists of multiple rounds. The auction favors bids for long
runs of consecutive slots when the service is provided by
the same coalition. This strategy is designed to exploit
temporal and spatial locality.
The auction starts with the longest runs and the low-
est price per slot and proceeds with increasingly shorter
runs and diminished incentives. Once a run of consec-
utive slots is the subject of a provisional winning bid,
all shorter runs of slots for that particular service are
removed from the coalition offerings.
During the first round only the longest run of consec-
utive slots for each one of the services offered by the par-
ticipating coalitions is auctioned and only bidders that
have committed to any of the slots of the run are allowed
to bid. The price per slot for the entire run is the lowest
price for any slot of the run the bidder has committed to
during the clock phase of the auction. If there are multi-
ple bids for service St
l the provisional winner is the one
providing the largest revenue for the coalition offering
the service.
If κt
l is the longest run of consecutive slots for service
St
l auctioned in the first round then, in the second round,
l − 1 slots is auctioned. The price for
a shorter run of κt
the entire run equals the second lowest price for any slot
of the run the bidder has committed to during the clock
phase of the auction times the number of the time slots
in the run.
The length of the consecutive slot runs auctioned de-
creases and the incentives diminish after each round.
The preliminary rounds end with the auction of a sin-
gle slot for each service. At the end of the preliminary
round each bidder is required to offer the price for the
slot committed to during the clock phase. Figure 5 de-
picts a plausible snapshot at the end of the preliminary
rounds of the proxy phase when four services S1,S2,S3
and S4, are offered and shows the provisional winners
for service S4.
timetimetsASt1AtAsτ t0τ s0ASt2AStκtASt3ASs1ASs2ASs3AStκsAuctioneerC1C2Cnqt2,j,lpc+ ID1D2Dmqt1,j,lqtn,jlrt1,j,lrt2,j,lrtm,j,lrti,j,lΣ Σ qtk,j,l>c Figure 5: A snapshot at the end of the preliminary rounds of the proxy phase when there are four services offered
and the auction covers 18 allocation slots. Dotted lines represent the quantity of service with provisional winners.
Only the provisional winers for S4 are shown, the clients labeled as C9, C13, C6, C1, C11, C7 and C9.
During the final round the bidders reveal the packages
they want to reserve; these packages include only the
provisional winners from the preliminary slots. Once
all provisional winning bids for services in a reservation
request are known, the auctioneer chooses the package
that best matches the consumer's needs and, at the same
time maximizes the profit for the cloud service provider.
The coalition for a reservation request consists of the
set of coalitions that provide the services in the winning
package.
In this auction all bids are firm, they cannot be with-
drawn. The auction is monotonic, the length of runs of
consecutive slots auctioned decreases continually; this
guarantees that the auction eventually terminates. Lin-
ear pricing guarantees that the price of any package can
be computed with ease.
The effectiveness of the protocol is captured by
several metrics including:
-1. The customer satisfaction index - percentage of reser-
vation requests fully or partially satisfied in each alloca-
tion slot given the total number of requests.
-2. The service mismatch index - percentage of services
requested but not offered in each allocation slot given
the total number of services in that slot.
-3. The service success index - percentage of services
used in each allocation slot given all services offered in
that slot.
-4. The capacity allocation index - percentage of the
capacity offered but not auctioned in each allocation slot
given the capacity offered in that slot.
-5. The overbidding factor - percentage of slots with a
provisional winner that have not been included in any
package given all slots offered at the beginning of the
auction.
-6. The temporal fragmentation index - percentage of
services successfully auctioned in non-consecutive slots
given all services successfully auctioned.
-7. The additional profit index - percentage of additional
profit of coalitions involved in the auction (the differ-
ence of the actual price obtained at the auction and the
price demanded by the coalition) relative to the price
demanded by the coalition.
Limitations and vulnerabilities. The protocol is
fairly complex and has at least one vulnerability. A bid-
der may be the provisional winner of services in slots
not included in its winning package; such services will
remain unassigned during the current auction. A solu-
tion is to penalizeexcess bidding activity and charge the
bidder a percentage of the costs for these services. An-
other alternative is to include, in a reservation request,
a set of "substitute services" for a service Si. Then,
during the last round of the proxy phase, the auctioneer
could try to match services having provisional winners
with unsatisfied requests for services.
The capacity offered, but not auctioned in each slot
is available for spot allocation thus, it has the potential
to be used, rather then being wasted. The capacity of
a coalition left uncommitted at the end of the auction
At for ASt
1, the first slot of the auction, is then available
for spot allocation at a price equal to pk,l, while the free
capacity in slots starting with ASt
2 can be offered at
the next auction if this auction takes place before the
beginning of the slot. This capacity is measured by the
spot allocation opportunity index.
9
AllocationslotsServiceS1ServiceS2ServiceS3ServiceS41234567891113141215161718AuctiontimeC9C13C6C1C11C7C9 6 Protocol Analysis and Evaluation
We report on the results of our simulation experiments
to gain some insight into the proxy phase of the clock-
proxy auction of the P C2P protocol. The system we
wish to evaluate requires the description of the environ-
ment in which the auction takes place, the reservation
requests, and the services offered:
1. The environment elements: n - the number of coali-
tions offering services in this round; m - the number
of clients; and κ - the number of slots auctioned.
2. The package j requested by client i: αn
i - the num-
ber of services in the package; the slots desired by
the service Sk, ordered by the length of the run of
consecutive slots; rk,j - the intensity of service Sk in
slot j; pi,j - the price per unit of service for slot j if
client i was a provisional winner of that slot during
the clock phase.
3. The service Sk provided by coalition Ck includes: γk
- the largest run of consecutive slots for each offered
service Sk; the profile of the service Sk - the slots
offered ordered by the length of consecutive slots,
when it is available; qk,j - the quantity of service
Sk offered in slot j; and pk - the price per unit of
service offered by coalition Ck.
For simplicity, we assume that a coalition offers one
service only and the number of services is ν < n. We
also assume that all platforms have a maximum capac-
ity of 100 vCPUs and that qk,j, the quantity of service
Sk offered for auction, and rk,j, the quantity of Sk re-
quested in slot j are the same for all the slots of an
offered/requested run. The number of slots auctioned is
fixed, κ = 50.
The range and the distribution of parameters for the
protocol evaluation are chosen to represent typical cases.
The parameters of the simulation are random variables
with a uniform distribution:
-a. The number of coalitions and clients requesting reser-
vations, n and m, respectively; the interval is [200−250].
-b. The number of services offered and requested ν; the
interval is [10 − 20].
-c. The number of clients bidding for each service in a
given slot; the interval is [0 − 4].
-d. The capacity offered for auction for a service in a
given slot; the interval is [60 − 90] vCPUs.
-e. The services offered by a coalition; the interval is
[1 − ν].
-f. The number of consecutive slots a service is offered
in; the interval is [1 − κ].
-g. The number of services in the package requested by
a client; the interval is [1 − 3].
-f. The number of consecutive slots of the services in the
package requested by a client; the interval is [1 − κ].
We also randomly choose the slots when the client is
the provisional winner. The evaluation process consists
of the following steps:
A. Initialization.
B. Preliminary rounds. Carry out γ preliminary rounds
with γ = maxk γk.
• In the first preliminary round auction auction κ1
slots of service S1, κ2 slots of service S2, and so on.
• Identify the first slot of each run and the reservation
request that best matches the offer.
• Identify the provisional winners if such matches ex-
ist and remove the corresponding runs from the set
of available runs. A match exists if the run consists
of the same number of slots or is one slot longer than
requested and if the capacity offered is at least the
one required by the reservation request. For services
without a match, remove the last slot, add both the
shorter run and the last slot to the list of available
runs.
• Continue this process until only single slots are
available.
C. Final round. In this round we:
• Identify the packages for each client and if mul-
tiple packages exist determine the one which best
matches the request.
• Compute the cost for the winning package for each
client.
Figures 6(a)-(e) show several performance metrics in-
cluding the customer satisfaction index, the service mis-
match index, the auction success ratio, the spot oppor-
tunity index, the temporal fragmentation index, and the
capacity allocation index. The simulation covers 50 time
slots.
The 5% confidence intervals for the mean of all per-
formance metrics are computed for 25 batches each one
of 200 realization of each random variable. The simula-
tion times are 6.4 seconds for 2, 000 runs and 11.7 sec-
onds for 5, 000 runs. The confidence intervals are rather
tight; this indicates that the performance of the protocol
is relatively stable for the range of parameters explored
in this evaluation.
The auction success rate is high, typically above 80%.
The initial low auction success rate is an artifact of the
manner we conducted the simulation; we picked up ran-
domly the service start up time. The spot allocation
opportunity index is in turn correlated with the auction
success rate and shows that a significant fraction of the
capacity is available for spot allocation. This result is
correlated with the one in Figure 6(f) which shows that
10
(a)
(b)
(c)
(d)
(e)
(f)
Figure 6: Proxy phase of an auction with 50 time slots. Indices of: (a) Customer satisfaction; (b) Service mismatch;
(c) Auction success; (d) Spot allocation opportunity; (e) Temporal fragmentation; (f) Capacity allocation.
11
01020304050 0%10%20%30%40%50%60%70%80%→ TimeSatisfied vs total number of requests01020304050 0% 5%10%15%20%25%30%35%40%45%50%→ TimeServices not offered but requested01020304050 0% 10% 20% 30% 40% 50% 60% 70% 80% 90%100%→ TimeAuction Success Ratio01020304050 0%10%20%30%40%50%60%70%80%90%→ TimeSpot Allocation Opportunity Index010203040500100200300400500600700800900→ TimeTemporal Fragmentation Index01020304050 0%10%20%30%40%50%60%→ TimeCapacity allocated by the auction on average some 50% of the server capacity is not allo-
cated by the reservation system and so is available for
spot contention.
A reservation system covering 50% of the server ca-
pacity is probably the most significant result; it shows
that self-management based on auctions can drastically
improve server utilization. We live in a world of limited
resources and cloud over-provisioning is not sustainable
either economically or environmentally.
The service mismatch index is fairly high, typically
in the 50% range and it is above 60% in a few slots.
The customer satisfaction is correlated with the service
mismatch and typically is in the region of 50%.
In a
realistic scenario, when coalitions maintain statistics re-
garding the services offered and avoid offering services
unlikely to be demanded by the cloud users, the service
mismatch would not affect the performance of the algo-
rithm. Temporal fragmentation, though rather low, is
undesirable. The overbidding factor 64 ± 2.93% is an-
other indication that the protocol needs to be fine tuned.
Self-organization cannot occur instantaneously in an
adaptive system and this simple observation has impor-
tant consequences.
It is critical to give autonomous
cloud platforms, interconnected by a hierarchy of net-
works, the time to form coalitions in response to ser-
vices demanded. Thus, self-management requires an ef-
fective reservation system and our results indicate that
the reservation protocol is working well.
7 Conclusions and Future Work
Self-organization and self-management offer an appeal-
ing alternative to existing cloud resource management
policies; they have the potential to significantly al-
ter the cloud computing landscape. So far, pragmatic
means for the adoption of self-organization principles for
large-scale computing and communication systems have
eluded us. A main reasons for this state of affairs is that
self-management has to be coupled with some mecha-
nisms for cooperation; these mechanisms should allow
autonomous servers, to act in concert towards global sys-
tem goals. Cooperation means that individual systems
have to partially surrender their autonomy. Striking a
balance between autonomy and cooperation is a chal-
lenging task, it requires a fresh look at the mechanics of
self-organization and the practical means to achieve it.
Practical implementation of cloud self-organization is
challenging for several reasons including the absence of
a technically suitable definition of self-organization, a
definition that could hint to practical design principles
for self-organizing systems and quantitative evaluation
of the results. Computer clouds exhibit the essential as-
pects of complexity and it is inherently difficult to con-
trol complex systems.
We started our investigation with a realistic model
of the cloud infrastructure, the hierarchical organiza-
tion reported in [5] which seems inherently tied to hi-
erarchical control. First, we compared hierarchical con-
trol which based on monitoring with a market model
in which the servers of a WSC place bids for service
requests and found out that the latter is much more
effective than hierarchical control [22].
In the simple
market model the servers act individually, rather than
cooperating with each other, a fundamental aspect of
self-organization. But cooperation is clearly needed be-
cause individual servers may not be able to supply the
resources demanded by many data-intensive application.
Thus, we concluded that servers have to form coalitions
to offer larger pools of resources. At the same time,
it seemed obvious to us that complex applications with
multiple phases would require packages of resources of-
fered by different coalitions.
Algorithms for coalition formation based on combi-
natorial auctions are at the heart of the cloud ecosys-
tem we propose. The path we chose seems logical as
auctions have been successfully used for resource man-
agement in the past. Auctions do not require a model
of the system, while traditional resource management
strategies do. The auction-based protocol is scalable,
and the computations can be done efficiently, though
the computational algorithms involved are often fairly
complex.
The results reported in Section 6 indicate that the per-
formance of the protocol is relatively stable for the range
of parameters explored in our evaluation. The protocol
leads to a higher server utilization and it seems reason-
able to expect that a fine-tuned version of the protocol
could further improve this critical performance measure.
Two basic strategies for coalition formation are possi-
ble: (1) the one described in this paper, when coalitions
are formed using information from previous combinato-
rial auctions, before knowing what the actual user de-
mands are; (2) the more natural one, when coalitions are
formed in response to user demands and then combinato-
rial auctions are organized. The former strategy, though
simple and elegant, has two obvious drawbacks: (a) it
needs the past history thus, starting from some arbitrary
initial state may not work very well; (b) there may be
cases when coalitions formed based on past history are
not useful to any package and, at the same time, one or
more packages could benefit from the available resources
of the same type, but in coalitions of different sizes. The
latter strategy, creating coalitions in response to known
user requests is more complex. It requires a protocol to
inform supply agents of the atrributes of desirable coali-
tions, a protocol for individual agents to express their
willingness to join a particular coalition, and, finally, a
coalition formation algorithm. Once the coalitions are
formed the bidding for service packages can take place.
These two strategies can be combined, the latter used
during the initial stages, when the historic information is
either missing or incomplete, and the former used when
history data is available. The second drawback of the
strategy described in this paper can be attenuated by
allowing a second round of coalitions formation. In this
12
second round the unsuccessful coalitions are disassem-
bled and new coalitions matching the needs of unsatis-
fied packages are formed. For example, if two coalitions
of size θ1 and θ2 were unsuccessful and a package P re-
quireing a coalition of size θ3 ≤ θ1 + θ2 was unsatisfied,
then a second round will guarantee that P is satisfied
and θ3 servers will be reserved, instead of being left idele
or offered to spot allocations.
Our future work will address this approach, as well
as, other problems revealed by this investigation, e.g.,
the effects of overbidding. Overbidding is the process
allowing a client to become a provisional winner of one
or more service slots and then, in the final round failing
to acquire some of them. This situation is critical for the
first slot of an auction as the next auctions could find
clients for these slots. A more difficult problem is the
temporal fragmentation which does not seem to have an
obvious solution.
References
[1] L. Ausubel and P. R. Milgrom. "Ascending auc-
tions with package bidding." Frontiers of Theoret-
ical Economics, 1(1):1 -- 42, 2002.
[2] L. M. Ausubel and P. Cramton. "Auctioning many
divisible goods." Journal European Economic As-
soc., 2(2-3):480-493, 2004.
[3] L. Ausubel, P. Cramton, and P. Milgrom. "The
clock-proxy auction: a practical combinatorial auc-
tion design." Chapter 5, in Combinatorial Auctions,
P. Cramton, Y. Shoham, and R. Steinberg, Eds.
MIT Press, 2006.
[4] L. A. Barroso and U. Hozle.
"The case for
energy-proportional computing." IEEE Computer,
40(12):33 -- 37, 2007.
[5] L. A. Barossso, J. Clidaras, and U.Hozle. The Dat-
acenter as a Computer; an Introduction to the De-
sign of Warehouse-Scale Machines. (Second Edi-
tion). Morgan & Claypool, 2013.
[6] M. Blackburn and A. Hawkins. "Unused server
www.thegreengrid.org/
survey results analysis."
media/WhitePapers/Unused %20Server%20Study
WP 101910 v1. ashx?lang=en (Accessed on De-
cember 6, 2013).
[7] D. Bruneo. "A stochastic model to investigate data
center performance and QoS in IAAS cloud com-
puting systems." IEEE Trans. on Parallel and Dis-
tributed Systems, 25(3):560 -- 569, 2014.
[9] S. Chaisiri, B. Lee, and D. Niyato. "Optimization
of resource provisioning cost in cloud computing."
IEEE Trans. on Services Computing, 5(2):164 -- 177,
2012.
[10] V. Chang, G. Wills, and D. De Roure. "A review
of cloud business models and sustainability." Proc.
IEEE 3rd Int. Conf. on Cloud Computing, pp. 43 --
50, 2010.
[11] E. H. Clarke. "Multipart Pricing of Public Goods."
Public Choice, IX:13 -- 33, 1971.
[12] T. Groves. "Incentives in teams." Econometrica,
41:617 -- 631, 1973.
[13] J. O. Gutierrez-Garcia and K.- M. Sim.
"Self-
organizing agents for service composition in cloud
computing." Proc IEEE 2nd Int. Conf. on Cloud
Computing Technology and Science, pp. 59 -- 66,
2010.
[14] L. He and T. R. Ioerger. "Forming resource-sharing
coalitions: a distributed resource allocation mech-
anism for self-interested agents in computational
grids." Proc. ACM Symp. on Applied Computing,
pp. 84 -- 91, 2005.
[15] S. U. Khan and I. Ahmad. "A cooperative game
theoretical technique for joint optimization of en-
ergy consumption and response time in computa-
tional grids." IEEE Trans. on Parallel and Dis-
tributed Systems, 20(3): 346-360, 2009.
[16] K. Lerman and O. Shehory. "Coalition formation
for large-scale electronic markets." Proc. ICMAS
2000 - 4th Int. Conf on Multiagent Systems, pp.
167 -- 174, 2000.
[17] C. Li and K. Sycara. "Algorithm for combinatorial
coalition formation and payoff division in an elec-
tronic marketplace." Proc. AAMAS02 - First Joint
Int. Conf. on Autonomous Agents and Multiagent
Systems, pp. 120 -- 127, 2002.
[18] H. Li, C. Wu, Z. Li, and F. Lau. "Profit-maximizing
virtual machine trading in a federation of selfish
clouds." Proc. of the IEEE INFOCOM, pp. 25 -- 29,
2013.
[19] H C. Lim, S. Babu, J. S. Chase, and S. S. Parekh.
"Automated control in cloud computing: challenges
and opportunities." Proc. First Workshop on Auto-
mated Control for Datacenters and Clouds,, ACM
Press, pp. 13 -- 18, 2009.
[8] T. E. Carroll and D. Grosu. "Formation of virtual
organizations in grids: a game-theoretic approach."
Concurrency and Computation: Practice and Expe-
rience, 22(14):1972 -- 1989, 2010.
[20] D. C. Marinescu, C. Yu, and G. M. Marinescu.
"Scale-free, self-organizing very large sensor net-
works." Journal of Parallel and Distributed Com-
puting (JPDC), 50(5):612 -- 622, 2010.
13
[21] D. C. Marinescu. Cloud Computing; Theory and
Practice. Morgan Kaufmann, a division of Elsevier,
Amsterdam, New York, 2013.
[22] D. C. Marinescu, A. Paya, J. P. Morrison, and
P. Healy.
"Distributed hierarchical control ver-
sus an economic model for cloud resource man-
agement." http://arxiv.org/pdf/1503.01061v1.pdf,
March 2015.
[23] L.Mashayekhy, M.M. Nejad, and D. Grosu. "Cloud
federations in the sky: formation game and mech-
anisms." IEEE Trans. on Cloud Computing, 2015
(to appear).
[24] I. Muller, R. Kowalczyk, and P. Braun. "Towards
agent-based coalition formation for service compo-
sition." Proc. IEEE/WIC/ACM Int. Conf. on In-
telligent Agent Technology, pp. 73-80, 2006.
[25] D.Niyato, A.Vasilakos, and Z.Kun.
"Resource
and revenue sharing with coalition formation of
cloud providers: Game theoretic approach." Proc.
IEEE/ACM Intl. Symp. on Cluster, Cloud and Grid
Comp., pp. 215 -- 224, 2011.
[26] A. Paya and D. C. Marinescu.
"Energy-aware
load balancing and application scaling for the cloud
ecosystem."
IEEE Trans. on Cloud Computing.
doi:10.1109/TCC.2015.2396059.
[27] S. Penmatsa and A. T. Chronopoulos.
"Price-
based user-optimal job allocation scheme for grid
systems." Proc of Parallel and Distributed Process-
ing Symposium, pp. 8-16, April 2006.
[28] T. Rahwan, S. D. Ramchurn, N. R. Jennings, and
A. Giovannucci. "An anytime algorithm for optimal
coalition structure generation." Journal of Artifi-
cial Intelligence Research, 34:521 -- 567, 2009.
[29] N. Samaan. "A novel economic sharing model in a
federation of selfish cloud providers." IEEE Trans.
on Parallel and Distributed Systems, 25(1):12 -- 21,
2014.
[30] T. W. Sandholm, K. S. Larson, M. Andersson, O.
Shehory, and F. Tohm. "Coalition structure gener-
ation with worst case guarantees." Artificial Intel-
ligence, 111(1-2):209 -- 238, 1999.
[31] S. Sen and P. S. Dutta. "Searching for optimal
coalition structures." Proc. ICMAS 2000 - 4th Int.
Conf on Multiagent Systems, pp. 287 -- 295, 2000.
[32] M. Sims, C. V. Goldman, and V. Lesser. "Self-
organization through bottom-up coalition forma-
tion." Proc. Int. Conf. on Autonomous Agents and
Multi Agent Systems, pp. 867 -- 874, 2003.
[33] R. Subrata, A. Y. Zomaya, and B. Landfeldt.
"Game-theoretic approach for load balancing in
computational grids." IEEE Trans. on Parallel and
Distributed Systems, 19(1):66 -- 76, 2008.
[34] S de Vries and R. Vohra.
"Combinatorial auc-
tions; a survey." INFORMS Journal of Computing,
15(3):284 -- 309, 2003.
[35] H-J. Zhang, Q-H. Li, and Y-L. Ruan. "Resource
co-allocation via agent-based coalition formation in
computational grids." Proc Second Int. Conf. on
Machine Learning and Cybernetics,, pp. 1936 -- 1940,
2003.
[36] G. Wei, A. Vasilakos, Y. Zheng, and N. Xiong. "A
game-theoretic method of fair resource allocation
for cloud computing services." The Journal of Su-
percomputing, 54(2):252 -- 269, 2010.
Dan C. Marinescu. During the period 1984-2001 Dan
Marinescu was an Associate and the Full Professor in
the Computer Science Department at Purdue University
in West Lafayette, Indiana. Since August 2001 he is
a Provost Professor of Computer Science at University
of Central Florida. He has published more than 220
papers in referred journals and conference proceedings
and several books including: Cloud Computing: Theory
and Practice, Morgan Kaufmann, 2013.
Ashkan Paya is a Ph.D. candidate in the Electrical
Engineering and Computer Science Department at Uni-
versity of Central Florida pursuing his degree in Com-
puter Science. He graduated from Sharif University of
Technology in Tehran, Iran, with a BS Degree in the
same major in 2011. His research interests are in the
area of resource management in large-scale systems and
cloud computing.
John Morrison is the founder and director of the Cen-
tre for Unified Computing. He is a co-founder and di-
rector of the Boole Centre for Research in Informatics, a
principle investigator in the Irish Centre for Cloud Com-
puting and Commerce and a co-founder and co-director
of Grid-Ireland. Prof. Morrison has held a Science Foun-
dation of Ireland Investigator award and has published
widely in the field of Parallel Distributed and Grid Com-
puting. He is a principle investigator in the Irish Centre
from Cloud Computing and Commerce, where he leads
the Service LifeCycle Group. He has been the guest
editor on many journals including the Journal of Super
Computing and the Journal of Scientific Computing. He
is on the Editorial Board of Multi-Agent and Grid Sys-
tems: An International Journal, published by ISO Press,
and the International Journal of Computational Intelli-
gence: Theory and Practice (IJCITP). He is a senior
member of the ACM and a senior member of the IEEE.
14
|
1110.3961 | 1 | 1110 | 2011-10-18T12:50:26 | A Dynamic Framework of Reputation Systems for an Agent Mediated e-market | [
"cs.MA",
"cs.AI",
"cs.IT",
"cs.SI",
"cs.IT"
] | The success of an agent mediated e-market system lies in the underlying reputation management system to improve the quality of services in an information asymmetric e-market. Reputation provides an operatable metric for establishing trustworthiness between mutually unknown online entities. Reputation systems encourage honest behaviour and discourage malicious behaviour of participating agents in the e-market. A dynamic reputation model would provide virtually instantaneous knowledge about the changing e-market environment and would utilise Internets' capacity for continuous interactivity for reputation computation. This paper proposes a dynamic reputation framework using reinforcement learning and fuzzy set theory that ensures judicious use of information sharing for inter-agent cooperation. This framework is sensitive to the changing parameters of e-market like the value of transaction and the varying experience of agents with the purpose of improving inbuilt defense mechanism of the reputation system against various attacks so that e-market reaches an equilibrium state and dishonest agents are weeded out of the market. | cs.MA | cs | IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
1
A Dynamic Framework of Reputation Systems for an
Agent Mediated e-market
Vibha Gaur1, Neeraj Kumar Sharma2
1 Department of Computer Science, University of Delhi
Delhi, India
[email protected]
2 Department of Computer Science, University of Delhi
Delhi, India
[email protected]
Abstract
The success of an agent mediated e-market system lies in the
underlying reputation management system to improve the quality
of services in an information asymmetric e-market. Reputation
provides an operatable metric for establishing trustworthiness
between mutually unknown online entities. Reputation systems
encourage honest behaviour and discourage malicious behaviour
of participating agents in the e-market. A dynamic reputation
model would provide virtually instantaneous knowledge about
the changing e-market environment and would utilise Internets’
capacity for continuous interactivity for reputation computation .
This paper proposes a dynamic reputation framework using
reinforcement
theory that ensures
learning and fuzzy set
judicious use of information sharing for inter-agent cooperation.
This framework is sensitive to the changing parameters of e-
market like the value of transaction and the varying experience of
agents with the purpose of improving inbuilt defense mechanism
of the reputation system against various attacks so that e-market
reaches an equilibrium state and dishonest agents are weeded out
of the market.
Keywords: Reputation, Reinforcement Learning, Fuzzy
attribute weights, e-market.
1. Introduction
With the growing popularity of e-commerce and amount of
information on WEB, users expect automated techniques
to assure the trustworthiness of information available on
internet. Software agents offer a promise to change e-
commerce trading by helping internet traders to purchase
products from online distributed resources based on their
interests and preferences [16]. Assuring the trustworthiness
of web products and services in such an environment
where actual traders may never meet each other is a
reputation systems.
challenging
task performed by
Reputation systems have a high utility
in
those
environments where entities are long lived, feedback about
the current interactions is captured and distributed, and
past feedback/experience guides buyer decisions [22].
These systems are oriented to develop trustworthiness or
the degree to which one agent has confidence in another
within the context of a given purpose or decision.
The definition and meaning of reputation varies with
applications and contexts. From an objective view,
reputation is expressed as “a quantity derived from the
underlying social network which is globally visible to all
members of the network” [25] or, “a perception that an
agent has of another’s intentions and norms” [17].
Reputation and Trust are often used in complementary
fashion as an agent expects positive outcomes when
interacting with another agent that has a reputation for
being trustworthy [8]. Some systems are described as trust
systems as therein agents determine whether another agent
will do what it says it will, whereas others are best
described as reputation systems because therein agents
compute and propagate their beliefs about other agents.
The e-market environment in which these agents operate is
generally open, that means agents can join or leave the
marketplace at any time; uncertain, i.e. the true worth of a
good can be judged only after its purchase; and un-trusted,
that is the e-market comprises of honest and dishonest
agents. The e-market is populated with self interested
buyer and seller agents that try to maximise their
respective gains. The e-market environment is itself
dynamic in nature as it undergoes continuous changes with
different agents joining and leaving the e-market at will.
The power of a reputation system in an agent mediated e-
commerce can be realized to the optimum if different
process models inherent to the e-transactions like deciding
about pricing of goods, computing and distributing
reputation of participants and selection of a seller for
purchasing a good are also dynamic [29]. A truly dynamic
the changing e-market
to
model must be sensitive
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
2
environment and must adapt to changing experience of
buyer/seller agents with each transaction. Dynamic e-
market models would provide virtually instantaneous
knowledge about the changing e-market environment and
would utilise
Internets’
capacity
for
continuous
interactivity. Designing efficient and robust reputation
systems that satisfy both the buyers as well as sellers is a
challenge for the research community.
The objective of this paper is to propose a framework for a
dynamic reputation system that is sensitive to the changing
parameters of the dynamic e-market environment like the
experience of agents involved in transactions, value of a
transaction and number of transactions between the same
buyer-seller pair. In the proposed model, each of the
individual process model of the dynamic reputation
framework is itself dynamic as selection of a seller for
buying a good depends on the changing experience
between a buyer-seller pair; computing sellers’ reputation
by a buyer depend on the experience of an agent in the e-
market, mutual experience of a buyer-seller pair and the
value of transaction. Further, incorporating value of
transaction in reputation computation affects the amount of
reputation that is to be enhanced or reduced after each
transaction. This makes the reward/penalty proportional to
the size of the transaction in which honest/dishonest
behavior is exhibited by seller agents, and negates any
benefit of a Value Imbalance attack where a seller agent
gains reputation by showing honesty for small value
transactions and then cheats for a large value transaction.
Making
the
reputation updation dependent on
the
experience of agents, by varying
the weightage of
individual experience and shared opinion from others,
reduces the effect of Ballot Stuffing attack where a number
of malicious agents artificially enhance or reduce the
reputation of another agent. Also, by making the reputation
updation sensitive to the fact that whether reputation is
earned from a single buyer or multiple buyers minimizes
the effect of collusion between a buyer-seller pair. The
proposed framework employs judicious use of information
sharing and thus reduces the associated cost by using
effective inter-agent communication.
The reputation computation strategy proposed in this paper
uses reinforcement learning (RL) techniques which provide
a general framework for sequential decision making
problems [10]. RL deals with what an agent should do in
every state that it can be and how to map situations to
action, in order to maximize the long term reward. The
learner must discover which actions yield the maximum
reward by trying them. Sometimes, actions may affect not
only the immediate reward, but also all subsequent
rewards. Hence, trial-and-error search and delayed reward
are the two most important distinguishing features of RL.
In the proposed strategy, for purchasing a good, the buyer
chooses a seller offering the highest expected value of the
good i.e. good with highest expected utility for the buyer .
Expected buyers’ requirement from a good constitutes
buyers’ estimation of goods’ attributes and is subjective
and fuzzy in nature. It is subjective as relative priority of
attributes of a good would vary with each good and with
each buyer. It is fuzzy as generally buyers’ expectations of
a particular attribute are specified in fuzzy terms like “low”
or “high”. Similarly, a buyer has to map linguistic
assessment of goods being offered by different sellers
based on their attributes to the fuzzy scale. Hence this
paper uses fuzzy set theory to allow a buyer agent to
compute attribute weights of a good and to select a seller
that offers the good with highest expected value.
The rest of this paper is organized as follows. Various
reputation models from literature and in commercial use
are introduced in section 2. Section 3 presents the
proposed dynamic reputation framework. To address
existing problems, section 4 illustrates the performance of
the proposed system against known attacks. A case study is
presented in section 5. Section 6 concludes the paper.
2. Related Work
Reputation models are an important component of e-
market, help building trust and elicit cooperation among
loosely connected and geographically dispersed economic
agents [12]. A number of reputation models described in
literature are discussed below.
The evidential model [2, 3] for reputation computation
assumes a distributed reputation environment and is based
on Dampster Shafer Theory. An agent
finds
the
trustworthiness of another agent [3] based on its direct
interaction and testimonies of other trustworthy agents.
Some reputation models [21, 26] from literature employ
reinforcement
individual
learning and are based on
experience only. In reputation model for increasing user
satisfaction [26], seller agents adjust the price and quality
of goods to maximise their profit. A multi-facet reputation
model [21] involves reputation computation of both buyer
and seller agents using quality, price and delivery time of
goods. But, these systems [21, 26] suffer heavily from re-
entry and multi-identity attacks as these use negative
reputation and new sellers do not start from minimum
reputation.
TRAVOS [15] employs Bayesian probability analysis and
computes trust of an agent by taking into account past
experience between two agents, and in case of lack of past
experience, this model utilizes the information collected
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
3
from third parties. To filter out unfair opinions, TRAVOS
uses an endogenous approach to filter out unfair opinions.
PeerTrust [19] is a reputation model that uses techniques
for resilient reputation management against vulnerabilities
like feedback sparsity and feedback manipulation. It talks
about dynamism in electronic communities from the
perspective of honest and dishonest behaviour of actors.
Reputation in Gregarious societies (REGRET) [16, 17]
employs fuzzy rules to find reliability of witness agents
based on their relationship with the target agent. REGRET
is a multi-facet reputation mechanism that models the
reliability of
reputation based on
the number of
interactions of witness agents with the target agent.
Another model “Trunits” [24] is based on accumulation of
trust units (trunits). A seller must possess sufficient number
of trunits before executing a transaction. To engage in a
transaction, seller must risk a particular quantity of trunits
which is put into an escrow with the market operator. After
a transaction, if buyer is satisfied, seller gets more trunits,
otherwise it loses risked ones.
Broker assisting TRS [4] is a flexible model based on
Artificial Neural Networks
(ANN)
that
employs
backpropagation algorithm. Use of ANN helps to reduce
noise data and supports incremental training, so each agent
requests for information only from those having a similar
reputation evaluation criterion.
In Reputation Dynamics and Convergence [8], authors
formalize the desiderata that from a dynamic systems’
perspective a reputation system should have the properties
of Monotonicity and Accuracy. As an example of
Monotonicity, agents who provide high quality goods at
low price should acquire better reputation and, in systems
with focus on Accuracy, the buyer should quickly learn
the accurate reputation value for the seller. The Dynamic
Framework proposed
in
this paper
incorporates
Monotonicity as the process of seller selection and also
updation of reputation are based on the presence of
favourable goods’ attribute like low price and high quality.
Further, a fraudulent seller is penalised immediately to
keep the reputation estimate accurate.
The P4P (Pervasive Platform for Privacy Preferences) [20]
system concentrates on privacy control in case of e-
transactions. The paper acknowledges the property of e-
market environment being dynamic and, the need that the
existing systems in this environment should also be
dynamic. It emphasizes importance of reputation by
allowing the clients, the freedom to not disclose personal
data according to the level of reputation.
A number of simple online reputation systems are in
commercial use. eBay [14] is the most popular auction site
that has feedback forum as a reputation system in which
after each transaction, a buyer rates a seller as positive,
negative or neutral i.e. +1, -1 or 0 respectively. The
reputation of a user is computed by subtracting total
number of negative feedbacks from the total number of
positive feedbacks obtained from distinct users [23].
Amazon [13] is America’s largest online retailer where
reviews include star ratings from 1 to 5 and a prose text.
Average of all ratings is used to assign reputation.
A limitation of the existing systems from literature [1, 3, 4,
8, 15, 16, 19, 20, 21] is that, during the process of
computing or updating of reputation values, these do not
take
into consideration
the changing parameters of
dynamic e-market environment like the varying experience
of agents and the value of a transaction which make them
vulnerable to different attacks. The proposed reputation
framework incorporates value of a transaction in the
strategy of reputation computation to remove the problem
of Value Imbalance attack and, varies the weightage of
individual and shared
reputation components with
changing experience of agents to minimise the effect of
Ballot Stuffing attack.
3. Dynamic Reputation System Framework
Reputation systems are oriented to encourage trustworthy
behaviour, increase user satisfaction and deter dishonest
participants by providing means through which reputation
could be computed and disseminated [22]. The e-market
environment in which reputation systems operate is
dynamic as it changes continuously in terms of agents
freely entering/exiting the market and also with the varying
experience of agents. Therefore, as a buyer gains
experience of a sellers’ behaviour with each repeated
transaction, the weightage of the individual experience of a
buyer-seller pair should increase as compared to the
opinion shared by other buyer agents. Moreover, economic
worth of being honest or dishonest in a transaction cannot
be judged without taking into account the value of a
transaction as honest behaviour in a large transaction is
more important than in a small transaction.
the
framework should base
A dynamic
reputation
reputation computation methodology
the
itself on
dynamics of the e-market environment to infuse some
inbuilt defense capability against possible attacks. In order
to have a robust and high utility reputation system,
different activities belonging to reputation computation
methodology it should be adaptive to the changing
environment and the experience of agents involved in a
transaction. The next section describes the proposed
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
4
dynamic reputation computation strategy that employs
reinforcement learning and fuzzy set theory.
3.1 Buyers’ Strategy for Reputation Computation
The proposed buyers’ strategy is based on the e-market
model having a set of buyers and sellers. In this model,
sellers are divided into four categories, namely, reputed,
non-reputed, dis-reputed and new sellers. The reputation of
seller s being computed by buyer b is composed of two
components: individual reputation and shared reputation.
These two components are combined to represent overall
reputation of a seller agent.
In this model, B represents the set of buyers, S represents
the set of sellers and, G the set of goods. Let
[0,1)
represents
individual
reputation
(IR)
component,
[0,1) represents shared reputation (SR) i.e. the
opinion of other buyers for seller s, and
[0,1)
represents overall reputation of seller s at time t, for the
buyer b. At time t+1, buyer b stores/remembers the overall
, with whom buyer b
reputation
of all sellers
has interacted at time t in the past. Each buyer maintains
four categories of sellers as defined below.
(i)
(ii)
: Sellers in the reputed list of buyer b, i.e.
, where
,
is the reputation
threshold of buyer b and
.
: Sellers in the non-reputed list of buyer b, i.e.
.
where
(iii)
: Sellers in the dis-reputed list of buyer b, i.e.
, where
,
is the dis-
reputation threshold and
.
(iv)
: Sellers that are new to buyer b in the market,
initially
A new seller s remains in this
list until its reputation crosses the dis-reputation
threshold
. Before crossing
, if a seller cheats
than it is moved to the list of dis-reputed sellers
and is never considered again for business.
The process of choosing a seller for purchasing a good
based on its expected value uses three important algebraic
operations on fuzzy numbers: inverse, addition and multi-
plication. If
and
are two positive trapezoidal fuzzy numbers then, the fuzzy
is given in (1) and inverse of a fuzzy
addition of
and
number
represented as
is shown in (2) below.
(1)
(2)
Unlike addition and subtraction, product of two trapezoidal
fuzzy numbers may not result into a trapezoidal number
[6, 7]. Therefore, this paper uses an approximation of the
product of two trapezoidal fuzzy numbers to a new
trapezoidal fuzzy number [7]. The product of
two
trapezoidal fuzzy numbers, A and B given above is
trapezoidal
approximated
by
the
fuzzy
number
as proposed in [6, 7] where,
,
,
,
(3)
For defuzzifying, Centre Of Area (COA) or Centroid
method is used. For a fuzzy number
, its
COA is computed as:
.
The proposed reputation computation methodology based
on the concept of reinforcement learning and fuzzy set
theory is divided into three phases. In Phase I, a buyer
expresses its willingness to buy a good and the set of
sellers’ who respond for selling that good are elicited and a
seller selection methodology using fuzzy arithmetic is
applied to select a seller for purchasing that good. Phase II
includes
reputation computation using reinforcement
learning. It begins after purchasing the good, where the
buyer updates
the sellers’ reputation based on
the
experience of the current transaction and the opinion from
others. Finally in Phase III, the buyer updates its list of
reputed, non-reputed, dis-reputed and new sellers. A
detailed description of this methodology divided into
Phase I, Phase II and Phase III is given below.
Phase I:
1. The process of buying and selling starts with a buyer b
announcing the need to buy a good g by sending
broadcast request to all sellers. Those sellers who are
willing to sell good g respond by submitting their bids.
At any given time, buyer b preferably purchases a
good from a reputed seller. If no seller from the
reputed list offers good g then the buyer b selects a
seller from the set of non-reputed sellers but in no case
the buyer would choose a dis-reputed seller [27]. In
addition, with a small probability ρ, buyer b would
choose a seller from the list of new sellers’ i.e.
.
Initially the value of ρ is 1 and it decreases over time
to some minimum value defined by buyer b.
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
2. After receiving sellers’ bids for good g, buyer b first
computes the expected value
of good g’s offer
from each seller and then selects a seller s that is
offering good g with highest expected value i.e. max.
based on the following strategy by computing
goods’ attribute weights using extent analysis method
[5, 28] and combining it with fuzzy AHP technique.
i. Obtain
the buyers’ assessment of pairwise
comparison of different attributes of a good in
linguistic
(E),
important
like Equally
terms
Moderately Important (M), Highly Important (H),
Very Highly
Important
(VH) or Extremely
Important (EI) as illustrated in Fig. 1.
Fig. 1 Fuzzy Scale for Relative Importance of Attributes
Using fuzzy scale of Fig. 1, map these linguistic
terms to trapezoidal fuzzy values. For example,
Highly Important (H) is mapped to trapezoidal
fuzzy number (3,5,5,7).
ii. Compute subjective fuzzy weights of different
attributes of good g from the buyer’s perspective by
combining extent analysis method [5] with fuzzy
AHP. Let
(Fuzzy Pairwise Matrix) represents
the fuzzy reciprocal n x n matrix representing all
pairwise comparisons
for all
as
illustrated in Eq (4) below.
Where
and all
and their inverse
are trapezoidal fuzzy numbers. The subjective
weight computation of attribute
denoted as
is given in Eq. (5).
Further, compute
for i = 1,2 ..., n, i.e. for all
attributes of a good represented by
Eq. (6).
is shown in
5
(6)
iii. Compute the empirical weight component
, i.e.
the average of fuzzy weight of each attribute, for i =
1, 2,..., n, in a maximum of k number of previous
transactions by the same buyer for the same good
represented by
below.
(7)
iv. Obtain the overall fuzzy attribute weight
of a
good by using Eq. (8) given below.
(8)
Similarly, compute
for i = 1,2,..,n, represented
by
as shown in Eq. (9).
(9)
In Eq. (8), the value of δ is zero in the case of a
buyer purchasing a good for the first time. With
each subsequent purchase of the same good by a
buyer, the value of δ increases by a small fraction.
This ensures that initially when a buyer has no
experience of buying a good, the overall weight of a
goods’ attributes depends only on subjective weight
component of each attribute of the good i.e.,
.
As buyer gains experience by buying a good
repeatedly, the importance of its empirical weight
component i.e.
increases and the importance
decreases
of subjective weight component i.e.
(4)
proportionately. This means that after participating
in sufficiently large number of transactions, say k
=100 transactions for an δ increment rate of 0.01,
by the same buyer for a particular good, it is not
necessary for a buyer to incur the overhead of
computing the subjective weights of the goods’
attributes
and
instead utilise
the previous
transactions weight information.
(5)
v. Solicit the buyers’ assessment of each seller’s offer
for the good in linguistic terms like Poor (P),
Average (A), High(H), Very High (VH) or
Excellent (EX) based on trapezoidal fuzzy scale of
Fig. 2.
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
6
computing the difference between the actual value and
the expected value of the good as given in Eq. (12)
below.
Δ =
(12)
5. If Δ > 0, then using reinforcement learning technique,
buyer b updates reputation of the seller s at time t+1 i.e.
with a value greater than its current value as
shown in Eq. (13) below.
(13)
Where μ represents effective reputation value increase
factor as shown in Eq. (14).
(14)
and,
(15)
Fig. 2 Fuzzy Scale for Linguistic Performance of Sellers
vi. Using fuzzy scale of Fig. 2, map these linguistic
terms into fuzzy performance ratings of good g’s
offers by different sellers. Let
represents fuzzy
performance ratings of seller i for attribute j. Fuzzy
performance of each seller i, for i = 1,2..,m and for
each attribute j, for j = 1,2,..,n is represented by
fuzzy attribute performance matrix
in Eq. (10).
As per Fig. 3, if seller 1’s goods’ performance for
attribute 2 is “VH” then its fuzzy performance
rating as per Eq. (10) is
.
vii. Compute the fuzzy value of the seller i’s good
as:
. The fuzzy value
matrix of each seller i’s good, for i = 1, 2, ..., m
represented by
is shown in Eq. (11) below.
Eq. (15) is used to map the value of a transaction x in
(10)
the range from 0 to 1 which in case of a single good
being purchased is equal to the price p of the good g.
Also λ is a constant in the range 0 to 1, and e is a
constant with a value of 1.01. The function to compute
η in Eq. (15) ensures that the value of μ in Eq. (14) and
hence the reputation
increases monotonically
with the value of transaction. In Eq. (14), β is a
constant with initial value 0 and its value increases by a
small
factor, say 0.001, with each successive
transaction between the same buyer seller pair. This
ensures that with increase in mutual experience of a
buyer-seller pair, reputation value i.e.
increases
at a relatively smaller rate for the same value
transaction according to the convention that reputation
earned from different buyers is more important than the
reputation earned from large number of repeated
transactions with the same buyer as shown in Table 1
below.
(11)
μ
Value of
Transaction
(x)
Table 1: Monotonic Increase of Reputation with value of transaction
but discounted with increase in number of transactions between the
same buyer-seller pair (For previous reputation i.e.
= 0.37)
For β = 0.5
For β = 0
% increase
Updated
% increase
Updated
Reputa-
in
Reputa-
in
Reputation
Reputation
tion
tion
0.113
0.169 0.0007 0.37
0.001 0.371
100
0.563
0.003 0.372
0.845
0.005 0.373
500
2.237
0.013 0.378
3.355
0.02 0.3824
2000
0.049 0.401
8.264
0.032 0.39
5.509
5000
0.063 0.4098 10.751
10000 0.095 0.4297 16.127
20000
0.18 0.4837
30.726 0.12 0.446
20.484
μ
It can also be observed from Eq. (13) that individual
reputation at time t+1 is based on overall reputation at
time t to impress upon the fact that in the next
viii. viii. Perform defuzzification on the fuzzy matrix
to
obtain crisp value matrix CVS using Centre of Area
approach (COA). CVS contains the crisp expected
value i.e.
of good g’s offer from each
seller.
ix. Select the seller s with the highest crisp expected
of the good g for placing
value i.e. max.
purchase order for the good g.
Phase II:
3. Once the buyer receives a good after purchase, it
computes the actual value of that good i.e.
,
reflecting whether the received good is satisfactory or
not as per the buyers’ assessment of the actual good by
again using step 2 of Phase I.
4. After computing the actual value of a good, buyer
updates the individual reputation of seller by first
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
7
Value of
Transaction
μ
(x)
Table 2: Monotonic Decrease of Reputation with value of a
transaction but discounted with increase in number of transactions
between the same buyer-seller (Previous reputation,
= 0.37)
For β = 0.5, γ = 2
For β = 0, γ = 2
% increase
Updated
% increase
Updated
Reputa-
in
Reputa-
in
Reputation
Reputation
tion
tion
-0.226
-0.339 0.0013 0.3692
0.002 0.3687
100
-1.127
0.006 0.3658
-1.69
500
0.01 0.3637
-4.473
-6.71
2000 0.039 0.3452
0.026 0.3534
5000 0.097 0.3088
-16.53 0.064 0.3292
-11.02
-21.5
-32.25 0.126 0.2904
10000 0.189 0.2507
20000 0.361 0.1426
-61.45
0.24 0.2184
-40.97
μ
The use of the penalty factor γ = 2 applied during
reputation computation ensures that the reputation
drops at
the
to
rate as compared
the
twice
corresponding rate of its increase for the same value
transaction. Comparison of relative
increase and
decrease in reputation corresponding to the changes in
the value of transaction and number of transactions
between a particular buyer-seller pair is shown in Fig. 4
below.
transaction, the overall reputation of a seller computed
by a buyer at the end of previous transaction becomes
the individual experience of that buyer agent. On
comparing the relative increase in percentage of
reputation in case of a buyer-seller pair having no
previous transaction represented by β = 0 and after
gaining experience of 500 transactions represented by β
= 0.5, it is found that relative increase in reputation is
less in case of β = 0.5 as compared to the situation
where β = 0 as illustrated in Fig. 3.
Fig. 3 Monotonic increase of reputation with value of transaction
but this increase is discounted/reduced with increase in number of
transactions between the same buyer-seller pair to minimise the
effect of collusion between a particular buyer and seller
6. If Δ < 0, which represents the fact that the purchased
good g has not been satisfactory as per buyer b’s
assessment, then using reinforcement learning, buyer b
updates the reputation of the seller s at time t+1 i.e.
by a value less than its current value as
described by Eq. (16).
(16)
Fig. 4 Reputation drops faster than its increase
to discourage dishonest sellers
Where ξ represents effective reputation value decrease
factor due to unsatisfactory or dishonest behaviour of a
seller agent and is illustrated in Eq. (17) below.
(17)
Where γ is the Penalty Factor and value of γ is kept
greater than 1 to ensure that reputation decreases at a
faster pace as compared to the rate of its increase. This
property is based on the convention that reputation is
difficult to build but easy to tear down. The underlying
purpose is to discourage dishonest behavior of seller
agents in e-market by slapping a higher penalty on
fraudulent sellers. Like μ, ξ is also dependent on the
value of a
transaction and the number of past
transactions between a particular buyer-seller pair.
Hence there is steep reputation drop for a large value
transaction as compared to a small value transaction as
described in Table 2.
7. After computing the individual reputation of a seller,
this model combines it with the shared reputation about
the seller s from other buyers to compute the overall
reputation of the seller agent s. The equation to
compute overall reputation function
is given
below in Eq. (18).
(18)
is the individual reputation of seller s
Where
that is computed by the buyer b itself and
is
the aggregate of the reputation rating of seller s that is
received from other buyer agents. Further, α is the
experience gain factor and
. The initial value
of α before the first transaction between a buyer-seller
pair is 0 and with each successive transaction, it is
incremented by a small factor of say 0.01 to ensure
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
8
To summarize, the main functions of dynamic reputation
framework are illustrated using flowcharts in Fig. 5 and
Fig. 6.
that with each successive transaction between a buyer-
seller pair, relative weight of Individual Reputation
increases and that of
(IR) component i.e.
Shared Reputation (SR) component i.e.
decreases. This implements the dynamic property that
with increasing mutual transactional experience, a
particular buyer-seller pair would depend more on their
past mutual experience and less on the opinion from
other agents. The actual rate at which the value of α
should increase depends on the good to be purchased
and is to be decided by domain experts. After
sufficiently large number of transactions, as value of α
would depend only on
approaches 1,
and the weightage of
would effectively
become zero. This ensures that initially when a buyer
agent has no experience of a seller, its dependence is
greater on the opinion from other buyers although it
means incurring some communication overhead. Once
a buyer gains sufficient experience of past transactions
with a particular buyer, it can avoid the overhead of
inter-agent communication as the computation of
overall reputation depends only on the individual
reputation component. Hence, this framework employs
judicious use of information sharing and thus reduces
its cost with effective inter-agent communication.
If a seller is new to a buyer b i.e.
then,
(19)
And, if a seller is new in the marketplace, i.e.
then,
(20)
Phase III:
8. Finally, on the basis of the overall reputation rating of a
seller s, sets of reputed, non-reputed, dis-reputed and
new sellers i.e. SR, SNR, SDR and SNewR are updated as:
If s is not a reputed seller, and
, then
. (21)
Fig. 5 Dynamic Reputation Framework for Reputation System
If s is a reputed seller, and
, then
the algorithm of seller
The flowchart summarizing
selection strategy for computing expected/actual value a
. (22)
product is given in Fig. 6 ahead.
If s is not a dis-reputed seller, and
, then
. (23)
If s is not non-reputed, and
,
. (24)
Finally, if s is a new seller, and, if
, then
. (25)
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
9
An attack in which two agents mutually rate each other
with abnormally high ratings is called RECiprocity (REC)
whereas in RETaliation (RET) both the agents rate each
other with abnormally low ratings.
Reputation-Lag (RL) takes advantage of the lag i.e. time
gap, before cheating results in reduced reputation. During
this period, an agent gets unlimited opportunities to cheat
before other agents become aware of its loss of reputation
due to malicious behaviour.
In Value-IMbalance (VIM) attack, reputation earned or
lost during a transaction is not related to value of a
transaction. The effect of showing honest behaviour by
selling a large number of high quality but low value goods
and, dishonest behaviour by selling a small number of low
quality but high value goods does not result into any
significant loss in reputation score. This helps a malicious
seller who behaves honestly for small transactions to gain
reputation and then cheats in large transactions.
If a seller agent has no further utility of good reputation, it
utilises its entire reputation to cheat and exits from e-
market. This attack is called Sudden-Exit (SE).
In Multiple-Identity (MI) or Sybil Attack, a seller is able to
open multiple accounts thereby increasing its probability to
sell a good. It continues selling the goods honestly through
some and dishonestly through others without facing any
significant penalty. It exits from the account with a low
reputation and opens another account.
Sometimes, a number of attackers employ a combination of
strategies to launch a multifaceted and coordinated attack.
This
is known as Orchestrated (ORC) attack [18].
Attackers change their behaviour overtime and divide
themselves into sub-groups where each group plays a
different role at different time.
4.1 Comparative Performance Analysis
to generate an accurate
Reputation systems seek
assessment of participants’ behaviour
in potentially
adversarial environments [18]. In uncertain and un-trusted
agent based environment of e-market, where the actual
buyers and sellers may never meet, absence of such
systems may lead to rampant cheating, fraud, mistrust and
eventual system failure. Hence, the success of a reputation
system
is measured by
the accuracy of computed
reputation that predicts the quality of future interactions in
an environment where a participant may try to exploit the
system to its own advantage. This section highlights the
performance of dynamic reputation framework based on its
relative strength as compared to other models from the
literature in Table 3.
Fig. 6 Seller Selection Strategy (SSS) for computing
expected / actual value of a good
4. Common Attacks and Proposed Defense
Reputation systems are different from general trust based
systems in a way that they include self interested actors or
agents who cheat and effectively launch various attacks to
defeat these systems. The impact of attacks against
reputation systems is much more than the manipulation of
reputation values as these result into money fraudulently
lost and ruined business reputations. This section discusses
different type of attacks classified in literature [2, 3, 18,
22, 24] and presents a comparative performance analysis
of the defense capability of the proposed system against
these attacks.
In Ballot Stuffing (BS), a group of agents collude to rate a
particular agent with abnormally high ratings, whereas in
Badmouthing (BM) an agent is rated abnormally low. In
this attack, colluding agents participate in events that lead
to allocation of reputation or feedback about that agent.
Re-ENtry (REN) is an attack where a low rated agent exits
the market and re-enters with a new identity. This attack is
facilitated by the availability of cheap pseudonyms in the
online environment. The reputation systems with negative
feedback are especially vulnerable to REN.
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
10
No model except
Truntis deals with
this problem.
Type of
Attack /
Problem
Value
Imbalance
(VIM)
Reputation
Lag (RL)
Table 3: Qualitative analysis of Dynamic Reputation System (DRS)
correctly reflect the behaviour of participants and their
based on known attacks/problems and defense mechanisms
capability to withstand different type of attacks launched
Defense
Defense Mechanisms in
by dishonest agents. Therefore, a good reputation system
Mechanisms in
proposed Dynamic Reputation
must incorporate some inbuilt defense mechanisms to
Framework
Other Models
ensure that honest participants are rewarded with economic
VIM is resolved as the amount
gains and cheaters are penalised with economic loss. The
of change in reputation is
proposed strategy incorporates inbuilt defense capability in
monotonically related to the
the
reputation computation methodology
itself by
value of a transaction.
RL is reduced as with increase
increasing its resilience against various attacks especially
in the mutual experience of a
Value Imbalance and Ballot Stuffing. It also discourages
buyer-seller pair, weightage of
fraudulent behaviour by slapping a higher penalty on
shared reputation reduces and
dishonest sellers than the corresponding reward for honest
finally becomes negligible after
behaviour.
large number of transactions.
Its effect is minimised as
reputation earned by a seller in
lieu of repeated transactions
with the same buyer is
discounted with each successive
transaction. Effect of REC/RET
is also limited by the value of
transaction.
REN attack is partially resolved
as to re-enter, an agent must
lose existing reputation and re-
start with minimum reputation.
Probability of SE attack is
reduced. As reputation earned is
proportional to value of
transaction, so to cheat and exit
from e- market, an agent has to
first earn sufficient reputation
by being honest for both large
value and large number of tran-
sactions. Losing hard earned
reputation is not viable unless
transaction value is very high.
the application of proposed reputation
illustrate
To
framework, a case study was conducted by simulating an
electronic marketplace with four users as buyers and six
users as sellers, i.e. B = {bi where i = 1...4} and S = {sj
where j = 1…6}, where B is the set of buyers and S is the
set of sellers in the marketplace for good g. Some
scenarios in the marketplace are shown below.
Models based only
on individual repu-
tation [21, 24, 26]
do not suffer from
RL, other models
have no solution.
Commercial
models like eBay
have a strong
presence of this
attack as 98% of
the eBay ratings
are positive due to
the fear of RET.
e-Bay and Truntis
deal with this
problem with
partial success.
Scenario 1: A situation was investigated where buyer b3
wanted to buy a good g. The sellers s1 to s6 were known to
buyer b3, although only three sellers were in its overall
= {sj where j = 3,4,5}. Further,
reputed list i.e.
=0.15, e = 1.01, α incremental rate of 0.01
=0.45,
and β incremental rate of 0.001 per transaction. Based on
buyer b3’s experience, existing overall reputation
No feasible
solution in any of
the proposed
model so far.
of each seller is depicted in Table 4.
5. Case Study
Reciprocity
(REC) and
Retaliation
(RET)
Re-entry
(REN)
Sudden Exit
(SE)
Multiple-
Identity (MI)
Ballot
Stuffing
(BS)/
Bad-
mouthing
(BM)
Orchestrated
(ORC)
No inbuilt feasible solution.
No feasible
solution provided.
The effect of BS/BM reduces
Evidential model,
with each successive transaction
TRAVOS,
between a buyer-seller pair as
REGRET and
weightage of shared reputation
Broker-Assisting
decreases and becomes
TRS try to deal
negligible when an agent gains
with this attack
sufficient experience of other
with varying
success.
trader agent.
Only partial solution to a subset
No known solution
for this type of
of attacks is possible as dealing
multifaceted
multiple attacks with actors
changing roles is very difficult.
attacks.
Table 4: Individual reputation ratings of different sellers to buyer b3
s6
s5
s4
s3
s2
s1
sj
0.25
0.48
0.50
0.37
0.57
0.20
The buyer b3 specified the pairwise importance of different
attributes of good g i.e. of Price (P), Quality (Q), Delivery
Period (DP) and Service Offered (SO) in linguistic terms.
Their equivalent fuzzy values based on the fuzzy scale of
Fig. 2 are shown as Fuzzy Pairwise Matrix (
) in (26).
Reputation systems foster good behaviour, punish bad
behaviour when it occurs, and reduces the risk of being
harmed by others’ bad behaviour. Strengths and
weaknesses of reputation systems are assessed qualitatively
on the basis of their ability to convert the experience of
buyer and seller agents into a reputation metric that
The average of the weights in the previous transactions
were
=
(0.0405,0.115,0.115,0.2435),
=
= (0.074,0.196,0.196,0.443)
(0.11,0.46,0.46,0.87),
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
11
= (0.0875,0.367,0.367,0.718). Hence, SW , EW,
and,
and W of different attributes of the good g as computed in
MATLAB with δ = 0.27 are illustrated in Fig. 5 below.
price of good g by seller s5 was 1500, so x = 1500. After
purchasing, and receiving the good g, buyer b3 computed
the actual value of the good g by again using step 2, Phase
I of Section 3 as
= 13.346.
Fig. 5 Overall weight computation of attributes of good g by buyer b
Sellers s1, s3, s4, s6, responded to sell good g to buyer b1.
Now, buyer b1 computed the expected value of the product
being offered by the four sellers as explained below.
After taking buyers’ assessment of each seller’s product
offer for the attributes Price (P), Quality (Q), Delivery
Period (DP) and Service Offered (SO) in linguistic terms,
the equivalent fuzzy performance matrix
representing
fuzzy performance of various sellers’ offer for good g is
shown in (27).
Using (11),
and
was computed as,
after defuzzifying the resultant crisp expected value (CVS)
representing the expected value of good g for each seller
, as computed using MATLAB is
illustrated in Fig. 6.
Using (12), ∆ = 13.146-12.2319 = 0.9141 > 0. (28)
As ∆ > 0, buyer b3 incremented the individual reputation
of seller s5 as shown below.
= 1 – (1.01) -0.001*1500 = 0.014815 (29)
and μ =
= 0.01373 (30)
Using (13),
=0.57+0.01373*(1-0.57)=0.576. (31)
The aggregated shared reputation value for seller s5 was
0.56. Therefore, b3 computed overall rating of seller s5 by
using Eq. (18) as:
combining
with
= 0.79*0.576+(1-0.79)*0.56 = 0.572. (32)
Scenario 2: Another situation was investigated where buyer
b2 wanted to buy good g. Sellers s1 to s4 and s6 were known
to buyer b2, whereas sellers s3 and s6 were in its overall
= {sj where j = 3,6}. Moreover,
reputed list, i.e.
=0.25, γ = 3, e = 1.01, α incremental rate of
=0.5,
0.01 per transaction and β incremental rate of 0.001 per
transaction. After previous transaction of buyer b2, overall
reputation ratings
for each seller are given in
Table 5.
Table 5: Reputation ratings of different sellers in buyer b2’s memory
s6
s4
s3
s2
s1
sj
0.312
0.43
0.51
0.39
0.53
Using step 2, Phase I of section 3, the expected value of
the good g equivalent to 11.65 was computed to be the
maximum for seller s3 so the buyer b2 chose seller s3 to buy
good g. Also, buyer b2 had 45 previous transactions with
the seller s3, therefore α = 0.45 and β = 0.045. As seller s3
offered the good g at a price of 6750, so x = 6750. After
purchasing, by again using step 2 of Phase I, buyer b2
computed the actual value of good g, i.e.
as
10.87.
Using (12), ∆ = 10.87 – 11.65 = - 0.78 < 0. (33)
As ∆ < 0, buyer b2 decremented the individual reputation
of seller s5 as shown below.
Fig. 6 Fuzzy (FVS) and Crisp (CVS) values of Sellers’ offers
= 1 – (1.01) -0.001*6750 = 0.064959 (34)
Based on Fig. 6, seller with the highest expected value
of the good g as 12.2319 is selected by buyer b3 for
purchase. Also, as buyer b3 had 79 previous transactions
with the seller s5, therefore α = 0.79 and β = 0.079. The
ξ = γ
= 0.18649 (35)
Using (16),
=0.51-0.18649(1-0.51))=0.4186. (36)
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
12
Further, the aggregated shared reputation value for seller s3
was 0.54. Therefore, b2 finally computed overall rating of
seller s3 by combining its individual rating of s3 with
using Eq. (18) as shown below in Eq. (37).
=0.45* 0.41862 + (1- 0.45)*0.54=0.4854 (37)
Scenario 3: In another case involving Ballot Stuffing
attack, buyer b4 needed a good g. The sellers s1 to s6 were
known to buyer b3 where s1, s2 and s4 are in its reputed list.
Further,
=0.18, e = 1.01, α incremental rate of
=0.4,
0.01 and β incremental rate of 0.001 per transaction. A
number of successive transactions between the buyer b4
and seller s2 were observed where Ballot Stuffing attack
was launched on buyer b4 after 20, 50, 75, 95 and 100
transactions between buyer b4 and seller s2. It was seen that
the increase in reputation due to BS reduced with the
increase in number of transactions as shown in Table 6.
Table 6: Effect of BS reduces with increase in number of
transactions between buyer b4 and seller s2
Value
Number
of Tran-
of Tran-
saction
sactions
Effect of
BS in %
Change of
Reputation
0.47
0.44
0.48
0.51
0.46
20
50
75
95
100
12000 0.528
1500 0.448
5300 0.505
3000 0.523
2700 0.473
0.94
0.93
0.95
0.94
0.95
0.858
0.689
0.616
0.565
0.473
62.29
53.83
22.01
3.98
0
It was also observed that the effect of Badmouthing would
also be reduced due to reduced weightage of shared
reputation with the increase in transactional experience of
a buyer-seller pair.
6. Conclusions
This paper proposed a framework for a dynamic reputation
system that is sensitive to the changing parameters of e-
market environment like experience of agents and the
value of a transaction in e-market environment. The
proposed system has improved inbuilt defense mechanisms
against various attacks especially against Ballot Stuffing
and Value Imbalance. In this framework, increase in
transactional experience leads to increased weightage of
individual reputation and honesty in a large transaction
leads to a greater increase in reputation as compared to a
small transaction. Further, non-satisfactory or fraudulent
large drop of
sellers are penalized with relatively
reputation
resulting
into
reduced
future business
opportunities. The proposed framework makes judicious
use of information sharing by adapting to the changing e-
market environment.
References
[1] A. Josang, R. Ismail, and C. Boyd, “A survey of Trust and
Reputation Systems
for Online service provision”,
Decision Support Systems, vol. 43, no. 2, 618 -644, (2007).
[2] B. Yu, M.P. Singh, “Detecting Deception in Reputation
Management”, In Proceedings of 2nd Int. Joint Conf. on
Autonomous Agents & Multiagent Systems, 73-80, (2003).
[3] B. Yu, M.P. Singh, “An evidential model of Distributed
Reputation management”, AAMAS’02, New York: ACM
Press, 294-301, (2002).
[4] B. Zong, F. Xu, J. Jiao, J. Lv, “A Broker-Assisting Trust
and Reputation System Based on Artificial Neural
Network”, IEEE, (2009).
[5] Chang, D.-Y., “Application of extent analysis method on
fuzzy AHP”, European Journal of Operations research,
Vol. 95, pp. 649-55, 2006.
[6] Chng-Hsing Yeh, Yu-Hern Chang, “Modelling subjective
evaluation for fuzzy group multi-criteria decision making”,
European Journal of Operational Research, pp. 464-473,
2009.
[7] Chou, C.C., “The representation of multiplication operation
on fuzzy numbers and application to solving fuzzy
multiple criteria decision making problems”, In Yang, Q.,
Webb, G. (eds.), LNCS (LNAI), vol. 4099, pp. 161 -169,
Springer –Heidelberg (2006).
[8] Christopher J. Hazard, M.P. Singh, “Reputation Dynamics
and Convergence: A Basis for Evaluating Reputation
Systems”.
[9] E. Shakshuki, S.A. Draz, “Agent-Mediated e-Commerce
System”, Proceedings of the 19th Int. Conference on
Advanced Information Networking and Application , vol. 2,
739–744, (2005).
[10] George Boulougaris, Kostas Kolomvatsos, and Stathes,
“Building the Knowledge Base of a Buyer Agent Using
Reinforcement Learning Techniques”, WCCI 2010 IEEE
World Congress on Computational Intelligence, July, 18 -
23, 2010 - CCIB, Spain.
[11] H. Huang, G. Zhu, S. Jin, “Revisiting Trust and Reputation
in Multi-Agent Systems”, Proceedings of
ISECS
International Colloquium on Computing, Communications,
Control, and Management, 424-429, (2008).
[12] H. Huang, G. Zhu, S.Jin, “Revisiting Trust and Reputation
in Multi-Agent Systems”, In Proceedings of ISECS Int.
Colloquium on Computing, Communications, Control, and
Management, 424-429, (2008).
[13] http://www.amazon.com/.
[14] http://www.ebay.com/.
[15] J. Patel, W.T.L. Teacy, N.R. Jennings, M. Luck, “A
Probabilistic Trust Model
Inaccurate
for Handling
Reputation Sources”, Springer-Verlag 2005, LNCS 377,
193-209, (2005).
[16] J. Sabater, C. Sierra, “REGRET: Reputation in Gregarious
garious Societies”, Proceedings of the Fifth Int. conference
on Autonomous Agents, Canada, 194-195.
[17] J. Sabater, C. Sierra, Reputation and Social Network
Analysis in MAS, AAMAS’02, Bologna, Italy, (2002).
[18] Kevin Hoffman, David Zage, and Cristina Nita-Rotaru, “A
Survey of Attack and Defence Techniques for Reputation
Systems”, ACM Computing Surveys, Vol. 42, Issue 1,
March 2010.
IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, July 2011
ISSN (Online): 1694-0814
www.IJCSI.org
13
[19] Li Ziong and ling Liu, “PeerTrust: Supporting Reputaiton -
Based Trust for P2P Electronic Communities”, IEEE
Transactions on Kowledge and Data Enineering, Vol. 16,
No. 7, 2004.
[20] Ohbyung Kwon, “A Pervasive P3P based negotiation
mechanism for privacy-aware pervasive e-commerce”,
Decision Support Systems 50, 213-221 (2011).
[21] Omid Roozmand, Mohammad Ali Bematbakshsh, Ahmad
Barrani, “An Electronic marketplace based on Reputational
and Learning”, Journal of Theorotical and Applied
Electronic Commerce Research,
ISSN
0718-1876
Electronic Version, VOL2/ISSUE 1/APRIL 2007/1-17.
[22] P. Resnick, R. Zeckhakuser R., E. Friedman, K. Kuwabara,
“Reputation Systems: Facilitating Trust
in
Internet
Interactions”, Comm. ACM 43-12, pp. 45-48, 2000.
[23] P. Resnick, R. Zeckhausser, J. Swanson, “Trust among
strangers in Internet Transactions: Empirical Analysis of
ebay’s Reputation System”, The Economics of the Internet
and E-Commerce, Advances in Applied microeconomics,
vol. 11, Elsevier Science, (2002).
[24] R. Kerr, R. Cohen, Modelling trust using transactional,
numerical Units, Proceedings of the conference on Privacy,
Security and Trust, Markham, Canada, (2006).
[25] T. Tran, and R. Cohen, “A Learning Algorithm for Buying
and Selling Agents in Electronic Marketplaces” Advances
in AI, LNCS, Vol. 2338, 31-43, Springer-Verlag (2002).
[26] T. Tran, and R. Cohen, “Improving user satisfaction in
agent-based
electronic marketplaces by
reputation
modelling and adjustable product quality”, AAMAS’04,
pp. 828-835 (2004).
[27] Vibha Gaur, Neeraj Kumar Sharma, Punam Bedi, “Evaluat-
ing Reputation Systems for Agent Mediated e-Commerce”,
ACEEE conference “International Conference on Advances
in Computer Science”, ACS 2010, Kerala, India, December
2010.
[28] Vibha Gaur, Neeraj Kumar Sharma, Punam Bedi, “A
Dynamic Learning Strategy of Reputation Systems for
Agent Mediated e-Commerce”, Int. J. on Recent Trends in
Engineering & Technology, Vol. 05, No. 01, Mar 2011.
[29] Vibha Gaur, Neeraj Kumar Sharma, “A Dynamic Seller
Seller Selection Model for Agent Mediated e-market”,
Springer
conference
“International Conference on
Advances in Computing and Communication”, ACC 2011,
Kochi, India, 2011.
Vibha Gaur: She is PhD from the department of Computer
Science, Delhi University. She is working as Reader in Delhi
University and has a teaching experience of about 12 years . She
has authored more than 18 papers in various international
conferences and journals. Her current research interests include
artificial
intelligence,
information
systems and
software
engineering.
Neeraj Kumar Sharma: PhD student at Delhi University. He is
also working as Assistant Professor in Delhi University and has a
teaching experience of about 8 years. He has published two
booklets pertaining to MCA syllabus of IGNOU in the subjects
Artificial Intelligence and Algorithms. He has presented two papers
in international conferences by ACEEE (ACS 2010) and Springer
(ACC 2011). He has also published a paper in International
Journal on Recent Trends in Engineering & Technology.
|
1806.04254 | 1 | 1806 | 2018-06-11T21:55:45 | Compositional Discovery of Workflow Nets from Event Logs Using Morphisms | [
"cs.MA",
"cs.LO"
] | This paper presents a modular approach to discover process models for multi-agent systems from event logs. System event logs are filtered according to individual agent behavior. We discover workflow nets for each agent using existing process discovery algorithms. We consider asynchronous interactions among agents. Given a specification of an interaction protocol, we propose a general scheme of workflow net composition. By using morphisms, we prove that this composition preserves soundness of components. A quality evaluation shows the increase in the precision of models discovered by the proposed approach. | cs.MA | cs |
Compositional Discovery of Workflow Nets from
Event Logs Using Morphisms(cid:63)
Luca Bernardinello2, Irina Lomazova1, Roman Nesterov1,2, and Lucia Pomello2
1 National Research University Higher School of Economics,
20 Myasnitskaya Ulitsa, 101000 Moscow, Russia
2 Dipartimento di Informatica Sistemistica e Comunicazione,
Universit`a degli Studi di Milano-Bicocca,
Viale Sarca 336 - Edificio U14, I-20126 Milano, Italia
Abstract. This paper presents a modular approach to discover process
models for multi-agent systems from event logs. System event logs are
filtered according to individual agent behavior. We discover workflow
nets for each agent using existing process discovery algorithms. We con-
sider asynchronous interactions among agents. Given a specification of
an interaction protocol, we propose a general scheme of workflow net
composition. By using morphisms, we prove that this composition pre-
serves soundness of components. A quality evaluation shows the increase
in the precision of models discovered by the proposed approach.
Keywords: Petri nets, workflow nets, multi-agent systems, morphisms,
composition, process discovery
1
Introduction
Process discovery focuses on the synthesis of process models from event logs con-
taining the observed record of an information system behavior. Process models
are usually developed at the design stage of the information system life-cycle.
However, the real observed behavior of the information system can eventually
differ from the designed one. In some cases, designers cannot develop precise
models describing all possible scenarios. That is why, process discovery is a topic
of great interest at the moment.
Many process discovery algorithms have been proposed over recent years.
They include Genetic algorithms, HeuristicMiner, Fuzzy miner, Inductive miner,
the algorithms based on integer linear programming (ILP), and on the theory of
regions (see [3] for a comprehensive review). They can be applied to solve typical
problems of event logs, e.g. incompleteness and noise.
Within multi-agent systems (MAS), models obtained by existing process dis-
covery tools can be incomprehensible since concurrent interacting agents pro-
duce rather sophisticated behavior. Process models of MAS, obtained by the
(cid:63) This work is supported by MIUR, Basic Research Program at the National Research
University Higher School of Economics, and Russian Foundation for Basic Research,
project No.16-01-00546.
algorithms mentioned above, are not structured in such a way that it is possible
to identify agent models as components as shown by the following example.
In order to give an intuition behind our approach, consider the system model
shown in Fig. 1(a). Two interacting agents can be clearly identified as well as
the way they interact by looking at the small grey places and arcs. Now take the
event log produced by this system as the input to a process discovery algorithm.
If we try to discover a model from this log directly, we can obtain, for instance,
the results shown in Fig. 1(b) by inductive miner and in Fig. 1(c) by ILP miner.
Although equivalent to the original one in their ability to reproduce the same
event sequences, their structure hides the fact the original system is made of two
agents, communicating through channels. More technically, the two agents cor-
respond to two S-invariants in the original model, while they are not "separable"
by means of S-invariants in the discovered models. It is true that we can improve
the overall structure of models by configuring algorithm parameters. However,
they still will not reflect the underlying MAS organization.
(a) an intitial system model
(b) inductive miner
(c) ILP miner
Fig. 1. Discovering process models of MAS from event logs
We propose a compositional approach to address the problem of discovering
process models of MAS from event logs clearly expressing agents as compo-
nents and their interactions. We assume that agents interact asynchronously
via message passing. System event logs are projected on each agent to discover
component models in terms of workflow nets by using existing process discovery
algorithms. By means of morphisms, we can construct abstract models of the
agents and compose them by adding channels for message passing. This com-
position models the protocol of agent interaction. In this way, it is possible to
check soundness of this simplified model and to compose the discovered agent
models on the basis of this protocol. We prove that this composition preserves
soundness of the protocol and components by construction. In this paper, we
start constructing a formal background to the general compositional approach
to discovering process models of multi-agent systems from event logs.
The problem of discovering structured models from logs is not new and has
been studied in several works based on composition. In [17] the authors have
designed a technique to discover readable models by decomposing transition
systems. A special method to deal with process cancellations has been studied
in [16]. Regular process behavior is composed with cancellations using reset arcs.
A rather large amount of literature has been devoted to Petri net compo-
sition (see, for example, [8,12,24]). In particular, regarding asynchronous com-
munication, several approaches have been proposed. Among the others, in [13]
asynchronous composition of Petri nets through channels has been studied con-
sidering preservation of various channel properties. A general approach to asyn-
chronous composition has also been discussed in [4], where open nets have been
introduced. The core problem of composition lies in preserving component prop-
erties. Open nets have been used by many researches especially focused on mod-
eling composite services (see, for example, [14,15]). They have proposed both
structural and behavioral techniques assuring the correctness of composition.
Using morphisms is another possible way to achieve inheritance of component
properties. Composition of Petri nets via morphisms has been studied in several
works [4,5,7,11,21,22,26]. We will use α-morphisms and the related composition
presented in [6]. In general, α-morphisms allow one to preserve/reflect properties
checking structural and only local behavioral constraints. In the particular case
considered here, α-morphisms preserve/reflect reachable markings and transition
firings as well as preserve soundness as it will be shown in Section 3.
From a more practical point of view, many researchers have considered work-
flow net composition. Workflow nets (WF-nets) form a class of Petri nets used
to model processes and services. Among the others, composition of WF-nets via
shared resources has been studied in [20,25] with a concern about soundness
preservation. In addition, other approaches have been used for modeling and
composing interacting workflow nets [19,23]. Many works have concerned the
web service composition (see the survey [10]), where the authors have stressed
that there is a lack of execution engines based on Petri nets.
The paper is organized as follows. Section 2 gives preliminary definitions
used in this paper as well as composition by channels and α-morphisms. Section
3 discusses properties preserved and reflected by α-morphisms, relevant to WF-
nets. Section 4 introduces a modular approach to construct models of MAS from
event logs, using the same illustrative example shown in Fig. 1. In Section 5, we
summarize the paper by discussing results and possible continuations.
x∈A •x, A• =(cid:83)
x∈A x•, •A• = •A ∪ A•.
elements.
of x ∈ X. Let A ⊆ X, then •A =(cid:83)
2 Preliminaries
N denotes the set of non-negative integers. A+ denotes the set of all finite non-
empty sequences over A, and A∗ = A+∪{}, is the empty sequence. A multiset
m over a set S is a function m : S → N. Let m1, m2 be two multisets over S.
Then m1 ⊆ m2 ⇔ m1(s) ≤ m2(s) for all s ∈ S. Also, m(cid:48) = m1 + m2 ⇔ m(cid:48)(s) =
m1(s) + m2(s), m(cid:48)(cid:48) = m1 − m2 ⇔ m(cid:48)(cid:48)(s) = max(m1(s) − m2(s), 0) for all s ∈ S.
A Petri net is a triple N = (P, T, F ), where P and T are two disjoint sets
of places and transitions, i.e. P ∩ T = ∅, and F ⊆ (P × T ) ∪ (T × P ) is a flow
relation, where dom(F )∪cod(F ) = P ∪T . We consider nets, s.t. ∀t ∈ T : •t ≥ 1
and t• ≥ 1.
Let N = (P, T, F ) be a Petri net, and X = P ∪ T . The set •x = {y ∈
X(y, x) ∈ F} denotes the preset of x ∈ X. The set x• = {y ∈ X(x, y) ∈ F}
denotes the postset of x ∈ X. The set •x• = •x ∪ x• denotes the neighborhood
By N (A) we denote a subnet of N generated by A, i.e. N (A) = (P ∩ A, T ∩
A, F ∩ (A × A)). Let N (A) be a subnet of N generated by A ⊆ X. The set
(cid:13)N (A) = {y ∈ A∃z ∈ X\A : (z, y) ∈ F or •y = ∅} denotes the input elements.
The set N (A)(cid:13) = {y ∈ A∃z ∈ X \ A : (y, z) ∈ F or y• = ∅} denotes the output
A marking (state) of a Petri net N = (P, T, F ) is a function m : P → N. It is a
multiset over a set of places P . A marked Petri net (N, m0) is a Petri net together
with its initial marking m0. A marking m enables a transition t ∈ T , denoted
m[t(cid:105), if •t ⊆ m. The firing of t at m leads to a new marking m(cid:48) = m − •t + t•,
denoted m[t(cid:105)m(cid:48).
A sequence w ∈ T ∗ is a firing sequence of N = (P, T, F, m0) if w = t1t2 . . . tn
and m0[t1(cid:105)m1[t2(cid:105) . . . mn−1[tn(cid:105)mn. Then we can write m0[w(cid:105)mn. The set of all
firing sequences of N is denoted by F S(N ).
A marking m of N = (P, T, F, m0) is reachable if ∃w ∈ F S(N ) : m0[w(cid:105)m.
The set of all markings of N reachable from m is denoted by [m(cid:105). A reachable
marking is dead if it does not enable any transition. N is deadlock-free if no
reachable marking is dead. N is safe if ∀p ∈ P ∀m ∈ [m0(cid:105) : m(p) ≤ 1. Then we
will specify reachable markings as subsets of places.
A state machine is a connected Petri net N = (P, T, F ), s.t. ∀t ∈ T : •t =
t• = 1. A subnet of a marked Petri net N = (P, T, F, m0) identified by a subset
of places A ⊆ P and its neighborhood, i.e. N (A∪•A•), is a sequential component
of N if it is a state machine and has a single token in the initial marking. N
is covered by sequential components if every place of N belongs to at least one
sequential component. Then N is said to be state machine decomposable (SMD).
Semantics of a marked Petri net is given by its unfolding as defined below.
Let N = (P, T, F ) be a Petri net, and F ∗ be the reflexive transitive closure
of F. Then ∀x, y ∈ P ∪ T : (a) x and y are in causal relation, denoted x < y, if
(x, y) ∈ F ∗; (b) x and y are in conflict relation, denoted x#y, if ∃tx, ty ∈ T , s.t.
tx (cid:54)= ty, •tx ∩ •ty (cid:54)= ∅, and tx < x, ty < y.
Definition 1. A Petri net O = (B, E, F ) is an occurrence net if:
1. ∀b ∈ B : •b ≤ 1.
2. F ∗ is a partial order.
3. ∀x ∈ B ∪ E : {y ∈ B ∪ Ey < x} is finite.
4. ∀x, y ∈ B ∪ E : x#y ⇒ x (cid:54)= y.
By definition, O is acyclic. M in(O) denotes the set of minimal nodes of O
w.r.t. F ∗, i.e. the elements having the empty preset. We only consider nets having
transitions with non-empty presets and postsets, then M in(O) ⊆ B.
Definition 2. Let N = (P, T, F, m0) be a marked safe Petri net, O = (B, E, F )
be an occurrence net, and π : B ∪ E → P ∪ T be a map. (O, π) is a branching
process of N if:
1. π(B) ⊆ P and π(E) ⊆ T .
2. π restricted to Min(O) is a bijection between Min(O) and m0.
3. ∀t ∈ T : π restricted to •t is a bijection between •t and •π(t), and similarly
4. ∀t1, t2 ∈ T :
if •t1 = •t2 and π(t1) = π(t2), then t1 = t2.
for t• and π(t)•.
The unfolding of N , denoted U(N ), is the maximal branching process of N ,
s.t. any other branching process of N is isomorphic to a subnet of U(N ) with
the map π restricted to the elements of this subnet. The map associated with
the unfolding is denoted u and called folding.
Workflow nets form a subclass of Petri nets used for process modeling. We
define generalized workflow nets having an initial state m0 and a final state mf .
Definition 3. A marked Petri net N = (P, T, F, m0, mf ) is a generalized work-
flow net (GWF-net) if and only if:
1. m0 = {s ∈ P •s = ∅} and m0 (cid:54)= ∅.
2. mf = {f ∈ P f• = ∅} and mf (cid:54)= ∅.
3. ∀x ∈ P ∪ T ∃s ∈ m0 ∃f ∈ mf : (s, x) ∈ F ∗ and (x, f ) ∈ F ∗.
If m0 = mf = 1, then a generalized workflow net is called just a workflow
net (WF-net, for short). State machine decomposable GWF-nets are safe. The
important correctness property of GWF-nets is soundness [2].
Definition 4. A GWF-net N = (P, T, F, m0, mf ) is sound if and only if:
1. ∀m ∈ [m0(cid:105) : mf ∈ [m(cid:105).
2. ∀m ∈ [m0(cid:105) : mf ⊆ m ⇒ m = mf .
3. ∀t ∈ T ∃m ∈ [m0(cid:105) : m[t(cid:105).
We consider composition of two GWF-nets by adding channels. A set of
channel places (channels, for short) is denoted by Pc. This approach is similar to
the one presented in [13]. They model asynchronous communication via message
passing. Some transitions of two nets can either send a message by an incoming
arc to a channel or receive a message by an outgoing arc from a channel. We
assume to know exactly which transitions send/receive messages to/from which
channel places. In order to simplify the notation, we will not introduce special
transition labels indicating sending/receiving.
Two GWF-nets can be composed via a set of channels Pc iff any channel
receiving a message from one GWF-net send it only to the other GWF-net.
0, mi
f ∪ m2
f .
Definition 5. Let Ni = (Pi, Ti, Fi, mi
f ) be a GWF-net for i = 1, 2, s.t. N1
and N2 are disjoint, where P1 ∩ P2 = ∅ and T1 ∩ T2 = ∅. Let Pc be a set of
channels. A channel-composition of N1 and N2, denoted N1 ⊕Pc N2, is a Petri
net N = (P, T, F, m0, mf ), where:
1. P = P1 ∪ P2 ∪ Pc, where Pc ∩ (P1 ∪ P2) = ∅.
2. T = T1 ∪ T2.
3. F = F1 ∪ F2 ∪ Fc, where Fc ⊆ (Pc × (T1 ∪ T2)) ∪ ((T1 ∪ T2) × Pc).
4. m0 = m1
0 ∪ m2
5. ∀p ∈ Pc :
0 and mf = m1
(a) (•p ⊆ T1 or •p ⊆ T2) and (p• ⊆ T1 or p• ⊆ T2),
(b) •p ⊆ Ti ⇔ p• ⊆ T(i+1) mod 2,
(c) •p (cid:54)= ∅ and p• (cid:54)= ∅.
Remark 1. The operation ⊕Pc is commutative, i.e. N1 ⊕Pc N2 = N2 ⊕Pc N1.
By the following proposition, the class of GWF-nets is closed under the
channel-composition. Figure 2(a) provides an example of channel-composition
of N1 and N2, where channels are indicated by small gray places.
Proposition 1. If N1 and N2 are GWF-nets, then N1 ⊕Pc N2 is a GWF-net.
Proof. We show that each channel lies on a path from a place in mi
0 to a place
in mj
f , where j = (i + 1) mod 2 and i = 1, 2. Take p ∈ Pc. By Def. 5.5, •p ⊆
Ti ⇔ p• ⊆ Tj. Take ti ∈ •p and tj ∈ p•. By Def. 3.3, ∃s ∈ mi
0, s.t. (s, ti) ∈ F ∗i .
Then (s, p) ∈ F ∗. By Def. 3.3, ∃f ∈ mj
f , s.t. (tj, f ) ∈ F ∗j . Then (p, f ) ∈ F ∗. (cid:117)(cid:116)
Further, we recall the definition of α-morphisms (see Definition 6 and Defi-
nition 7 in [6]) supporting abstraction and refinement for Petri nets. An exam-
ple of α-morphism is shown in Fig. 2(b), where the refinement is given by the
shaded ovals and by the transition labeling explicitly. Refinement can also require
splitting transitions of the abstract model. After giving the formal definition of
α-morphisms, we will provide an intuition behind them.
Definition 6. Let Ni = (Pi, Ti, Fi, mi
0) be a marked SMD safe Petri net, Xi =
Pi∪Ti, i = 1, 2. An α-morphism from N1 to N2 is a total surjective map ϕ : X1 →
X2, also denoted ϕ : N1 → N2, such that:
1. ϕ(P1) = P2.
2. ϕ(m1
0) = m2
0.
if ϕ(t) ∈ T2, then ϕ(•t) = •ϕ(t) and ϕ(t•) = ϕ(t)•.
3. ∀t ∈ T1 :
if ϕ(t) ∈ P2, then ϕ(•t•) = {ϕ(t)}.
4. ∀t ∈ T1 :
5. ∀p2 ∈ P2 :
(a) N1(ϕ−1(p2)) is an acyclic net;
(b) ∀p ∈ (cid:13)N1(ϕ−1(p2)) : ϕ(•p) ⊆ •p2, and if •p2 (cid:54)= ∅, then •p (cid:54)= ∅;
(c) ∀p ∈ N1(ϕ−1(p2))(cid:13) : ϕ(p•) = p2•;
(d) ∀p ∈ P1 ∩ ϕ−1(p2) : p /∈ (cid:13)N1(ϕ−1(p2)) ⇒ ϕ(•p) = p2 and
p /∈ N1(ϕ−1(p2))(cid:13) ⇒ ϕ(p•) = p2;
(e) ∀p ∈ P1 ∩ ϕ−1(p2) :
N1, s.t. p ∈ P (cid:48) and ϕ−1(•p2•) ⊆ T (cid:48).
there is a sequential component N(cid:48) = (P (cid:48), T (cid:48), F (cid:48)) of
(a) N1 ⊕Pc N2, Pc = {x, y, z}
(b) α-morphism ϕ : N(cid:48)
2 → N2
Fig. 2. Introductory examples
By definition, α-morphisms allow us to refine places by replacing them with
subnets. Thus, if a transition is mapped on a place, then its neighborhood should
be mapped on the same place (4). If a transition is mapped on a transition, it
should have the corresponding neighborhood (3).
Indeed, α-morphisms are motivated by the attempt to ensure that properties
of an abstract model hold in its refinement. Each output place of a subnet should
have the same choices as its abstraction does (5c). However, input places do not
need this constraint (5b), because the choice between them is done before, since
there are no concurrent events in the neighborhood of the subnet (5e). Moreover,
5d guarantees that presets and postsets of places, internal to the subnet, are
mapped to the same place as the subnet. Conditions 5a-5e together ensure the
intuition behind α-morphisms. If a subnet of N1 refines a place in N2, then it
behaves "in the same way". More precisely, by Lemma 1 of [6], (a) no input
transition of the subnet is enabled whenever a token is inside the subnet; and
(b) firing an output transition of the subnet empties it.
3 Properties Preserved and Reflected by α-Morphisms
In this section, we study properties preserved and reflected by α-morphisms. In
[6] several properties of α-morphisms have already been studied. Here we will
mention some of them and consider properties of α-morphisms for GWF-nets.
In the following propositions, we assume Ni = (Pi, Ti, Fi, mi
0) to be a marked
SMD safe Petri net for i = 1, 2, s.t. there is an α-morphism ϕ : X1 → X2, where
Xi = Pi ∪ Ti.
To begin with, α-morphisms preserve the structure of GWF-nets.
Proposition 2. If N1 is a GWF-net, then N2 is a GWF-net.
Proof. We prove that N2 satisfies three structural conditions of Def. 3.
1. By Def. 6.2, ϕ(m1
0. Suppose ∃p2 ∈ m2
2. By Def. 3.2, m1
f ) by m2
f : p• = ∅.
0 : •p2 (cid:54)= ∅. By Def. 6.5b,
0, s.t. ϕ(p1) = p2.
0 : •p = ∅. Then,
3. Suppose ∃x2 ∈ X2, s.t. ∀p ∈ m2
1, . . . , xk
f and p1• = ∅ (by Def. 3.2). Then p2• = ∅ and ∀p ∈ m2
0) = m2
if •p2 (cid:54)= ∅, then •p (cid:54)= ∅. Take p1 ∈ m1
∀p ∈ (cid:13)N1(ϕ−1(p2)) :
Since p1 ∈ (cid:13)N1(ϕ−1(p2)), then •p1 (cid:54)= ∅. By Def. 3.1, ∀p ∈ m1
•p2 = ∅ and ∀p ∈ m2
0 : •p = ∅.
f : p• = ∅. Denote ϕ(m1
f ⊆ P1, s.t. ∀p ∈ m1
f ⊆ P2.
f : p2• (cid:54)= ∅. Take p1 ∈ m1
Suppose ∃p2 ∈ m2
f , s.t. ϕ(p1) = p2. By Def. 6.5c,
∀p ∈ N (ϕ−1(p2))(cid:13) : ϕ(p•) = p2•. Since p1 ∈ N (ϕ−1(p2))(cid:13), then p1• (cid:54)= ∅. But
p1 ∈ m1
0 : (p, x2) /∈ F ∗2 . By Def. 6, ϕ−1(x2) =
1} ⊆ X1. If x2 ∈ T2, then ϕ−1(x2) ⊆ T1, and take x1 ∈ ϕ−1(x2). If
{x1
0 : (s, x1) ∈ F ∗1 .
x2 ∈ P2, then take x1 ∈ (cid:13)N (ϕ−1(x2)). By Def. 3.3, ∃s ∈ m1
Then, ϕ(•x1) ∈ •x2 or ϕ(•x1) = x2. We come backward through the whole path
from s to x1 mapping it on N2 with ϕ. We obtain that ∃x(cid:48) ∈ X2 : (x(cid:48), x2) ∈ F ∗2 ,
s.t. ϕ(s) = x(cid:48).
f : (x2, p) /∈ F ∗2 . By Def. 6, ϕ−1(x2) =
1} ⊆ X1. If x2 ∈ T2, then ϕ−1(x2) ⊆ T1, and take x1 ∈ ϕ−1(x2).
{x1
If x2 ∈ P2, then take x1 ∈ N (ϕ−1(x2))(cid:13). By Def. 3.3, ∃f ∈ m1
f : (x1, f ) ∈ F ∗1 .
Then, ϕ(x1•) ∈ x2• or ϕ(x1•) = x2. We come forward through the whole path
from x1 to f mapping it on N2 with ϕ. We obtain that ∃x(cid:48) ∈ X2 : (x2, x(cid:48)) ∈ F ∗2 ,
s.t. ϕ(f ) = x(cid:48).
(cid:117)(cid:116)
Remark 2. It follows from Proposition 2 that ϕ(m1
f . In the general case
the converse of Proposition 2 is not true. In fact, α-morphisms do not reflect the
initial state of GWF-nets properly (see Fig. 3(a)).
Suppose ∃x2 ∈ X2, s.t. ∀p ∈ m2
1, . . . , xk
f ) = m2
Recall from [6] that N1 is well marked w.r.t. ϕ if each input place of a subnet
in N1, refining a marked place in N2, is marked. Consider the α-morphism shown
in Fig. 3(a), the token of the shaded subnet must be placed into p to make N1
well marked w.r.t. to ϕ. The structure of GWF-nets is reflected under the well-
markedness of N1 (see the following Proposition). However, if N1 is a GWF-net
related to N2 by an α-morphism, then N1 is well marked w.r.t. ϕ.
(a)
(b)
Fig. 3. An α-morphism: two examples
Proposition 3. If N2 is a GWF-net and N1 is well marked w.r.t. ϕ, then N1
is a GWF-net.
Proof. We prove that N1 satisfies three structural conditions of Def. 3.
1. By Def. 3.1, ∀s2 ∈ m2
0}. Take N1(ϕ−1(s2)) corresponding to s2 ∈ m2
f} and ∀f1 ∈ m1
3. Suppose ∃x1 ∈ X1, s.t. ∀s1 ∈ m1
0 : •s2 = ∅. Since N1 is well marked w.r.t. ϕ, m1
0 =
{(cid:13)N1(ϕ−1(s2)) s2 ∈ m2
0. Assume
∃p ∈ (cid:13)N1(ϕ−1(s2)), s.t. •p (cid:54)= ∅. Note that ϕ(p) = s2. Then ϕ(•p) = s2 (by Def.
6.4), and p /∈ (cid:13)N1(ϕ−1(s2)).
f : f2• = ∅. Take N1(ϕ−1(f2)) corresponding to
2. By Def. 3.2, ∀f2 ∈ m2
f and p ∈ N1(ϕ−1(f2))(cid:13). Note that ϕ(p) = f2. Assume p• (cid:54)= ∅.
f2 ∈ m2
Then ϕ(p•) = p2 (by Def. 6.4) and p /∈ N1(ϕ−1(f2))(cid:13). We obtain that m1
f =
f : f1• = ∅.
{N1(ϕ−1(f2))(cid:13) f2 ∈ m2
0 : (s1, x1) /∈ F ∗1 . If (x1, x1) /∈ F ∗1 , we
go backward through the path from x1 to the first node x(cid:48)1 ∈ X1, s.t. •x(cid:48)1 = ∅.
Since ∀t ∈ T1 : •t ≥ 1, then x(cid:48)1 ∈ P1. If x(cid:48)1 /∈ m1
0, then N1 is not well marked
w.r.t. ϕ. If (x1, x1) ∈ F ∗1 , we know by Def. 6.5a that there is a corresponding
0 : (s2, x2) ∈ F ∗2 .
cycle in N2. Take x2 ∈ X2, s.t. ϕ(x1) = x2. By Def. 3.3, ∃s2 ∈ m2
Take x(cid:48)2 ∈ X2 belonging to this cycle, s.t. at least one element in •x(cid:48)2 is not in
the cycle. Since ϕ is surjective, ∃x(cid:48)1 ∈ X1 : ϕ(x(cid:48)1) = x(cid:48)2, belonging to the cycle
(x1, x1) ∈ F ∗1 . If x(cid:48)2 ∈ T2, then ϕ−1(x(cid:48)2) ⊆ T1. By Def. 6.3, the neighborhood of
transitions is preserved by ϕ. Then, ∀t1 ∈ ϕ−1(x(cid:48)2) : ϕ(•t1) = •x(cid:48)2, i.e. there must
be a place in •ϕ−1(x(cid:48)2) which is not in the cycle (x1, x1) ∈ F ∗1 . If x(cid:48)2 ∈ P2, then
take (cid:13)N1(ϕ−1(x(cid:48)2)). At least one place in (cid:13)N1(ϕ−1(x(cid:48)2)) should have an input
transition which is not in the cycle (x1, x1) ∈ F ∗1 , since there exists an element
in •x(cid:48)2 which is not in the cycle in N2. We have shown that ∃x ∈ •x(cid:48)1, s.t. x is
not in the cycle (x1, x1) ∈ F ∗1 . Now either there is a path from x to x∗ in N1,
s.t. •x∗ = ∅, or there is another cycle (x∗, x∗) ∈ F ∗1 .
f : (x1, f1) ∈
F ∗1 . The only difference is that we need to go forward through the paths.
(cid:117)(cid:116)
We recall in the following proposition that α-morphisms preserve reachable
By using a similar reasoning, we prove that ∀x1 ∈ X1 ∃f1 ∈ m1
markings and firing of transitions.
0(cid:105). If m1[t(cid:105)m(cid:48)1, where
0(cid:105). Then ϕ(m1) ∈ [m2
Proposition 4 ([6]). Let m1 ∈ [m1
t ∈ T1, then:
1. ϕ(t) ∈ T2 ⇒ ϕ(m1)[ϕ(t)(cid:105)ϕ(m(cid:48)1).
2. ϕ(t) ∈ P2 ⇒ ϕ(m1) = ϕ(m(cid:48)1).
Remark 3 ([6]). In the general case α-morphisms do not reflect reachable mark-
0(cid:105) and m2[t2(cid:105) do not imply that there exists m1 = ϕ−1(m2) ∈
ings, i.e. m2 ∈ [m2
[m1
0(cid:105), s.t. ∀t ∈ ϕ−1(t2) : m1[t(cid:105).
To reflect reachable markings, we need to check local conditions as shown
in [6]. For any place p2 ∈ P2, refined by a subnet of N1, we construct a "local"
net, denoted S2(p2), of N2 by taking the neighborhood transitions of p2 plus
an artificial input and output place if needed. At the same time, we construct
the corresponding "local" net, denoted S1(p2), of N1 by taking the subnet of
N1 refining p2, i.e. N1(ϕ−1(p2)), and the transitions ϕ−1(•p2), ϕ−1(p2•) plus
an artificial input and output place if needed. The details are given in [6] (see
Definition 9 there).
There is an α-morphism ϕS from S1(p2) to S2(p2) which is a restriction of
ϕ on the places and transitions of S1(p2). In the following Lemma, taking the
unfolding of S1(p2), we assure that the final marking of the subnet enables the
same set of transitions which are enabled by its image. By Definition 6.5a, since
S1(p2) is acyclic, its unfolding is finite.
Lemma 1. Let U(S1(p2)) be the unfolding of S1(p2) with the folding function u,
and ϕS be an α-morphism from S1(p2) to S2(p2). Let N1 be a sound GWF-net.
Then, the map from U(S1(p2)) to S2(p2) obtained as ϕS ◦ u is an α-morphism.
Proof. Since N1 is a GWF-net, S1(p2) is a GWF-net. By [6] (see Lemma 1 there),
when a transition in ϕ−1(p2•) fires, it empties the subnet N1(ϕ−1(p2)). Then
S1(p2) is sound, and by Def. 4.3 ∀t ∈ T1∃m ∈ [m1
0(cid:105) : m[t(cid:105). So, each transition of
S1(p2) will occur at least once. Then, the folding u is a surjective function from
U(S1(p2)) to S1(p2), and ϕS ◦ u is an α-morphism from U(S1(p2)) to S2(p2). (cid:117)(cid:116)
As for the main results, we obtain that under the assumption of soundness
of the GWF-net N1, α-morphisms not only preserve, but also reflect reachable
markings and transition firings (see Proposition 5). Moreover, α-morphisms pre-
serve soundness as shown in Proposition 6.
Proposition 5. If N1 is a sound GWF-net, then ∀m2 ∈ [m2
ϕ(m1) = m2, and if ∃t2 ∈ T2 : m2[t2(cid:105), then ∀t1 ∈ ϕ−1(t2) : m1[t1(cid:105).
Proof. Follows from Lemma 1 and from [6] (see Proposition 5 there).
0(cid:105)∃m1 ∈ [m1
0(cid:105) :
Proposition 6. If N1 is a sound GWF-net, then N2 is a sound GWF-net.
Proof. We prove that N2 satisfies three behavioral conditions of Def. 4.
f ∈ [m1(cid:105). Then, ∃w ∈ T ∗1 : m1[w(cid:105)m1
[tn(cid:105)m1
f , i.e. w =
f . By Prop. 4, we can simulate w on N2,
1. By Def. 4.1, ∀m1 ∈ [m1
t1t2 . . . tn and m1[t1(cid:105)m1
0(cid:105) : m1
1 . . . mn−1
1
0(cid:105) : m2
f ∈ [m2(cid:105).
f . Then we can write m(cid:48)2 = m2
0(cid:105), s.t. ϕ−1(m1) = m(cid:48)2 and m1
f ) = m2
f . Now assume ∃m2 ∈ [m2
2. Suppose ∃m(cid:48)2 ∈ [m2
0(cid:105) : m(cid:48)2 ⊇ m2
and by Prop. 2, ϕ(m1
∃m(cid:48)1 ∈ [m1
By Prop. 4, we again simulate w(cid:48) on N2. Then m2
0(cid:105) : ϕ−1(m2) = m(cid:48)1. By Def. 4.1, m1
f . By Prop. 5, take m1 ∈ [m1
1 ∈ [m1(cid:105), i.e. ∃w ∈ T ∗1 : m1[w(cid:105)mf
∀p ∈ P (cid:48) : p /∈ m2
By Def. 4.1, mf
N2. By Prop. 2, ϕ(m1
is to consume at least one token from m2
contradiction.
f /∈ [m2(cid:105). By Prop. 5,
f ∈ [m(cid:48)1(cid:105), i.e. ∃w(cid:48) ∈ T ∗ : m(cid:48)1[w(cid:48)(cid:105)m1
f .
f ∪ P (cid:48), where
f (cid:54)⊆ m1.
1 . By Prop. 4, we simulate w on
f . The only way to completely empty places in P (cid:48)
f : p• (cid:54)= ∅ which is a
0(cid:105) : m1[t1(cid:105). The map ϕ is surjective, i.e.
∀t2 ∈ T2 : ∃t1 ∈ T1 : ϕ(t1) = t2. By Prop. 4, m1[t1(cid:105)m(cid:48)1 ⇒ ϕ(m1)[ϕ(t1)(cid:105)ϕ(m(cid:48)1).
Then, ∀t2 ∈ T2 ∃m2 ∈ [m2
(cid:117)(cid:116)
Remark 4. In the general case the converse of Proposition 6 is not true. Consider
the example shown in Fig. 3(b), where N2 is sound and N1 is not sound, since
transitions y1 and y2 are dead. Thus, α-morphisms do not reflect soundness.
Note that reachable markings are also not reflected in this example. However, in
the next section we will provide conditions under which soundness is reflected.
3. By Def. 4.3, ∀t1 ∈ T1 ∃m1 ∈ [m1
f ) = m2
f . Then ∃p ∈ m2
0(cid:105) : m2[t2(cid:105).
4 From Event Logs to Structured and Sound Models of
Multi-Agent Systems
In this section, we present our approach to process discovery by composing
individual agent models through α-morphisms. An event log L is a finite multiset
of finite non-empty sequences (traces) over a set of observable actions A.
We assume to have an event log L of two interacting agents. For instance,
we will further work with the same event log obtained from the MAS shown in
Fig. 1(a) which we have used in Section 1. We assume to know what actions are
executed by which agent, A = A1∪A2, s.t. A1∩A2 = ∅. Also, we assume to know
actions corresponding to their asynchronous "message-passing" interaction.
Instead of discovering the model directly from L, we propose to filter the log
according to the agent actions A1 and A2 producing two new logs L1 and L2.
Traces of L1 and L2 contain actions done only by a corresponding agent. By
using, for example, inductive miner [18], from L1 and L2 we obtain two GWF-
nets N(cid:48)1 and N(cid:48)2 modeling the two agents. By construction, N(cid:48)1 and N(cid:48)2 are well-
structured and sound, which implies that they are state machine decomposable
(see Corollary 4 in [1]). Well-structured models are recursively built from blocks
representing basic control flow constructs, e.g. choice, concurrency or cycle.
It is possible to compose N(cid:48)1 and N(cid:48)2 using the channel-composition as in
Definition 5 obtaining a new GWF-net N(cid:48)1 ⊕Pc N(cid:48)2, where channels are defined
according to the specification of agent interaction we have assumed to know.
However, it is obvious that N(cid:48)1 ⊕Pc N(cid:48)2 might not be sound. In order to avoid the
verification of N(cid:48)1 ⊕Pc N(cid:48)2, we apply the following approach to get its soundness
by construction.
(a) F c
= F ∩ [((Pc ∪ P2) × T2) ∪ (T2 × (Pc ∪ P2))];
N2
1 ⊆ (Pc × T (cid:48)1) ∪ (T (cid:48)1 × Pc).
(b) F c
N(cid:48)
4. m(cid:48)0 = m(cid:48)01 ∪ m2
5. ∀p ∈ Pc, ∀t ∈ T1 :
0 and m(cid:48)f = m(cid:48)f 1 ∪ m2
f .
We can abstract the discovered nets N(cid:48)1 and N(cid:48)2 w.r.t. interacting actions, thus
producing two GWF-nets N1 and N2, s.t. there is an α-morphism ϕi : N(cid:48)i → Ni
for i = 1, 2. According to Proposition 6, Ni is sound. The abstract models
can also be composed via the same channels obtaining N1 ⊕Pc N2. This abstract
model represents the interaction protocol between the agents. Obviously, N1⊕Pc
N2 is less complex than N(cid:48)1⊕Pc N(cid:48)2, and its soundness can be much easier verified.
Given N1 ⊕Pc N2 and the two α-morphisms ϕi : N(cid:48)i → Ni for i = 1, 2, we
can construct two new GWF-nets: N(cid:48)1 ⊕Pc N2 and N1 ⊕Pc N(cid:48)2, which actually
represent different abstractions of the same MAS. This construction is formally
defined in the following definition. We also show in Remark 5 that there is an
α-morphism from N(cid:48)1⊕Pc N2 (by symmetry, from N1⊕Pc N(cid:48)2) towards N1⊕Pc N2.
Definition 7. Let Ni = (Pi, Ti, Fi, mi
f ) be a GWF-net for i = 1, 2, and Pc
be a set of channels. Let N1⊕Pc N2 = (P, T, F, m0, mf ) be a channel-composition
of N1 and N2. Let N(cid:48)1 = (P (cid:48)1, T (cid:48)1, F (cid:48)1, m(cid:48)01, m(cid:48)f 1) be a GWF-net, s.t. there is an
α-morphism ϕ1 : N(cid:48)1 → N1. Construct N(cid:48)1 ⊕Pc N2 = (P (cid:48), T (cid:48), F (cid:48), m(cid:48)0, m(cid:48)f ), where:
1. P (cid:48) = P (cid:48)1 ∪ P2 ∪ Pc.
2. T (cid:48) = T (cid:48)1 ∪ T2.
3. F (cid:48) = F (cid:48)1 ∪ F c
N2 ∪ F c
N(cid:48)
, where:
0, mi
1
((t, p) ∈ F ⇒ ϕ−1(t) × {p} ⊆ F c
N(cid:48)
1
) and ((p, t) ∈ F ⇒ {p} × ϕ−1(t) ⊆ F c
N(cid:48)
1
).
Remark 5. Let N1, N(cid:48)1 and N2 be GWF-nets, and ϕ1 : N(cid:48)1 → N1 be an α-
morphism. Then there is an α-morphism ϕ(cid:48)1 : N(cid:48)1 ⊕Pc N2 → N1 ⊕Pc N2. In fact,
by construction, ϕ(cid:48)1 is given by ϕ1 plus the identity mapping of places and tran-
sitions of N2 together with the identity mapping of channel places.
Example 1. Here we consider as N1 ⊕Pc N2 the GWF-net shown in Fig.2(a). We
can refine it by models N(cid:48)1 or N(cid:48)2, discovered from filtered event logs L1 and L2
as shown in Fig. 4. The α-morphisms between the discovered models and the
abstract ones are indicated by the shaded ovals and the transition labeling. The
α-morphism between N(cid:48)2 and N2 has also been shown in Fig. 2(b).
As for the main result, we will prove in Proposition 7 that an α-morphism
from N(cid:48)1 ⊕Pc N2 (by symmetry, from N1 ⊕Pc N(cid:48)2) to N1 ⊕Pc N2 reflects the
soundness of N1⊕Pc N2. To prove this fact, we will use the property of reachable
markings of a channel-composition stated in the following Lemma.
Lemma 2. Let Ni = (Pi, Ti, Fi, mi
(P, T, F, m0, mf ). Then, ∀m ∈ [m0(cid:105) : m = m1∪ mc∪ m2, where m1 ∈ [m1
[m2
f ) a GWF-net for i = 1, 2, N1 ⊕Pc N2 =
0(cid:105), m2 ∈
0, mi
0(cid:105), and mc ⊆ Pc.
Proof. By Def. 5.4, m0 = m1
f . Take m ∈ [m0(cid:105), then
∃w ∈ F S(N1 ⊕Pc N2) : m0[w(cid:105)m, where w ∈ T ∗. By Def. 5.2, T = T1 ∪ T2.
Restricting w to T1 and T2 produces two firing sequences of N1 and N2 leading
from m1
0 to the reachable markings m1 and m2 which constitute m. (cid:117)(cid:116)
0 and mf = m1
f ∪ m2
0 ∪ m2
0 and m2
(a) N(cid:48)
1 ⊕Pc N2
(b) N1 ⊕Pc N(cid:48)
2
Fig. 4. Refining N1 ⊕Pc N2 with agent nets N(cid:48)
1 and N(cid:48)
2 discovered from filtered logs
Proposition 7. Let N1, N(cid:48)1 and N2 be sound GWF-nets, and ϕ1 : N(cid:48)1 → N1 be
an α-morphism. If N1 ⊕Pc N2 is sound, then N(cid:48)1 ⊕Pc N2 is sound.
Proof. By Rem. 5, there is an α-morphism ϕ(cid:48)1 from N(cid:48)1 ⊕Pc N2 to N1 ⊕Pc N2.
Assume N(cid:48)1 ⊕Pc N2 = (P (cid:48), T (cid:48), F (cid:48), m(cid:48)0, m(cid:48)f ) and N1 ⊕Pc N2 = (P, T, F, m0, mf ).
We prove that N(cid:48)1 ⊕Pc N2 satisfies the three conditions of soundness of Def. 4.
1. Take m(cid:48) ∈ [m(cid:48)0(cid:105). By Lemma 2, m(cid:48) = m(cid:48)1 ∪ m2 ∪ mc. By Prop. 4 for ϕ(cid:48)1,
ϕ(cid:48)1(m(cid:48)) = m ∈ [m0(cid:105). By Lemma 2, m = m1 ∪ m2 ∪ mc, where m2 and mc are
the same as in m(cid:48) and ϕ1(m(cid:48)1) = m1 (by Prop. 4 for ϕ1). Since N1 ⊕Pc N2 is
sound, ∃w ∈ F S(N1 ⊕Pc N2) : m[w(cid:105)mf , where w ∈ T ∗. It is possible to write
2v, where v = or v = t1
w = w1
1 ∈ T1, s.t. i ≥ 1.
2 can be obviously executed on the component N2 of N(cid:48)1 ⊕Pc N2 as well,
Then wi
because ϕ(cid:48)1 reflects the connection to channels (by Def. 7.5 and Rem. 5). Since
N(cid:48)1 is sound, ϕ1 reflects reachable markings and firing of transitions (by Prop. 5)
between N(cid:48)1 and N1. Thus, there is a reachable marking m(cid:48)1i in N(cid:48)1, belonging
to ϕ−1
1) is enabled
at m(cid:48)1i in N(cid:48)1. Moreover, these transitions are enabled in N(cid:48)1 ⊕Pc N2, because ϕ(cid:48)1
reflects connection to channels (by Def. 7.5 and Rem. 5). Hence, the sequence
w ∈ F S(N1 ⊕Pc N2) is reflected in N(cid:48)1 ⊕Pc N2 reaching its final marking m(cid:48)f .
m(cid:48)f = m(cid:48)f 1 ∪ m2
In this way, ϕ(cid:48)1(m(cid:48)) = ϕ(cid:48)1(m(cid:48)f 1) ∪ ϕ(cid:48)1(m2
2. Suppose by contradiction ∃m(cid:48) ∈ [m(cid:48)0(cid:105) : m(cid:48) ⊇ m(cid:48)f and m(cid:48) (cid:54)= m(cid:48)f . By Def. 7,
f ∪ m3. By Prop. 4 for ϕ(cid:48)1, ϕ(cid:48)1(m(cid:48)) ∈ [m0(cid:105).
f ∪ m3 =
1(cid:105) in N1, then any transition in ϕ−1
f ) ∪ ϕ(cid:48)1(m3) = ϕ1(m(cid:48)f 1) ∪ m2
f . Then m(cid:48) = m(cid:48)f 1 ∪ m2
2 ∈ T ∗2 and ti
1) in N1, s.t. if mi
1w2
2t2
1 . . . with wi
1 (ti
1 (mi
1[ti
m1
f ∪ m2
f ∪ m3 = mf ∪ m3 (by Rem. 5 and Def. 7.4). Thus, this marking strictly
covers the final marking mf of N1 ⊕Pc N2 which contradicts its soundness.
3. We prove that ∀t(cid:48) ∈ T (cid:48) ∃m(cid:48) ∈ [m(cid:48)0(cid:105) : m(cid:48)[t(cid:48)(cid:105). By Lemma 2, m(cid:48) = m(cid:48)1 ∪ m2 ∪
mc. By Def. 7.2, ∀t(cid:48) ∈ T (cid:48) : t(cid:48) ∈ T (cid:48)1 or t(cid:48) ∈ T2. If t(cid:48) ∈ T2, then since N1 ⊕Pc N2
is sound, ∃m ∈ [m0(cid:105), s.t. m[t(cid:48)(cid:105). By Rem. 5 and by Def. 7.5, m2 ∪ mc enables
t(cid:48). If t(cid:48) ∈ T (cid:48)1, then there are two cases. If ϕ(cid:48)1(t(cid:48)) ∈ P , then t(cid:48) is not connected
to channels. Since N(cid:48)1 is sound, m(cid:48)1 ⊆ m(cid:48) enables t(cid:48). If ϕ(cid:48)1(t(cid:48)) ∈ T , then take
t ∈ T , s.t. ϕ(cid:48)1(t(cid:48)) = t. Since N1⊕Pc N2 is sound, ∃m ∈ [m0(cid:105) : m[t(cid:105). By Prop. 5, ϕ1
reflects reachable markings and firings of transitions on N(cid:48)1. Moreover, ϕ(cid:48)1 reflects
connection to channels (by Def. 7.5 and Rem. 5). Then m(cid:48)1 ∪ mc enables t(cid:48). (cid:117)(cid:116)
Since there are two α-morphisms from N1 ⊕Pc N(cid:48)2 and N(cid:48)1 ⊕Pc N2 towards
N1 ⊕Pc N2, we can compose them by using the composition defined in [6] (see
Definition 12) obtaining as a result N with other two α-morphisms from N
towards N1⊕Pc N(cid:48)2 and N(cid:48)1⊕Pc N2, s.t. the diagram shown in Fig. 5(b) commutes.
Alternatively, by applying a similar construction as the one given in Defi-
nition 7, we can refine N2 in N(cid:48)1 ⊕Pc N2 by N(cid:48)2 obtaining N(cid:48)1 ⊕Pc N(cid:48)2 which is
isomorphic to the previously obtained composition N . Symmetrically, it is pos-
sible to refine N1 in N1 ⊕Pc N(cid:48)2 by N(cid:48)1 coming to the same result. According to
Remark 5, there are α-morphisms from N(cid:48)1⊕Pc N(cid:48)2 to N(cid:48)1⊕Pc N2 and to N1⊕Pc N(cid:48)2.
According to Proposition 7, N(cid:48)1 ⊕Pc N2 (by symmetry, N1 ⊕Pc N(cid:48)2) is sound, and
then N(cid:48)1 ⊕Pc N(cid:48)2 is also sound.
Thus, we have shown that it is possible to even simultaneously refine N1 and
N2 in the sound abstract model N1⊕Pc N2 by sound models N(cid:48)1 and N(cid:48)2 obtained
from filtered event logs, s.t. the result N(cid:48)1 ⊕Pc N(cid:48)2 is also sound.
Example 2. In Fig. 5(a), we show the result of composing, by means of α-
morphisms, N(cid:48)1 ⊕Pc N2 and N1 ⊕Pc N(cid:48)2 constructed in Example 1. This com-
position can also be obtained directly by the channel-composition of N(cid:48)1 and N(cid:48)2.
The obtained model meets the desired requirement stated in Section 1: we can
identify the agents as subnets and explicitly see their asynchronous interaction.
(a) composition N isomorphic to N(cid:48)
1 ⊕Pc N(cid:48)
2
(b) diagram
Fig. 5. Composition of N(cid:48)
1 ⊕Pc N2 and N1 ⊕Pc N(cid:48)
2 based on α-morphisms
N1⊕PcN2N01⊕PcN2N1⊕PcN02N 4.1 Comparing Quality of Presented MAS Process Models
Process models of MAS presented in Fig. 1(b) and in Fig. 1(c) are discovered
directly from the log produced by simulating the model shown in Fig. 1(a). The
composition shown in Fig. 5(a), in fact, is obtained from the same log. That is
why, we have compared their quality using the standard process discovery metrics
[9]. Fitness measures how accurately a model can replay traces of an initial event
log. Intuitively, precision indicates a ratio between the behavior given by the log
and the one allowed by the model. If it is low, then a model allows for too much
additional behavior. Table 1 provides the result of this quality analysis.
Table 1. Fitness and precision of MAS models presented in the paper
Model
Algorithm
Discovery Fitness Precision
Figure 1(b) Inductive miner Direct
Figure 1(c) ILP miner
Direct
Figure 5(a) Inductive miner Composed
1,0000
1,0000
1,0000
0,1732
0,8516
0,8690
The result of the compositional process discovery shows the increase in preci-
sion resulting from the separate analysis of agent behavior. However, we can see
that the precision of the composed model is close to that of the model obtained
by ILP miner. Thus, the model can be appropriate in terms of precision, but not
in terms of the MAS structure.
5 Conclusion and Future Work
In this paper, we have proposed a compositional approach to discover process
models of multi-agent systems from event logs. We have considered asynchronous
agent interactions which are modeled by a channel-composition of two nets. We
assume that events can be partitioned, s.t. system event logs are filtered accord-
ing to agent behavior to discover their detailed models. In order to guarantee
that their composition is sound, we have proposed to abstract agent models,
by using α-morphisms, w.r.t. interacting actions thus constructing an abstract
model of a communication protocol. The general algorithm of constructing α-
morphisms is one of the open problems which is a subject for further research.
We have proven that, when this abstract protocol is sound, the direct composi-
tion of the detailed models is also sound. In proving this fact, we have used two
intermediate models corresponding to a detailed view of one agent composed
with the abstract view of the other.
The obtained system model is structured in such a way that it is possible to
identify agent models as components and their interactions are clearly expressed.
We have compared the quality of directly obtained models with the quality of
models obtained by the proposed approach. The quality of a composed model
is seen to be at the appropriate level in terms of fitness and precision being the
general process discovery quality dimensions. In the future, we suggest intro-
ducing a structural indicator showing the extent to which it is easy to identify
agents as parts of the model as well as their interaction.
We plan to explore more general asynchronous interactions, e.g. when the
same channel can be used by several agents as well as the possibility to have
both asynchronous and synchronous communication. Moreover, we would like to
generalize the proposed approach to other classes of nets. Another line of research
will be focused on dealing with dynamic changes. In particular, intermediate
models, mentioned above, can be used to check if changes in the behavior of an
agent affect the soundness of the whole system.
References
1. van der Aalst, W.M.P.: Workflow verification: Finding control-flow errors using
petri-net-based techniques. In: Business Process Management: Models, Techniques,
and Empirical Studies. LNCS, vol. 1806, pp. 161–183. Springer, Heidelberg (2000)
2. van der Aalst, W.M.P., van Hee, K.M., ter Hofstede, A.H.M., Sidorova, N., Verbeek,
H.M.W., Voorhoeve, M., Wynn, M.T.: Soundness of workflow nets: classification,
decidability, and analysis. Formal Aspects of Computing 23(3), 333–363 (2011)
3. Augusto, A., Conforti, R., Dumas, M., Rosa, M.L., Maggi, F.M., Marrella, A.,
Mecella, M., Soo, A.: Automated discovery of process models from event logs:
Review and benchmark. CoRR abs/1705.02288 (2017)
4. Baldan, P., Corradini, A., Ehrig, H., Heckel, R.: Compositional modeling of reactive
systems using open nets. In: Larsen, K.G., Nielsen, M. (eds.) CONCUR 2001.
LNCS, vol. 2154, pp. 502–518. Springer, Heidelberg (2001)
5. Bednarczyk, M.A., Bernardinello, L., Caillaud, B., Paw(cid:32)lowski, W., Pomello, L.:
Modular system development with pullbacks. In: ICATPN 2003. LNCS, vol. 2679,
pp. 140–160. Springer, Heidelberg (2003)
6. Bernardinello, L., Mangioni, E., Pomello, L.: Local state refinement and composi-
tion of elementary net systems: An approach based on morphisms. In: Transactions
on Petri Nets and Other Models of Concurrency VIII. LNCS, vol. 8100, pp. 48–70.
Springer, Heidelberg (2013)
7. Bernardinello, L., Monticelli, E., Pomello, L.: On preserving structural and be-
havioural properties by composing net systems on interfaces. Fundamenta Infor-
maticae 80(1-3), 31–47 (2007)
8. Best, E., Devillers, R., Hall, J.G.: The box calculus: A new causal algebra with
multi-label communication. In: APN 1992. LNCS, vol. 609, pp. 21–69. Springer,
Heidelberg (1992)
9. Buijs, J.C.A.M., van Dongen, B.F., van der Aalst, W.M.P.: On the role of fitness,
precision, generalization and simplicity in process discovery. In: OTM 2012. LNCS,
vol. 7565, pp. 305–322. Springer, Heidelberg (2012)
10. Cardinale, Y., El Haddad, J., Manouvrier, M., Rukoz, M.: Web service composition
based on petri nets: Review and contribution. In: RED 2012. LNCS, vol. 8194, pp.
83–122. Springer, Heidelberg (2013)
11. Fabre, E.: On the construction of pullbacks for safe petri nets. In: Donatelli, S.,
Thiagarajan, P.S. (eds.) ICATPN 2006. LNCS, vol. 4024, pp. 166–180. Springer,
Heidelberg (2006)
12. Girault, C., Valk, R.: Petri Nets for Systems Engineering: A Guide to Modeling,
Verification, and Applications. Springer, Heidelberg (2003)
13. Haddad, S., Hennicker, R., Møller, M.H.: Channel properties of asynchronously
composed petri nets. In: ICATPN 2013. LNCS, vol. 7927, pp. 369–388. Springer,
Heidelberg (2013)
14. van Hee, K.M., Mooij, A.J., Sidorova, N., van der Werf, J.M.: Soundness-preserving
refinements of service compositions. In: Web Services and Formal Methods. LNCS,
vol. 6551, pp. 131–145. Springer, Heidelberg (2011)
15. van Hee, K.M., Sidorova, N., van der Werf, J.M.: Construction of asynchronous
communicating systems: Weak termination guaranteed! In: Software Composition.
LNCS, vol. 6144, pp. 106–121. Springer, Heidelberg (2010)
16. Kalenkova, A.A., Lomazova, I.A.: Discovery of cancellation regions within process
mining techniques. Fundamenta Informaticae 133(2-3), 192–209 (2014)
17. Kalenkova, A.A., Lomazova, I.A., van der Aalst, W.M.P.: Process model discovery:
A method based on transition system decomposition. In: ICATPN 2014. LNCS,
vol. 8489, pp. 71–90. Springer, Heidelberg (2014)
18. Leemans, S.J.J., Fahland, D., van der Aalst, W.M.P.: Discovering block-structured
process models from event logs - a constructive approach. In: ICATPN 2013. LNCS,
vol. 7927, pp. 311–329. Springer, Heidelberg (2013)
19. Lomazova, I.A.: Interacting workflow nets for workflow process re-engineering. Fun-
damenta Informaticae 101(1-2), 59–70 (2010)
20. Lomazova, I.A., Romanov, I.V.: Analyzing compatibility of services via resource
conformance. Fundamenta Informaticae 128(1-2), 129–141 (2013)
21. Nielsen, M., Rozenberg, G., Thiagarajan, P.: Elementary transition systems. The-
oretical Computer Science 96(1), 3–33 (1992)
22. Padberg, J., Urb´asek, M.: Rule-based refinement of petri nets: A survey. In: Petri
Net Technology for Communication-Based Systems: Advances in Petri Nets. LNCS,
vol. 2472, pp. 161–196. Springer, Heidelberg (2003)
23. Pankratius, V., Stucky, W.: A formal foundation for workflow composition, work-
flow view definition, and workflow normalization based on petri nets. In: APCCM
2005, vol. 43. pp. 79–88. Australian Computer Society, Inc. (2005)
24. Reisig, W.: Understanding Petri Nets: Modeling Techniques, Analysis Methods,
Case Studies. Springer, Heidelberg (2013)
25. Siegeris, J., Zimmermann, A.: Workflow model compositions preserving relaxed
soundness. In: BPM 2006. LNCS, vol. 4102, pp. 177–192. Springer (2006)
26. Winskel, G.: Petri nets, morphisms and compositionality. In: APN 1985. LNCS,
vol. 222, pp. 453–477. Springer, Heidelberg (1986)
|
1311.6870 | 1 | 1311 | 2013-11-27T05:26:03 | Multi-agent based protection system for distribution system with DG | [
"cs.MA"
] | This paper introduces the basic structure of multi-agent based protection system for distribution system with DGs. The entire system consists of intelligent agents and communication system. Intelligent agents can be divided into three layers, the bottom layer, the middle layer and the upper layer. The design of the agent in different layer is analyzed in detail. Communication system is the bridge of multi-agent system (MAS). The transmission mode, selective communication and other principles are discussed to improve the transmission efficiency. Finally, some evaluations are proposed, which provides the design of MAS with reference. | cs.MA | cs | MULTI-AGENT BASED PROTECTION SYSTEM FOR
DISTRIBUTION SYSTEM WITH DG
JIN SHANG, NENGLING TAI and QI LIU
Department of Electrical Power Engineering, Shanghai JiaoTong University,
Shanghai,200240,China
[email protected]
This paper introduces the basic structure of multi-agent based protection system for distribution system with DGs. The
entire system consists of intelligent agents and communication system . Intelligent agents can be divided into three layers,
the bottom layer, the middle layer and the upper layer. The design of the agent in different layer is analyzed in detail.
Communication system is the bridge of multi-agent system (MAS). The transmission mode, selective communication and
other principles are discussed to improve the transmission efficiency. Finally, some evaluations are proposed, which
provides the design of MAS with reference.
Keywords: Multi-agent; protection; distribution system; DG; communication
1. Background and Introduction
The environmental problems have aroused worldwide attention, and the energy crisis is getting more and
more serious. With the development of renewable sources and its controlling technology, a number of
distributed generations (DG) have been connected to the distribution system, which provides a new method
to save the limited energy and reduce carbon dioxide emissions.
However, DGs also bring some problems. Protection coordination is one of the most important problem s.
There’re lots of researches that analyze the problems from different aspects [1-4]. The key point is that,
there is only single power supply in the original system, but DGs add new power supplies to the grid, which
make it into a complex multi power supplies network. The new power supplies will significantly affect the
power flow. When a fault occurs, they will provide additional current to the fault point, which will greatly
affect the reliability of protection.
Since the original protection of distribution system is no longer as reliable as before, there’re many
improved methods proposed to solve this problem. It is very convenient to disconnect the DG unit when a
fault occurs [4]. And most standards for DG also require like this (IEEE-Std.1547-2003). In this case, DG
will not affect the fault current. However, for the system with large DG penetration level, it will not support
the grid voltage and frequency during and immediately after the fault. Restricting the capacity and access
point of DG is another method [5-7]. Then the influence of DGs is greatly reduced. But this will greatly
stunt the development and application of DG. The fault current limiter (FCL) has also been popular for a
while [2,8]. There’re also many other methods that proposed to improve the original protection scheme to
make the protection system more reliable [9-12]. However, all these methods have some limitation. They
may have solved the problems in some degree, but, they’re not systemic and effective enough. The
protection scheme should be modified when the structure of the system changes.
The multi-agent technology was first applied to the power system in the early 1990s. And there’re lots of
researches after that, such as optimal power flow algorithm, power market, voltage controlling and so on.
The protection system based on multi-agent was proposed not long ago. It has been studied a lot recently
[13-16]. However, the application of MAS for the protection of distribution system with DGs has been
rarely analyzed before. This paper presents an effective and extensible scheme. The basic structure of MAS
is introduced, the design of the agents and communication system are analyzed.
2. Structure of MAS
The multi-agent system is composed of different intelligent agents, which cooperate with others as
individual units. Any electrical element that contains necessary information can be represented as an
element agent, such as the line breaker, transformer and so on. There’re also integrated agents, they do not
collect information from certain element directly, but collect information from other agents and draw
conclusions. Fig. 1 shows the schematic diagram of a typical distribution system and the corresponding
protection system.
(a) Schematic diagram of distribution system
(b) Structure of corresponding MAS
Fig.1 Schematic diagram of distribution system and MAS
The distribution system in sub-graph (a) consists of two feeder lines and three energy sources, the
Photovoltaic (PV), Combined Cooling Heating and Power (CCHP), and Composite Energy Storage System
(CESS). In sub-graph (b), the seven orange boxes (B1 to B7) represent the branch agents. which gather
information from the distribution system directly. The three blue boxes represent DG agents, which gather
the real-time information of DGs. The two green ellipses represent the regional agents, which are the
middle segments between terminal agents and central agent. The red ellipse represents the central agent.
The arrows represent the information flow, two -way arrow means two-way flow. The arrows can be
considered to make up the communication system.
According to the structure introduced above, there’re mainly two aspects for the implementation of MAS.
One aspect is the design of various kinds of agents. The other is the construction of communication system.
The two aspects are both very important, and they influence each other at the same time. The model of
agent affects the structure of communication system. Meanwhile, the restriction of communication limits
the information transmission between different agents. It’s of great importance to design both of them
properly.
3. Intelligent Agents
In order to realize the function of different levels and make the structure of MAS clear, the whole MAS can
be divided into three layers [17], the bottom layer, the middle layer and the upper layer, as shown in Fig.2.
PVCCHPCESSB1B2B3B4B5B6B7LoadLoadMVB1B2B3B4B5B6B7StatusSwitch timesCurrentVoltageFrequencyLoadDistribution System StatusPV Electrical quantities Other information StatusCCHP Electrical quantities Other information StatusCESS Electrical quantities Other informationRegional Unit1Regional Unit2Central UnitInstructions Fig.2 Hierarchy chart of agents
All the terminal agents that directly interact with the distribution system (including different DGs), make
up the bottom layer. Regional agents that connect bottom layer and the upper layer make up the middle
layer. The central agent gathers information of the entire system, and it forms the upper layer. Three layers
play different roles. Agents are designed according to the functional requirements.
3.1. Terminal agents
The agents for breakers, lines, DGs and other elements in the bottom layer are classified as terminal agents.
They not only gather information from the power system, but also work as actuators. The tripping,
reclosing, load shedding and other instructions are all implemented through them. Since their functions are
complicated, the model should be well designed.
Fig.3 shows the structure of the branch agent. The branch agent is the integration of line agent and breaker
agent. Since the breaker agent provides the line with protection and the line agent provides the relay with
information, the two agents are closely related with each other. Thus they’re integrated to reduce the
investment and make the communication between them easier.
Fig.3 Structure of branch agent
As shown in Fig.3, the dashed box represents for the branch agent. Data collection module is responsible
for the data acquisition from the sensors, which directly measure the electrical qualities of the power
system. The data preprocess module will fulfill the data conversion and filtering, Then, the data with
prescribed format will be stored in the database. The analysis module reads the data from the database and
conducts the fault analysis [18]. If the breaker should act to clear the fault, the instruction will be sent out.
The action will also be recorded in the database. Once the instruction is successfully executed, the agent
will get the new status. The data interface is a bi-directional port connected to the database. The agent will
communicate with other agents through this port.
Database is the core of the agent, all the electrical data collected from the power system are stored here,
such as the status, the setting values of protection, and other parameters that related to the line and the
Terminal Agent 1Terminal Agent 2Terminal Agent 3Terminal Agent 4Regional Agent 1Regional Agent 2CentralAgent Bottom layerMiddle layerUpper layerData collection moduleInstructions DatabaseData preprocessAnalysis moduleStatusElectrical dataSetting value of protectionOther parametersSensorsBreakerData interface… …Other agentBranch agent equipment (e.g. relay, breaker). The model of the database will directly affect the efficiency of the agent.
Fig.4 shows the structure of the database model.
There are more details in the practical model. Each property has its sub-properties. For example, if the
“Breaker status” shows that the breaker is closed, there is a sub-property to tell whether the current stratus
is normal state, reclosing or manual operation. Besides, the analog properties (current, voltage) should be
stored separately according to phase a, b and c. “Set value” is concerned with the protection principle and
the status of DGs. Several groups of set value are stored for different situations.
Fig.4 Structure of a database model
Besides the branch agent analyzed above, there’re also transformer agent, busbar agent and so on. The
design of this kind of agent is similar with the branch agent. The design of DG agent is also similar,
although the internal control of DG is very complex, only the external information is the key for protection
system. The external information presented by DG is a little different from that of the ordinary element , but
the design idea is the same. In a word, the agent should collect information from the power system, take
analysis, and communicate with other agent.
3.2. Regional agents
Regional agents play the role as a gateway. They do not connect with the distribution system directly , but
collect information from a certain area through several terminal agents. The information is preprocessed
here. The existing of regional agents can avoid the data flood. Although each regional agent covers a
certain area, adjacent areas should overlap with each other to ensure the comprehensiveness and reliability
of the protection system. Since the regional agents exchange information widely with others, its database,
algorithm, interface and the whole structure are all very important. The structure of regional agent is shown
in Fig.5.
Branch ID #1Status propertyBreaker statusAnalog propertyPower (real)Current (real)Voltage (real)Frequency (real)Set value Set value 1Tripping value (real)Delay (real)Set value 2Other parametersSwitching times (integer)Length of line (real)Thermal stable current (real) Fig.5 Structure of regional agent
When the regional agent is to be designed, the function of the agent is defined at first. Then, the
information required and the instructions sent out are confirmed. The database and interfaces are designed
according to this. Database lays a solid foundation for the function modules . For different function modules,
it’s very important to choose a proper algorithm to achieve the function quickly and accurately.
Since the influences of DGs to the protection system are complicated, MAS should be adaptive in different
situations. However, the speed of online adaptive system can hardly meet the requirement of protection.
But the possible running status of DGs in a certain distribution system are fixed, which provides possibility
to build an offline adaptive system. An expert system is built to support the offline adaptive system. When
the running status of the network changes, the expert system will provide appropriate protection knowledge,
including protection configuration, coordination, setting values and so on. The knowledge as well as the
real-time information will be read by the function module. Then, the new protection scheme can be figured
out and instructions will be sent to the terminal agents.
3.3. Central agents
Since MAS has good abilities of information interaction and controlling [19-20], besides working as
protection system, it is also responsible for the optimal operation of the entire system.
Some kinds of DG have a random power output, such as PV wind turbine and CCHP . Fuel cell and CESS
are flexible units, which can be adjusted as source or load. Since the central agent can obtain the
information of the entire system, it’s able to arrange the running of each DG to achieve the optimal and
stable running of distribution system. DG agents will implement the instructions. The concrete
implementations are carried out by the controlling module inside of DG, which is not concerned by MAS.
Once the voltage or frequency of the distribution system is abnormal, the central agent can make
adjustments by controlling the active and reactive power of DGs, transferring load or even shedding load to
keep the system stable.
The goal of central agent involves the entire distribution system. The information it requires mainly comes
from the regional agents. Naturally, only the conclusive and necessary messages after preprocessing are
sent here. Since the controlling target of central agent is complex, the algorithm is much more complicated.
Different algorithms are applied under different situations. Using appropriate algorithms for different
application is of great significance.
DatabaseInformation collection interfaceInstruction interfaceanalysis unit algorithmCriterionFunction module1. . .analysis unit algorithmCriterionFunction module2Information interfaceAdjacent regional Agent Information sendinginterfaceInstruction interfaceTerminal Agent 1Terminal Agent 2Terminal Agent 3Central Agent ExpertSystem 4. Communication System
Communication system is the bridge of MAS. The transmission mode, efficiency, synchronization, cost and
information securities are important aspects to evaluate the system. They not only challenge the
communication technology, but also put forward higher demands on the configuration scheme of MAS.
4.1. Transmission mode
Using proper transmission mode will significantly improve the performance of MAS. There’re three
communication modes that commonly used, direct communication, radio communication, and blackboard
system. Direct communication is applied for the situation, in which both sides know each other in advance.
Radio communication enables one agent to broadcast information to all the members in the same group ,
even without knowing the address. Blackboard system is a centralized controlling system, in which all the
group members upload their information to a public area and the others can share the information as they
need. The three modes are applied in different situations according to their characteristics.
Terminal agents directly contact with the power system. They demand little of information analysis. So, it’s
unnecessary for them to communicate widely with others . They only communicate with adjacent agents,
which is fixed and regular. It’s better to use direct communication. Terminal agents also send information
to regional agent. Since each regional agent covers a certain area, it will be determined which few terminal
agents are in this area. In turn, these terminal agents will clearly know their communication object. So, it’s
also suitable to use direct communication. There’re lots of terminal agents in bottom layer, direct
communication reduces the cost and make the communication more safe and effective.
A regional agent is the core of the certain area. Besides receiving information from the terminal agents, it
also sends instructions to them. Since there is more than one terminal agent in the group, the instructions
may be sent to any one of them. It’s better to use radio communication, then, all the agents in this group
can receive the information and the instruction will be executed by the corresponded one after further
confirmation. Regional agents also communicate with the other agents in the same layer. Blackboard
system is applied in this situation. All the regional agents put their information in the public area. Other s
can easily get what they need for the further analysis. Because the public area is accessible for all the
regional agents, it requires strong fault tolerance. When the messages uploaded by different agents are in
conflict, only the right one or the one with higher priority will be recorded. The conclusions made by the
regional agents should be sent to the central agent, which is similar to the information transmission from
terminal agent to regional agent. The direct communication is suitable in this case.
The central agent gets information from all the regional agents for global optimization. The results and
instructions will then be sent to the regional agents by radio communication.
The communication modes between different objects in MAS are listed in Tab.1.
Tab. 1 Communication mode of MAS
Sender
Terminal
Regional
Receiver
agent
agent
Direct
Radio
Terminal
agent
Regional
agent
Central
agent
Central
agent
——
Direct
Blackboard
Radio
——
Direct
——
4.2. Selective communication
In order to save costs and increase speed, the communication system should be effective. When the agent
communicates with others, the objects should be selective. In the bottom layer, terminal agent
communicates with adjacent agents. As shown in Fig.1, line B7 is far from line B4, whether breaker B7
should act or not almost has nothing to do with line B4. So it’s unnecessary to connect agent B7 with agent
B4. On the contrary, line B1 and B2 are adjacent, which make them closely related . It’s very important to
acquire the information of the adjacent line to make fault analysis. So the two agents are connected to share
the information.
The information transferred should also be selective. For the communication between terminal agents, the
breaker status and current information is useful for the others, while the voltage and switch times are
useless. It’s of more importance to make it selective when the terminal agents send information to the
regional agents and the regional agents send informat ion to the central agent. Regional agents and central
agent get a large amount of information from more than one agent, it must be ensured that only the
information necessary for the further analysis are received. What’s more, central agent is more likely to
receive the analysis results from the regional agents to achieve the rapid analysis of optimization operation.
4.3. Principles
There’re some other principles that should be followed to improve the efficiency and reduce the cost of the
communication system.
Firstly, the task of local agent should be arranged reasonably. Since all the agents are intelligent , they
should carry out analysis as much as possible within their capability. This will reduce the information
transmission from the source.
Secondly, it’s better to transport status information instead of analog information [21]. For example,
directional overcurrent protection needs current direction as criterion, while low voltage start overcurrent
protection needs the value of voltage. Direction can be easily represented by a status property, which takes
up only 1 byte. But the value of voltage takes up more memory and it’s more demanding on the
synchronization of information. The transmission of status information is easier and more reliable.
The transmission distance should be as short as possible. A terminal agent should send information to the
nearest regional agent rather than the other one. Meanwhile, terminal agents do not send information
directly to central agent. The information is sent to regional agent and preprocessed to get simplified
conclusions. The conclusions will then be sent to the central agent.
5. Conclusions
The design of MAS can be more flexible on the premise of reliability. An excellent MAS implementation
should at least meet four requirements as following:
First of all, it must satisfy the reliability, sensitivity, rapidity, selectivity requirements of protection and
achieve the optimal controlling of the distribution system.
Second, the modules of MAS should be standardized to ensure the compatibility of differ ent systems and
good scalability to the expansion of grid.
Third, the construction and running of MAS should be economical.
Fourth, the information security should be guaranteed.
The four requirements define the optimal goals and provide the design of MAS with reference. Generally,
the future protection of distribution system should be intelligent based on the multi-agent technology and
the communication network. With the development of DGs and its controlling technology, a
communication network for DGs should be built to achieve the optimal running of power system. This will
also lay foundation for the multi-agent based protection system and makes it an inexorable trend.
References
1. Yuan Chao, Wu Gang, Protection technology for distributed generation systems, Power System Protection and
Control, 37(2009) 99-105.
2. K. Kauhaniemi, L. Kumnpulainen, Impact of distributed generation on the protection of distribution networks, in
Proc. et lEE Int. Conf.on Developments in Power System Protection . Vol. 1(2004), pp. 315 - 318.
3. Vandevelde, L, Melkebeek, J, Protection of Power Distribution Network with High Penetration Level of
Distributed Generation, Fifth FTW PhD Symposium, Faculty of Engineering, Ghent University, 2004, paper No.
024.
4. Morren, J., Haan, S.W.H. Impact of distributed generation units with power electronic converters on distribution
network protection, IET 9th International Conference on Power Energy, & Industry Applications, Glasgow, UK ,
2008
5. Wang Jianghai, Tai Nengling. Penetration level permission of DG in distributed network considering relay
protection, Proceedings of the CSEE, 30(2010) 37-43.
6. Feng Xike, Tai Nengling. Research on the impact of DG capacity on the distribution network curren t protection
and countermeasure, Power System Protection and Control, 38(2010) 156-165.
7. Lei Jinyong，Huang Wei，Xia Xiang．Penetration level calculation with considerations of phase-to-phase short
circuit fault, Automation of Electric Power Systems，32(2008) 82-86
8. YE Lin, LIN Liang-zhen. Superconducting Fault Current Limiter Applications in Electric Power Systems,
Proceedings of the CSEE, 20(2000) 1-5
9. Lin Xia, Lu Yuping, Wang Lianhe. New Current Protection Scheme Considering Distributed Generation Impact,
Automation of Electric Power System, 32(2008) 50-56.
10. Zhang Yanxia, Dai Fengxian. New Schemes of Feeder Protection for Distribution Networks Including Distributed
Generation, Automation of Electric Power System, 33(2009) 71-74.
11. Li Yongli, Jin Qiang, Li Botong. Application of inverse-time overcurrent protection based on low voltage
acceleration in Micro-Grid, Journal of Tianjin University, 44(2011) 955-960.
12. Sun Jingliao, Li Yongli. Study on Adaptive Current Instantaneous Trip Protection Scheme for Distribution
Network with Inverter Interfaced DG, Automation of Electric Power System, 33(2009) 71-76.
13. Renan Giovanini, Kenneth Hopkinson, Denis V. Coury, and James S. Thorp, A Primary and Backup Cooperative
Protection System Based on Wide Area Agents, IEEE Transactions on Power Delivery, 21(2006) 1222-1230
14. Wang Huifang, He Benteng, Shi Hongyu, Applications of multi-Agent technology in protection system, Electric
Power Automation Equipment. 27(2007) 102-106.
15. COURY D V, THORP J S, HOPKINSIN K M. An agent- based current differential relay for use with a utility
Intranet. IEEE Transactions on Power Delivery, 17(2002) 47-53.
16. HOSSACK J A, MENAL J, STEPHEN D J, et al. A multi-agent architecture for protection engineering diagnostic
assistance. IEEE Transactions on Power Systems, 18(2003) 639-647.
17. Xu Tianqi, Yin Xianggen, Analysis on functionality and feasible structure of wide area protection system, Power
System Protection and Control, 37(2009) 93-97.
18. K.K. Li, W.L. Chan, Xiangjun Zeng and Xianzhong Duan, Agent-based self-healing protection system, IEEE
Transactions on Power Delivery, 21(2006) 610-618
19. I. Zabet and M. Montazeri, Implementing Cooperative Agent-based Protection and Outage anagement System for
Power Distribution Network Control, 4th International Power Engineering and Optimization Conf, Shah Alam,
Selangor, MALAYSIA, 2010.
20. H. N. Aung, A. M. Khambadkone, D. Srinivasan, and T. Logenthiran, Agent-based Intelligent Control for
Real-time Operation of a Microgrid, 2010 Joint International Conf. Power Electronics, Drives and Energy
Systems (PEDES) & 2010 Power India, New Delhi, India, 2010
21. Sun Hui, Liu Qianjin, Power Distribution Network Protection and Control Using Mas, Proceedings of the
CSU-EPSA. 23(2011) 135-141.
|
1910.03640 | 1 | 1910 | 2019-10-08T18:51:15 | MAMS-A*: Multi-Agent Multi-Scale A* | [
"cs.MA",
"cs.DS"
] | We present a multi-scale forward search algorithm for distributed agents to solve single-query shortest path planning problems. Each agent first builds a representation of its own search space of the common environment as a multi-resolution graph, it communicates with the other agents the result of its local search, and it uses received information from other agents to refine its own graph and update the local inconsistency conditions. As a result, all agents attain a common subgraph that includes a provably optimal path in the most informative graph available among all agents, if one exists, without necessarily communicating the entire graph. We prove the completeness and optimality of the proposed algorithm, and present numerical results supporting the advantages of the proposed approach. | cs.MA | cs |
MAMS-A*: Multi-Agent Multi-Scale A*
Jaein Lim1 and Panagiotis Tsiotras2 ∗†
Abstract
We present a multi-scale forward search algorithm for distributed agents to solve single-query
shortest path planning problems. Each agent first builds a representation of its own search space of
the common environment as a multi-resolution graph, it communicates with the other agents the
result of its local search, and it uses received information from other agents to refine its own graph
and update the local inconsistency conditions. As a result, all agents attain a common subgraph
that includes a provably optimal path in the most informative graph available among all agents, if
one exists, without necessarily communicating the entire graph. We prove the completeness and
optimality of the proposed algorithm, and present numerical results supporting the advantages of
the proposed approach.
1
Introduction
Humans appear to rely heavily on hierarchical structures for decision-making, especially for complex
tasks (e.g., planning a route from work to home, navigating traffic, etc). Despite the fact that any
plan is physically realizable as a sequence of refined actions (e.g., walking to the office door, opening
the door, walking down the hallway, etc), most humans do not plan at this low level set of actions.
Instead, planning is achieved using a variety of abstraction levels, by aggregating the sequence of
actions into high-level macro-actions (i.e., exiting the office building, changing lanes in traffic, etc),
and by executing this sequence of high-level actions by refining them down [1,2]. This is done because
acquiring perfect knowledge about the environment is often prohibitive, and hence acting on the right
level of granularity of pertinent information is imperative in order to operate in dynamic and uncertain
environments. This observation has led many researchers to investigate multi-scale representations of
the underlying search space for planning [3 -- 8].
Kambhampati and Davis [4] incorporated a top-down refinement scheme of the environment ab-
stracted in a quad-tree for path-planning problems. The obstacle region starting from the coarsest
resolution level is excluded from the refinement in order to reduce the search space. Likewise, Pai and
Reissell [5] utilized top-down refinement for path-planning, specifically using the wavelet transform
of a 2-D environment. As the wavelet difference coefficients inform how smooth the refined path will
be in the next refined level, the magnitude of the coefficients serve as an inconsistent heuristic for
refinement. Marthi et al [1] introduced the angelic semantic for a top-down hierarchical planning,
which specified for each high-level action the set of reachable states by refinement, along with the
associated upper and lower bounds on the cost. They showed that some pruning rules based on the
upper and lower bounds on the value of the high-level action plans preserve global optimality, while
the optimistic evaluation of an abstract plan serves as a heuristic for refinement [1].
More recently, the construction of multi-resolution graphs of the environment has been studied
for path-planning problems to efficiently solve for a locally optimal solution [9 -- 12]. These references
∗1Jaein Lim is a graduate student at the School of Aerospace Engineering, Georgia Institute of Technology, Atlanta.
†2 Panagiotis Tsiotras is a Professor and David and Andrew Lewis Chair at the School of Aerospace Engineering
and Associate Director at the Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta.
GA 30332-0150, USA. Email: [email protected]
GA 30332-0150, USA. Email: [email protected]
Figure 1: Top: Three agents' abstract graphs of the same environment with gray scale representing
risk level. Bottom: The individual agents' abstract graphs and optimal paths (green) from bottom left
initial vertex (purple) to the goal vertex in the top right. After the search, the agents share a common
subgraph that consists of the optimal path in the finest resolution abstraction available. Unexpanded
vertices were not communicated and are left blank.
construct fine resolution graph representations near the agent to accurately represent the environment
in the vicinity of the agent, and coarser resolution representations far away from the agent to reduce
the dimensionality of the search space. As a result, a multi-resolution framework allows a natural
formulation of the perception-action loop, while still maintaining practical efficacy by utilizing a
reduced search space.
Different approaches have been studied for planning problems with multiple agents, using dis-
tributed or parallel processing techniques to increase the solution quality via inter-agent communi-
cation. The C-FOREST algorithm [13] employs multiple CPU units in parallel using the same start
and goal states for randomized motion planning. Each computing agent stores its own search tree,
and communicates with the other agents to restrict the sampling space to focus the search so as to
increase solution quality. Botea et al [7] built a hierarchical abstraction of the map, in which local
roadmaps are computed in parallel with fine abstraction. Then, whenever a new set of start and
goal locations is given, a solution is computed holistically using the coarser resolution abstraction. A
somewhat similar approach was used in [14] where a multi-scale algorithm was proposed to accelerate
A*. By pre-processing the environment in a hierarchical dyadic decomposition, the overall problem is
divided to a nested sequence of smaller problems. The solutions of these smaller problems were then
merged together using a "bottom-up" fusion algorithm to get the globally optimal path much faster
than existing methods. Finally, Nissim and Brafman [15] formulated a distributed agent planning
problem where each agent maintains a separate search space and expands the state from its OPEN
list similar to the regular A* algorithm. Each agent then informs the relevant agents (that is, those
who require the current state as a precondition for their actions) of the best cost-to-come and the
heuristic estimate of that state by sending a message. Once the agent receives the message from
another agent, this agent adds the message to its OPEN list only if the cost-to-come of the state is
better than its own. In essence, this is an optimality-preserving pruning technique, where multiple
agents communicate local information in an attempt to build their own optimal solution [15].
In this paper, we extend the distributed forward search algorithm proposed in [15] to incorporate
the multi-resolution framework of [9] for path-planning with multiple agents. Unlike other top-down
pre search agent1pre search agent2pre search agent3post search agent1post search agent2post search agent3 refinement schemes mentioned above, which rely on the agent's own abstraction hierarchy, we refine
each agent's abstract path using the information provided by the other agents.
2 Problem Formulation
2.1 Multiresolution World Representation
Without loss of generality, we assume that the environment W ⊂ Rd is given as a hypercube of
side length 2(cid:96) for some positive integer (cid:96). The hypercube W is hierarchically abstracted as a 2d-
tree T = (N ,R) using a recursive dyadic partition, where N is the node set and R is the edge set
describing the relations of the nodes in N , such that each node in the tree encodes the information
contained in a subset of W. Specifically, each node nk,p ∈ N at depth k abstracts information of the
world contained in the hypercube H(nk,p) ⊆ W of side length 2k centered at p ∈ W. The function
V : N → R+ maps each node to some non-negative real value V (nk,p), for example, the probability
of occupancy of H(nk,p) or a cost measure to the same region. We assume that T is a full tree, that
is, each nk,p ∈ N has either 2d children or none. The children of nk,p are denoted by {nk−1,qi}i∈[1,2d]
where qi = p + 2k−2ei and ei is a vector in the set {[±1, ...,±1] ∈ Rd}.
2.2 Abstract Graph Construction
Let Φ = {ϕi}i∈I be a finite set of agents, where I = {1, 2, ..., n} is the agent index set. Each
agent ϕi ∈ Φ builds a non-empty multi-resolution graph Gi = (Vi,Ei) from T , such that Gi spatially
represents W, where Vi is the vertex set and Ei is the edge set. The agent ϕi selects some nodes from
N as its vertices in a top-down fashion, and for each node nk,p ∈ N , it selects either all of its children
or none of them to maintain a dyadic representation of W using the selected vertices in Vi. More
precisely, when the agent selects the children of a node in N , it excludes this node from Vi. This rule
ensures that the union of the regions corresponding to the vertices in Vi sufficiently covers W. Hence,
Vi ⊆ N , and each v ∈ Vi corresponds to a node of N , but the converse is not true. The node nk,p of
T is selected as a vertex of Gi if
√
(cid:107)p − pi(cid:107)2 −
d > α2k,
(1)
define H(W ) be the hypercube covered by W , that is, H(W ) =(cid:83)
where pi is the position of agent ϕi and α > 0 is a user-specified parameter.
Similarly to the nodes, we denote the hypercube covered by vertex v ∈ Vi as H(v), such that if
vertex v ∈ Vi corresponds to node n ∈ N , then H(v) = H(n). Let a set of vertices W ⊆ N . We
v∈W H(v). Two vertices v and v(cid:48)
in Vi are neighbors if the union of the boundaries of the corresponding hypercubes is neither empty
nor a singleton. Each edge e ∈ Ei assigns a non-negative real value E(v, v(cid:48)) to a pair of vertices
(v, v(cid:48)) ∈ Vi × Vi if and only if the corresponding vertices are neighbors. The edge value E(v, v(cid:48)) is the
cost to traverse from v to v(cid:48), and is defined as [9]
(2)
where λ1, λ2 ∈ (0, 1] are weights used to penalize the content and the abstraction level, respectively,
of the corresponding node nk,p of v(cid:48).
E(v, v(cid:48)) = 2dk(λ1V (nk,p) + λ2),
2.3 Merged Graph
Consider an arbitrary vertex set W which includes the vertices u and v. A vertex u is defined to be
a child vertex of v in W , if H(u) ⊂ H(v). We define a vertex w ∈ W to be a fine vertex in W if there
are no children vertices of w in W . Otherwise, w is defined to be a coarse vertex in W . Hence, every
vertex in the individual agent's graph Gi = (Vi,Ei) is finest by construction, as the agent excludes the
parent node from Vi if the agent chooses the children nodes to include in Vi.
only the fine vertices of the union VI =(cid:83)
For the set of agents Φ, we define the merged graph GΦ = (VΦ,EΦ) with vertices consisting of
i∈I Vi of the vertices of all agents' graphs Gi, and edge set
EΦ that assigns the edge cost E(v, v(cid:48)) to every pair of neighboring (v, v(cid:48)) ∈ VΦ × VΦ. The merged
graph excludes a vertex v from VI if there exist children vertices of v that cover the region covered
by v. Note that VΦ ⊆ VI is sufficient to spatially represent W compactly. Figure 2 illustrates the
construction of the graph GΦ.
Figure 2: Example of problem formulation for W ⊂ R2, from left to right: a) T - abstracted world
encoded in a quad-tree; b) G1 - agent ϕ1's graph construction from T ; c) G2 - agent ϕ2's graph
construction from T ; d) VI={1,2} - hashed blue vertex of V1 is coarse and solid green vertices of V2
are fine in V1 ∪ V2; and e) GΦ - merged graph from both Gi and Gj
2.4 Path-Planning Problem
A path P = (v0, v1, . . . , vk) on the graph G = (V,E) is an ordered set of vertices vi ∈ V, i = 0, . . . , k
such that for any two consecutive vertices there exists an edge e ∈ E. Given G, an initial vertex
vinit ∈ V and a goal vertex vgoal ∈ V, we define the path-planning problem Π, as the problem to find a
path P = (vinit, . . . , vgoal) in G. An optimal path PG(Π) or PG for a problem instance Π = Π(vinit, vgoal)
with initial and goal vertices vinit to vgoal is a path having the smallest cost over all paths from vinit
to vgoal. To simplify the notation, we use PΦ instead of PGΦ to denote an optimal path in the merged
graph GΦ.
We denote the cost-to-come with g : V → R+ and the heuristic cost-to-go with h : V → R+, which
both assign each vertex in V a non-negative real value. For each individual agent, these functions
may assign different values to the same vertex. Hence, we use a subscript, namely gi(vk), to denote
the cost accumulated from the initial vertex vinit = v0 to vk in the agent ϕi's graph, and use hi(vk)
as the heuristic estimate from vk to the goal vertex vgoal = vg in the agent ϕi's graph representation.
For each agent, we define the evaluation function fi(vk) = gi(vk) + hi(vk). For the single agent case
with uniform resolution, the problem formulation reduces to the well-known shortest-path-planning
problem on a graph.
3 The Multi-Agent Multi-Scale A* (MAMS-A*) Algorithm
Some agents may have a finer resolution abstraction in a region than others, and we are interested
in finding an optimal path that would be optimal in the finest resolution abstraction of the world if
agreed upon by all agents. One naıve way to achieve this objective would be to build a single merged
graph abstraction GΦ from all agents in Φ first, and then search for an optimal path PΦ within this new
graph. Consequently, all agents will perform a search in the most informative graph available among
them. However, as much this approach refines the abstraction level of the graph, it also enlarges the
search space. In addition, the amount of information communicated among the agents is increased
unnecessarily by broadcasting the entire graph, whether or not the received information can improve
the solution quality. Thus, construction of such a graph dilutes the benefit of using multi-resolution
abstractions during the search.
To alleviate the mentioned issues, we propose an algorithm in which each agent ϕi broadcasts a
message containing the vertex information (i.e., (cid:104)s, gi(s), hi(s)(cid:105)) only when ϕi expands its own vertex
s. Also, each agent processes the received message only if it contains finer resolution information about
the environment or if the messaged vertex has not been expanded or it has a better cost-to-come than
the agent's own cost-to-come value. Hence, we limit communication only to the expanded vertices.
If the heuristic used by the agents is consistent, then the algorithm expands and broadcasts only the
vertices with the lowest possible total cost. In addition, we assume that every broadcasted message
reaches all agents in Φ. We present the pseudo-algorithm below.
Algorithm 1 MAMS-A* for ϕi
1: while ∃ active agent ∈ Φ do
2:
3:
4:
5:
6:
ProcessMessage
if ϕi is active then
s ← pop front OPEN
Expand(s)
Publish(s)
Algorithm 2 ProcessMessage
1: for message = ((cid:104)s, gj(s), hj(s)(cid:105)) ∈ message que do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
switch message do
case ∃ v ∈ Vi such that H(v) = H(s)
if v not expanded or gi(v) > gj(s) then
Adopt(s)
case ∃ v ∈ Vi such that H(s) ⊂ H(v)
add s to Vi
for v(cid:48) ∈ neighbors(v) do
remove e(v, v(cid:48)) from Ei
add e(s, v(cid:48)) to Ei if exists
remove v from Vi
remove v from OPEN or CLOSED
Adopt(s)
case ∃ v ∈ Vi such that H(v) ⊂ H(s)
else
discard message
add s to Vi
for v(cid:48) ∈ Vi do
add e(s, v(cid:48)) to Ei if exists
Adopt(s)
The proposed multi-agent multi-scale A* algorithm comprises of three major procedures: 1) Pro-
cessMessage, 2) Expand, and 3) Publish, and is summarized in Algorithm 1. Each individual agent
ϕi ∈ Φ repeats these procedures until the finest goal vertex is expanded and processed among the
agents in Φ. Once agent ϕi expands the goal vertex of its graph Gi, this agent is inactivated by Line 2
of Algorithm 3. Inactive agents do not expand any vertices nor publish any messages until a message
from another agent reactivates them by Line 6 of Algorithm 4. The algorithm starts with all agents
being active, and terminates when all agents become inactive (Line 1 of Algorithm 3). Suppose ui
and uj are the goal vertices of agents ϕi and ϕj respectively, such that H(uj) ⊂ H(ui). If agent ϕi
expands its goal vertex ui ∈ Vi (which is coarse in the vertex set VI since H(uj) ⊂ H(ui)) before the
vertex uj ∈ Vj is broadcasted by another agent ϕj, then ϕi will not expand any vertices nor publish
any messages until this agent is reactivated by an incoming message containing the vertex uj (Line 6
of Algorithm 4). On the other hand, if a message containing vertex uj is broadcasted to agent ϕi
Algorithm 3 Expand(s)
1: if s is a goal vertex or OPEN = ∅ then
inactivate ϕi
2:
3: else
4:
5:
6:
7:
8:
9:
10:
11:
move s to CLOSED
for s(cid:48) ∈ neighbors(s) do
if s(cid:48) /∈ CLOSED then
if gi(s(cid:48)) > gi(s) + E(s, s(cid:48)) then
gi(s(cid:48)) ← gi(s) + E(s, s(cid:48))
fi(s(cid:48)) ← gi(s(cid:48)) + hi(s(cid:48))
add s(cid:48) to OPEN
predecessor(s(cid:48)) ← s
2:
3:
Algorithm 4 Adopt(s)
1: if ∃ predecessor(s) in Vi then
gi(s) ← gj(s)
hi(s) ← max(hi(s), hj(s))
fi(s) ← gi(s) + hi(s)
add s to OPEN
reactivate ϕi
4:
5:
6:
before ϕi expands its goal vertex ui, then ui will be simply removed from the agent ϕi's graph. Hence,
every agent will expand the fine goal vertex in VΦ at least once before termination.
In case an agent receives finer resolution information about one of its vertices (Line 6 of Algo-
rithm 2), the agent removes this vertex from its graph and from its OPEN or CLOSED lists. When
agent ϕi receives the expansion result of the vertex from agent ϕj that already exists in its graph Gi,
Line 3 of Algorithm 2 puts the expansion result in the OPEN list only if the vertex has not been
expanded or if it has better cost-to-come value. When a new vertex is received that has no related
vertex in Gi because the coarse vertex has already been removed, then Line 16 of Algorithm 2 simply
updates the local graph and puts the expansion result in the OPEN list.
Figure 3: Illustrative example of Algorithm 2, from left to right: a) Gi - agent ϕi's original graph; b)
Gj - agent ϕj's graph; c) Gk - agent ϕk's graph; d) agent ϕi's modified graph after receiving messages
from ϕj and ϕk
By removing a vertex from its graph upon receiving fine resolution information about this vertex,
we are preemptively preventing each agent to construct a path that includes coarse resolution vertices
from VI. Instead, if necessary, the agent will receive fine vertices passing through the region origi-
nally covered by the removed vertex, as the other agents with fine resolution information available
expand and broadcast (see Figure 3, for instance). Indeed, if an optimal path PΦ passes through the
region covered by the removed vertices, then the corresponding segment of PΦ will be expanded and
broadcasted by other agents. Hence, path connectivity for the agent will be restored despite of the
removal of coarse vertices. We give the formal proof of optimality and completeness of the MAMS-A*
algorithm in the next section.
4 Analysis
We prove the completeness of the MAMS-A* algorithm and the optimality of the solution with respect
to the merged graph by extending a well known result of A* [16]. We will use Lemma 1 and its corollary
to show the completeness of the algorithm regardless of the removal of coarse vertices, assuming that
every broadcasted message arrives to all the agents.
Lemma 1. Let VI be the collection of all agents' vertices, that is, VI = (cid:83)
i∈I Vi. For any coarse
vertex u in VI of an agent ϕi, there exists a set of the fine vertices W ⊆ VI such that H(W ) = H(u).
Proof. Let k be some integer such that vertex vk has corresponding region H(vk) with side length 2k.
Suppose vertex vk of agent ϕi is coarse in VI, that is, there exists at least one vertex vm ∈ VI such
that H(vm) ⊂ H(vk) for m < k. Let M = {m : H(vm) ⊂ H(vk), vm ∈ VI}. Then M is non-empty,
closed and bounded, since vk is coarse and there is only a finite number of vertices in VI. Let n be the
minimum of M . Then vn is a fine vertex corresponding to vk, that is, H(vn) ⊂ H(vk) with n ≤ m < k.
Without loss of generality, let Vj be the vertex set of agent ϕj which includes vn. Since H(Vj) = W,
and Vj is the dyadic partition of W, there exists a subset Vn ⊆ Vj of cardinality 2d containing vn
whose elements have the same corresponding parent node, namely, µn+1, and H(Vn) = H(µn+1). If
n = k − 1, then H(Vk−1) = H(vk) and Vk−1 ⊆ Vj ⊆ VI. Since there cannot exist a finer vertex than
vn, the claim holds.
Consider now n < k − 1. Since the agent ϕj has the fine resolution vertex vn, the agent had once
selected the node µn+1 ∈ N , the parent node of µn ∈ N , where H(µn) = H(vn), from tree T as
its vertex during the construction of its abstract graph Gj. Hence, there exists a set W1 ⊆ Vj which
contains either siblings of vn+1 or their children, such that H(W1) = H(vn+2). In the same way, since
agent ϕj had once selected µn+1 ∈ N , it also had selected the parent node µn+2 ∈ N during the
construction of its abstract graph Gj. Hence, there exists a set W2 ⊆ Vj which contains either siblings
of vn+2 or their children, such that H(W2) = H(vn+3). Repeating the same argument k− n− 1 times,
we have a set Wk−n−1 ⊆ Vj such that H(Wk−n−1) = H(vk), and Wk−n−1 has vertices with side
length not greater than level k − 1. Repeating the entire argument for all coarse vertices of Wk−n−1
completes the proof.
Figure 4: Illustrative example Lemma 1, from left to right: a) agent ϕ1's vertex set V1; b) agent ϕ2's
vertex set V2; c) agent ϕ3's vertex set V3; d) VI={1,2,3} - ϕ1's hashed red vertex is coarse in VI={1,2,3},
and there exists a set of the fine vertices (green solid) covering the same region.
Corollary 1. For any coarse vertex u of an agent ϕi, there exists a set of fine vertices W ⊆ VΦ such
that H(W ) = H(u).
Proof. By Lemma 1, for any coarse vertex u ∈ VI, there exists a set of the fine vertices W ⊆ VI and
H(W ) = H(u). Since VΦ ⊆ VI is the largest finest resolution vertex set in VI, and W consists only
the fine vertices, it follows that W ⊆ VΦ.
The following lemma and the corollary will be used to prove the optimality of the path with respect
to the merged graph GΦ.
Lemma 2. Let v be a vertex that has not been expanded by any agent. For any optimal path PΦ from
s to v, there exists an agent ϕi ∈ Φ that has either an open vertex v(cid:48) or has an incoming message
(cid:104)v(cid:48), gj(v(cid:48)), hj(v(cid:48))(cid:105) from ϕj, i (cid:54)= j such that v(cid:48) is on PΦ and gi(v(cid:48)) = gΦ(v(cid:48)).
Proof. Let PΦ = (s = v0, v1, ..., vk = v). If v(cid:48) = s, then the lemma is trivially true, since gi(s) =
gΦ(s) = 0. Suppose now that s is closed by all agents, and let ∆ be the set of closed vertices by some
agents in PΦ, such that gi(δ) = gΦ(δ) for all δ ∈ ∆. Then ∆ is not empty, since s ∈ ∆. Let v∗ be
the element of ∆ with the highest index closed by agent ϕi, that is v∗ = argminδ∈∆ fi(δ). Clearly,
v∗ (cid:54)= v, as v is not closed. Let v(cid:48) be the successor of v∗ on PΦ (possibly v(cid:48) = v). If v(cid:48) is reachable
by ϕi, then gi(v(cid:48)) = gi(v∗) + E(v∗, v(cid:48)) = gΦ(v(cid:48)), because v(cid:48) is on PΦ. Otherwise, there exists ϕj who
can reach v(cid:48) via processing the message (cid:104)v∗, gi(v∗), hi(v∗)(cid:105), because no vertices in PΦ can be removed.
By Line 8 of Algorithm 3, gj(v(cid:48)) = gj(v∗) + E(v∗, v(cid:48)) = gΦ(v(cid:48)). Then, the message (cid:104)v(cid:48), gj(v(cid:48)), hj(v(cid:48))(cid:105)
is broadcasted, and eventually gj(v(cid:48)) = gi(v(cid:48)) = gΦ(v(cid:48)) by Line 2 of Algorithm 4. Hence the claim
holds.
Corollary 2. Suppose hk is admissible for all k ∈ I, that is, hk ≤ hΦ, where hΦ is the true cost-to-go
in GΦ, and suppose the algorithm has not terminated. Then, for any optimal path PΦ from the initial
vertex s to any goal vertex, there exists an agent ϕi which either has an open vertex v(cid:48) or has an
incoming message containing v(cid:48), such that v(cid:48) is on PΦ and fi(v(cid:48)) ≤ fΦ(s).
Proof. By Lemma 2, there exists an agent ϕi which either has an open vertex v(cid:48) or has an incoming
message containing v(cid:48) on PΦ with gi(v(cid:48)) = gΦ(v(cid:48)). Then
fi(v(cid:48)) = gi(v(cid:48)) + hi(v(cid:48))
= gΦ(v(cid:48)) + hi(v(cid:48))
≤ gΦ(v(cid:48)) + hΦ(v(cid:48)) = fΦ(v(cid:48)).
Since v(cid:48) is on the optimal path PΦ, fΦ(v(cid:48)) = fΦ(s), which completes the proof.
Theorem 1. With admissible heuristisc hi(s), i ∈ I, MAMS-A* terminates in a finite number of
iterations by finding an optimal solution, if one exists, in the merged graph GΦ.
Proof. We prove this theorem by contradiction. Suppose the algorithm does not terminate by finding
an optimal path to a goal vertex in the merged graph GΦ. There are three cases to consider:
1. The algorithm terminates at a non-goal. This contradicts the termination condition (Line 1 of
Algorithm 1) since the agents become inactive only if they expand a goal vertex or the OPEN list
is empty (Line 1 of Algorithm 3). At least one agent has non-empty OPEN if a goal vertex has
not been expanded.
2. The algorithm fails to terminate. Since there is a finite number of non-goal vertices, a finite number
of agents, and a finite number of non-cyclic paths from the start vertex s to any vertex v with
non-negative edge cost, a vertex will be closed forever by all agents or removed permanently from
the search space by Line 6 of Algorithm 2. Hence, the only possibility left for an agent to remain
active without reaching the goal vertex is when the graph is disconnected along the path from the
initial vertex to the goal vertex by Line 9 of Algorithm 2. Suppose a coarse vertex u of Gi was
removed by Line 9 of Algorithm 2. By Corollary 1, there exists a set W of fine vertices such that
H(W ) = H(u). Hence, if H(u) contains the part of the path to the goal, the vertices of W will be
added to graph Gi by Line 17 of Algorithm 2. Moreover, any vertex w in W cannot be removed
because it is a fine vertex. Hence, a path to the goal vertex that may have become disconnected
by the removal of the coarse vertex u will be eventually restored by the set of the fine vertices in
W . This contradicts the assumption that the algorithm failed to terminate.
3. The algorithm terminates at a fine resolution goal without achieving the minimum cost in the merged
abstraction Φ. Suppose the algorithm terminates at some goal vertex v with fj(v) > fΦ(v). By
Corollary 2, just before termination, there existed an agent ϕi which had an open node v(cid:48), or had
an incoming message containing v(cid:48), such that v(cid:48) is on an optimal path and fi(v(cid:48)) ≤ fΦ(s). Thus,
at this stage, v(cid:48) would have been selected for expansion rather than v, or at least one agent would
have been reactivated by the message containing v(cid:48), contradicting the assumption the algorithm
terminated without achieving the minimum cost.
5 Discussion
Note that since the attained solution does not necessarily imply resolution completeness, it may
be prudent to alleviate the computational burden of message broadcasting by sequentially sending
messages from one agent to another at the expense of optimality. Let the sequence of agents be
(ϕ1, ϕ2, . . . , ϕn), such that ϕ1 passes its expansion result to ϕ2, and ϕ2 passes it to ϕ3, and so forth.
The agents will not loose any information that has been deemed optimal by the previous agents, and
therefore the next agent ϕi+1 will always make decisions based on augmented information provided
by the previous agent ϕi, resulting in an optimal path in the partially merged graph. Nonetheless, the
last agent ϕn's path may not be necessarily the same as PΦ, since a segment of PΦ could be ignored
by an agent ϕi without the information of some agent ϕj for j > i, and thus this segment may not
be passed onto the next agent ϕi+1. However, in our experiments the results were shown to be very
close to the optimal one.
We also adopted the backtracking algorithm presented in [9] to incorporate a further refinement
scheme for resolution completeness, i.e., to find a feasible path in the fine resolution space. At each
iteration, the agent traverses along the abstract solution path and stores only the fine resolution
path segment to its memory, then re-solves the shortest path problem at a new vertex with different
abstraction until the goal vertex is reached. If no solution can be found from the current vertex v, the
agent backtracks to the previous vertex u in the accumulated path and removes the edge e = (u, v)
from the graph to avoid a cyclic iteration. The proof of completeness is omitted for brevity, and
instead we refer the reader to [9]. Note that it is desirable to backtrack as early as possible, however
this depends on the environment and the choice of α. A larger value of α makes the algorithm easier
to backtrack earlier, as the agent attains more information far away from the current location than
with a smaller value of α. However, the cardinality of the search space also increases with increasing
α, and therefore the right granularity of the abstract graph for optimal performance is not known a
priori.
6 Numerical Results
The MAMS-A* algorithm was implemented within the Robot Operating System (ROS) framework
for modularized agents. In a single workstation, multiple ROS nodes were generated to solve a single
query planning problem cooperatively. The communication among the nodes was made via ROS
messaging.
Different number of agents were generated at different locations including the start and goal
positions in a 2D labeled map; see Figure 1. In our experiments, we chose λ1 and λ2 to be 0.999
and 0.001, respectively, and we set V (nk,p) ∈ [0, 1] to be the risk level averaged over the region
H(nk,p). Different values of α were chosen with varying size of the search space represented as depth,
e.g., depth 7 corresponds to 27d nodes. Each result was then normalized by the resolution complete
solution computed with a regular A* in the corresponding depths. Figure 5 shows the result after one
iteration.
The computational advantage of the proposed algorithm is most prominent when the original
search space is large and α is small (e.g., α = 1 depth 7), as the abstraction reduces the search space
Figure 5: Computation time and solution cost of abstract path for different number of agents and
parameters normalized by regular A* result. Left: normalized path cost. Right: normalized time.
most significantly. The algorithm finds an abstract path three orders of magnitude faster compared to
the regular A*. As we penalize the abstraction in the cost function defined in equation (2), the cost of
the abstracted path is substantially worse compared to the resolution optimal solution especially for
the single agent case. The increased number of agents improves the solution quality of the abstract
path.
A similar comparison was made and plotted in Figure 6, but using instead the backtracking
algorithm to solve for a fine resolution path. At each iteration, only one agent was allowed to move,
while the other agents remained stationary. This was done to demonstrate the application of the
algorithm for cooperative agents with different goal locations.
Figure 6: Computation time and solution cost of fine path for different number of agents and param-
eters normalized by regular A* result. Left: normalized path cost. Right: normalized time.
In the single agent case, the performance was highly sensitive to the choice of α. At some abstrac-
tion structure, the single agent could still find near optimal solutions, but in the worst case, the path
length was 16 times longer than the optimal. In contrast, as a better heuristic for refinement was
used, the multi-agent cases resulted in more consistent computation times and cost performance that
was less sensitive to the particular choice of α.
0246810121416SingleTwoThreeFourNormalized Path CostNumber of Agents𝛼=1 depth 7𝛼=2 depth 7𝛼=3 depth 7𝛼=1 depth 6𝛼=2 depth 6𝛼=3 depth 6𝛼=1 depth 5𝛼=2 depth 5𝛼=3 depth 50.00010.0010.010.1110SingleTwoThreeFourNormalized TimeNumber of Agents𝛼=1 depth 7𝛼=2 depth 7𝛼=3 depth 7𝛼=1 depth 6𝛼=2 depth 6𝛼=3 depth 6𝛼=1 depth 5𝛼=2 depth 5𝛼=3 depth 5024681012141618SingleTwoThreeFourNormalized Path CostNumber of Agents𝛼=1 depth 7𝛼=2 depth 7𝛼=3 depth 7𝛼=1 depth 6𝛼=2 depth 6𝛼=3 depth 6𝛼=1 depth 5𝛼=2 depth 5𝛼=3 depth 50.0010.010.1110SingleTwoThreeFourNormalized TimeNumber of Agents𝛼=1 depth 7𝛼=2 depth 7𝛼=3 depth 7𝛼=1 depth 6𝛼=2 depth 6𝛼=3 depth 6𝛼=1 depth 5𝛼=2 depth 5𝛼=3 depth 5 7 Conclusion
In this paper, we propose a new algorithm to solve a single query shortest path planning problem us-
ing multiple multi-resolution graphs representing the same search space. The solution quality and the
speeds up from abstraction is balanced efficiently by using multiple agents distributed in the search
space, as they communicate only the expansion result of A* to avoid unnecessary communication.
The completeness and optimality of the algorithm are shown. The proposed scheme was applied to a
backtracking algorithm and demonstrated the advantages of using selective and distributed informa-
tion provided by other agents for refinement heuristics in terms of computational time and solution
quality.
Acknowledgement
This work has been supported by ARL under DCIST CRA W911NF-17-2-0181.
References
[1] B. Marthi, S. Russell, and J. Wolfe, "Angelic hierarchical planning: optimal and online algorithms
(revised)." University of California, Berkeley" Technical Report, 2009.
[2] S. J. Kiebel, J. Daunizeau, and K. J. Friston, "A hierarchy of time-scales and the brain," PLoS
Computational Biology, vol. 4, no. 11, p. e1000209, 2008.
[3] M. Helmert, P. Haslum, J. Hoffmann, and R. Nissim, "Merge-and-shrink abstraction: A method
for generating lower bounds in factored state spaces," Journal of the ACM, vol. 3, 05 2014.
[4] S. Kambhampati and L. S. Davis., "Multiresolution path planning for mobile robots," IEEE
Journal of Robotics and Automation, pp. 135 -- 145, 1986.
[5] D. K. Pai and L. . Reissell, "Multiresolution rough terrain motion planning," IEEE Transactions
on Robotics and Automation, vol. 14, no. 1, pp. 19 -- 33, Feb 1998.
[6] R. Holte, M. Perez, R. Zimmer, and A. MacDonald, "Hierarchical A*: Searching abstraction
hierarchies efficiently." in AAAI, 1996, p. 530535.
[7] A. Botea, M. Muller, and J. Schaeffer, "Near optimal hierarchical path-finding," Journal of Game
Development, vol. 1, pp. 7 -- 28, 2004.
[8] H. A. Simon, "The architecture of complexity," Proceedings of the American Philosophical Society,
vol. 106, no. 6, pp. 467 -- 482, 1962.
[9] F. Hauer, A. Kundu, J. M. Rehg, and P. Tsiotras, "Multi-scale perception and path planning
on probabilistic obstacle maps," in IEEE International Conference on Robotics and Automation
(ICRA), Seattle, WA, May 2015, pp. 4210 -- 4215.
[10] D. Jung, "Hierarchical path planning and control of a small fixed-wing uav: theory and experi-
mental validation." Ph.D. dissertation, Georgia Institute of Technology, Atlanta, GA, 2007.
[11] R. V. Cowlagi and P. Tsiotras, "Multiresolution motion planning for autonomous agents via
wavelet-based cell decompositions," IEEE Transactions on Systems, Man, and Cybernetics, Part
B (Cybernetics), vol. 42, no. 5, pp. 1455 -- 1469, Oct 2012.
[12] P. Tsiotras and E. Bakolas, "A hierarchical on-line path planning scheme using wavelets," in 2007
European Control Conference (ECC), Kos, Greece, July 2007, pp. 2806 -- 2812.
[13] M. Otte and N. Correll, "C-FOREST: parallel shortest path planning with superlinear speedup,"
IEEE Transactions on Robotics, vol. 29, no. 3, pp. 798 -- 806, June 2013.
[14] Y. Lu, Y. Huo, and P. Tsiotras, "A beamlet-based graph structure for path planning using
multiscale information," IEEE Transactions on Automatic Control, vol. 57, no. 5, pp. 1166 -- 1178,
May 2012.
[15] R. Nissim and R. Brafman, "Distributed heuristic forward search for multi-agent planning,"
Journal of Artificial Intelligence Research, vol. 51, no. 1, pp. 293 -- 332, 2014.
[16] P. E. Hart, N. J. Nilsson, and B. Raphael, "A formal basis for the heuristic determination of
minimum cost paths," IEEE Transactions on Systems Science and Cybernetics, vol. 4, no. 2, pp.
100 -- 107, July 1968.
|
1302.0488 | 1 | 1302 | 2013-02-03T13:34:08 | A multi-lane traffic simulation model via continuous cellular automata | [
"cs.MA",
"nlin.CG"
] | Traffic models based on cellular automata have high computational efficiency because of their simplicity in describing unrealistic vehicular behavior and the versatility of cellular automata to be implemented on parallel processing. On the other hand, the other microscopic traffic models such as car-following models are computationally more expensive, but they have more realistic driver behaviors and detailed vehicle characteristics. We propose a new class between these two categories, defining a traffic model based on continuous cellular automata where we combine the efficiency of cellular automata models with the accuracy of the other microscopic models. More precisely, we introduce a stochastic cellular automata traffic model in which the space is not coarse-grain but continuous. The continuity also allows us to embed a multi-agent fuzzy system proposed to handle uncertainties in decision making on road traffic. Therefore, we simulate different driver behaviors and study the effect of various compositions of vehicles within the traffic stream from the macroscopic point of view. The experimental results show that our model is able to reproduce the typical traffic flow phenomena showing a variety of effects due to the heterogeneity of traffic. | cs.MA | cs |
A multi-lane traffic simulation model via continuous
cellular automata
Emanuele Rodaro
Centro de Matem´atica, Faculdade de Ciencias, Universidade do Porto,
R. Campo Alegre 687, 4169-007 Porto, Portugal
e-mail: [email protected]
Oznur Yeldan
Dipartimento di Matematica, Politecnico di Milano,
Piazza Leonardo da Vinci 32, 20133 Milano, Italy
e-mail: [email protected]
September 27, 2018
Abstract
Traffic models based on cellular automata have high computational
efficiency because of their simplicity in describing unrealistic vehicular
behavior and the versatility of cellular automata to be implemented on
parallel processing. On the other hand, the other microscopic traffic
models such as car-following models are computationally more expen-
sive, but they have more realistic driver behaviors and detailed vehicle
characteristics. We propose a new class between these two categories,
defining a traffic model based on continuous cellular automata where
we combine the efficiency of cellular automata models with the accu-
racy of the other microscopic models. More precisely, we introduce
a stochastic cellular automata traffic model in which the space is not
coarse-grain but continuous. The continuity also allows us to embed a
multi-agent fuzzy system proposed to handle uncertainties in decision
making on road traffic. Therefore, we simulate different driver behav-
iors and study the effect of various compositions of vehicles within the
traffic stream from the macroscopic point of view. The experimental
results show that our model is able to reproduce the typical traffic
flow phenomena showing a variety of effects due to the heterogeneity
of traffic.
1
1
Introduction
Traffic models are fundamental resources in the management of road net-
work. A real progress in the study of traffic has obtained with the introduc-
tion of the models based on cellular automata (CA)1. A cellular automaton is
a collection of cells (sites) on a grid of specified shape (lattice) that evolves
through a number of discrete time steps according to a set of local rules
based on the states of neighboring cells, for more details see [9]. CA models
(CAMs) have the ability of being easily implemented for parallel computing
because of their intrinsic synchronous behavior. A set of simple rules can
be used to simulate a complex behavior, thus these models are conceptu-
ally simple (for the theory and applications of CA as models for complex
systems in several scientific fields, see [2, 3, 5]). The traffic models based
on CA are capable of capturing micro-level dynamics and relating these to
macro-level traffic flow behavior. However, they are not as accurate as the
usual microscopic traffic models2 such as the time-continuous car-following
ones [1]. A basic one-dimensional CAM for highway traffic flow was first
introduced by Wolfram, where he gave an extensive classification of CAMs
as mathematical models for self-organizing dynamic systems [23, 24]. Nagel
and Schreckenberg proposed the first nontrivial traffic model (the NaSch
model) based on CA for single-lane highway in 1992 [17]. This is a time-
and space-discrete model where the traffic road is divided into cells of 7.5
m and it is defined on a one-dimensional array of a fixed number of sites
with closed (periodic) boundary conditions. The velocity is expressed as the
number of cells that a vehicle advances in one time step. The maximum
velocity (vmax) is assumed to be 5 cells/sec (135 km/h) and every vehicle
has the same target velocity vmax. Each update of the movements of the ve-
hicles is determined by four consecutive rules that are performed in parallel
to each vehicle at each second. In literature, the CA traffic models are up-
dated similarly where they only differ on the randomization step introduced
to add some randomness, for a survey see [15]. Note that in all of them the
cells represent the space as it is in the NaSch model, so in the sequel we call
them as "NaSch-type" models [12, 18, 20, 21].
In this paper, we propose a new traffic model via continuous cellular
automata (CCA) which gets closer to the car-following models by introduc-
1Throughout this paper, abbreviation CA refers to both cellular automata (plural) and
cellular automaton (singular).
2We use the term "usual microscopic traffic models" to refer to the microscopic traffic
models other than CA, which are in general defined by means of a system of differential
equations.
2
ing some continuity without losing the computational advantages typical of
CAMs. Continuous cellular automata (or coupled map lattices) are cellular
automata where the states of the cells are real values in [0, 1], and the local
transition rule is a real function. They have have been introduced by Kaneko
as simple models with the features of spatiotemporal chaos, and have appli-
cations in many different areas like fluid dynamics, biology, chemistry, etc.
(for more details on couple map lattices, see [8, 10]). In this work we drop
the restriction that the set of states is [0, 1], and we use the name CCA to
consider CA where the set of state is in general an infinite set of the same
cardinality of the continuum. Therefore, we consider a hybrid between the
usual microscopic traffic models which are very accurate in predicting gen-
eral traffic behavior but computationally expensive, and the CAMs which
are very efficient due to their simplicity and intrinsic parallelism making
them natural to be implemented for parallel computing. This process of
passing from the typical coarse-granularity of CAMs to the continuity in
space is done with a change of vision where we abandon the cell-space cor-
respondence and assume that cells represent vehicles. Thus we consider a
model with open boundary conditions where the number of cells is equal to
the number of vehicles. In this way, we obtain the immediate advantage of
having less cells to compute (just the number of vehicles). The continuity
also gives us the possibility to refine the microscopic rules that govern the
traffic dynamics, using fuzzy reasoning to mimic different real-world driver
behaviors. All parameters of the decision process of the drivers are modeled
individually by means of fuzzy subsets so that various types of drivers can be
taken into consideration. Hence, we are able to study how the heterogene-
ity influences the traffic macroscopically. The CCA model proposed here is
defined first for a single-lane road and then we extend it to the multi-lane
case. The assumption that cells represent vehicles makes this extension not
as natural as it is in the NaSch-type models.
2 The single-lane model
In our CCA model, for the purpose of simulating different driver (vehicle)
behaviors, we apply a multi-agent fuzzy system where the environmental
information is transformed into the decision for the next time step action
for each vehicle. Using this fuzzy logic-based system, we categorize the
vehicles into types (kinds) where they have common characteristics such
as the same perception of time or distance. In this way, we aim to study
the effect of different composition of vehicles, i.e., the heterogeneity in road
3
traffic. Let us now describe our model for a single-lane road. Note that for
the sake of simplicity, we fix the unit of time as one second. Consider the one
dimensional CCA SL = (Z, Σ, N , δ) where the lattice is the set of integers
and the set of cell states Σ = (K ×R+
0 ×R×{L, 0, R}×{L, 0, R})∪{⊥}.
A cell with the state ⊥ represents a cell without a vehicle. The generic i-th
non-empty cell is in the state σi(t) = (ki, xi(t), vi(t), si(t), di(t), d′
i(t)) where
0 ×R+
• ki represents the kind consisting of all the information (parameters)
specified differently for each kind of vehicle, such as: the maximum
velocity vmaxi, the optimal velocity vopti (the velocity which each kind
of vehicle feels comfortable in traffic stream), the length li, the fuzzy
membership functions, the maximum stress smaxi, the minimum stress
smini, the probability functions of lane-changing to the right lane
PRi(x) and to the left lane PLi(x).
• xi(t) is the position, defined as the distance from the origin of the
road to the middle point of this vehicle, vi(t) is the velocity, and si(t)
is the stress, a variable to keep track of how much the driver is above
or below of his optimal velocity. The stress parameter is introduced
to implement a more realistic driver behavior since we make the as-
sumption that drivers usually tend to decelerate when they are moving
with a velocity higher than their optimal velocity. We assume that if
si(t) < 0 then the driver is stressed and wants to go faster, instead
if si(t) ≥ 0 then his stress derives from the desire to go slower. Note
that the main usage of si(t) is related to the lane-changing process
described in Section 3.
• di(t) is the variable describing the desire for lane-changing to the left
"L", to the right "R" and staying on the same lane "0". d′
i(t) is
the variable showing from which lane the i-th vehicle is transferred:
from the left lane "L", from the right lane "R" or not transferred
"0". Although these variables are used in the multi-lane model, the
evaluations are done in the single-lane part.
N is a kind of one-dimensional extended Moore neighborhood defined by
N (i) = (i − 1, i, i + 1, i + 2), and δ : Σ4 → Σ is the local transition function
(local rule) acting upon a cell and its direct neighborhoods. The local rule
is defined componentwise by δ(σi−1(t), σi(t), σi+1(t), σi+2(t)) = (ki, xi(t +
1), vi(t + 1), si(t + 1), di(t + 1), 0). The variables si(t + 1) and di(t + 1) are de-
scribed in Section 3. The space is updated as usual by xi(t+1) = xi(t)+vi(t+
1) and the velocity by vi(t + 1) = min(vmaxi, ∆x+
i (t), max(0, vi(t) + Ai(t)))
4
where Ai(t) is the acceleration calculated using the fuzzy decision modules
and ∆x+
i (t) is the distance with the front vehicle from front bumper to rear
bumper defined in (1). This front distance constraint in the choice of the
updated velocity is introduced to make the model collision-free as in the
NaSch model. However, since it is unrealistic, it is applied in our model
only in the borderline situations where extreme decelerations are involved.
Indeed, we have tried to avoid this constraint as much as possible by intro-
ducing enough fuzzy rules to make the system more reactive to reduce these
extreme situations. It is worth noting that in the case Ai(t) = 7.5 m/s2, we
essentially obtain the open boundary, space-continuous and deterministic
version of the NaSch model [17].
Ai(t) is depending on the kind ki, the velocity vi(t) and the variables
defined as following.
1. Back Distance (BD): ∆x−
i (t) = xi(t) − xi−1(t) − li
2 − li−1
2 ,
2. Front Distance (F D):
∆x+
i (t) = xi+1(t) − xi(t) −
li+1
2
−
li
2
,
(1)
3. Next Front Distance (N F D): ∆x+
2 − li
2 .
This is introduced to have a better perception of the driver behavior
of the next front vehicle since in reality drivers observe not only the
vehicle just in front of them but also the vehicles ahead.
i,N (t) = xi+2(t) − xi(t) − li+2
4. Perceived Front Collision Time (P F CT ):
i,P (t) = (cid:26) ζi(t),
τ +
min(ζi(t), τ +
i (t)),
i (t) < 0,
if τ +
otherwise.
smaxi
−si(t)
vi(t)
and τ +
where ζi(t) is the parameter introduced for slowing down depending on
the stress calculated by ζi(t) =
i (t) is Front Collision
Time (F CT ), the time that passes for the i-th vehicle to reach to
(to collide with) the front vehicle, defined as τ +
vi(t)−vi+1(t) .
P F CT is a parameter which is a combination between FCT and an
auxiliary time defined to keep the velocity close to the optimal velocity
using a parameter depending on the stress variable.
If the FCT is
strictly negative, then the driver perceives an artificial vehicle with the
FCT ζi(t) to keep the vehicle close to its optimal velocity. Otherwise,
P F CT takes the smallest FCT value between ζi(t) and τ +
i (t) to avoid
collisions.
i (t) =
∆x+
i (t)
5
5. Worst Front Collision Time (W F CT ): τ +
. The colli-
sion time with the front vehicle in the case it suddenly stops. This
parameter is introduced for safety reasons.
vi(t)
i,W (t) = ∆x+
i (t)
∆x+
i,N (t)
i,N (t) =
6. Next Front Collision Time (N F CT ): τ +
vi(t)−vi+2(t) . The colli-
sion time with the next front vehicle. It is clear that a vehicle never
reaches to its next front vehicle, but N F CT is introduced to anticipate
the braking maneuver of next front vehicle.
i (t) = ∆x−
vi−1(t)−vi(t) . The collision time
of the back vehicle. This is introduced to take into account the phe-
nomenon where the back driver comes closer the front driver and forces
him to accelerate. We call this situation as pushing effect.
7. Back Collision Time (BCT ): τ −
i (t)
The approach of embedding fuzzy logic while dealing with a system de-
scribed by continuous variables has been already introduced, indeed, there
are several works based on fuzzy logic systems in car-following models [1].
In our model, we use a fuzzy rule-based system by means of fuzzy IF-THEN
rules [16] that determines the behavior of the vehicles in traffic stream. In
this fuzzy system we introduce two fuzzy modules, formed by two sets of
fuzzy rules, according to their significance. The first module has more im-
portance since it is based on the information related to the front and the
back vehicles while the second one has less importance since the rules in that
module are related to the next front vehicle. As a consequence of having
two different fuzzy modules, we have two outputs: Ai,1 is the output coming
out from the first fuzzy module and Ai,2 is the output coming out from the
second fuzzy module. The first and the seconds sets of fuzzy rules used in
the simulator are summarized in Table 1 and Table 2, respectively. After re-
ceiving the inputs τ +
i (t),
vi(t) determined by the environment, the final decision of the acceleration
Ai(t) is returned by the function (to avoid cumbersome notation, we omit
the dependency of t):
i,N (t), ∆x−
i (t), ∆x+
i (t), ∆x+
i,W (t), τ +
i,N (t), τ −
i,P (t), τ +
F (Ai,1, Ai,2) =
min(Ai,1, Ai,2)
Ai,1+Ai,2
2
Ai,1
if Ai,1 ≤ 0,
if Ai,1 > 0 ∧ Ai,2 ≤ −0.25,
otherwise.
(2)
where we give more weight to the decision taken by the first module. Indeed,
in the simulation, we have noticed that without these kind of constraints
the vehicles were tending to slow down too much. This is also the reason of
splitting the fuzzy module into two parts. Moreover, the second module and
6
consequently the second acceleration gains significance only in the case of
emergencies such as a sudden breakdown or deceleration of the next front ve-
hicle. The linguistic terms (properties) used for the acceleration outputs Ai
are: PB=Positive Big, PM=Positive Medium, PS=Positive Small, Z=Zero,
NS=Negative Small, NM=Negative Medium, and NB=Negative Big.
Table 1: First fuzzy set of rules
PFCT
BIG
MEDIUM
SMALL
VERY SMALL
WCT
SMALL
Front Distance (FD)
Velocity (vi)
BIG
MEDIUM
SMALL VERY SMALL
SMALL
(vi NOT SMALL)
(vi NOT SMALL)
ZERO
ZERO
PB
PM
ZERO
NM
NB
PS
ZERO
NM
NB
NS
NS
NM
NB
NM
NS
NM
NB
NM
(Jam Situation)
In the first fuzzy set of rules we have also the ones related to the back
vehicle, thus we consider the pushing effect where BCT is "Very Small"
and BD is "Very Small". In this case, we apply the following rule formed
by taking into consideration the safety with the front vehicle: (IF P F CT is
BIG AND F D is BIG) OR (IF P F CT is BIG AND F D is M EDIU M ) OR
(IF P F CT is M EDIU M AND F D is BIG) OR (IF P F CT is M EDIU M
AND F D is M EDIU M ) then the acceleration choice is P S.
Table 2: Second fuzzy set of rules
Next Front Distance (NFD)
VERY SMALL SMALL MEDIUM BIG
NFCT
VERY SMALL
SMALL
MEDIUM
BIG
NB
NM
NS
NS
NB
NM
NM
NS
NM
NS
These rules are based on some common sense of driver behaviors in-
cluding some experiences. Although the rules are the same for each kind
of vehicle, they have different "weights" depending on the definition of the
membership functions of different kinds. In this way, it is possible to give a
description of a variety of behaviors, such as the behavior of a long vehicle
driver or a driver with low reflexes. For instance, a person with low reflexes
7
perceives a time of collision of 5 seconds as a very short time, however a
person with higher reflexes probably would feel comfortable with that time
of collision.
After the fuzzy system receives all the inputs, we first determine the
degree of membership of each input. Then by taking the minimum value of
the images of the inputs (corresponding to the fuzzy logic AND operation)
in each rule, we obtain the weights of the fuzzy rules. This evaluation is
described in general terms as following.
and C j be fuzzy subsets with the membership functions
and µC j , respectively, and let Rj be the fuzzy rules defined
1, ..., Bj
kj
Let Bj
, . . . , µBj
kj
µBj
as:
1
Rj : IF xj
1 is Bj
1 AN D . . . AN D xj
kj
is Bj
kj
T HEN y is C j
with
µBj
1
(xj
1) ∧ . . . ∧ µBj
kj
(xj
kj
) = min(cid:26)µBj
1
(xj
1), . . . , µBj
kj
(xj
kj
)(cid:27)
for 1 ≤ j ≤ m, where m is the number of fuzzy rules.
In our fuzzy system, there are also rules including the form: xj
r is not Bj
r
r) for some r ∈ [1, kj ]. These rules
with the degree of membership µ∼Bj
which are the first two rules of the first fuzzy decision module, are formed
with the purpose of emphasizing the vehicle is not in a jam situation (see
Table 1 where vi is not small). For an input xj
r, the degree of membership
of not having a property Bj
r ") can be written as:
r ("not being Bj
(xj
r
µ∼Bj
r
(xj
r) = 1 − µBj
r
(xj
r).
The output of the previous process so far is a fuzzy set, so we should con-
vert our fuzzy output set into one single number as the output of the fuzzy
system, which is the "acceleration" in our case. Recall that since we have
two fuzzy modules, there are two outputs of the system: Ai,1 and Ai,2 which
are used to evaluate Ai(t) at time t by the function F , see Equation 2. The
conversion process is called as defuzzification and there are many defuzzifica-
tion techniques to obtain a crisp value, such as the center-of-gravity (COG)
and the weighted average formula (WAF) method. In our model, since we
would like to use a simple method that does not require too much compu-
tational power, we define a new method and call it as generalized weighted
average formula (GWAF), i.e., we generalize the WAF to the case where the
membership functions are not necessarily symmetric. We describe it now in
8
general terms. Suppose that each j-th rule receives the values xj
inputs. Let
1, ..., xj
kj
as
wj = min(cid:26)µBj
1
(xj
1), ... , µBj
kj
(xj
kj
)(cid:27)
be the weights for each rule j and let P j = µ−1
C j (wj ) be the preimage of the
weight of j-th rule. We assume that µC j does not have any plateau, since
we do not want P j to contain intervals (this is done to have a discrete set).
The defuzzified output is thus calculated by:
y = Pm
j=1 wj Pz∈P j z
Pm
j=1 P j wj
.
3 The multi-lane model
In this section, we extend the single-lane model to the multi-lane case. This
extension is not trivial as it is in the NaSch-type models where adding a
lane simply means adding an array of cells and where the local transition
function can naturally be extended. This is a consequence of having a clear
physical interpretation of the model given by the fact that space is rep-
resented by cells. However, as a consequence of not having the cell-space
correspondence, first of all we do not have the natural order which makes the
extension of the local transition function so easy to achieve, and secondly
not all the configurations c : Z → Σ of our model represent a physical situa-
tion. For this reason, we consider just the configurations that have physical
meaning, i.e., two adjacent cells n, n + 1 are in the states where the positions
fulfill the constraint xn(t) < xn+1(t). We denote the set consisting of such
configurations by Confp(SL). The support of a configuration c ∈ Confp(SL)
is a maximal interval [i, j] of integers with i ≤ j such that for all k < i and
k > j, c(k) =⊥.
The union of two arrays of cells representing a road with two lanes is
a natural candidate for a multi-lane model. Therefore, we first present
our multi-lane model as a union of interacting single-lane CCA where the
interaction is given by a transfer operation, and then we prove in Prop. 3.1
that this model can actually be simulated by a CCA.
The process of lane-changing behavior consists of several steps. First the
driver feels the desire of lane-changing, then if the conditions of the place
in the lane where the driver wants to move respect some safety constraints,
the driver can finally perform the maneuver. Regarding the decision of
performing a lane-changing, we first check if the i-th vehicle with the state
9
σi(t) = (ki, xi(t), vi(t), si(t), di(t), d′
i(t)) desires to change lanes, i.e., di(t) 6=
0. The desire of lane-changing is depending on the stress variable si(t).
Having positive stress means that the driver is above of his optimal velocity,
thus he tends to change lane to the right (slower lane). Otherwise, the driver
is nervous and he desires to go faster, thus he tends to change lane to the
left (faster lane). The stress cannot be increased (or decreased) arbitrarily,
so we define the two parameters smini and smaxi which are the maximum
amounts of negative and positive stress that a driver can tolerate. Thus
if the updated stress si(t + 1) exceeds the boundaries smini and smaxi we
simply assign it to these limit values. The stress is updated by
si(t + 1) =
sacc(t)/2
sacc(t) · (1 + Φ)
sacc(t)
i (t) < 0 and
i (t) ≥ 0 and
if τ +
if τ +
otherwise.
smini
2 < sacc(t) < 0,(3)
2 < sacc(t) < 0,(4)
smini
where the accumulated stress is defined by sacc(t) = si(t)+(vi(t+1)−vopti )·
X(t) and X(t) is a random variable distributed uniformly, X(t) ∼ U (0, 1).
Condition (3) is used to avoid frequent lane-changing in the case of a jam
situation, if the queue is moving. We also embed the strategy of trying to
change lanes instead of braking in the case the front vehicle is close and
tends to brake. We model this effect with the condition (4) considering a
factor Φ representing how much the i-th vehicle is in the situation where
the front vehicle is close and going much more slower than the i-th vehicle.
Φ is calculated by the membership value of (τ +
i (t) is V ERY SM ALL ∧
∆x+
i (t) is SM ALL) ∨
(τ +
i (t) is
SM ALL). It is used to increase the amount of stress and consequently to
have more probability of changing lanes.
i (t) is M EDIU M ) ∨ (τ +
i (t) is SM ALL∧∆x+
i (t) is V ERY SM ALL ∧ ∆x+
i (t) is M EDIU M )∨(τ +
i (t) is SM ALL∧∆x+
We now describe the stochastic process that evaluates the desire of lane-
changing for the next time step di(t+1) = Eval(L,R)(ki, vi(t), si(t)) described
in Algorithm 1. The decision of this action is made by means of a Bernoulli
process B(2, p). The probabilities of changing lane to the left and to the
right, which are in general different, are calculated by two functions PLi(x) :
[0, 1] → [0, 1] and PRi(x) : [0, 1] → [0, 1] contained in the kind ki. The reason
is that some kinds of vehicles tend to move more to the left lane while some
others tend to move more to the right. For instance, long vehicles prefer to
go to the right lane more than left, i.e., they tend to stay more on the right
lane. The variable used to calculate such probabilities is the normalized
stress nsi(t) = max(si(t)/smini , si(t)/smaxi ). Thus if si(t) ≥ 0 then the
choice to stay or go to the right is done by B(2, PRi (nsi(t))), otherwise
10
else
di(t + 1) = R
di(t + 1) = 0
if si(t) ≥ 0 then
end if
else
Execute the Bernoulli Trial Z ∼ B(2, µvels(vi(t)))
if Z = 1 then
nsi(t) = si(t)/smini
Execute the Bernoulli Trial Y ∼ B(2, PLi (nsi(t))
if Y = 1 then
nsi(t) = si(t)/smaxi
Execute the Bernoulli Trial X ∼ B(2, PRi (nsi(t))
if X = 1 then
Algorithm 1 The pseudo-code for evaluating Eval(L,R).
1: procedure Eval(L,R)(ki, vi(t), si(t))
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35: end procedure
end if
Execute the Bernoulli Trial W ∼ B(2, 0.7)
if W = 1 then
if R = 0 then
di(t + 1) = L
end if
if L = 0 then
di(t + 1) = R
end if
end if
else
end if
else
end if
di(t + 1) = L
di(t + 1) = R
di(t + 1) = L
else
di(t + 1) = 0
by B(2, PLi(nsi(t))) to go to the left.
If the latter process is resulted as
going to the left, we take into consideration also the jam situation where the
drivers randomly move to the left or right to find an emptier lane. For this
purpose we use the velocity parameter and we need the fuzzy set of "Velocity
SMALL" with the membership function denoted by µvels to perform this
evaluation which is done by the Bernoulli process B(2, µvels(vi(t))). If the
result of the outcome is positive, meaning that the traffic is indeed jammed,
the decision for lane-changing depends on the relative position of the lane,
i.e., if the vehicle is in the most left (right) lane then it can only move to
the right (left). For this reason, the single-lane model SL(L,R) depends on
two parameters L, R which are Boolean variables used to describe whether
or not there exists a lane on the left or on the right, respectively. Thus if
there is no lane on the left (right) of the driver, i.e., L = 0 (R = 0), he can
only move to the right (left) lane. If there are lanes both on the left and
11
right, the choice is obtained using the Bernoulli process B(2, 0.7) where the
decision is "moving to the left lane" with the probability of 0.7 and "moving
to the right lane" otherwise. These probabilities are decided to be different
since we make the assumption that the drivers desiring to go faster usually
tend to move to the left lane more than the right even in a situation of traffic
congestion.
After the desire of lane-changing, we perform the transfer of a vehicle.
The transfer clearly changes the configuration of the single-lane CCA, thus
we need to introduce some operations to describe this process. Given a vehi-
cle represented by a state σ = (k, x, v, s, d, d′) ∈ Σ \ {⊥} and a configuration
c ∈ Confp(SL), the inserting operator at position n ∈ Z is the function
Insn : Σ \ {⊥} × Confp(SL) → Confp(SL) that changes the configuration
from c into Insn(σ, c) by shifting the states of all cells one step starting
from the n-th position and setting the state of the n-th cell to the value σ.
The right-inverse of Insn is the deleting operator at position n denoted by
Deln : Confp(SL) → Confp(SL). The index n gives the position where we
insert or delete the content of a cell. The function Insn(σ, c) is well defined
only if the insertion of the vehicle still generates a physical configuration.
However, before the insertion, we check what is the relative position of the
vehicle that wants to perform a lane-changing with respect to the lane that
it is going to enter. For this reason, we need to define an index operator
Indx : Σ \ {⊥} × Confp(SL) → Z such that Indx(σ, c) represents the index
where the vehicle with state σ would go if we would try to insert it into the
configuration c.
In our model, the process of lane-changing is based on some safety crite-
rions which check the possibility of executing a lane-changing by considering
the situation in the target lane. These criterions, obtained from some simu-
lation experiments, guarantee that after the lane-changing there will be no
danger (avoidance of collision) with the back and the following vehicle on
the target lane. Consider the vehicle represented by σ that wants to enter
in a lane with configuration c, thus the corresponding position in the target
lane is between the j − 1-th and j-th vehicle with velocities vj−1 and vj
where j = Indx(σ, c). Suppose that the back and front distances of this
entering vehicle with respect to the j − 1-th and j-th vehicle are denoted by
∆− and ∆+, respectively, then the safety constraints are:
∆− > max(0, v1.2
j−1 − v + vj−1 − v + 3), ∆+ > max(0, v1.25 − vj + 3)
(5)
The condition for performing a lane-changing to the left (right) lane can
be summarized by the Boolean variable transL(σ, c) (transR(σ, c)) which
12
consists of the safety criterion (5), the desire to change lane to the left
(right), i.e., d = L (d = R) and not being already transferred, i.e., d′ = 0 (to
avoid multiple transfers in one step of simulation). In the process of lane-
changing, we also need to keep trace where the vehicle comes from because
the transfer of a vehicle is seen as a process consisting of copying and erasing
steps. For this purpose we define the copy of σ as:
σcp = (cid:26) (k, x, v, s/5, L, R)
(k, x, v, s/5, R, L)
if σ = (k, x, v, s, L, 0),
if σ = (k, x, v, s, R, 0).
where we decrease the stress from s to s/5, firstly because we make the
assumption that when a vehicle changes lanes there is a sense of satisfaction
reducing the stress, and secondly this change helps to reduce the ping-pong
phenomenon [12, 20] of vehicles continuing to move from one lane to another
in case of traffic congestion. If a vehicle with state σ desires to go to the left
(right) lane, then we copy σ into c, i.e., we update c into a new configuration
denoted by σ L c (σ R c)∈ Confp(SL) defined by the following equation
(X ∈ {L, R}):
σ X c = (cid:26) Insj(σcp, c)
c
if transX(σ, c), where j = Indx(σ, c)
otherwise.
Besides, we need an operator to update a configuration c′ in which we have
to erase the state that has already been copied into c. The erasing procedure
is accomplished by changing c′ into the new configuration c′ \ω ∈ Confp(SL)
defined as:
Deli(c′)
if (c′(i) = (h, y, w, 5r, R, 0) ∧ ω = (h, y, w, r, R, L))
∨(c′(i) = (h, y, w, 5r, L, 0) ∧ ω = (h, y, w, r, L, R)),
where i = Indx(ω, c′)
c′
otherwise.
c′ \ ω =
We extend these functions to the binary operations L, R and \ on
Confp(SL) to describe the transfer process of all the vehicles from one
lane to another. We call L and R as the left and right copying op-
eration, respectively, and \ as the erasing operation. The extension is made
in the following way. Let c, c′ ∈ Confp(SL) be two configurations with
c(i) = ωi 6=⊥ for i = m, . . . , M and c′(i) = σi 6=⊥ for i = n, . . . , N , where
[m, M ], [n, N ] are the supports of c, c′, respectively. We define the config-
urations c′
let e0 = c and ek := σk L ek−1
for k = 1, . . . , N − n + 1 then we define c′
L c := eN −n+1 (analogously
for R). On the other hand, for c′ \ c, let g0 = c′ and gk := gk−1 \ ωjk for
L c inductively as follows:
13
for i = 0 → M − 1 do
if i = 0 then
c1 := (c0 R c1)
c0 := c0 \ (c0 R c1)
end if
if 0 < i < M − 1 then
ci−1 := (ci L ci−1)
ci := ci \ (ci L ci−1)
ci+1 := (ci R ci+1)
ci := ci \ (ci R ci+1)
if i = 1 then
Algorithm 2 The pseudo-code for one time step evolution of the multi-lane
model
1: procedure Update(c0, . . . , cM −1)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25: end procedure
cM −2 := (cM −1 L cM −2)
cM −1 := cM −1 \ (cM −1 L cM −2)
cM −2 := δ∗
cM −1 := δ∗
end if
if i = M − 1 then
(1,1)(cM −2)
(1,0)(cM −1)
ci−1 := δ∗
(1,1)(ci−1)
c0 := δ∗
(0,1)(c0)
else
end if
end if
end for
k = 1, . . . , M − m + 1 then c′ \ c := gM −m+1. Using these operators it is easy
to see that the process of the transfer of vehicles from c′ to c, where c is the
configuration of a lane on the left of the lane with configuration c′, can be
easily described by transforming c into c′
L c and c′ into c′ \ (c′
L c).
We now present our multi-lane model using these copying and erasing
operations. Suppose that there are M ≥ 2 lanes on a highway. We model
each lane using the single-lane CCA model SL. Recall that SL depends on
the parameters L, R, thus we rewrite it as SL(L,R) = (Z, Σ, N , δ(L,R)). In
this way, we can associate to the road the ordered M -tuple:
(SL(0,1), SL1
(1,1), . . . , SLM −2
(1,1) , SL(1,0))
where if M ≥ 3 we have M − 2 copies of SL(1,1), and where SL(0,1) and
SL(1,0) represent the left- and the right-most lane, respectively. In the case
M = 2, we consider just the pair (SL(0,1), SL(1,0)). Suppose that these M
number of CCA are in the configurations (c0, . . . , cM −1) ∈ Confp(SL)M . In
our multi-lane model, we scan each lane and we transfer the vehicles to the
adjacent lanes. After this process, for each lane we apply the single-lane
CCA model to update the configuration and this update is done by means
14
of the global transition function of SL(a,b) denoted by δ∗
(a,b), a, b ∈ {0, 1}.
In this way, we have a new array of configurations (c′
M −1), and this
process represents 1 sec. of simulation. Moreover, the order with which the
transfer is performed is from the left-most lane to the right-most one, and
this is done to satisfy the precedence requirement in European roads. In
Algorithm 2, it is described the updating process U pdate : (c0, . . . , cM −1) 7→
(c′
M −1). We now show that Algorithm 2 can be simulated by a CCA
which implies that our multi-lane model is framed as a CCA model.
0, . . . , c′
0, . . . , c′
Proposition 3.1. There is a stochastic CCA multi-lane model which sim-
ulates Algorithm 2.
Proof. We define the stochastic CCA:
ML = (Z, Ω, M, ∆)
where
• Ω = (Confp(SL) × {copy, erase} × N3) ∪ {⊥}, where ⊥ is the state
associated to the empty cells representing cells with no lane.
• M(i) = (i − 1, i, i + 1) is the von Neumann neighborhood.
• ∆ : Ω3 → Ω is the local transition function such that
∆(ω−1, ω0, ω1) = ω′
0
defined in Algorithm 3 with inputs
ωj = (cj, Xj, Mj, Pj , Kj), j = −1, 0, 1.
If we consider M lanes with the configurations c0, . . . , cM −1, we associate to
ML = (Z, Ω, M, ∆) the configuration
C = (ω0, . . . , ωM −1)
where ωi = (ci, copy, M, i, 0) for i = 0, . . . , M − 1.
It is easy to see that
applying 2M times the global transition function ∆∗ to C, we obtain a new
configuration
∆∗2M
(C) = (ω′
0, . . . , ω′
M −1)
with ω′
i = (c′
i, copy, M, i, 0) for i = 0, . . . , M − 1 where
(c′
0, . . . , c′
M −1) = U pdate(c0, . . . , cM −1),
and U pdate(c0, . . . , cM −1) is the function defined in Algorithm 2.
15
end if
if X0 = copy then
end if
X0 := erase
else
end if
X0 := erase
else
if K0 = 0 then
if X0 = copy then
if ω0 =⊥ then
ω′
else
0 =⊥
if P0 = K0 then
c0 := (c0 \ c1)
if P0 = K0 + 1 then
c0 := (c−1 R c0)
if P0 = K0 − 1 then
c0 := (c1 L c0)
end if
if P0 = K0 + 1 then
c0 := (c−1 R c0)
end if
if 0 < K0 < M0 − 1 then
end if
K0 := K0 + 1 mod M0
X0 := copy
exit
Algorithm 3 The pseudo-code to compute the local transition function ∆.
1: procedure ∆ (ω−1, ω0, ω1)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
52:
53:
54:
55:
56:
57:
58:
59:
60:
61:
62:
63: end procedure
end if
K0 := K0 + 1 mod M0
X0 := copy
exit
end if
if K0 = M0 − 1 then
if X0 = copy then
end if
K0 := K0 + 1 mod M0
X0 := copy
end if
if 0 < P0 < M0 − 1 then
end if
if P0 = M0 − 1 then
c0 := δ∗
(1,0)(c0)
if P0 = K0 − 1 then
c0 := (c1 L c0)
if P0 = K0 then
c0 := (c0 \ c−1)
end if
if P0 = 0 then
c0 := δ∗
(0,1)(c0)
if P0 = K0 then
c0 := (c0 \ c−1)
c0 := (c0 \ c1)
end if
end if
end if
end if
X0 := erase
else
end if
c0 := δ∗
(1,1)(c0)
16
4 The experiments and the simulation results
The model presented here is a CAM which is intrinsically parallel. We first
implement the model using the programming language Python since it is
a high level language making the implementation of the model faster and
easier. Indeed, during the phase of the development of writing the code, we
use an object-oriented philosophy, especially while passing from the single-
lane to multi-lane case. Thus, it is easier to modify and to rewrite some
parts of the code to tune it better. Using this code we run a series of
experiments for a first evaluation of the model. Part of the data obtained
by these experiments have been presented in the EWGT 2012 conference
[26], and the complete version is contained in the PhD thesis [25]. We also
implement the model using CUDA to take advantage of the power of the
modern graphic cards to perform parallel computation. Indeed, the single-
lane model can be naturally parallelize simply by giving a cell (a vehicle in
our case) to each thread of the GPU. In Section 5 we give a brief description
of this process.
4.1 Setting the kinds of vehicles and the experiment scenar-
ios
In this section, we describe the setting up conditions for the kinds of vehicles
used in a series of simulations. In our experiments, we use just two kinds of
vehicles which we call as passenger vehicles and long vehicles. The vehicles
have the following parameters:
• passenger vehicles: vmax = 36 m/s, vopt = 28 m/s, the length 4 m,
the natural acceleration noise: 0.2 m/s2 (see [7]), smax = 500 m,
smin = −450 m, the function of the probability of lane-changing to the
right lane PR(x) = x, the function of the probability of lane-changing
to the left lane PL(x) = x.
• long vehicles: vmax = 25 m/s, vopt = 20 m/s, the length 9 m, the
natural acceleration noise: 0.1 m/s2 (see [7]), smax = 300 m, smin =
−700 m, the function of the probability of lane-changing to the right
lane PR(x) = x, the function of the probability of lane-changing to the
left lane PL(x) = x1.25.
17
Figure 1: The membership functions of the passenger vehicles for the front
distance
Regarding the fuzzy membership functions, we tune them according to a
questionary posed to a group of drivers. The results represent how much the
claims are related to each interviewee's perception while driving. We then
consider the corresponding membership functions by interpolating linearly
these data. However, in the simulations, we do not use these functions
based on the data. Instead, we approximate them with the triangular and
trapezoidal membership functions which are more common and easier from
the computational point of view. We notice from the experiments that this
kind of approximation do not alter the results of the simulation significantly.
Firstly, because the sample is too small so the data obtained is subject to
errors, and secondly, the obtained membership functions are not so different
from their approximated versions and the simulation can gain a boost of
a factor 1.5.
In Figure 1, we see one of the membership functions of the
passenger vehicles for the perception of front distance adjusted according to
the collected data and the approximated version, respectively.
For the first evaluation of our model, we simulate a piece of highway
near an off-toll plaza. The parameters used in the simulations are:
• Length of the road L
• Number of lane M
• Number of iterations (the simulation time)
• Number of repetitions of the same experiment
• Emission rate λ: the number of vehicles entering this piece of highway
• Influence radius ρ of the off-toll plaza, note that ρ = −1 means that
there is an open road tolling
18
• Obstacle: An obstacle can be placed on the left-most lane or on the
right-most lane
We model the entrance of the each lane using a Poisson stochastic process.
The probability of emitting at least one vehicle in a lane in the time interval
[0,t] is calculated by a Poisson distribution. If the random test succeeds,
i.e., if the result is "emit a vehicle" then we randomly choose a kind of
vehicle where the choice is a "passenger vehicle" with the probability of
1 − p and a "long vehicle" with the probability of p, and in the experiments
we consider p = 0%, 10%, 20%, 30%. The emission rate is set to λ/M for
each lane where we consider λ = 0.25, 0.5, 1, 1.5, 2 veh/s. For instance, if
λ = 1.5 and M = 3, the average number of vehicles entering the piece
of highway is 90 veh/min which means that each lane at the beginning of
this road is charged of around 30 veh/min corresponding to a situation of
heavy traffic since the maximum capacity of a lane is considered around
40 veh/min [14]. Regarding the off-toll plaza, we introduce a parameter
called influence radius which is the interval of the road segment in which
the vehicles are captured when they are exiting. This parameter is used to
control the processing time of each vehicle which influences the throughput
(the number of vehicles processed in every 10 seconds). The influence radii
considered are ρ = 10, 25, 50 m and "-1" which means that the off-toll plaza
is not visible to the vehicles, thus they simply act as if there is no any off-
toll plaza. This situation corresponds to the case of an open road tolling
payment system where the vehicles do not need to slow down and stop to
make payment. Every 10 seconds the simulator checks the number of vehicles
processed (exited) and it calculates the average latency, i.e., the average of
the time spent to travel from the entrance to the exit. We also place an
obstacle in some experiments either on the left- or on the right-most lane
with a dimension of 2L/5 in the middle of the piece of highway to analyze
a bottleneck phenomenon.
In the simulations, each experiment has 1000
iterations and it is repeated 100 times. The execution time takes around
6.6 hours using one core of a computer equipped with a 16 core Xeon at 2.7
GHz (X5550) with 16 GB of RAM running Debian Linux (kernel 2.6.32).
4.2 Analysis of the experimental results
The macroscopic parameters considered for the purpose of analyzing the
general behavior of traffic are evaluated as following. The density describing
the number of vehicles per unit length of the piece of highway (measured
in vehicles per meter) at time t is calculated by D(t) = N (t)
L where N (t)
19
is the total number of vehicles at time t and L is the length of the road
representing a piece of highway (5 km long) near the off-toll plaza. The
number of vehicles is not constant since the entering rate of the vehicles
(depending on a Poisson stochastic process) is different from the exiting
one, so we obtain different densities. In our experiments, the density can
reach at 0.25 veh/m/lane as maximum since the length of a passenger vehicle
is assumed to be 4 m long. The average velocity, i.e., the averaged sum of
i=1 vi/N (t), and finally,
the velocities at time t is calculated by vav(t) = PN (t)
the flow is evaluated by q(t) = D(t) · vav(t).
The analysis of traffic flow is typically performed by constructing the
fundamental diagram (the flow-density diagram) that determines the traffic
state of a roadway by showing the relation between flow and density. There-
fore, the computations described above are used to plot the fundamental di-
agrams where we analyze different traffic phenomena, see for instance Figure
2.
In the simulation results, we notice that the different composition of
vehicles obtained by changing the percentages of the long vehicles is an im-
portant factor in influencing the flow. It is seen in Figure 2 that adding even
a small percentage of long vehicles to the traffic stream changes the funda-
mental diagram significantly. The heterogeneity effects also the throughput
as it is predictable since long vehicles are slower and so in the queue near
the off-toll plaza it takes more time to move and to get processed.
The experiments show that the model is able to reproduce the typical
traffic flow physical phenomena such as the three phases of traffic flow [11]:
Free flow, synchronized flow and wide-moving jam, see Figure 3. Free flow
corresponds to the region of low to medium density and weak interaction
between vehicles. In general, the slope of the fundamental diagram in the
free flow phase is related to the speed limit, meaning that in this phase
vehicles can move almost at the speed limit. Instead, in the free flow phase of
the fundamental diagram in our experimental results, the slope is related to
the optimal velocity, meaning that in this phase the vehicles can move almost
at the optimal velocity. The reason is that in our model the vehicles do not
aim to reach to the maximum velocity as it is in the NaSch-type models,
but they tend to go with their optimal velocity. Free flow is characterized
with a strong correlation and quasi-linear relation between the local flow
and the local density [19]. The synchronized flow presents medium and
high density while the flow can behave free or jammed. In other words, it
is defined by the interaction between the vehicles and is characterized by
an uncorrelated flow-density diagram. However, this phase is not clearly
20
Figure 2: Comparison of fundamental diagrams with 3 and 4 lanes, respec-
tively
21
Figure 3: Traffic phases in the fundamental diagram: Free flow, synchronized
flow and wide-moving jam, and the cross-covariance between the flow and
density
understood in the context of CA and not observed in most of the NaSch-
type models. It probably requires the presence of sources (on-ramps, on-toll
plaza) and sinks (off-ramps, off-toll plaza), see [22]. The wide-moving jam
phase represents the situation where the traffic is jammed (congested). In
this phase, an increase in the density results with a decrease in the flow. Let
us now consider the cross-covariance between the flow q(t) and the density
D(t), see Figure 3:
cc(q, D) =
hq(t)D(t)i − hq(t)ihD(t)i
phq(t)2i − hq(t)i2phD(t)2i − hD(t)i2
where the brackets h·i indicate averaging the values obtained in all the ex-
periments at time t. In the free flow phase, the flow is strongly related to
the density indicating that the average velocity is nearly constant. For large
densities, in the wide-moving jam phase, the flow is mainly controlled by
density fluctuations. There is a transition between these two phases where
the fundamental diagram shows a plateau when the cross-variance is close
to zero. This situation with cc(q, D) ≈ 0 is identified as the synchronized
flow [13, 19].
22
It is seen in Figure 3 that we reproduce these relations where the free
flow phase initiates with the situation of cross-covariance close to 1 and
continues with a positive cross-covariance, the synchronized flow phase is in
the region where the cross-covariance is close to zero and the wide-moving
jam phase corresponds to the situation where the cross-covariance is negative
(anticorrelation).
The plateau formation has the dependency also on the throughput. In
Figure 4, it is seen this situation where we have compared the fundamen-
tal diagrams for different influence radii which are related to the average
throughputs. Observing any of these three plots, we see that when the aver-
age throughput is increased, there occurs a more immediate transition from
the synchronized flow phase to the wide-moving jam phase. In other words,
the phase-change between these two flows occurs with a higher density with
the decrease of the average throughput. Furthermore, when we compare the
three plots in Figure 4, we see that with more long vehicle percentage we
have the phase-change with a higher density. The same phenomenon is also
shown with arrows in Figure 2.
This phenomenon is inverted in the presence of an obstacle (on the right-
most lane3). More precisely, when an obstacle is set we observe that the
phase-change occurs with a lower density, see Figure 5. This phenomenon is
more clear with the presence of long vehicles, probably because when long
vehicles get stuck in traffic this phase emerges faster. Besides, without an
obstacle there is an increase at the slope (absolute value) of the fundamental
diagram where there is the wide-moving jam phase, when the percentage of
long vehicles is increased (see Figure 2). In other words, the flow decreases
faster in the presence of long vehicles with the same increment of density,
which is not observed in the case of placing an obstacle (the lines are almost
parallel, see Figure 5). However, this parallelism occurs only when the road
is enough saturated. For instance in Figure 6 we see that the parallelism
occurs in the first and second plots, but not in the third plot since the lane
number is increased and so the emission rate per lane is decreased, i.e., there
is not enough saturation.
Another effect of setting an obstacle is a reduction on the traffic capacity
as it is expected. For instance, in the experiments plotted in Figure 5,
the maximum flow that can be reached is around 1.33 veh/sec without an
obstacle and around 1.2 veh/sec with an obstacle placed on the right-most
3When we place an obstacle on the left-most lane, we observe that the result is almost
the same with placing it on the right-most lane, so when we make comparisons of having
or not having obstacle we use the result of placing it on the right-most lane.
23
Figure 4: The fundamental diagrams depending on the different average
throughputs
24
Figure 5: The fundamental and the cross-covariance diagrams without and
with obstacle, respectively
25
26
Figure 6: The slope in the wide-moving jam phase with the obstacles
Figure 7: The effect of open road tolling on the flow phases with 3 lanes and
4 lanes
27
lane. In this figure, we also see that this reduction is more evident with the
presence of long vehicles.
In the case that there is an open road tolling, the wide-moving jam
phase does not take place, so there is no anticorrelation situation. As it is
seen in Figure 7, there is just a transition between the free flow and the
synchronized flow. The absence of wide-moving jam phase occurs in general
with the situations where the emission rate is low with respect to the rate
that the vehicles are processed (the influence radius or throughput) at the
off-toll plaza as it is expected.
The experiments also show that the model is able to reproduce the hys-
teresis phenomena in transition between free flow and synchronized flow
phases (see [11]) and in transition between synchronized flow and wide-
moving jam phases. The regions where there are saddles followed by a
capacity drop, as in the Fig.9(e) of [6], are the regions of metastability in
the phase-changes (from free flow to synchronized flow and from synchro-
nized flow to wide-moving jam). This metastability phenomenon is not so
evident when there are long vehicles. More precisely, the heterogeneity of
traffic effects the formation of the plateaus. In Figure 2 and 4, we see that in
the case of the presence of long vehicles, the bumpy plateaus in the synchro-
nized flow phase are replaced by more flattened plateaus, so the saddles in
the phase-changes are not observed as they are in the absence of long vehi-
cles. This is probably due to the fact that passenger vehicles are faster and
so the flow of a traffic without long vehicles changes its phase more sharply.
However, we observe that the presence of obstacles increases the number
of the saddles even when there are long vehicles, making the fundamental
diagrams more bumpy especially in the synchronized flow phase (see Figure
5 and 6).
5 Conclusion
It is introduced a new model for multi-lane traffic. This model is an attempt
to define a new class of traffic models which is a hybrid between the usual
microscopic traffic models, like car-following models, and the usual CAMs
which we have identified as the NaSch-type models. In the process of the
extension of the single-lane model to the multi-lane case, we have first pre-
sented the model as an array of communicating one-dimensional CCA, and
then we have proved that this model can be simulated by a suitable CCA.
In this way, we have framed our multi-lane model inside the class of CCA.
28
For a first test, we have implemented the model using Python4 with an
object-oriented philosophy of programming. Using a questionnaire we have
set up two kinds of vehicles which we have used to run a series of experiments.
Analyzing the experimental results, we have studied the influence of different
composition of vehicles on the macroscopic behavior of the traffic in order to
observe the typical traffic flow phenomena. We have found that the results
are promising, indeed we reproduce and enrich the fundamental diagrams
of traffic flow.
Table 3: Computation Time Comparison between CPU and GPU
Number of Vehicles CPU GPU
(sec)
45,8
104
194
556
1288
(per lane)
0; 0; 1500
0; 0; 3000
0; 0; 5000
0; 0; 10000
(sec)
484
958
1608
4270
9679
5000; 5000; 5000
The code written in Python does not take advantage of CA and its
typical synchronous behavior. For this reason, we have adapted the code
using PyCuda5 to partially parallelize Algorithm 2 on GPU's and we have
seen that it is possible to boost the speed of execution by a factor of ∼ 10.
Indeed, in Table 3 we see the computation time comparison between the
two codes made by running 1000 steps of simulation for a road with 3 lanes
and where it is given the initial numbers of the vehicles distributed per lane
(on a laptop equipped with a processor i7 intel and with a graphic card
NVIDIA GeForce GT 555M). The increased speed of execution by higher
factors of simulation is important also to adapt the model for a forecasting
usage. By partial parallelization, we mean that the lane-changing process
performed using the operations L, R, \ is done sequentially, and the
only parts that can be run in parallel are the global transition functions
δ∗
(a,b), a, b ∈ {0, 1}. Therefore Algorithm 2 can be completely parallelized
only if we decide not to apply the precedence rules and assume a concurrent
strategy of lane-changing.
The analysis we perform here is not conclusive, but gives an insight of
the potentiality of our model. For this reason, we suggest the following tasks
4The
code
can
in
https://dl.dropbox.com/u/50248089/Ozsim.py.
in
5The
found
found
be
be
code
can
https://dl.dropbox.com/u/50248089/Cozsim.py.
29
the
public
Dropbox
link
the
public
Dropbox
link
as future works and research directions to improve and validate the model
and the simulator:
• We did not use all the potentiality of the code since our aim was to
give a first evaluation to our model. However, it would be interesting
to consider also the experiments involving on- and off-ramps and loop-
detectors to analyze different and more realistic situations.
• The heterogeneity considered is reduced to two kinds of vehicles. A
natural question is how the system reacts introducing other kinds. For
instance, in highway environments, motorcycles or sport vehicles can
be added to the mixed traffic of passenger vehicles and long vehicles.
This also brings with it the interesting issue of how to tune the mem-
bership functions for such new kinds.
• The process of lane-changing is purely stochastic. However, in litera-
ture there are some attempts in microscopic models where the process
of lane-changing is described by using a fuzzy logic-based system [1, 4],
thus it would be interesting to extend our model to a model in which it
is implemented a fuzzy logic-based system to refine the lane-changing
rules.
• The model has to be compared with real data. In other words, a careful
case study on specific scenarios with the data available is necessary
for the validation by the community of people working on traffic flow
theory, granular flow theory and traffic (transportation) engineering.
Acknowledgments
The author acknowledges the support from the European Regional Devel-
opment Fund through the programme COMPETE and by the Portuguese
Government through the FCT -- Funda¸cao para a Ciencia e a Tecnologia
under the project PEst-C/MAT/UI0144/2011. The second author also ac-
knowledges the support of the FCT project SFRH/BPD/65428/2009.
References
[1] M. Brackstone and M. McDonald. Car-following: A historical review.
Transportation Research Part F, 2:181 -- 196, 1999.
30
[2] P. Chaudhuri, D. Chowdhury, S. Nandi, and S. Chattopadhyay. Ad-
IEEE Computer
ditive cellular automata: Theory and applications.
Society Press, New York, 1, 1997.
[3] B. Chopard. Modelling physical systems by cellular automata. In: G.
Rozenberg et al (eds) Handbook of Natural Computing: Theory, Exper-
iments, and Applications. Springer, Heidelberg, 2010.
[4] M. Errampalli, M. Okushima, and T. Akiyama. Fuzzy logic based
lane change model for microscopic traffic flow simulation. Journal
of Advanced Computational Intelligence and Intelligent Informatics
(JACIII), 12(2):172 -- 181, 2008.
[5] F. Farina and A. Dennunzio. A predator-prey cellular automaton with
parasitic interactions and environmental effects. Fundamenta Informat-
icae, 83:337 -- 353, 2008.
[6] Nikolas Geroliminis and Jie Sun. Hysteresis phenomena of a macro-
scopic fundamental diagram in freeway networks. Transportation Re-
search Part A, 45:966 -- 979, 2011.
[7] I.D. Greenwood. A new approach to estimate congestion impacts for
highway evaluation-effects on fuel consumption and vehicle emissions.
PhD thesis, University of Auckland, New Zealand, 2003.
[8] K. Kaneko. Theory and application of coupled map lattices. John Wiley
and Sons Ltd, 1993.
[9] J. Kari. Theory of cellular automata: A survey. Theoretical Computer
Science, 334:3 -- 33, 2005.
[10] G. Keller, M. Kunzle, and T. Nowiki. Some phase transitions in coupled
map lattices. Physica D, 59:39 -- 51, 1992.
[11] B.S. Kerner and H. Rehborn. Experimental properties of phase transi-
tions in traffic flow. Phys. Rev. Lett., 79:4030 -- 4033, 1997.
[12] W. Knospe, L. Santen, A. Schadschneider, and M. Schreckenberg. Dis-
order effects in cellular automata for two-lane traffic. Physica A: Sta-
tistical and Theoretical Physics, 265(3 -- 4):614 -- 633, 1999.
[13] W. Knospe, L. Santen, A. Schadschneider, and M. Schreckenberg. To-
wards a realistic microscopic description of highway traffic. Journal of
Physics A, 33:477 -- 485, 2000.
31
[14] Myer Kutz. Handbook of transportation engineering. Myer Kutz Assoc.
Inc., pages 8.1 -- 8.17, 2004.
[15] S. Maerivoet and B. De Moor. Cellular automata models of road traffic.
Physics Reports, 419(1):1 -- 64, November 2005.
[16] E.H. Mamdani. Applications of fuzzy algorithm for control a simple
dynamic plant. Proceedings of the IEEE, 121(12):1585 -- 1588, 1974.
[17] K. Nagel and M. Schreckenberg. A cellular automaton model for freeway
traffic. Journal de Physique I, 2(12):2221 -- 2229, 1992.
[18] K. Nagel, D. E. Wolf, P. Wagner, and P. Simon. Two-lane traffic
rules for cellular automata: A systematic approach. Phys. Rev. E,
58(2):1425 -- 1437, 1998.
[19] L. Neubert, L. Santen, A. Schadschneider, and M. Schreckenberg.
Single-vehicle data of highway trac: A statistical analysis. Phys. Rev.
E, 60:6480 -- 6490, 1999.
[20] M. Rickert, K. Nagel, M. Schreckenberg, and A. Latour. Two lane
traffic simulations using cellular automata. Physica A: Statistical and
Theoretical Physics, 231(4):534 -- 550, 1996.
[21] P. Wagner, K. Nagel, and D.E. Wolf. Realistic multi-lane traffic rules
for cellular automata. Physica A: Statistical and Theoretical Physics,
234(3 -- 4):687 -- 698, 1997.
[22] Dietrich E. Wolf. Cellular automata for traffic simulations. Physica A,
263:438 -- 451, 1999.
[23] S. Wolfram. Theory and applications of cellular automata. World
Scientific Press, Singapore, 1986.
[24] S. Wolfram. A new kind of science. Wolfram Media, Inc., 2002.
[25] O. Yeldan. A stochastic continuous cellular automata traffic model with
fuzzy decision rules. PhD thesis, Politecnico di Milano, Milan, March
2012.
[26] O. Yeldan, A. Colorni, A. Lu`e, and E. Rodaro. A stochastic continuous
cellular automata traffic flow model with a multi-agent fuzzy system.
Procedia -- Social and Behavioral Sciences, 54:1350 -- 1359, 2012.
32
|
0910.3580 | 3 | 0910 | 2010-02-24T14:52:26 | Set-Rationalizable Choice and Self-Stability | [
"cs.MA"
] | A common assumption in modern microeconomic theory is that choice should be rationalizable via a binary preference relation, which \citeauthor{Sen71a} showed to be equivalent to two consistency conditions, namely $\alpha$ (contraction) and $\gamma$ (expansion). Within the context of \emph{social} choice, however, rationalizability and similar notions of consistency have proved to be highly problematic, as witnessed by a range of impossibility results, among which Arrow's is the most prominent. Since choice functions select \emph{sets} of alternatives rather than single alternatives, we propose to rationalize choice functions by preference relations over sets (set-rationalizability). We also introduce two consistency conditions, $\hat\alpha$ and $\hat\gamma$, which are defined in analogy to $\alpha$ and $\gamma$, and find that a choice function is set-rationalizable if and only if it satisfies $\hat\alpha$. Moreover, a choice function satisfies $\hat\alpha$ and $\hat\gamma$ if and only if it is \emph{self-stable}, a new concept based on earlier work by \citeauthor{Dutt88a}. The class of self-stable social choice functions contains a number of appealing Condorcet extensions such as the minimal covering set and the essential set. | cs.MA | cs |
Set-Rationalizable Choice and Self-Stability
Felix Brandt
and Paul Harrenstein
Ludwig-Maximilians-Universitat Munchen
80538 Munich, Germany
{brandtf,harrenst}@tcs.ifi.lmu.de
A common assumption in modern microeconomic theory is that choice
should be rationalizable via a binary preference relation, which Sen showed
to be equivalent to two consistency conditions, namely α (contraction) and γ
(expansion). Within the context of social choice, however, rationalizability
and similar notions of consistency have proved to be highly problematic,
as witnessed by a range of impossibility results, among which Arrow's is the
most prominent. Since choice functions select sets of alternatives rather than
single alternatives, we propose to rationalize choice functions by preference
relations over sets (set-rationalizability). We also introduce two consistency
conditions, bα and bγ, which are defined in analogy to α and γ, and find that
a choice function is set-rationalizable if and only if it satisfies bα. Moreover, a
choice function satisfies bα and bγ if and only if it is self-stable, a new concept
based on earlier work by Dutta. The class of self-stable social choice functions
contains a number of appealing Condorcet extensions such as the minimal
covering set and the essential set.
1 Introduction
Arguably the most basic model of individual and collective choice is a choice function,
which associates with each set A of feasible alternatives a non-empty subset S(A) ⊆ A.
Apparently, not every choice function complies with our intuitive understanding of
rationality. Consider, for example, the choice function S with S({a, b}) = {a} and
S({a, b, c}) = {b}. Doubts as to an agent's rationality could be raised, if, when offered
the choice between apple pie and brownies, he were to choose the former, but the lat-
ter, when told that chocolate mousse is also an option.1 In microeconomic theory, the
1Sen (1993, 1997) gives examples where rational choosers actually make choices as described and gener-
ally argues against imposing internal consistency conditions on rational choice. His examples usually
involve a kind of context-dependence. For instance, a modest person may be unwilling to take the
largest piece of cake and thus his choice depends on the other pieces that are available. Our view is
1
existence of a binary relation R on all alternatives such that S returns precisely the
maximal elements according to R from any feasible set is commonly taken as a minimal
rationality condition on choice functions. Choice functions for which this is the case are
called rationalizable (see, e.g., Richter, 1966; Herzberger, 1973; Blair et al., 1976; Moulin,
1985a).2 Rationalizable choice functions have been characterized using two consistency
conditions that relate choices within feasible sets of variable size, namely conditions α
and γ (Sen, 1971). Clearly, acyclicity of the strict part P of R is necessary and sufficient
for S to be rationalizable if every finite set of alternatives is feasible. Stronger ratio-
nality conditions can be obtained by requiring the rationalizing relation R to satisfy
certain structural restrictions, such as completeness, transitivity, or quasi-transitivity
(i.e., transitivity of P ).
The above considerations have had a profound impact on the theory of social choice, in
particular on the interpretation of Arrow's general impossibility theorem (Arrow, 1951),
which states the impossibility of social choice functions that satisfy four intuitive criteria,
including rationalizability via a transitive preference relation. An obvious way around
Arrow's disturbing result is to try to relax this condition, e.g., by requiring social choice
functions to be merely rationalizable. Although this approach does allow for some social
choice functions that also meet the remaining three criteria, these functions turned out
to be highly objectionable, usually on grounds of involving a weak kind of dictatorship
or violating other conditions deemed to be indispensable for rational social choice (for an
overview of the extensive literature, see Blair et al., 1976; Kelly, 1978; Schwartz, 1986;
Sen, 1977, 1986; Campbell and Kelly, 2002). Sen (1995, page 5) concludes that
[. . . ] the arbitrariness of power of which Arrow's case of dictatorship is
an extreme example, lingers in one form or another even when transitivity
is dropped, so long as some regularity is demanded (such as the absence of
cycles).
One possibility to escape the haunting impossibility of rationalizable social choice
It turns out that α (and
is to require only α or γ but not both at the same time.
even substantially weakened versions of α) give rise to impossibility results that retain
Arrow's spirit (Sen, 1977). By contrast, there are a number of social choice functions
that satisfy γ. An attractive one among these based on majority rule is the uncovered
set (Fishburn, 1977; Miller, 1980; Moulin, 1986).
In this paper, we approach the matter from a slightly different angle. Choice functions
are defined so as to select subsets of alternatives from each feasible set, rather than a
single alternative. Still, the consistency and rationality conditions on choice functions
that violation of internal consistency conditions need not necessarily point at irrational behavior as
such but can just as well indicate the presence of situational features that affect choice but are not
(appropriately) represented in the mathematical model. In some cases the set of alternatives can be
redefined so as to capture all aspects that affect their choice. For instance, the choice of the modest
person above is arguably not between mere pieces of cake, but rather between tuples that consist of
a piece of cake and the pieces left for others to choose from.
2Rationalizable choice functions have also been referred to as binary (Schwartz, 1976), normal (Sen,
1977), and reasonable (Allingham, 1999).
2
have been defined in terms of alternatives. Taking cue from this observation, we propose
an alternative notion of rationality called set-rationalizability. A choice function S is
set-rationalizable if a binary relation R on all non-empty subsets of alternatives can be
found such that for each feasible subset A, S(A) is maximal with respect to R among
all non-empty subsets of A.
ural variant of α defined in terms of sets rather than alternatives. Despite its intu-
We find that set-rationalizable choice functions can be characterized by bα, a nat-
itive appeal, bα has played a remarkably small role in (social) choice theory (Chernoff,
1954; Aizerman and Aleskerov, 1995). Yet, it differentiates quite a number of well-
known choice functions. In particular, we will show that various prominent social choice
functions -- such as all scoring rules, all scoring runoff rules, and all weak Condorcet
sure maximality, the minimal covering set, and the essential set -- do.
extensions -- do not satisfy bα, whereas several Condorcet extensions -- such as weak clo-
For our second result, we introduce a new property bγ, which varies on γ in an analogous
way as bα varies on α. It turns out that bα and bγ characterize the class of self-stable choice
Despite the logical independence of bα and bγ, the class of self-stable social choice functions
functions, whose definition is inspired by earlier work of Dutta (1988) and Brandt (2009).
also contains the Condorcet extensions mentioned above. These Condorcet extensions
also satisfy all conditions typically appearing in Arrovian impossibility results except
rationalizability, i.e., α and γ. Accordingly, by replacing α and γ with bα and bγ, the
impossibility of rationalizable social choice can be avoided and turned into a possibility
result.
2 Preliminaries
We assume there to be a universe U of at least three alternatives. Any subset of U
from which alternatives are to be chosen is a feasible set (sometimes also called an issue
or agenda). Throughout this paper we assume the set of feasible subsets of U to be
given by F(U ), the set of finite and non-empty subsets of U , and generally refer to finite
non-empty subsets of U as feasible sets. Our central object of study are choice functions,
i.e., functions S : F(U ) → F(U ) such that S(A) ⊆ A for all feasible sets A.
A choice function S is called rationalizable if there exists a binary relation R on U
such that for each feasible set A,
S(A) = {a ∈ A : x P a for no x ∈ A}
where P is the strict part of R. Observe that acyclicity of P is required to guarantee
that S invariably returns a non-empty set.
Two typical candidates
the rationalizing relation are the base relation
RS (Herzberger, 1973) and the revealed preference relation RS (Samuelson, 1938), which,
for all alternatives x and y, are given by
for
a RS b if and only if a ∈ S({a, b}), and
a RS b if and only if a ∈ S(X) for some X with b ∈ X.
3
Thus, the revealed preference relation relates a to b if a is chosen in the presence of b and
possibly other alternatives, whereas the base relation only relates a to b if a is chosen in
the exclusive presence of b.
Rationalizable choice functions are characterized by a consistency axiom, which
Schwartz (1976) defined such that for all feasible sets A and B and all alterna-
tives x ∈ A ∩ B,
x ∈ S(A ∪ B) if and only if x ∈ S(A) and x ∈ S(B).
The above equivalence can be factorized into two implications, viz. the conditions α and
γ (Sen, 1971) for feasible sets A and B and alternatives x ∈ A ∩ B,3
if x ∈ S(A ∪ B) then x ∈ S(A) and x ∈ S(B),
if x ∈ S(A) and x ∈ S(B) then x ∈ S(A ∪ B).
(α)
(γ)
Axiom α is a contraction consistency property, which states that alternatives that are
chosen in a feasible set are still chosen in feasible subsets. By contrast, γ is an expansion
consistency property, which states that alternatives chosen in two feasible sets are also
chosen in their union. Sen (1971) proved that a choice function S is rationalizable if and
only if it satisfies both α and γ, with the witnessing relations RS and RS, which are
identical in the presence of α.
Theorem 1 (Sen, 1971). A choice function is rationalizable if and only if it satisfies
both α and γ.
Similar results can also be obtained if stronger requirements are imposed on the ra-
tionalizing relation (see, e.g., Sen, 1977; Moulin, 1985a; Schwartz, 1976). For instance,
Arrow (1959) showed that a choice function can be rationalized by a complete and tran-
sitive relation if and only if it satisfies the weak axiom of revealed preference (WARP) -- a
consistency condition, first proposed by Samuelson (1938), which is stronger than the
conjunction of α and γ and central to large parts of microeconomic theory. Formally,
WARP is defined such that for all feasible sets A and B with B ⊆ A,
if S(A) ∩ B 6= ∅ then S(A) ∩ B = S(B).
(WARP)
3 Set-Rationalizable Choice
In analogy to the definitions of Section 2, we now define the concept of set-
rationalizability, the base and revealed preference relations over sets of alternatives,
and properties bα and bγ. The main result of this section is that set-rationalizable choice
is completely characterized by bα.
We say a choice function is set-rationalizable if it can be rationalized via a preference
relation on sets of alternatives.
3The definitions of α and γ given here are equivalent, but not syntactically identical, to Sen's original
ones. They are chosen so as they reveal their similarity to bα and bγ below.
4
Definition 1. A choice function S is set-rationalizable if there exists a binary re-
lation R ⊆ F(U ) × F(U ) such that for each feasible set A there is no X ∈ F(A)
with X P S(A) where P is the strict part of R.
We define the base relation RS and the revealed preference relation bRS of a choice
function S on sets as follows:4
A RS B if and only if A = S(A ∪ B),
A bRS B if and only if A = S(X) for some X with B ⊆ X.
3.1 Set-contraction consistency
X ⊆ A ∩ B,
if X = S(A ∪ B) then X = S(A) and X = S(B).
chosen alternatives rather than its individual elements.
Condition bα is defined as a natural variant of α that makes reference to the entire set of
Definition 2. A choice function S satisfies bα, if for all feasible sets A, B, and X with
(bα)
bα is not implied by the standard contraction consistency condition α (see Example 2).
Moreover, bα is a not a contraction consistency property according to Sen's original ter-
minology (see, e.g., Sen, 1977). It does not only require that chosen alternatives remain
in the choice set when the feasible set is reduced, but also that unchosen alternatives
remain outside the choice set. Thus, it has the flavor of both contraction and expansion
consistency. bα can be split into two conditions that fall in Sen's categories: an expansion
condition known as ǫ+ (Bordes, 1983) and Aizerman (Moulin, 1986), which requires that
S(B) ⊆ S(A) for all S(A) ⊆ B ⊆ A, and a corresponding expansion condition. Similarly,
bγ can be factorized into two conditions.
In this paper, however, we are concerned with the choice set as a whole and bα merely
says that the set S(A) chosen from a feasible set A is also chosen from any subset B
of A, provided the former contains S(A). This reading is reflected by the useful char-
acterization of bα given in the following lemma, which reveals that bα is equivalent to
such established notions as Chernoff's postulate 5∗ (Chernoff, 1954), the strong superset
property (Bordes, 1979), and outcast (Aizerman and Aleskerov, 1995).
Lemma 1. A choice function S satisfies bα if and only if for all feasible sets A and B,
if S(A) ⊆ B ⊆ A then S(A) = S(B).
Proof. For the direction from left to right, let S(A) ⊆ B ⊆ A. Then, both A∪B = A and
B = A ∩ B. Hence, S(A ∪ B) = S(A) ⊆ B = A ∩ B. Since S satisfies bα, S(A) = S(B).
For the opposite direction, assume for an arbitrary non-empty set X, both X ⊆ A ∩ B
and X = S(A ∪ B). Then, obviously, both S(A ∪ B) ⊆ A ⊆ A ∪ B and S(A ∪ B) ⊆ B ⊆
A ∪ B. It follows that S(A ∪ B) = S(A) and S(A ∪ B) = S(B).
4Given a choice function S, the base relation on sets is a natural extension of the base relation on
alternatives and, hence, both are denoted by RS.
5
As a corollary of Lemma 1, we have that choice functions S satisfying bα, like those
satisfying α, are idempotent, i.e., S(S(A)) = S(A) for all feasible sets A.
An influential and natural consistency condition that also has the flavor of both con-
traction and expansion is path independence (Plott, 1973). Choice function S satis-
fies path independence if S(A ∪ B) = S(S(A) ∪ S(B)) for all feasible sets A and B.
Aizerman and Malishevski (1981) have shown that path independence is equivalent to
the conjunction of α and ǫ+. Since α is the strongest contraction consistency property
and implies the contraction part of bα, we obtain the following alternative characterization
of path independence.
Proposition 1. A choice function satisfies path independence if and only if it satisfies
α and bα.
It can easily be verified that the revealed preference relation on sets bRS of any choice
function S that satisfies bα is closed under intersection, i.e., for all feasible sets X, Y ,
and Z such that Y ∩ Z 6= ∅,
X bRS Y and X bRS Z imply X bRS Y ∩ Z.
3.2 Set-expansion Consistency
We define bγ in analogy to γ as follows.
Definition 3. A choice function S satisfies bγ if for all feasible sets A, B, and X,
it also chooses X from their union.
if X = S(A) and X = S(B) then X = S(A ∪ B).
(bγ)
Thus, a choice function satisfies bγ, if whenever it chooses X from two different sets,
Example 2 shows that bα is not a weakening of α (and not even of the conjunction of
α and γ). To see that bγ is not implied by γ, consider the following choice function over
the universe {a, b, c}, which satisfies γ but not bγ:
X
S(X)
{a, b}
{a, c}
{b, c}
{a, b, c} {a, b, c}
{a}
{a}
{b}
However, bγ is implied by the conjunction of α and γ.
Proposition 2. Every rationalizable choice function satisfies bγ.
Proof. Assume both α and γ to hold for an arbitrary choice function S and consider
an arbitrary feasible sets X, A, and B with X = S(A) and X = S(B). The inclusion
of X in S(A ∪ B) follows immediately from γ. To appreciate that also S(A ∪ B) ⊆ X,
consider an arbitrary x /∈ X and assume for contradiction that x ∈ S(A ∪ B). Then,
either x ∈ A or x ∈ B. Without loss of generality we may assume the former. Clearly,
x ∈ (A ∪ B) ∩ A and α now implies that x ∈ S(A), a contradiction.
6
Schwartz (1976) has shown that quasi-transitive rationalizability is equivalent to the
conjunction of α, γ, and ǫ+. Since α implies the contraction part of bα and α and γ imply
bγ, we obtain the following alternative characterization of quasi-transitive rationalizabil-
ity. As a consequence, WARP implies both bα and bγ.
Proposition 3. A choice function is quasi-transitively rationalizable if and only if it
satisfies α, bα, and bγ.
bγ is reminiscent of the generalized Condorcet condition (see, e.g., Blair et al., 1976),
which requires that for all feasible sets A and all a ∈ A,
if S({a, b}) = {a} for all b ∈ A then S(A) = {a}.
Choice functions that satisfy this condition we will refer to as generalized Condorcet
extensions. It is easily appreciated that bγ implies the generalized Condorcet condition.
In the setting of social choice, Condorcet extensions are commonly understood to be
social choice functions for which additionally choice over pairs is determined by majority
rule (see Section 5.3.1).
In analogy to the relationship between closure under intersection of bRS and bα, bRS of
a choice function S that satisfies bγ is closed under union,5 i.e., for all feasible sets X, Y ,
and Z,
X bRS Y and X bRS Z imply X bRS Y ∪ Z.
3.3 Set-Rationalizability
As in the case of α and γ, a single intuitive consistency condition summarizes the con-
junction of bα and bγ: for all feasible sets A, B, and X with X ⊆ A ∩ B,
X = S(A) and X = S(B) if and only if X = S(A ∪ B).
For illustrative purposes consider the following examples.
Example 1. Let the choice function S over the universe {a, b, c} be given by the table
sets RS coincide and are depicted in the graph on the right. A routine check reveals that
in Figure 1. For S the revealed preference relation on sets bRS and the base relation on
this choice function satisfies both bα and bγ (while it fails to satisfy α). Also observe that
each feasible set X contains a subset that is maximal (with respect to bRS) among the
non-empty subsets of X, e.g., {a, b, c} in {a, b, c} and {a} in {a, b}. Theorem 2, below,
shows that this is no coincidence.
Example 1 also shows that the revealed preference relation over sets need not be
complete. Some reflection reveals that the relation is always incomplete.
Example 2. The table in Figure 2 summarizes a choice function that is rationalizable
by any acyclic relation P with a P c P b. Nevertheless, the revealed preference relation
5This condition is also known as robustness (Arlegi, 2003, see also Barber`a et al. (2004)).
7
X
S(X)
{a, b}
{a, c}
{b, c}
{a, b, c} {a, b, c}
{a}
{c}
{b}
{a,b,c}
{a,b}
{b,c}
{a,c}
{a}
{b}
{c}
Figure 1: The revealed preference relation bRS of the choice function S as in Example 1.
X
S(X)
{a, b}
{a, b}
{a}
{a, c}
{c}
{b, c}
{a, b, c} {a}
{a,b,c}
{a,b}
{b,c}
{a,c}
{a}
{b}
{c}
Figure 2: The revealed preference relation bRS of the choice function S as in Example 2.
8
over sets does not set-rationalize this choice function. Observe that both {a} and {a, b}
are maximal in {a, b} with respect to the strict part of bRS. As S({a, b, c}) = {a} and
S({a, b}) = {a, b}, S clearly does not satisfy bα. Again Theorem 2, below, shows that
this is no coincidence.
By definition, the base relation RS of any choice function S is anti-symmetric, i.e.,
thus both anti-symmetric.
X RS Y and Y RS X imply X = Y . In the presence of bα, bRS and RS coincide and are
Set-rationalizable choice functions are characterized by bα.6
Theorem 2. A choice function is set-rationalizable if and only if it satisfies bα.
Proof. For the direction from left to right, assume S is set-rationalizable and let R be
the witnessing binary relation on sets. Now consider arbitrary feasible sets A, B and X
with X ⊆ A ∩ B and assume that X = S(A ∪ B). Then, S(A ∪ B) ⊆ A ∩ B. Hence,
Y P S(A ∪ B), for no non-empty subset Y ⊆ A ∪ B. It follows that there is no non-empty
subset Y of A such that Y P S(A ∪ B) either. Hence, S(A ∪ B) is maximal with respect
to R within A. As S is set-rationalizable, S(A ∪ B) has also to be the unique such
subset in A. The argument that S(A ∪ B) is also the unique maximal element of R in B
runs along analogous lines. It follows that both S(A) = S(A ∪ B) and S(B) = S(A ∪ B).
For the opposite direction assume S to satisfy bα and consider arbitrary feasible sets A
and B such that B ⊆ A and S(A) 6= B. Then,
Hence, by virtue of Lemma 1,
S(A) ⊆ B ∪ S(A) ⊆ A.
S(A) = S(B ∪ S(A)),
which implies that S(A) RS B. Since RS is anti-symmetric, it is thus impossible that
B RS S(A). A similar argument holds for bRS, which coincides with RS in the presence
of bα.
In the proof of Theorem 2, it is precisely the revealed preference relation on sets that
is witness to the fact that choice functions satisfying bα are set-rationalizable. In contrast
to Sen's Theorem 1, however, the revealed preference relation on sets is not the unique
relation that can achieve this. It is also worth observing that the proof shows that for
each feasible set X and choice function S satisfying bα, the selected set S(X) is not merely
a maximal set but also the unique maximum set within X given bRS, i.e., S(X) bRS Y for
all non-empty subsets Y of X.
4 Self-Stability
It turns out that the notions of set consistency introduced in the previous section bear
a strong relationship to the stability of choice sets as introduced by Dutta (1988) and
6Moulin shows a similar statement for single-valued choice functions (Moulin, 1985a).
9
generalized by Brandt (2009). Stability of choice sets is based on the notions of internal
and external stability by von Neumann and Morgenstern (1944), which can be merged
in the following fixed-point characterization.
Definition 4. Let A, X be feasible sets and S a choice function. X is S-stable in A if
X = {a ∈ A : a ∈ S(X ∪ {a})}.
Alternatively, X is S-stable in A if it satisfies both internal and external S-stability
in A:
S(X) = X,
(internal S-stability)
a 6∈ S(X ∪ {a}) for all a ∈ A \ X.
(external S-stability)
The intuition underlying stable sets is that there should be no reason to restrict the se-
lection by excluding some alternative from it and, secondly, there should be an argument
against each proposal to include an outside alternative into the selection.
For some choice functions S, a unique inclusion-minimal S-stable set generally exists.
If that is the case, we use bS to denote the choice function that returns the unique minimal
S-stable set in each feasible set and say that bS is well-defined. Within the setting of
social choice, a prominent example is Dutta's minimal covering set MC (Dutta, 1988;
Dutta and Laslier, 1999), which is defined as the unique minimal stable set with respect
frequently turns out to be highly non-trivial (Brandt, 2009).
to the uncovered set UC , i.e., MC = dUC . Proving that a choice function bS is well-defined
We find that there is a close connection between bγ and minimal S-stable sets.
Lemma 2. Let S be a choice function such that bS is well-defined. Then bS satisfies bγ.
Proof. Consider arbitrary feasible sets A, B, X and assume that bS(A) = bS(B) = X.
Trivially, as X is internally S-stable in A, so is X in A ∪ B. To appreciate that X is also
externally S-stable in A ∪ B, consider an arbitrary x ∈ (A ∪ B) \ X. Then, x ∈ A \ X
or x ∈ B \ X. In either case, S(X ∪ {x}) = X, by external S-stability of X in A if
the former, and by external S-stability of X in b if the latter. Also observe that any
subset of X that is S-stable in A ∪ B, would also have been S-stable in both A and B.
Hence, X is minimal S-stable in A ∪ B. Having assumed that bS is well-defined, we may
conclude that bS(A ∪ B) = X.
We now introduce the notion of self-stability. A choice function S is said to be self-
stable if for each feasible set A, S(A) is the unique (minimal) S-stable set in A.
Definition 5. A choice function S is self-stable if bS is well-defined and S = bS.
In the next section we argue that self-stability defines an interesting class of choice
functions, containing a number of well-known and important social choice functions.
First, and on a more abstract level, however, we establish that the class of self-stable
choice functions is characterized by the conjunction of bα and bγ.
10
Theorem 3. A choice function is self-stable if and only if it satisfies both bα and bγ.
Proof. For the direction from left to right, assume S to be self-stable. Lemma 2 implies
that S satisfies bγ. For bα, consider arbitrary feasible sets A, B such that S(A) ⊆ B ⊆ A.
By virtue of Lemma 1, it suffices to show that S(B) = S(A). As, moreover, bS(B) = S(B)
and bS is well-defined, S(B) is the unique S-stable set in B. Hence, it suffices to show
that S(A) is both internally and externally S-stable in B. Internal S-stability of S(A)
in B is trivial since S(S(A)) = S(A) by S(A)'s being internally S-stable in A. To
appreciate that S(A) is also externally S-stable in B, consider an arbitrary x ∈ B \S(A).
Then also x ∈ A \ S(A) and by S(A)'s being externally S-stable in A, we obtain S(A) =
S(S(A) ∪ {x}). It follows that S(A) is also externally S-stable in B.
For the other direction, assume S satisfies both bα and bγ and consider an arbitrary
feasible set A. To show that S satisfies internal S-stability, observe that trivially S(A) ⊆
S(A) ⊆ A. Hence, by Lemma 1, S(A) = S(S(A)). To appreciate that S also satisfies
external S-stability, let x ∈ A \ S(A). Then, S(A) ⊆ S(A) ∪ {x} ⊆ A and, again by
Lemma 1, S(A) = S(S(A) ∪ {x}). To see that bS is well-defined, consider an arbitrary
S-stable set Y in A and let A \ Y = {x1, . . . , xk}, i.e., A = Y ∪ {x1, . . . , xk}. By external
S-stability of Y , then S(Y ∪ {xi}) = S(Y ) for each i with 1 ≤ i ≤ k. Thus by k − 1
applications of bγ, we obtain S(Y ) = S(Y ∪ {x1, . . . , xk}) = S(A).
As an immediate consequence of Theorem 3 and the observation that bγ implies the
generalized Condorcet condition, we have the following corollary.
Corollary 1. Every self-stable choice function is a generalized Condorcet extension.
Examples of self-stable Condorcet extensions will be given in Section 5.3.1.
5 Social Choice
In this section, we assess the consequences of the reflections in the previous two sections
on the theory of social choice. Before we do so, however, we introduce some additional
terminology and notation.
5.1 Social Choice Functions
We consider a finite set N = {1, . . . , n} of at least two agents. Each agent i entertains
preferences over the alternatives in U , which are represented by a transitive and complete
preference relation Ri. In some cases, we will assume preferences to be linear, i.e., also
satisfying anti-symmetry, but otherwise we impose no further restrictions on preference
relations. We have a Ri b denote that agent i values alternative a at least as much as
alternative b. We write Pi for the strict part of Ri, i.e., a Pi b if a Ri b but not b Ri a.
Similarly, Ii denotes i's indifference relation, i.e., a Ii b if both a Ri b and b Ri a. The
set of all preference relations over the universal set of alternatives U will be denoted
by R(U ). The set of preference profiles, with typical element R = (R1, . . . , Rn), is then
given by R(U )N .
11
The central object of study in this section are social choice functions, i.e., functions
that map the individual preferences of the agents and a feasible set to a set of socially
preferred alternatives.
Definition 6. A social choice function (SCF) is a function f : R(U )N × F(U ) → F(U )
such that f (R, A) ⊆ A for all preference profiles R and feasible sets A.
Clearly, every SCF f together with a preference profile R in R(U ) defines a choice
function Sf,R on U in a natural way by letting for each feasible set A, Sf,R(A) = f (R, A).
We say that f satisfies WARP, rationalizability, or any other condition defined for choice
functions, if Sf,R does for every preference profile R. Pareto-optimality, independence
of irrelevant alternatives, and non-dictatorship are conditions that are more specifically
defined for SCFs.
Pareto-optimality requires that an alternative should not be chosen if there exists
another alternative that all agents unanimously prefer to the former.
Definition 7. An SCF f satisfies (pairwise) Pareto-optimality if for all preference pro-
files R and all alternatives a, b, if b Pi a for all i ∈ N then a 6∈ f (R, {a, b}).
Independence of irrelevant alternatives reflects the idea that choices from a set of
feasible alternatives should not depend on preferences over alternatives that are not
contained in this set.
Definition 8. An SCF f satisfies independence of irrelevant alternatives (IIA) if
f (R, A) = f (R′, A) for all feasible sets A and preference profiles R, R′ such that
RA = R′A.
In the context of SCFs, IIA constitutes no more than a framework requirement for
social choice.
Another minimal requirement for any SCF is that it should be sensitive to the pref-
erences of more than one agent. In particular, there should not be a single agent who
can enforce the inclusion of alternatives in the choice set no matter which preferences
the other agents have. Such an agent is usually called a (weak) dictator.7
Definition 9. An SCF f is (pairwise) non-dictatorial if there is no agent i such that
for all preference profiles R and alternatives a, b, if a Pi b then a ∈ f (R, {a, b}).
Definition 7 through Definition 9 are also referred to as the Arrovian conditions.
Other useful and frequently imposed requirements on SCFs are neutrality, anonymity,
and positive responsiveness.
Neutrality can be seen as a strengthening of IIA and requires SCFs to be invariant
under renaming alternatives, i.e., all alternatives are to be treated equally.
Definition 10. An SCF f is neutral if π(f (R, A)) = f (R′, A) for all feasible sets A,
preference profiles R, R′, and permutations π : A → A such that a R′
i b if and only if
π(a) Ri π(b) for all alternatives a, b and agents i.
7For presentational purposes we employ the notion of a weak dictator or vetoer in all impossibility
theorems, although Theorem 4 holds for an even weaker notion of non-dictatorship.
12
By contrast, anonymity says that SCFs be invariant under renaming agents and as
such is a strong variant of non-dictatorship.
Definition 11. An SCF f is anonymous if f (R, A) = f (R′, A) for all feasible sets A,
preference profiles R and R′, and permutations π : N → N such that R′
i = Rπ(i) for all
agents i.
It also appears reasonable to demand that SCFs are monotonic in the sense that
increased support may not hurt an alternative.
Definition 12. An SCF f is (pairwise) positive responsive if for all alternatives a, b and
all preference profiles R, R′, there is some agent i such that Rj = R′
j for all agents j 6= i
and either both a Ii b and a P ′
i b or both b Pi a and a R′
i b,
if a ∈ f (R, {a, b}) then f (R′, {a, b}) = {a}.
5.2 Impossibility Results
Famously, Arrow's general impossibility theorem, as formulated for SCFs, states that no
SCF that satisfies all of the Arrovian conditions exists.
Theorem 4 (Arrow, 1951; 1959). No SCF satisfies Pareto-optimality, IIA, WARP, and
non-dictatorship.
As the Arrovian conditions cannot be satisfied by any SCF, at least one of them needs
to be excluded or relaxed to obtain positive results. Clearly, dropping non-dictatorship is
unacceptable and, as already mentioned, IIA merely states that the SCF represents a rea-
sonable model of preference aggregation (see, e.g., Schwartz, 1986; Bordes and Tideman,
1991). Wilson (1972) has shown that without Pareto-optimality only SCFs that are con-
stant (i.e., completely unresponsive) or fully determined by the preferences of a single
agent are possible. Moreover, it could be argued that not requiring Pareto-optimality
runs counter to the very idea of social choice. Accordingly, the only remaining possibility
is to exclude WARP.
Imposing weaker rationality conditions than WARP, however, offers little relief as it
turns out that the vicious essence of Arrow's impossibility remains. There is a range of
results stating the impossibility of SCFs satisfying weaker versions of WARP in a satisfac-
tory way (see, e.g., Kelly, 1978; Schwartz, 1986; Campbell and Kelly, 2002; Banks, 1995).
Among these, the results by Mas-Colell and Sonnenschein (1972) and Blau and Deb
(1977) deserve special mention as they concern rationalizability instead of WARP. We
will employ a variant of Blau and Deb's theorem due to Austen-Smith and Banks (2000).
Theorem 5 (Mas-Colell and Sonnenschein, 1972). No SCF satisfies Pareto-optimality,
positive responsiveness, IIA, rationalizability, and non-dictatorship, provided that n > 3.
By strengthening IIA to neutrality and assuming that the number of alternatives
exceeds the number of agents, positive responsiveness is no longer required.
13
1 1 1
a c
b
b a c
b a
c
{x, y} f (R, {x, y})
{a, b}
{b, c}
{a, c}
{a}
{b}
{c}
Table 1: On the left a preference profile (figures indicate numbers of agents) leading to
the Condorcet paradox. On the right the corresponding choice function on pairs
if determined by majority rule.
Theorem 6 (Austen-Smith and Banks, 2000). No SCF satisfies Pareto-optimality, neu-
trality, rationalizability, and non-dictatorship, provided that U > n.
For further characterizations of rationalizable social choice the reader be referred to
Moulin (1985b), Banks (1995), and Austen-Smith and Banks (2000).
5.3 Condorcet Extensions and Scoring Rules
In light of the severe problems that α and γ entail in social choice, we now investigate
which of the well-known SCFs satisfy bα and bγ. We focus on two types of SCFs, namely
Condorcet extensions and scoring rules.
5.3.1 Condorcet Extensions
Despite the Arrovian impossibility results, social choice over two alternatives is unprob-
lematic. May (1952) has shown that the simple majority rule -- choosing the alternative
that a majority prefers to the other alternative, and in case of a tie, choose both -- can
be characterized by neutrality, anonymity, and positive responsiveness. Thus, it seems
reasonable to require of SCFs f that they reflect majority rule on pairs. Extending any
such SCF f to feasible sets with more than two alternatives, one immediately runs into
arguably one of the earliest Arrovian impossibility results, viz. the Condorcet paradox
(de Condorcet, 1785). Consider the preference profile depicted in Table 1. Then, if f
on pairs is determined by majority rule, the base relation RSf,R fails to be acyclic, and
therefore Sf,R does not satisfy α. Observe, moreover, that f (R, {a, b, c}) = {a, b, c},
if f satisfies bα. For suppose otherwise, then, without loss of generality, we may assume
fies bα, however, the former is at variance with f (R, {a, c}) = {c}, and the latter with
f (R, {a, b}) = {a}. Thus, Sf,R coincides with the choice function S of Example 1 and
Figure 1 depicts its weak revealed preference relation bRSf,R.
By Theorem 3, the class of SCFs that satisfy both bα and bγ consists precisely of all
that either f (R, {a, b, c}) = {a} or f (R, {a, b, c}) = {a, b}. Assuming that f satis-
self-stable SCFs. By virtue of Theorem 2 and Corollary 1 the SCFs in this class are all
set-rationalizable generalized Condorcet extensions. May's characterization furthermore
implies that all self-stable SCFs that satisfy anonymity, neutrality, and positive respon-
siveness are Condorcet extensions. Among them are well-known rules like weak closure
14
3 2 1
a b
c
c a b
c a
b
Table 2: A preference profile (figures indicate numbers of agents) showing that no weak
Condorcet extension and scoring rule satisfies bα. For every weak Condorcet
extension and every scoring rule the choice function for this profile is as in
Example 2 (also compare Figure 2).
maximality (also known as the top cycle, GETCHA, or the Smith set), the minimal cov-
ering set, the essential set, and their generalizations (Bordes, 1976; Dutta and Laslier,
1999; Laslier, 2000).8
Interestingly, well-known Condorcet extensions that satisfy only one of bα and bγ appear
he refers to as GOCHA, is an example of an SCF that satisfies bγ but not bα. By contrast,
the iterated uncovered set (e.g., Dutta, 1988) satisfies bα but not bγ.
to be less common. Still, Schwartz's strong closure maximality (Schwartz, 1972), which
When pairwise choice is determined via majority rule, the set of weak Condorcet
winners for a given preference profile R and feasible set A is defined as {a ∈ A : a ∈
f (R, {a, b}) for all b ∈ A}. An SCF is called a weak Condorcet extension if it returns
the set of weak Condorcet winners whenever this set is non-empty. Clearly, every weak
Condorcet extension is a Condorcet extension. The converse is not generally the case,
but many Condorcet extensions (such as Kemeny's rule, Dodgson's rule, Nanson's rule,
and the minimax rule) are also weak Condorcet extensions (see Fishburn, 1977). It turns
out that no weak Condorcet extension is set-rationalizable.
Theorem 7. No weak Condorcet extension satisfies bα.
Proof. Let f be a weak Condorcet extension and consider the linear preference profile R
with preferences over A = {a, b, c} as given in Table 2. Since alternative a is the unique
weak Condorcet winner -- three out of six agents prefer it to b and all agents but one
prefer it over c -- f (R, {a, b, c}) = {a}. Now observe that the preferences of the same
agents over the subset {a, b} are such that three agents prefer a to b and three b to a.
Accordingly, f (R, {a, b}) = {a, b}. As c /∈ f (R, {a, b, c}) but f (R, {a, b, c}) 6= f (R, {a, b}),
we may conclude that f does not satisfy bα.
5.3.2 Scoring Rules
Scoring rules are based on the idea that the voters each rank the alternatives in a feasible
set according to their preferences, which, for technical convenience we will here assume
8Brandt (2009) defines an infinite hierarchy of self-stable SCFs. If we assume an odd number of agents
with linear preferences, the class of self-stable SCFs is also conjectured to contain the tournament
equilibrium set (Schwartz, 1990) and the minimal extending set. Whether this is indeed the case
depends on a certain graph-theoretic conjecture (Laffond et al., 1993; Brandt, 2009).
15
to be linear. Each time an alternative is ranked mth by some voter it gets a particular
score sm. The scores of each alternative are then added and the alternatives with the
highest cumulative score are selected. The class of scoring rules includes several well-
known SCFs, like the Borda rule -- alternative a gets k points from agent i if i prefers a
to k other alternatives -- and the plurality rule -- the cumulative score of an alternative
equals the number of agents by which it is ranked first.
Formally, we define a score vector of length k as a vector s = (s1, . . . , sk) in Rk such
that s1 ≥ · · · ≥ sk and s1 > sk. For example, (1, 0, 0), (2, 1, 0), and (1, 1, 0) are the
score vectors of length 3 for the plurality rule, the Borda rule, and the anti-plurality
rule, respectively. Given a feasible set X of k alternatives, an x ∈ X, and a linear
preference profile R, we have s(x, i) denote the score alternative x obtains from voter i,
i.e., s(x, i) = sm if and only if x is ranked mth by i within X. Then, the (cumulative)
score s(x) of an alternative x within X given R is then defined such that
s(x) = X
i∈N
s(x, i).
A scoring rule is an SCF that selects from each feasible set X for each preferences
profile the set of alternatives x in X with the highest score s(x) according to some score
vector s of length X. Observe that no restrictions are imposed on how the score vectors
for different lengths are to be related.
As every scoring rule fails to select the Condorcet winner for some preference profile
(Fishburn, 1973) and coincides with majority rule on two alternatives, they generally do
rule is set-rationalizable.
not satisfy bγ. We find that no scoring rule can satisfy bα either. It follows that no scoring
Theorem 8. No scoring rule satisfies bα.
Proof. Let f be a scoring rule. Let further s = (s1, s2, s3) and s′ = (s′
2) be its
associated score vectors of lengths 3 and 2, respectively. Without loss of generality we
may assume that s1 = 1 and s3 = 0. Consider the preference profile R with preferences
over A = {a, b, c} as depicted in Table 2. Then, s(a) = 3 + 2s2, s(b) = 2 + s2 and
s(c) = 1 + 3s2. Since 1 ≤ s2 ≤ 0, it can easily be appreciated that s(a) > s(b)
as well as s(a) > s(c). Hence, f (R, {a, b, c}) = {a}. As in the proof of Theorem 7,
1, s′
f (R, {a, b}) = {a, b} and we may conclude that f does not satisfy bα.
Using the same example as in the proof of Theorem 8, the reader can easily verify that
all scoring run-off rules -- such as single transferable vote (STV) -- also fail to satisfy bα
and as such are not set-rationalizable.
An interesting question in this context is whether the impossibility shown in Theorem 8
can be generalized to rank-based SCFs, i.e., SCFs that merely take into account the
positions of alternatives in the individual rankings (Laslier, 1996).
It turns out this
is not the case because, for instance, the (rather unattractive) SCF that chooses all
alternatives that are ranked first by at least one voter is rank-based and satisfies bα.9 It
9This SCF has also been mentioned by Gardenfors (1976) and Kelly (1977). Taylor (2005) calls it the
omninomination rule.
16
might also be worth observing that this SCF happens to satisfy bγ and Pareto-optimality
as well.
6 Summary and Conclusion
Problems relating to the possibility of reasonable social choice functions have proved to
be rather tenacious. In particular, attempts to circumvent Arrow's impossibility result
by replacing the weak axiom of revealed preference (WARP), which requires a transitive
and complete preference relation on alternatives underlying choice, by weaker conditions
on the underlying preference relation have generally failed to deliver.
By weakening WARP to set-rationalizability, we have shown that social choice func-
tions that also satisfy the other Arrovian postulates do exist. These social choice func-
tions are generally characterized by their satisfying bα. This condition, which also goes
by the names of strong superset property, 5∗, and outcast, is no stranger in choice
theory, but nevertheless has played a surprisingly small role therein. In an early pub-
lication, Chernoff writes that "postulate 5 ∗ is not imposed in our definition of a ra-
tional solution" (Chernoff, 1954, page 430) and Aizerman and Aleskerov chime in by
(Aizerman and Aleskerov, 1995, page 21). The characterization of set-rationalizable
stating that "this property [bα] did not find wide use in the choice theory literature"
choice functions via bα can be interpreted as a strong argument for this postulate.
By also imposing bγ, the set-expansion property that forms the counterpart of bα, we
obtain the class of self-stable generalized Condorcet extensions. In the context of social
choice, this class comprises appealing social choice functions like the minimal covering
set and the essential set, yet excludes other well-known rules like all scoring rules, all
scoring runoff rules, and all weak Condorcet extensions. As such, self-stability defines
a fascinating class of social choice functions, which offers an interesting way around the
impossibility results that have haunted social choice theory for such a long time.
Acknowledgements
We thank Nicholas Baigent and Cristopher Tyson for helpful suggestions.
This material is based upon work supported by the Deutsche Forschungsgemeinschaft
under grant BR 2312/3-3.
References
M. Aizerman and F. Aleskerov. Theory of Choice, volume 38 of Studies in Mathematical
and Managerial Economics. North-Holland, 1995.
M. Aizerman and A. Malishevski. General theory of best variants choice: Some aspects.
IEEE Transactions on Automatic Control, 26(5):1030 -- 1040, 1981.
M Allingham. Rational Choice. Macmillan, 1999.
17
R. Arlegi. A note on Bossert, Pattanaik and Xu's "Choice under complete uncertainty:
Axiomatic characterization of some decision rules". Economic Theory, 22:219 -- 225,
2003.
K. J. Arrow. Social Choice and Individual Values. New Haven: Cowles Foundation, 1st
edition, 1951. 2nd edition 1963.
K. J. Arrow. Rational choice functions and orderings. Economica, 26:121 -- 127, 1959.
D. Austen-Smith and J. S. Banks. Positive Political Theory I: Collective Preference.
University of Michigan Press, 2000.
J. S. Banks. Acyclic social choice from finite sets. Social Choice and Welfare, 12:293 -- 310,
1995.
S. Barber`a, W. Bossert, and P. K. Pattanaik. Ranking sets of objects. In S. Barber`a, P. J.
Hammond, and C. Seidl, editors, Handbook of Utility Theory, volume II, chapter 17,
pages 893 -- 977. Kluwer Academic Publishers, 2004.
D. H. Blair, G. A. Bordes, J. S. Kelly, and K. Suzumura. Impossibility theorems without
collective rationality. Journal of Economic Theory, 13:361 -- 379, 1976.
J. H. Blau and R. Deb. Social decision functions and the veto. Econometrica, 45(4):
871 -- 879, 1977.
G. Bordes. Consistency, rationality and collective choice. Review of Economic Studies,
43(3):451 -- 457, 1976.
G. Bordes. Some more results on consistency, rationality and collective choice. In J. J.
Laffont, editor, Aggregation and Revelation of Preferences, chapter 10, pages 175 -- 197.
1979.
G. Bordes. On the possibility of reasonable consistent majoritarian choice: Some positive
results. Journal of Economic Theory, 31:122 -- 132, 1983.
G. Bordes and N. Tideman.
Independence of irrelevant alternatives in the theory of
voting. Theory and Decision, 30(2):163 -- 186, 1991.
F. Brandt.
report,
http://arxiv.org/abs/0803.2138, 2009. Presented at the 9th International Meeting of
the Society of Social Choice and Welfare.
tournaments.
Technical
Minimal
stable
sets
in
D. E. Campbell and J. S. Kelly. Impossibility theorems in the Arrovian framework. In
K. J. Arrow, A. K. Sen, and K. Suzumura, editors, Handbook of Social Choice and
Welfare, volume 1, chapter 1. Elsevier, 2002.
H. Chernoff. Rational selection of decision functions. Econometrica, 22:422 -- 443, 1954.
18
Marquis de Condorcet. Essai sur l'application de l'analyse `a la probabilit´e des d´ecisions
rendues `a la pluralit´e des voix. Imprimerie Royale, 1785. Facsimile published in 1972
by Chelsea Publishing Company, New York.
B. Dutta. Covering sets and a new Condorcet choice correspondence. Journal of Eco-
nomic Theory, 44:63 -- 80, 1988.
B. Dutta and J.-F. Laslier. Comparison functions and choice correspondences. Social
Choice and Welfare, 16(4):513 -- 532, 1999.
P. C. Fishburn. The Theory of Social Choice. Princeton University Press, 1973.
P. C. Fishburn. Condorcet social choice functions. SIAM Journal on Applied Mathe-
matics, 33(3):469 -- 489, 1977.
P. Gardenfors. Manipulation of social choice functions. Journal of Economic Theory, 13
(2):217 -- 228, 1976.
H. G. Herzberger. Ordinal preference and rational choice. Econometrica, 41:187 -- 237,
1973.
J. Kelly. Strategy-proofness and social choice functions without single-valuedness. Econo-
metrica, 45(2):439 -- 446, 1977.
J. S. Kelly. Arrow Impossibility Theorems. Economic Theory and Mathematical Eco-
nomics. Academic Press, 1978.
G. Laffond, J.-F. Laslier, and M. Le Breton. More on the tournament equilibrium set.
Math´ematiques et sciences humaines, 31(123):37 -- 44, 1993.
J.-F. Laslier. Rank-based choice correspondences. Economics Letters, 52:279 -- 286, 1996.
J.-F. Laslier. Aggregation of preferences with a variable set of alternatives. Social Choice
and Welfare, 17:269 -- 282, 2000.
A. Mas-Colell and H. Sonnenschein. General possibility theorems for group decisions.
Review of Economic Studies, 39(2):185 -- 192, 1972.
K. May. A set of independent, necessary and sufficient conditions for simple majority
decisions. Econometrica, 20:680 -- 684, 1952.
N. R. Miller. A new solution set for tournaments and majority voting: Further graph-
theoretical approaches to the theory of voting. American Journal of Political Science,
24(1):68 -- 96, 1980.
H. Moulin. Choice functions over a finite set: A summary. Social Choice and Welfare,
2:147 -- 160, 1985a.
H. Moulin. From social welfare ordering to acyclic aggregation of preferences. Mathe-
matical Social Sciences, 9(1):1 -- 17, 1985b.
19
H. Moulin. Choosing from a tournament. Social Choice and Welfare, 3:271 -- 291, 1986.
C. R. Plott. Path independence, rationality, and social choice. Econometrica, 41(6):
1075 -- 1091, 1973.
M. K. Richter. Revealed preference theory. Econometrica, 34(3):635 -- 645, 1966.
P. A. Samuelson. A note on the pure theory of consumers' behaviour. Econometrica, 5:
61 -- 71, 1938.
T. Schwartz. Rationality and the myth of the maximum. Nous, 6(2):97 -- 117, 1972.
T. Schwartz. Choice functions, "rationality" conditions, and variations of the weak
axiom of revealed preference. Journal of Economic Theory, 14:414 -- 427, 1976.
T. Schwartz. The Logic of Collective Choice. Columbia University Press, 1986.
T. Schwartz. Cyclic tournaments and cooperative majority voting: A solution. Social
Choice and Welfare, 7:19 -- 29, 1990.
A. K. Sen. Choice functions and revealed preference. Review of Economic Studies, 38
(3):307 -- 317, 1971.
A. K. Sen. Social choice theory: A re-examination. Econometrica, 45(1):53 -- 89, 1977.
A. K. Sen. Social choice theory. In K. J. Arrow and M. D. Intriligator, editors, Handbook
of Mathematical Economics, volume 3, chapter 22, pages 1073 -- 1181. Elsevier, 1986.
A. K. Sen. Internal consistency of choice. Econometrica, 61:495 -- 521, 1993.
A. K. Sen. Rationality and social choice. The American Economic Review, 85(1):1 -- 24,
1995.
A. K. Sen. Maximization and the act of choice. Econometrica, 65(4):745 -- 779, 1997.
A. D. Taylor. Social Choice and the Mathematics of Manipulation. Cambridge University
Press, 2005.
J. von Neumann and O. Morgenstern. Theory of Games and Economic Behavior. Prince-
ton University Press, 1944.
R. B. Wilson. Social choice theory without the Pareto principle. Journal of Economic
Theory, 5:478 -- 486, 1972.
20
|
1102.1261 | 2 | 1102 | 2011-08-18T09:56:19 | Symmetry in behavior of complex social systems - discussion of models of crowd evacuation organized in agreement with symmetry conditions | [
"cs.MA"
] | The evacuation of football stadium scenarios are discussed as model realizing ordered states, described as movements of individuals according to fields of displacements, calculated correspondingly to given scenario. The symmetry of the evacuation space is taken into account in calculation of displacements field - the displacements related to every point of this space are presented in the coordinate frame in the best way adapted to given symmetry space group, which the set of basic vectors of irreducible representation of given group is. The speeds of individuals at every point in the presented model have the same quantity. As the results the times of evacuation and average forces acting on individuals during the evacuation are given. Both parameters are compared with the same parameters got without symmetry considerations. They are calculated in the simulation procedure. The new program (using modified Helbing model) has been elaborated and presented in this work for realization the simulation tasks the. | cs.MA | cs | Discussion of crowd evacuation models organized according to symmetry
analysis approach
W. Sikora, J. Malinowski
Faculty of Physics and Applied Computer Science, University of Science and Technology,
al.Mickiewicza 30, 30-059 Krakow, Poland
Abstract The evacuation of football stadium scenarios are discussed as model realizing ordered states, described as
movements of individuals according to fields of displacements, calculated correspondingly to given scenario. The symmetry
of the evacuation space is taken into account in calculation of displacements field - the displacements related to every point
of this space are presented in the coordinate frame in the best way adapted to given symmetry space group, which the set of
basic vectors of irreducible representation of given group is. The speeds of individuals at every point in the presented model
have the same quantity. As the results the times of evacuation and average forces acting on individuals during the evacuation
are given. Both parameters are compared with the same ones got without symmetry considerations. They are calculated in the
simulation procedure. For realization the simulation tasks the new program (using modified Helbing model) was elaborated.
Introduction
The behavior of social systems under the action of some external conditions may be
considered in analogy to the behavior of solid states under the action of temperature or
external electric or magnetic fields. Both systems are complex, containing many interacting
elements, and are realized in strictly defined spaces. Very often these spaces are strongly
restricted, and because of these restrictions not any types of evolutions of these systems are
allowed. When these spaces are symmetrical (crystals are good examples of such situation)
the symmetry considerations conducting in the frame of theory of groups and representations
are able to predict the types of behavior of the systems, which are permitted by the symmetry
of these spaces. The symmetry analysis method (SAM) [1] from many years successfully was
used for significant simplification of descriptions of different type of phase transitions in
crystals. The first trial of application the SAM to modeling the evacuation of one floor of
skyscraper, in which the symmetry of allowed space had been broken because of some
accident, is presented in our work at KES AMSTA [2] It had been done, that evacuation along
the paths related to the field of displacements calculated by SAM in agreement with
symmetry of allowed space leads to a little longer time of evacuation (because they are not the
shortest ways to the exits) but significantly lower average force acting on evacuated
individuals. . It is known, that to big press acting on evacuated agents during evacuation occurs
very often the cause of tragic accidents, so the possibility to lower this press seems to be important.
.Taken into account the crowd psychology is rather not realistic assumption, that in empty space of
the floor evacuated individuals choose the paths to the allowed exits non along shortest way, but
along the line indicated by SAM, even they would be drown in some maner by some AMI devices.
There are such constructions, where many people are keeping together, where exists the possibility
to build the paths of evacuation in agreement with SAM field of displacements. Such construction is
for example football stadium. In this work the model of evacuation of one floor of skyscraper
regarded in [2] is applied to one sector of football stadium. The football stadium build from
many similar sectors is a very good approximation of the structure fulfilling the periodic
boundary conditions. The results of symmetry calculations done for one floor of skyscraper
may be applied also to planning the evacuation paths on stadium sector. We discuss the
evacuation of stadium which architecture takes into account the symmetry calculations and
compare it with evacuation of traditional stadium.
Elements of the symmetry analysis method
The symmetry group of some object is defined as some special set of transformations
which leave this object unchanged. As the elements of space groups appear translations,
mirror planes, inversion and some special rotations around the axis, as good as all
combinations of them. Translational symmetry gives the possibility to choose the small block
named elementary cell. The repetition of it reconstructs all system. In this way description of
all system is reduced to description of elementary cell. Such procedure reduces the number of
parameters needed to description of the system, but steel this number may be decreased by
using full symmetry restrictions, not only translational one. The properties of given system in
mathematical language are presented as functions defined on allowed space. In crystals it may
be the displacements of atoms from initial equilibrium positions, magnetic moments localized
on given positions or probability of sites occupations for example. The transformation
properties of such functions are given by matrix representations of allowed space symmetry
group. One may find the books where the representations of different groups (also
crystallographic 32 point groups and 230 space groups) are calculated and listed.
Theory of group representations had been applied for simplification of description of
many-body, complex physical systems many years ago. E. Wigner, G.J.Lubarski and A.P.
Cracknell introduced as "symmetric coordinates" set of basis vectors of irreducible
representations of molecules symmetry groups in calculations of molecules vibrations.
Lubarski discussed also the role of irreducible representations (IR) of crystal symmetry group
in crystallographic second order phase transitions. This method to the description of magnetic
ordering in crystals was at first introduced by E.F. Bertaut. Later that line of analysis has been
developed by many other theoreticians, like Izyumov and others [1]. The presentation of the
functions describing interesting system properties, localized on given set of symmetry
equivalent positions, in the usually used frame of coordinates related to the edges of
elementary cell (crystallographic system), as was mentioned previously, takes advantage of
translation symmetry only. The other symmetry relations are lost in this description and as a
consequence the description of many system properties is not as simple as possible. From
rules of the theory of group representations follows that any function defined on the space
with given symmetry group may be presented as linear combination of basis vectors of
irreducible representations (BV-s) Ψ of this group:
S= ∑
l , ν , λ
(l - number of k vectors describing the translational symmetry of ~S, ν- number of IR's, λ -
number of dimensions of ν's IR).
From equation (1) follows, that the function S may be treated as some vector in the functional
space, the BV-s Ψ play the same role in the presentation of function S, as axis of coordinate
system, and coefficients ckl, ν λ
, as components of it. When irreducible representation is taken
into account the dimension of this functional space invariant under all group transformations
is the smallest one. It leads to smallest number of ckl, ν λ
parameters, which in the frame of BV-s
Ψ presents the function S. Each choice of these free parameters ckl, ν λ
uniquely determines one
of the possible models of new structure that may be realized after the phase transition. The
c λ
k l , ν Ψ λ
k l , ν
(1)
presentation of model structures in the frame of basis vectors of irreducible representations of
the initial symmetry group instead of that in the frame of crystallographic system (x; y; z), is
the best matching to the symmetry of the problem and it provides the simplest (requiring the
lowest number of independent parameters) form of the structure description.
The form of the basis vectors and the information which of the representations take part in the
phase transition under consideration are directly given by the theory of groups and
representations. It is important to note that the basis vectors have the same translational
properties as Bloch functions. Therefore, the basis vectors may be defined on positions in
given elementary cell of the crystal as well as in the elementary cell translated by a lattice
vector t, which just corresponds to a multiplication by eikt . Not all from the possible ckl, ν λ
are
allowed, because the parameters should be selected in such way that the resulting properties
related to all atoms have real values. This condition influences on the set of equations which
ckl, ν λ
have to satisfy and as a result the number of independent free parameters is reduced and
strictly determined. After such operation the final model contains clearly defined minimum
number of free parameters and presents strictly defined relations between localized on
different crystal sites quantities describing considered property. The choice of representation
τν and the coefficients ckl, ν λ
uniquely determines the symmetry of the new, ordered structure,
independently on the kind of the property taken into account. The type of phase transition and
the property under consideration is included in the form of basis vectors. In the frame of SAM
also discussion of connections between different types of transformations is possible.
For realization of the mentioned problems the computer program MODY based on the
representations of symmetry space groups [2], and another associated programs - SPLIT,
Tensor-vis, Tensor_OpenGL, had been elaborated. They offer to find all possibilities of new
scalar, vector and tensor type structure properties allowed by the initial symmetry of the
structure.
The discussion of coexistence of different types of system behavior, leading to different
properties of the system are based on the assumption, that different functions describing these
properties of the system should have the same symmetry. From theory of representation it is
known, that these different functions should belong to the same irreducible representation of
system symmetry group.
Application the SAM to stadium evacuation scenario
The scenario of evacuation of football stadium organized according to SAM approach is
discussed in this chapter and compared with traditional model of stadium. By using the
simulations with and without implementation of SAM results the time of evacuations and
average forces acting on the pedestrians during the evacuation are calculated.
The football stadium, similar as existing in Munich for example, are the space where the big
number of individuals are localized, which with quite good approximation presents the
translational symmetry with periodic boundary conditions (usually taken into account in
crystals). This is the reason that as the first step of stadium evacuation scenario the evacuation
of one stadium sector may be discussed. It significantly reduces the number of evacuated
individuals and, as consequence, the number of parameters used for scenario description. The
real symmetry of the sector is described by P1m1 space group. For designation the ordering of
chairs and evacuation paths guaranteeing quick and safe evacuation, the field of
displacements which agree with full symmetry of available space should be calculated by
using the vector type SAM. On the base of P1m1 symmetry group the number of positions in
the sector, which should be regarded is reduced to ½, but the directions of displacements
which should be defined on these positions remain free. The best choice of displacement
directions from evacuation point of view can be made by regarding the P1m1 group as the
broken symmetry of Pmm2 group generated by τ4 representation of k=(0,0,0), which lead to
field of displacements on the xy plane (which are the polar vector quantities), and
corresponding field of speeds (which are the scalar quantities) as given in the Table1 quoted
below. Because the steps going up to exits, the displacemenst at each positions should have
the z components, directed to the up. As it is shown in the Tab1, such relation between z
components belongs to the τ1 representation. From SAM follows that τ1 may be associated
with any representation, because it doesn’t change the initial symmetry. Taking into account
the vector field of displacements got by mentioned above treatment the paths for evacuation
are design in one stadium sector.
Table1
Representation
and
destination
group
Tau1
destination
group Pmm2
Tau2
destination
group P112
Tau3
destination
group Pm11
Tau4
destination
group P1m1
Version
vectorI
vectorII
vectorIII
scalar
vectorI
vectorII
vectorIII
scalar
vectorI
vectorII
vectorIII
scalar
vectorI
vectorII
vectorIII
scalar
A1
B1
C1
P1
A2
B2
C2
P2
A3
B3
C3
P3
A4
B4
C4
P4
The polar vector type and scalar type basic vector functions of representations of Pmm2 space
group with k=(0,0,0), defined at positions 4i in the elementary cell
Pos. 4i
3: (1-x, y, 0)
2: (1-x, 1-y,0)
4: (x, 1-y,0)
Free parameter
1: (x, y, 0)
(1,0,0)
(0,1,0)
(0,0,1)
1
(1,0,0)
(0,1,0)
(0,0,1)
1
(1,0,0)
(0,1,0)
(0,0,1)
1
(1,0,0)
(0,1,0)
(0,0,1)
1
(-1,0,0)
(0,-1,0)
(0,0,1)
1
(-1,0,0)
(0,-1,0)
(0,0,1)
1
(1,0,0)
(0,1,0)
(0,0,-1)
-1
(1,0,0)
(0,1,0)
(0,0,-1)
-1
(-1,0,0)
(0,1,0)
(0,0,1)
1
(1,0,0)
(0,-1,0)
(0,0,-1)
-1
(-1,0,0)
(0,1,0)
(0,0,1)
1
(1,0,0)
(0,-1,0)
(0,0,-1)
-1
(1,0,0)
(0,-1,0)
(0,0,1)
1
(-1,0,0)
(0,1,0)
(0,0,-1)
-1
(-1,0,0)
(0,1,0)
(0,0,-1)
-1
(1,0,0)
(0,-1,0)
(0,0,1)
1
Simulation of stadium evacuation
The evacuation of the stadium sector is simulated by using the Helbing ”social forces” model
[4], which was taken as a starting point to model presented in this paper.
m i
d ⃗v i
dt
=m i
0 ⃗e i
v i
0 ( t )− ⃗v i ( t )
τ
+ ∑
j (≠i )
⃗f ij +∑
W
⃗f iW
(2)
Equation (2) on right side has a sum of three forces acting on human during evacuation simulation.
Second and third one describe sum of repel forces from all other mans and sum of repel forces from
all walls and all obstacles inside the evacuated space. First force is the most interesting at this stage.
In general it describes in which direction each person should go. In simple case like empty room with
some doors, this force can be easily calculated as a sum of attract forces from all doors. Here is also
the place for the implementation of vector field calculated by symmetry considerations. This
assumption is inadequate with more complicated geometry of the building. Figure (1) shows the
example when such approach leads to situation when pedestrian will stick for eternity in dead end.
Fig 1. The „dead-end”.
In approach presented in this paper floor space was divided into cells (this got nothing in common
with “cellular automata” approach to such kind of simulations!). Every cell can contain obstacle, free
space or exit. Every cell that contain free space have vector of desirable velocity connected with it,
calculated according to schema described below. Any person that stands inside particular cell takes
this vector as his desirable velocity. Unlike in cellular automata approach in this approach ”cell” can
contain one or more persons and any direction of velocity. Single cell size depends on building
geometry only. To make a simulation of an evacuation of single room, small number (like 7x6 for
example) of cells is more than enough. Figure (2) show example of such table presenting the final
field of velocities. These vectors can be calculated using “ray casting” method as shown in paper [5].
Fig 2. Final field of velocities
The vector field calculated for stadium sector by SAM is presented on the figure 3 and the
sector architecture which takes into account these SAM calculations for building the
evacuation paths is presented on the figure 4. It gives the map of free spaces, obstacles and
exits needed for calculation of forces acting on individuals as mentioned in the equation 2 and
velocities field used in the simulation (figure 5). The traditional stadium sector architecture is
given on the figure 6.
Figure 3 The vector field calculated by SAM related to one stadium sector for Pmm2 space
group and τ4 representation of k=(0,0,0)
Figure 4 Construction of evacuation paths in stadium sector with agreement to SAM vector
field calculations
Figure 5 Table of final field of velocities calculated for simulation of stadium sector
evacuation for SAM model
Figure 6 Table of final field of velocities calculated for simulation of stadium sector
evacuation for traditional model
The time of evacuation and average forces acting on the individuals are calculated during the
simulation by using SAM related and traditional architectures of stadium. The number of
evacuated individuals – 408 - taken for simulation agrees with number of sites in the sector of
Munchen football stadium and is the same for both analysed architectures. This two models
differ only in distribution of chairs and evacuation paths inside the sector. The results for both
models – SAM related and traditional - are given as the plots presenting the average force
acting on one individual as the function of time from the beginning to the end of evacuation.
The comparison of these functions is seen on the figure 7.
Figure 7. Average forces acting on pedestrian at each step of evacuation process calculated for
traditional and SAM stadium models.
Conclusions
The new model of architecture of sectors of football stadium is proposed, which leads
to the best sector evacuation. As the criterion of goodness of evacuation two parameters - the
time of evacuation te (expressed by the steps of simulation), and average force acting on the
individuals during the evacuation Fa – are taken into account. The model, in which the
evacuation paths are related to stadium symmetry considerations, as may be seen on the figure
(7) significantly minimalize the Fa while the te remain almost the same as for the traditional
construction.
Acknowledgement
This work is partially supported by EU program SOCIONICAL (FP7 No 231288).
References
[1]. W. Sikora, L. Pytlik, “Application of Symmetry Analysis to Description of Ordered
Structures in Crystals”, Copyright @ 2010 by Nova Science Publishers , NY, Chapter 1 (pp 1-
40) in “Group Theory: Classes, Representations and Connections, and Applications.”
[2]. W. Sikora, J. Malinowski “Symmetry Approach to Evacuation Scenarios”, 4th KES
International Symposium, KES-AMSTA 2010, Gdynia, Poland, June 23-25, 2010,
Proceedings Part II, Lecture Notes in Artificial Intelligence ( Subseries of Lecture Notes in
Computer Science, Edited by R Goebel, J. Siekmann, and W. Wahlster, ) Springer-Verlag
Berlin Heidelberg 2010, p 229-241
[3]. D.Helbing, et al., Simulating dynamical features of escape panic, Nature, vol 407,
2000, p. 487
[4] D.Helbing, P.Molnár “Social force model for pedestrian dynamics”,
Physical Review E 51, (1995) 4282-4286
[5] T.Kretz, C.Bonish, P.Vortish, „Comparision of Various Methods for the Calculation of the
Distance Potential Field”, arXiv:0804.3868v1 [physics.comp-ph]
|
1609.01808 | 1 | 1609 | 2016-09-07T02:34:45 | Review on Microscopic Pedestrian Simulation Model | [
"cs.MA"
] | Microscopic Pedestrian Simulation Model is computer simulation model of pedestrian movement where every pedestrian in the model is treated as individual. Most of pedestrian researches have been done on macroscopic level, which does not consider the interaction between pedestrians and does not well suited for prediction of pedestrian flow performance in pedestrian areas or building with some objects that reduce the effective width of it. In the other hand, microscopic level has more general usage and considers detail of the design. Tough the analytical model for microscopic pedestrian model is existed exist, the numerical solution of the model is very difficult and simulation is favorable. The model has practical application of Evacuation from building, Design of pedestrian area, and Experimental & Optimization Design Tool. In general, Microscopic Pedestrian Simulation Model consist of two terms, that make the pedestrian moving toward the destination and make repulsive effect toward other pedestrian or obstacles. | cs.MA | cs | Teknomo, Kardi; Takeyama, Yasushi; Inamura, Hajime, Review on Microscopic Pedestrian Simulation
Model, Proceedings Japan Society of Civil Engineering Conference March 2000, Morioka, Japan, March
2000
Review on Microscopic Pedestrian Simulation Model
Kardi Teknomo1, Yasushi Takeyama2, Hajime Inamura3
collisions. Two forces are work on each pedestrian.
First, magnetic force as formulated by Coulomb's
law, which is depend on the intensity of magnetic
load of a pedestrian and distance between
pedestrians. Another force acts on a pedestrian to
avoid the collision with another pedestrian or
obstacle exerts acceleration a is calculated as:
a = V. cos (alpha).tan (beta)
(2)
A
alpha
V
a
RV
beta
C
B
Figure 1. Additional Repulsive Force on Magnetic
Force Model
Total of forces from goals, walls and other
pedestrians act on each pedestrian, and its decides
the velocity of each pedestrian each time.
c. Social Force Model
Helbing (1991-99) was developed Social Force
Model with Molnar, and Vicsek which has similar
principles of both Benefit Cost cellular Model and
Magnetic Force Model. A pedestrian is subjected to
social forces that motivate the pedestrian. The
summation of these forces act upon a pedestrian
create acceleration dv/dt as:
(cid:71)
)t(v
)t(vdm
i
i
=
dt
∑
+
(cid:71)
evm
io
(cid:71)
(cid:71)
))t(x),t(x(f
ij
(cid:71)
))t(x(f
b
(cid:71)
ξ+
(3)
)t(
+
−
(cid:71)
(cid:71)
(cid:71)
τ
i
j
i
i
The first term in the right hand of eq.(3) represent
the motivation to reach the goal. The model based
on assumption that every pedestrian has intention to
reach certain destination at a certain target time.
The direction is a unit vector from a particular
location to the destination point.
is existed exist [Henderson
Microscopic Pedestrian Simulation Model is
computer
simulation model of pedestrian
movement where every pedestrian in the model is
treated as
individual. Most of pedestrian
researches have been done on macroscopic level
[for best classical examples: Fruin (1971), HCM
(1985)], which does not consider the interaction
between pedestrians and does not well suited for
prediction of pedestrian flow performance in
pedestrian areas or building with some objects
that reduce the effective width of it. In the other
hand, microscopic level has more general usage
and considers detail of the design. Tough the
analytical model for microscopic pedestrian
model
(1974),
Helbing (1992)], the numerical solution of the
model
is
favorable. The model has practical application of
Evacuation from building, Design of pedestrian
area, and Experimental & Optimization Design
Tool. There are several microscopic pedestrian
simulation models:
a. Benefit Cost Cellular Model
Gipps and Marksjo (1985) propose this model. It
simulates the pedestrian as particle in a cell. Each
cell can be occupied by at most one pedestrian
and a score assigned to each cell on the basis of
proximity to pedestrians. The score represent the
gain made by the pedestrian when moving toward
his destination. The repulsive effect of the nearby
pedestrians, and formulated as:
X-
is very difficult and simulation
(1)
(S ).X-
)X-
(S .K
X-
i
Score
=
).(D
i
X-S
i
i
i
i
2
i
i
X-D .
i
i
).(D
2
i
i
-
1
)
αΔ
-
(
2
β+
≠
(j
)i
the
in
(including
the nine-cell neighbor of
location of
Where the field of two pedestrian overlap, the
score in each cell is the sum of the score
generated by pedestrian individually. The score is
the
calculated
pedestrian
the
pedestrian). Pedestrian will move to the next cell
that has maximum net benefit.
b. Magnetic Force Model
Prof. Okazaki (1979-93) developed this model
with Matsushita. The application of magnetic
models and equation of motion in the magnetic
field cause pedestrian movement. Each pedestrian
and obstacle have positive pole. Negative pole is
assumed to be located at the goal of pedestrians.
Pedestrian moves to their goals and avoids
1 Doctoral Student, Graduate School of Information Science, Tohoku University Japan
2 Associate Professor, Graduate School of Information Science, Tohoku University Japan
3 Professor, Graduate School of Information Science, Tohoku University Japan
1
Table 1. Comparison Microscopic Pedestrian Simulation Models
Movement to goal
Benefit Cost Cellular
Gain Score
and positive
Magnetic Force
Positive and negative
magnetic force
Positive
magnetic forces
continuous
physical meaning
queuing, way finding in
maze, evacuation
heuristic
possible
by inspection
Social Force
Intended velocity
Interaction forces
continuous
physical meaning
queuing,
-
self
organization, oscillatory
change
mathematics
not possible
by inspection
Flows. Mathematics and Computers in
Simulation 27, pp. 95-105.
3. Helbing, D. (1991) A mathematical model
for the behavior of pedestrians. Behavioral
Science 36, pp. 298-310.
4. Helbing, D. (1992) A fluid-dynamic model
for the movement of pedestrians. Complex
Systems 6, pp. 391-415.
5. Helbing, D. and Vicsek, T. (1999) Optimal
Self-Organization. New Journal of Physics
1, 13.1-13.17.
6. Henderson, L. F. (1974) On the Fluid
Mechanic of Human Crowd Motions,
Transportation Research 8, pp. 509-515.
7. Matsushita and Okazaki (1991) A study of
Simulation Model for Way Finding Behavior
in Mazes. Journal of
by Experiments
Architecture, Planning. Environment
Engineering. AIJ, No.429, pp. 51-59.
8. Okazaki, S (1979) A Study of Pedestrian
Movement in Architectural Space, Part 1:
Pedestrian Movement by the Application on
of Magnetic Models. Trans. of A.I.J.,
No.283, pp. 111-119.
9. Transportation Research Board
(1985)
Highway Capacity Manual, Special Report
204 TRB, Washington D.C.
m,K,
Notation
Si = vector location of target cell
Xi = vector location of the subject
Di = vector location of destination
Δ = Distance between cell and the pedestrian.
,βα
V = velocity of pedestrian A
RV = relative velocity
xi (t) =Location of pedestrian i at time t.
vi(t) = velocity of pedestrian i at time t = dxi(t)/dt
vo = intended velocity.
ei = direction of pedestrian i ∈{(1,0),(=1,0)}
iξ = Fluctuation of individual velocities
= Constants
Repulsive Effect
Cost Score
Pedestrian movement
Value of variables
Phenomena explained
discreet
arbitrary
queuing
cellular based
by inspection
Higher programming
orientation in
Evacuation Application possible
Parameter Calibration
The ideal speed is equal to remaining distance per
remaining time. The remaining distance is the
different between the destination point and the
location at that time, while the remaining time is
different between target time and the simulation
time. Intended Velocity is the ideal speed times
the unit vector of direction. Speed limitation
(maximum and minimum) can be put to make the
speed more realistic. The second and last term of
the right hand side of eq (3) designate for
interaction between pedestrians and pedestrian to
obstacles (i.e. column) and interaction pedestrian
with the boundaries
Comparison of the Models
In general, Microscopic Pedestrian Simulation
Model consist of two terms, that make the
pedestrian moving toward the destination and
make repulsive effect toward other pedestrian or
obstacles. Benefit Cost cellular also uses arbitrary
scores while Magnetic and Social Force model
has more variables with physical meaning. Social
Force model is highly mathematics evolved to
explain
the behavior of pedestrian, while
Magnetic Force Model is more developed in
heuristic approach.
None of the microscopic pedestrian simulation
models have calibrated and validated
their
constant parameters based on real pedestrian
movement data. It has no statistical guarantee that
the parameters will work for general cases or
even for specific region. Therefore, in order to
use
research
direction is proposed to be done to automate data
collection of individual movement data, and to
set up statistical procedures for calibration and
calibration of the parameters.
REFERENCES
1. Fruin, J.J. (1971). Designing for Pedestrians:
A Level of Service Concept. Highway
research Record 355, 1-15.
the model practically,
2. Gipps, P.G. and Marksjo, B. (1985) A
for Pedestrian
Micro-Simulation Model
future
2
|
1706.05839 | 1 | 1706 | 2017-06-19T09:09:50 | On Optimal Group Claims at Voting in a Stochastic Environment | [
"cs.MA",
"cs.SI",
"eess.SY",
"math.OC",
"physics.soc-ph"
] | There is a paradox in the model of social dynamics determined by voting in a stochastic environment (the ViSE model) called "pit of losses." It consists in the fact that a series of democratic decisions may systematically lead the society to the states unacceptable for all the voters. The paper examines how this paradox can be neutralized by the presence in society of a group that votes for its benefit and can regulate the threshold of its claims. We obtain and analyze analytical results characterizing the welfare of the whole society, the group, and the other participants as functions of the said claims threshold. | cs.MA | cs |
ISSN 0005-1179, Automation and Remote Control, 2017, Vol. 78, No. 6, p. 1087. c(cid:13) Pleiades Publishing, Ltd., 2017.
Original Russian Text c(cid:13) V.A. Malyshev, P.Yu. Chebotarev, 2017, published in Avtomatika i Telemekhanika, 2017, No. 6, 157 -- 172.
INTELLECTUAL CONTROL SYSTEMS, DATA ANALYSIS
On Optimal Group Claims at Voting
in a Stochastic Environment
V.A. Malyshev∗,∗∗,a and P.Yu. Chebotarev∗,∗∗,b
∗Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia
∗∗Moscow Institute of Physics and Technology, Moscow, Russia
e-mail: [email protected], [email protected]
Received October 28, 2016
Abstract -- There is a paradox in the model of social dynamics determined by voting in a stochastic
environment (the ViSE model) called "pit of losses." It consists in the fact that a series of democratic
decisions may systematically lead the society to the states unacceptable for all the voters. The paper
examines how this paradox can be neutralized by the presence in society of a group that votes for
its benefit and can regulate the threshold of its claims. We obtain and analyze analytical results
characterizing the welfare of the whole society, the group, and the other participants as functions of
the said claims threshold.
Keywords: ViSE model, social dynamics, voting, stochastic environment, pit of losses.
DOI: 10.1134/S0005117917060091
1. INTRODUCTION
It is known that decisions made by the majority of voters on the proposals generated by a stochastic
environment may systematically lead to situations unacceptable for all the voters. This voting paradox
(the "pit of losses" paradox) has been described in [1]. A natural approach to neutralizing this paradox
is increasing the majority threshold (i.e., the number of votes necessary to accept a proposal).
However, there is another approach that deserves consideration. Suppose that there is a "group" in
the society whose participants pursue group rather than individual interests. Moreover, the group has
a "claims threshold," i.e., the minimum profitability of proposals the group considers acceptable for it.
Which group claims threshold is profitable (optimal) for the whole society, for the group and for the
other participants? In this paper, we give the answers to these and several other questions.
Consider the main features of the ViSE (Voting in Stochastic Environment) model [2]. Let society
consist of n members. ℓ members are egoists and g = n − ℓ are group members; therefore, δ = ℓ/n is
the proportion of egoists. The ViSE model extends the voting model introduced by A.V. Malishevsky
(see [3]), another extension of which is the well-known dynamic multidimensional voting model [4].
Each participant is characterized by the current value of his/her capital (an alternative interpretation
of it is utility). A proposal (of the environment) is a vector of proposed participants' capital increments. In
the present paper, these increments are realizations of independent identically distributed normal random
variables. The parameters of the corresponding normal distribution N (µ, σ2) can be easily interpreted:
the cases of µ > 0, µ = 0, and µ < 0 correspond to a favorable, neutral, and unfavorable environment,
respectively; σ characterizes the scatter of the proposals.
In the basic version of the ViSE model, the only stochastic agent is the environment. The behavior
of the voters is deterministic in contrast to the random voting models [5]. An egoist votes for those and
only those proposals that increase his/her capital. All members of a group vote for the proposals that
are beneficial to this group and vote against the other proposals. The group can consider a proposal as
1087
1088
MALYSHEV, CHEBOTAREV
a favorable one in the following cases: (a) the average increase (which can be zero, positive, or negative)
of a group member's capital exceeds the chosen claims threshold t [6], or (b) the percentage of the group
members receiving positive capital increments exceeds some threshold. In this paper, we consider the
first case. The group supports those and only those proposals in which the average capital increment of
its members is greater2 than a selected threshold t. Let α ∈ [0, 1] be the strict relative majority threshold
for all votes of the society.
If the proportion of the society supporting a proposal is greater than α,
then the proposal is accepted and then implemented. The implemented proposals constitute the voting
trajectory. The subject of our study is the statistical dependence of such trajectories (which express the
dynamics of social welfare) on the parameters of the model.
It should be noted that the behavior of the voters in the ViSE model corresponds to the Downsian
concept [8], more precisely, to its operationalization that is based on the comparison of the proposed state
with status quo [9]. Dynamic voting models have been studied in the theory of legislative bargaining [10],
where in some cases [11 -- 13] stochastic generation of proposals has been assumed. In this type of models,
voters have "ideal states" that maximize their individual utility and the central problem is searching
for equilibrium. Such kind of problems are trivial or even pointless in the ViSE model because the
participants do not have finite ideal states. However, in this model, individual utilities (capitals) can be
naturally aggregated into group utilities3. This makes it possible to study collectivistic and altruistic
voting strategies. Because of this, the ViSE model gives the researcher a tool for studying cooperation
and egoism, which provides an alternative to the simple games (such as Prisoner's Dilemma, Ultimatum,
Avatamsaka, Public Goods, etc.) traditionally used for these purposes.
The ViSE model allows one to identify a number of social phenomena, mechanisms, and relationships
associated with collective decisions. Their presence in practice (and degree) is a subject of special study.
The model also allows to test various approaches to bringing society to the desired or optimal state. Of
course, this model can not reflect all aspects of social reality, however, all its predictions are characterized
by mathematical transparency, therefore, the conditions for the manifestation of these phenomena are
amenable to verification.
As indicated above, the paper examines the social dynamics in the presence of a group with a claims
threshold. A moderate group of this kind can be interpreted as an elite.
We obtain the expressions for the mathematical expectations of egoists' and group members' capital
increments in the described society. These expressions are presented in Section 2.
In Section 3, it is
shown that the group can choose such a claims threshold (it can be called optimal) that the expected
capital increment of the society is maximal. We provide a closed form expression for this threshold.
The dependence of the participants' expected capital increments on the majority threshold is studied in
Section 4. The influence of choosing the optimal claims threshold on the size of the "pit of losses" is
examined in Section 5.
2. THE CAPITAL INCREMENTS OF EGOISTS AND GROUP MEMBERS
If the group does not support a proposal, then for its acceptance, the number of egoists' votes cast
for it must exceed αn. Otherwise, if the group supports the proposal, then it is necessary and sufficient
that the egoists give more then
αn − g = αn − n + ℓ = (α + δ − 1)n
votes, i.e., the share (in the society) of the egoists supporting the proposal should exceed the value
2 This voting rule can be formulated in terms of the deterministic version of the model used in [7].
3 Such aggregation is natural because of the assumption of the transferability of utilities.
γ = α + δ − 1.
(1)
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
ON OPTIMAL GROUP CLAIMS AT VOTING
1089
Fig. 1. The expected one-step capital increments of a group member (1), an egoist (2), and a randomly selected
participant (3), where n = 100, ℓ = 50, µ = −1, σ = 10, and α = 0.5.
Proposition 1. The mathematical expectation of egoist's capital increment edE in one voting step is
equal to
M(edE ) =hµ+(µ, σ, ℓ, γn) µ+(µ, σ, ℓ, αn)i"P
Q#,
(2)
(3)
where a row vector is multiplied by a column vector , µ+(µ, σ, ℓ, ℓ0) is the expectation of the normal voting
sample of size ℓ with parameters (µ, σ) and voting threshold ℓ0 [14], P = F(cid:16) (µ−t)√g
F (·) is the standard normal distribution function.
σ
(cid:17) , Q = 1 − P, and
The proofs of all results are given in the Appendix.
σf
normal density.
−1#!,
Proposition 2. The mathematical expectation of a group member's capital increment edG in one voting
step is equal to
x=[ξ]+1b(x ℓ) ≈ F(cid:18)− [ξ]+0.5−pℓ
σ(cid:1), q = 1 − p, P and Q are defined in Proposition 1, f = f(cid:16) (µ−t)√g
M(edG ) =hFγn Fαni µ"P
Q# +
√g" 1
√pqℓ (cid:19), b(x ℓ) =(cid:0)ℓ
x(cid:1)pxqℓ−x is the binomial probability function,
where Fξ =Pℓ
(cid:17), and f (·) is the standard
p = F(cid:0) µ
Dependence of expected capital increments M(edE ) and M(edG ) of the participants on the claims thresh-
Fig. 1 demonstrates that in a moderately unfavorable environment (µ/σ = −0.1), the group (half of
the society in the present case) has a maximum income at zero claims threshold t. However, the egoists
jointly lose more than the group wins, therefore, the expected capital increment of the whole society is
negative, and so the society is gradually losing welfare. If the group's claims increase, then its income
decreases more slowly than the loss of egoists. When the claims threshold t is slightly higher than 0.2,
the expected capital increment of the whole society becomes positive. When t = 1, the expected capital
increment of the society reaches a maximum. A group with higher claims is blocking more and more
profitable proposals, and the expected capital increment of the society is decreasing to zero.
old t are shown in Fig. 1.
σ
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
1090
MALYSHEV, CHEBOTAREV
Fig. 2. The expected one-step capital increments of an egoist (a), a group member (b), a randomly selected partici-
pant (c), and the difference between the expected capital increments of a group member and an egoist (d ) as functions
of the adjusted claims threshold of the group t/σ and the proportion of the egoists δ, when n = 100, µ = −0.1, σ = 1,
and α = 0.5.
Fig. 3. The expected one-step capital increments of an egoist (a), a group member (b), a randomly selected participant
(c), and the difference between the expected capital increments of a group member and an egoist (d ) as functions of
the adjusted claims threshold of the group t/σ and the proportion of the egoists δ, when n = 100, µ = 0.1, σ = 1,
and α = 0.5.
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
ON OPTIMAL GROUP CLAIMS AT VOTING
1091
Fig. 4. The expected one-step capital increments of an egoist (1), a group member (2), and a randomly selected
participant (3): sections of the surfaces in Fig. 2; n = 100, µ = −0.1, σ = 1, and α = 0.5.
Dependencies of the expected one-step capital increments of the participants on the adjusted claims
threshold of the group t/σ and the proportion of the egoists δ are presented in Figures 2 and 3 for the
cases of moderately unfavorable (µ/σ = −0.1) and favorable (µ/σ = 0.1) environments. Fig. 4 shows
the sections of the surfaces presented in Fig. 2. It allows to compare the capital increments of an egoist,
a group member, and a randomly selected participant. These dependencies are characterized by the
following regularities:
1. For µ < 0 and any proportion of the egoists, the average capital increment of an egoist monotonically
increases with the claims threshold of the group, because the number of accepted proposals that are
unprofitable for egoists decreases. The fastest growth of the capital increment is observed at t close
to µ;
2. When the proportion of egoists is small and the claims threshold of the group is low, the group accepts
many unfavorable proposals. This causes a fast loss of egoists' welfare;
3. There is a maximum (with respect to the proportion δ of egoists) of the expected capital increment of
an egoist in the domain of sufficiently high group's claims thresholds t and a high proportion of egoists
(δ > 0.9). For large t, this maximum exceeds the average capital increment in the society consisting of
egoists only. This is caused by the fact that a small group with a high t votes against most proposals.
Therefore, the proposals must have a high support from the egoists to be accepted. Such an actual
increase of the majority threshold for egoists (for the approvement of a proposal, 51 votes are needed,
but of 90 -- 98 votes instead of 100) radically reduces the "pit of losses" [1], and the average capital
increment of egoists reaches the largest values;
4. According to Theorem 1 in [1], for the society consisting of egoists and the environment with the
parameters under consideration, the optimal majority threshold is α0 ≈ 0.52;
5. The group and the egoists lose welfare when the claims threshold and the proportion of egoists are
low. When t is close to µ, the expected capital increment of a group member increases rapidly with the
growth of the claims threshold. However, in contrast to the average capital increments of the egoists,
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
1092
MALYSHEV, CHEBOTAREV
for any proportion of the group, the expected capital increment of a group member has a maximum
at t = 0 (since this threshold optimizes the group benefit). This maximum takes the greatest value
when the group is small (cf.
[15]);
6. In the domain of high thresholds t and the proportion of egoists δ > 0.9, the expected capital increment
of a group member has a minimum with respect to δ, which is lower than the corresponding value for
egoists. Thus, excessive claims of a small group worsen its position in comparison with the position
of egoists;
7. The expected capital increment of a group member is higher than that of an egoist everywhere, except
for the domains of very high and very low claims thresholds and small group sizes.
The expected capital increment of a randomly selected participant (surfaces (c) in Figures 2 and 3)
is expressed by the weighted average of the functions (a) and (b) with weights equal to the proportions
of egoists and the group, respectively. The surface (c) has a combination of properties of the averaged
surfaces. A randomly chosen participant loses capital when the group's claims threshold is low and the
proportion of egoists is small. If the share of egoists is moderate, then the expected capital increment of
an average participant (and of the whole society) has a maximum with respect to t/σ; the corresponding
optimal group's claims threshold is found in Section 3. The value of the maximum is higher for a larger
group. In comparison to the maximum for the group, this maximum is shifted to the domain of larger
t/σ, because if t/σ grows from the group's optimum, then the expected capital increment of a group
member decreases slower than the expected capital increment of an egoist increases. This leads to the
growth of the expected capital increment of the whole society.
The established relationships allow us to draw the following conclusions.
The zero claims threshold is optimal for the group in any environment. It is advantageous for the
group to be relatively small (about 15 members out of 100 participants) in this case. In a moderately
unfavorable environment, such a group becomes richer, whereas egoists lose welfare.
In an unfavorable environment, the optimal group's claims threshold is positive. The decrease of this
threshold from the optimum causes a faster loss for the society then the increase from the optimum.
In a favorable environment, on the contrary, the optimal group's claims threshold is negative, and
the expected capital increment of the society decreases faster when the threshold is deviated from the
optimal value to the positive direction. This conclusion is analogous to the one of [1] saying that the
more favorable environment is encountered, the lower majority threshold is optimal.
3. OPTIMAL GROUP'S CLAIMS THRESHOLD
In this section, we obtain an analytical expression for the optimal threshold, i.e., for the group's claims
threshold that maximizes the capital increment of the society. Let β be the ratio of the number of egoists
to the number of group members:
β =
δ
.
1 − δ
ℓ
n − ℓ
=
Theorem 1. The expected one-step capital increment of the society reaches its maximum at the group's
claims threshold
t0 =
β
Fγn − Fαn(cid:0)µ+(µ, σ, ℓ, αn) − µ+(µ, σ, ℓ, γn)(cid:1),
(4)
where the notations is introduced in Section 1 and Propositions 1 and 2.
Let us find out how the optimal threshold t0 depends on the model parameters. Since γ < α (see (1)),
it follows from the definition of Fξ that Fγn − Fαn > 0 (and the smaller the share of the group 1 − δ,
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
ON OPTIMAL GROUP CLAIMS AT VOTING
1093
the smaller this difference); β is nonnegative. Consequently, the sign of t0 coincides with the sign of the
difference µ+(µ, σ, ℓ, αn) − µ+(µ, σ, ℓ, γn).
Fig. 5,a demonstrates the dependence of t0 on the proportion of egoists δ and the majority threshold α
for n = 100 and µ/σ = 0.1.
We note that for a large group, the absolute value of threshold t0 is small because the factors β and
(Fγn − Fαn)−1 are close to zero. The decrease of t0 with the increase of the majority threshold α becomes
faster with the growth of the number of egoists. The shape of this surface does not substantially depend
on the sign of µ.
Fig. 5,b shows the dependence of t0 on the proportion of egoists δ and the mean of the environment
proposals µ for n = 100, α = 0.5, and σ = 1.
The two surfaces in Fig. 5 have the same shape. When µ increases, t0 decreases and the smaller the
group is, the faster the change is.
Fig. 6 shows various sections of the surface presented in Fig. 5,a by the planes of fixed majority
thresholds. For α = 0.46, we have t0 ≈ 0; for the smaller α's, the optimal claims threshold t0 is positive
and increases with δ; for the lager α's, it is negative and decreases.
Corollary. If the votes of a group are sufficient to accept a proposal (α < 1 − δ), then
t0 =
β
1 − Fαn
(µ+(µ, σ, ℓ, αn) − µ).
If the votes of the egoists are insufficient to accept a proposal (δ 6 α), then
t0 = −
β
Fγn
µ+(µ, σ, ℓ, γn).
If both of the above conditions are met (δ 6 α < 1 − δ), then
t0 = −β µ.
In the latter case, the optimal group's claims threshold t0 has an extremely simple expression. If the
number of votes necessary for a proposal to be accepted exceeds the number of egoists and does not
exceed the size of the group, then t0 is in the same ratio to (−µ) as the number of egoists to the number
of group members is, and the larger the group is, the closer to zero t0 is.
Fig. 5. Dependence of the optimal claims threshold t0 of the group on: (a) the proportion of egoists δ and the majority
threshold α for µ = 0.1, n = 100, and σ = 1; (b) the proportion of egoists δ and the mathematical expectation of the
proposals µ for α = 0.5, n = 100, and σ = 1.
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
1094
MALYSHEV, CHEBOTAREV
Fig. 6. Dependence of group's optimal claims threshold t0 on the proportion of egoists δ for various majority thresholds
(α = 0.15 (1), α = 0.46 (2), α = 0.5 (3), α = 0.6 (4), and α = 0.9 (5)); n = 100, µ = 0.1, σ = 1.
4. DEPENDENCE OF CAPITAL INCREMENTS ON THE MAJORITY THRESHOLD
The expression of the optimal majority threshold for a society consisting of egoists has been obtained
in [1]. Now we consider the problem of optimizing the majority threshold α for the society that consists
of egoists and a group. Recall that the optimal threshold is the threshold that maximizes the increment
of the total capital of the society.
Using Propositions 1 and 2 consider the analytical dependencies of the expected capital increments
of an egoist, a group member, and a randomly selected participant on the majority threshold α and the
proportion of egoists δ in the moderately unfavorable environment (µ/σ = −0.1) for the zero group's
claims threshold t.
The sections of the corresponding surfaces by the planes of equal egoists shares are shown in Fig. 7. If ℓ
is not very high, then sections (a) have two maxima with respect to α; the curves of maxima arguments
intersect at the point (δ = 1, α = α0), where α0 is the optimal majority threshold in the society consisting
of egoists. Consider the reasons for the appearance of these "ridges."
We start with the relationship for egoists (Fig. 7,a). At a very low proportion of egoists and a low
majority threshold, an egoist has a high expected capital increment, because the probability of the event
that all the egoists are satisfied with a proposal is markedly different from zero, while their votes are
sufficient to accept such a proposal. Therefore, their expected one-step capital increment is comparable
in the absolute value to µ. When the majority threshold is higher than the proportion of egoists, then
they are unable to accept a proposal by their votes only, and their expected capital increment is much
lower. The second maximum is realized at the majority threshold close to 1. Its appearance is caused by
the following fact. When the adoption of a proposal requires both the votes of the group and the egoists
(i.e., overall, the egoists have a veto), then egoists' influence on decision-making is quite noticeable, and
their interests are substantially taken into account. However, the support of the proposal by the group
is mandatory for the realization of this scenario, while with µ < 0, it does not happen very often, so this
maximum is much lower than the first one.
If the number of egoists increases, then both maxima still exist, but they shift to α0 (with a linear
dependence of their arguments on δ) and become lower. It can be simply explained. Here, one egoist
capable of accepting a proposal is replaced by a "clique" which less often provides the required number
of votes. The expected capital increment of a "clique" member (when the clique is satisfied with the
proposal) is lower than that of the above egoist due to the law of large numbers. The part of the curve
behind the second maximum lies somewhat lower, since an excessive majority threshold does not even
allow to accept proposals that are beneficial to both the group and most egoists.
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
ON OPTIMAL GROUP CLAIMS AT VOTING
1095
The dependence for the group (Fig. 7,b) has two "slopes" on the δ-sections. Between them, there is
a "plateau." It shows that the majority threshold located between the maxima of the expected capital
increment of the egoists is quite beneficial for the group. Lower thresholds allow the egoists to ignore the
interest of the group, which reduces the group welfare. Excessive thresholds do not even allow to accept
highly profitable proposals. As can be expected, the "plateau" is higher when the proportion of egoists
is larger, because a smaller group has a greater benefit per participant.
Fig. 7,c depicts the expected capital increment of a randomly selected participant. As in the case of
the egoists, the corresponding surface in the coordinates α, δ has two "ridges." When the proportion of
egoists δ is high, the surface of the expected capital increment of a randomly selected participant is close
to that of an egoist. The height of the maxima decrease as the number of egoists increases. At a low δ,
the surface of a randomly selected participant is close to the group surface and does not have high peaks.
Fig. 7,d shows the difference between the expected capital increments of a group member and an egoist.
This difference is positive except for the domain near the first "ridge" of the egoists' surface, where they
win, because the egoists can "dictate their terms" in this case.
5. CLAIMS THRESHOLD ADJUSTMENT TO REDUCE THE "PIT OF LOSSES"
The group can reduce the size of the "pit of losses" (the area with negative expected capital increments
of the society) by choosing an optimal claims threshold. If the group has a sufficient size, then it can
Fig. 7. The expected one-step capital increments of an egoist (a), a group member (b), a randomly selected participant
(c), and the difference between the expected capital increments of a group member and an egoist (d ) as functions
of the majority threshold α for n = 100, µ = −0.1, σ = 1, t = 0, and various proportions of egoists in the society
(δ = 0.02 (1), δ = 0.1 (2), δ = 0.15 (3), δ = 0.3 (4), δ = 0.5 (5), δ = 0.7 (6), and δ = 0.95 (7)).
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
1096
MALYSHEV, CHEBOTAREV
Fig. 8. The influence of adopting the optimal claims threshold on the presence of the "pit of losses" for n = 100.
(a) α = 0.4; (b) α = 0.5; (c) α = 0.6. The medium gray area is the area of the presence of "pit of losses" for t = 0,
the dark gray area corresponds to the optimal claims threshold t = t0.
Fig. 9. Dependence of the maximum proportion of egoists for which the choice of group's optimal claims threshold
neutralizes the "pit of losses" on the majority threshold α. Gray points correspond to n = 100 (1), black points to
n = 10 (2).
completely remove the pit. Fig. 8 illustrates this possibility. In other words, a sufficiently large group can
help the society to overcome the paradoxical capital reduction caused by democratic decision-making. It
is important to note that a higher majority threshold allows a smaller group to do so. For example, if
µ/σ > −1, then for α = 0.4 the "pit of losses" does not appear for the societies with the proportion of
egoists up to 44 %; for α = 0.5, up to 56 %, and for α = 0.6, up to 83 %.
Fig. 9 shows the dependence on α of the maximum share of egoists for which the choice of the optimal
group's claims threshold t = t0 removes the "pit of losses." This relationship is basically expressed by
the function y = x (after the correction taking into account the equivalence of majority thresholds that
determine the same winning coalitions).
Thus, it is established that the choice of the optimal group's claims threshold can be considered as a
tool to neutralize the "pit of losses" paradox.
6. CONCLUSION
In this paper, we considered the dependence of the group's claims threshold that maximizes the capital
increment of the society on the parameters of the ViSE model for a society consisting of a group and
egoists. The absolute value of this threshold (called optimal) increases as the proportion of egoists
in the society grows.
If the percentage of egoists is fixed, then the optimal group's claims threshold
decreases with the grow of the majority threshold and becomes negative. Moreover, the optimal group's
claims threshold decreases as the mean µ of the proposals grows. The latter relationship is essentially
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
ON OPTIMAL GROUP CLAIMS AT VOTING
1097
a counterpart of one of the conclusions of [1], which recommends increasing the majority threshold in
deteriorating environments and reducing it in improving environments.
We have shown that the choice of the optimum group's claims threshold helps to neutralize the "pit of
losses" paradox in the societies consisting of a group and egoists. For a complete success, the proportion
of egoists should not exceed a certain value depending on the majority threshold. Thereby, an increase
of the majority threshold expands the zone of neutralization (by means of choosing the optimal group's
claims threshold) of the "pit of losses" paradox.
This work was supported by the Russian Science Foundation, project 16-11-00063 granted to IRE RAS.
ACKNOWLEDGMENTS
Proof of Proposition 1. Let G be the event that the group supports a proposal, G is the opposite
APPENDIX
event, P = P (G), and Q = P (G) = 1 − P.
When G occurs, it is necessary and sufficient for the proposal acceptance that egoists add more than
γn votes (case (a)).
If the group votes against the proposal, then more than αn votes of egoists are
required to accept the proposal (case (b)). Since the average of the independent capital increments of
the group members is distributed as N(cid:0)µ, σ2
P = F (µ − t)√g
g (cid:1), the probabilities of events G and G are given by
! and Q = 1 − P = F (t − µ)√g
! .
σ
σ
(A.1)
Formulas for the expected capital increments of an egoist in cases (a) and (b) can be obtained using
Lemma on "normal voting samples" [14] and (A.1), which leads to the expression (2):
M(edE ) = M(cid:0)edE G(cid:1)P (G) + M(cid:0)edE G(cid:1)P (G)
= µ+(µ, σ, ℓ, γn) F (µ − t)√g
σ
! + µ+(µ, σ, ℓ, αn) F (t − µ)√g
σ
! .
Proposition 1 is proved.
Proof of Proposition 2. Let Eξ be the event that the number of egoists' votes cast for a proposal is
higher than ξ. Let
Fξ = P (Eξ).
(A.2)
[ξ] + 0.5 − pℓ
√pqℓ
(cid:19),
Then (see [6])
Fξ =
ℓXx=[ξ]+1
σ(cid:1) , q = F(cid:0)− µ
σ(cid:1) , and b(x ℓ) =(cid:0)ℓ
b(x ℓ) ≈ F(cid:18)−
where p = F(cid:0) µ
x(cid:1)pxqℓ−x.
Using the notation MP(A) = M(cid:0)edG A(cid:1)P (A) we obtain
= MP(cid:0)G ∧ Eγn(cid:1) + MP(cid:0)G ∧ Eαn(cid:1),
(µ − t)√g
t =
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
σ
M(cid:0)edG(cid:1) = MP(cid:0)G∧ Eγn(cid:1)+ MP(cid:0)G∧ Eγn(cid:1) + MP(cid:0)G∧ Eαn(cid:1) + MP(cid:0)G∧ Eαn(cid:1)
since the second and fourth terms of the original expression are zero (as sunder the corresponding condi-
tions, the proposal is rejected and the participants' capital does not change). Introducing the notation
(A.3)
(A.4)
1098
MALYSHEV, CHEBOTAREV
and using (A.1), (A.2), independence of the proposal components, the fact that edG is distributed as
N(cid:0)µ, σ2
g (cid:1), and formula (A.9) in [14] expressing M(ζ ζ > t) for a normal ζ, we obtain:
MP(cid:0)G∧ Eγn(cid:1) = M(cid:0)edG G∧ Eγn(cid:1) P (G∧ Eγn) = M(cid:0)edG G(cid:1)P (G) P (Eγn)
√gF(cid:0)t(cid:1)! F(cid:0)t(cid:1) Fγn
σf(cid:0)t(cid:1)
(A.5)
= M(cid:0)edG edG > t(cid:1)F(cid:0)t(cid:1) Fγn = µ +
= µP +
√g!Fγn.
σf
Similarly,
MP(cid:0)G ∧ Eαn(cid:1) = µQ −
σf
√g!Fαn.
(A.6)
Substituting (A.5) and (A.6) into (A.3), we obtain (3). Proposition 2 is proved.
Proof of Theorem. The expected capital increment of a randomly selected participant is equal to
Using Propositions 1 and 2 and the notation (A.4) we now equate the derivative of the expression
M( d) = δM(edE ) + (1 − δ)M(edG ).
(A.7)
(A.7) with respect to t to zero:
δ(cid:2)µ+(µ, σ, ℓ, αn) − µ+(µ, σ, ℓ, γn)(cid:3) f (t)
√g
σ
+ (1 − δ)(Fγn − Fαn)f (t)
σ "−µ +
√g
t# = 0,
σ
√g
which implies
δ [µ+(µ, σ, ℓ, αn) − µ+(µ, σ, ℓ, γn)] − (1 − δ)(Fγn − Fαn) t = 0,
(A.8)
and consequently (4) takes place. Using (A.8), calculate the second derivative of (A.7) at t0 denoting by
t0 the result of substituting t0 into (A.4):
f′(t0)
√g
σ (cid:0)δ(cid:2)µ+(µ, σ, ℓ, αn) − µ+(µ, σ, ℓ, γn)] − (1 − δ)(Fγn − Fαn) t0(cid:3)(cid:1) + f (t0)
= f (t0)
√g
σ
(1 − δ)(Fαn − Fγn).
√g
σ
(1 − δ)(Fαn − Fγn)
This expression is negative. Consequently, t0 is a point of maximum. Theorem is proved.
Proof of Corollary. If the first condition is satisfied, then γ < 0, consequently, Fγn = 1 and
µ+(µ, σ, ℓ, γn) = µ. If the second condition is true, then Fαn = µ+(µ, σ, ℓ, αn) = 0. Substituting these
expressions into (4) we complete the proof of Corollary 3.
REFERENCES
1. Chebotarev, P.Yu., Malyshev, V.A., Tsodikova, Ya.Yu., Loginov, A.K., Lezina, Z.M., and Afonkin, V.A., The
Optimal Majority Threshold as a Function of the Coefficient of Variation of the Environment, Large-scale
Systems Control , 2016, vol. 62, pp. 169 -- 187.
2. Borzenko, V.I., Lezina, Z.M., Loginov, A.K., Tsodikova, Ya.Yu., and Chebotarev, P.Yu., Strategies of Voting in
Stochastic Environment: Egoism and Collectivism, Autom. Remote Control , 2006, vol. 67, no. 2, pp. 311 -- 328.
3. Mirkin, B.G., Group Choice, P.C. Fishburn, Ed., New York: Wiley, 1979.
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
ON OPTIMAL GROUP CLAIMS AT VOTING
1099
4. McKelvey, R.D., Intransitivities in Multidimensional Voting Models and Some Implications for Agenda Control,
J. Econom. Theory, 1976, vol. 12, no. 3, pp. 472 -- 482.
5. Coughlin, P. and Nitzan, S., Directional and Local Electoral Equilibria with Probabilistic Voting, J. Econom.
Theory, 1981, vol. 24, no. 2, pp. 226 -- 239.
6. Chebotarev, P.Yu., Loginov, A.K., Tsodikova, Ya.Yu., Lezina, Z.M., and Borzenko, V.I., Voting in a Stochastic
Environment: The Case of Two Groups, Autom. Remote Control , 2011, vol. 72, no. 7, pp. 1537 -- 1547.
7. Palfrey, T.R. and Rosenthal, H., Voter Participation and Strategic Uncertainty, Amer. Political Sci. Rev., 1985,
vol. 79, no. 1, pp. 62 -- 78.
8. Downs, A., An Economic Theory of Democracy, N.Y.: Harper and Brothers, 1957.
9. Grofman, B., The Neglected Role of the Status Quo in Models of Issue Voting, J. Politics, 1985, vol. 47, no. 1,
pp. 229 -- 237.
10. Duggan, J. and Kalandrakis, T., Dynamic Legislative Policy Making, J. Econom. Theory, 2012, vol. 147, no. 5,
pp. 1653 -- 1688.
11. Penn, E.M., A Model of Farsighted Voting, Amer. J. Political Sci., 2009, vol. 53, no. 1, pp. 36 -- 54.
12. Dziuda, W. and Loeper, A., Voting Rules in a Changing Environment , SSRN paper 2500777, 2014. DOI:
10.2139/ssrn.2500777.
13. Dziuda, W. and Loeper, A., Dynamic Collective Choice with Endogenous Status Quo, J. Polit. Economy, 2016,
vol. 124, no. 4, pp. 1148 -- 1186.
14. Chebotarev, P.Yu., Analytical Expression of the Expected Values of Capital at Voting in the Stochastic Envi-
ronment, Autom. Remote Control , 2006, vol. 67, no. 3, pp. 480 -- 492.
15. Chebotarev, P.Yu., Loginov, A.K., Tsodikova, Ya.Yu., Lezina, Z.M., and Borzenko, V.I., Snowball of Cooper-
ation and Snowball Communism, Proceedings of the Fourth Int. Conf. on Control Sciences, ISC RAS, 2009,
pp. 687 -- 699.
This paper was recommended for publication by A.L. Fradkov, a member of the Editorial Board
AUTOMATION AND REMOTE CONTROL Vol. 78 No. 6
2017
|
1602.08150 | 4 | 1602 | 2017-01-31T17:45:24 | Reachability-Based Safety and Goal Satisfaction of Unmanned Aerial Platoons on Air Highways | [
"cs.MA",
"eess.SY"
] | Recently, there has been immense interest in using unmanned aerial vehicles (UAVs) for civilian operations. As a result, unmanned aerial systems traffic management is needed to ensure the safety and goal satisfaction of potentially thousands of UAVs flying simultaneously. Currently, the analysis of large multi-agent systems cannot tractably provide these guarantees if the agents' set of maneuvers is unrestricted. In this paper, platoons of UAVs flying on air highways is proposed to impose an airspace structure that allows for tractable analysis. For the air highway placement problem, the fast marching method is used to produce a sequence of air highways that minimizes the cost of flying from an origin to any destination. The placement of air highways can be updated in real-time to accommodate sudden airspace changes. Within platoons traveling on air highways, each vehicle is modeled as a hybrid system. Using Hamilton-Jacobi reachability, safety and goal satisfaction are guaranteed for all mode transitions. For a single altitude range, the proposed approach guarantees safety for one safety breach per vehicle, in the unlikely event of multiple safety breaches, safety can be guaranteed over multiple altitude ranges. We demonstrate the platooning concept through simulations of three representative scenarios. | cs.MA | cs |
Reachability-Based Safety and Goal Satisfaction of Unmanned Aerial
Platoons on Air Highways
Mo Chen∗, Qie Hu∗, Jaime F. Fisac∗, Kene Akametalu∗, Casey Mackin†, Claire J. Tomlin‡
University of California, Berkeley
Recently, there has been immense interest in using unmanned aerial vehicles (UAVs) for civilian operations. As
a result, unmanned aerial systems traffic management is needed to ensure the safety and goal satisfaction of
potentially thousands of UAVs flying simultaneously. Currently, the analysis of large multi-agent systems cannot
tractably provide these guarantees if the agents' set of maneuvers is unrestricted. In this paper, platoons of UAVs
flying on air highways is proposed to impose an airspace structure that allows for tractable analysis. For the
air highway placement problem, the fast marching method is used to produce a sequence of air highways that
minimizes the cost of flying from an origin to any destination. The placement of air highways can be updated in
real-time to accommodate sudden airspace changes. Within platoons traveling on air highways, each vehicle is
modeled as a hybrid system. Using Hamilton-Jacobi reachability, safety and goal satisfaction are guaranteed for
all mode transitions. For a single altitude range, the proposed approach guarantees safety for one safety breach
per vehicle; in the unlikely event of multiple safety breaches, safety can be guaranteed over multiple altitude
ranges. We demonstrate the platooning concept through simulations of three representative scenarios.
Nomenclature
c
P
C
V
H
d
S
W
x
p = (px, py)
v = (vx, vy)
¯p
¯v
dsep
tfaulty
Qi
Qi
= Cost map
= A path between two points
= Cumulative cost of a path
= Value function of partial differential equations
= Air highway
= Direction of travel of air highway
= A sequence of air highways
= Waypoint
=
= Horizontal position
= Horizontal velocity
=
=
=
=
=
=
Target position
Target velocity
Separation distance of vehicles within a platoon
Time limit for descent during potential conflict
ith vehicle
Set of vehicles for vehicle Qi to consider for safety
System state (of a vehicle)
I.
Introduction
Unmanned aerial vehicles (UAVs) have in the past been mainly
used for military operations [1, 2]; however, recently there has been
an immense surge of interest in using UAVs for civil applications.
Through projects such as Amazon Prime Air [3] and Google Project
Wing [4], companies are looking to send UAVs into the airspace to not
only deliver commercial packages, but also for important tasks such
as aerial surveillance, emergency supply delivery, videography, and
search and rescue [5]. In the future, the use of UAVs is likely to become
more and more prevalent.
As a rough estimate, suppose in a city of 2 million people, each
person requests a drone delivery every 2 months on average and each
delivery requires a 30-minute trip for a UAV. This would equate to thou-
sands of UAVs simultaneously in the air just from package delivery ser-
vices. Applications of UAVs extend beyond package delivery; they can
also be used, for example, to provide supplies or to respond to disasters
in areas that are difficult to reach but require prompt response [6, 7].
As a result, government agencies such as the Federal Aviation Admin-
istration (FAA) and National Aeronautics and Space Administration
(NASA) are also investigating unmanned aerial systems (UAS) traffic
∗PhD Candidate, Department of Electrical Engineering and Computer Sci-
†PhD Student, Department of Electrical Engineering and Computer Sciences
‡Professor, Department of Electrical Engineering and Computer Sciences,
ences
Member AIAA
management (UTM) in order to prevent collisions among potentially
numerous UAVs [5, 8, 9].
Optimal control and game theory present powerful tools for pro-
viding safety and goal satisfaction guarantees to controlled dynamical
systems under bounded disturbances, and various formulations [10–12]
have been successfully used to analyze problems involving small num-
bers of vehicles [13–16]. These formulations are based on Hamilton-
Jacobi (HJ) reachability, which can compute the backward reachable
set (BRS), defined as the set of states from which a system is guaran-
teed to have a control strategy to reach a target set of states. HJ reach-
ability is a powerful tool because BRS can be used for synthesizing
both controllers that steer the system away from a set of unsafe states
("safety controllers") to guarantee safety, and controllers that steer the
system into a set of goal states ("goal satisfaction controllers") to guar-
antee goal satisfaction. Unlike many formulations of reachability, the
HJ formulations are flexible in terms of system dynamics, enabling
the analysis of controlled nonlinear systems under disturbances. Fur-
thermore, HJ reachability analysis is complemented by many numeri-
cal tools readily available to solve the associated HJ partial differential
equation (PDE) [17–19]. However, the computation is done on a grid,
making the problem complexity scale exponentially with the number
of states, and therefore with the number of vehicles. Consequently, HJ
reachability computations are intractable for large numbers of vehicles.
In order to accommodate potentially thousands of vehicles simul-
taneously flying in the air, additional structure is needed to allow for
tractable analysis and intuitive monitoring by human beings. An air
highway system on which platoons of vehicles travel accomplishes
both goals. However, many details of such a concept need to be ad-
dressed. Due to the flexibility of placing air highways compared to
building ground highways in terms of highway location, even the prob-
lem of air highway placement can be a daunting task. To address this,
in the first part of this paper, we propose a flexible and computationally
efficient method based on [19] to perform optimal air highway place-
ment given an arbitrary cost map that captures the desirability of having
UAVs fly over any geographical location. We demonstrate our method
using the San Francisco Bay Area as an example. Once air highways
are in place, platoons of UAVs can then fly in fixed formations along
the highway to get from origin to destination. The air highway struc-
ture greatly simplifies safety analysis, while at the same time allows
intuitive human participation in unmanned airspace management.
A considerable body of work has been done on the platooning of
ground vehicles [20]. For example, [21] investigated the feasibility of
vehicle platooning in terms of tracking errors in the presence of dis-
turbances, taking into account complex nonlinear dynamics of each
vehicle. [22] explored several control techniques for performing var-
ious platoon maneuvers such as lane changes, merge procedures, and
1
split procedures. In [23], the authors modeled vehicles in platoons as
hybrid systems, synthesized safety controllers, and analyzed through-
put. Reachability analysis was used in [24] to analyze a platoon of two
trucks in order to minimize drag by minimizing the following distance
while maintaining collision avoidance safety guarantees. Finally, [25]
provided a method for guaranteeing string stability and eliminating
accordion effects for a heterogeneous platoon of vehicles with linear
time-invariant dynamics.
Previous analyses of a large number of vehicles typically do not
provide safety and goal satisfaction guarantees to the extent that HJ
reachability does; however, HJ reachability typically cannot be used
to tractably analyze a large number of vehicles. In the second part of
this paper, we propose organizing UAVs into platoons, which provides
a structure that allows pairwise safety guarantees from HJ reachability
to better translate to safety guarantees for the whole platoon. With re-
spect to platooning, we first propose a hybrid systems model of UAVs
in platoons to establish the modes of operation needed for our platoon-
ing concept. Then, we show how reachability-based controllers can be
synthesized to enable UAVs to successfully perform mode switching,
as well as prevent dangerous configurations such as collisions. Finally,
we show several simulations to illustrate the behavior of UAVs in vari-
ous scenarios.
Overall, this paper is not meant to provide an exhaustive solution
to the unmanned airspace management problem. Instead, this paper
illustrates that the computation intractability of HJ reachability can be
overcome using an air highway structure with UAVs flying in platoons.
In addition, the results are intuitive, which can facilitate human partic-
ipation in managing the airspace. Although many challenges not ad-
dressed in this paper still need to be overcome, this paper can provide
a starting point for future research in large-scale UASs with safety and
goal satisfaction guarantees.
II. Air Highways
We consider air highways to be virtual highways in the airspace
on which a number of UAV platoons may be present. UAVs seek to
arrive at some desired destination starting from their origin by traveling
along a sequence of air highways. Air highways are intended to be the
common pathways for many UAV platoons, whose members may have
different origins and destinations. By routing platoons of UAVs onto a
few common pathways, the airspace becomes more tractable to analyze
and intuitive to monitor. The concept of platoons will be proposed in
Section III; in this section, we focus on air highways.
Let an air highway be denoted by the continuous function H :
[0, 1] → R2. Such a highway lies in a horizontal plane of fixed al-
titude, with start and end points given by H(0) ∈ R2 and H(1) ∈ R2
respectively. For simplicity, we assume that the highway segment is
a straight line segment, and the parameter s indicates the position in
some fixed altitude as follows: H(s) = H(0) + s(H(1) − H(0)). To
each highway, we assign a speed of travel vH and specify the direction
of travel to be the direction from H(0) to H(1), denoted using a unit
vector d =
. As we will show in Section III, UAVs use
simple controllers to track the highway.
H(1)−H(0)
(cid:107)H(1)−H(0)(cid:107)2
Air highways must not only provide structure to make the analy-
sis of a large number of vehicles tractable, but also allow vehicles to
reach their destinations while minimizing any relevant costs to the ve-
hicles and to the surrounding regions. Such costs can for example take
into account people, assets on the ground, and manned aviation, enti-
ties to which UAVs pose the biggest risks [5]. Thus, given an origin-
destination pair (eg.
two cities), air highways must connect the two
points while potentially satisfying other criteria. In addition, optimal
air highway locations should ideally be able to be recomputed in real-
time when necessary in order to update airspace constraints on-the-fly,
in case, for example, airport configurations change or certain airspaces
have to be closed [5]. With this in mind, we now define the air highway
placement problem, and propose a simple and fast way to approximate
its solution that allows for real-time recomputation. Our solution based
on solving the Eikonal equation can be thought of as converting a cost
map over a geographic area in continuous space into a discrete graph
whose nodes are waypoints joined by edges which are the air highways.
Note that the primary purpose of this section is to provide a method
2
for the real-time placement of air highways. The specifics of deter-
mining the cost map based on population density, geography, weather
forecast information, etc., as well as the criteria for when air highway
locations need to be updated, is beyond the scope of this paper.
In addition, if vehicles in the airspace are far away from each other,
it may be reasonable for all vehicles to fly in an unstructured man-
ner. As long as multiple-way conflicts do not occur, pairwise collision
avoidance maneuvers would be sufficient to ensure safety. Unstruc-
tured flight is likely to result in more efficient trajectories for each indi-
vidual vehicle. However, whether multiple-way conflicts occur cannot
be predicted ahead of time, and are not guaranteed to be resolvable
when they occur. By organizing vehicles into platoons, the likelihood
of multiple-way conflicts is vastly reduced. Structured flight is in gen-
eral less efficient for the individual vehicle, and this loss of efficiency
can be thought of as a cost incurred by the vehicles in order ensure
higher levels of safety.
In general, there may be many different levels of abstractions in
the airspace. For larger regions such as cities, air highways may prove
beneficial, and for a small region such as a neighborhood, perhaps un-
structured flight is sufficiently safe. Further research is needed to better
understand parameters such as the density of vehicles above which un-
structured flight is no longer manageable, and other details like platoon
size.
A. The Air Highway Placement Problem
Consider a map c : R2 → R which defines the cost c(p) incurred when
a UAV flies over the position p = (px, py) ∈ R2. Given any posi-
tion p, a large value of c(p) indicates that the position p is costly or
undesirable for a UAV to fly over. Locations with high cost could, for
example, include densely populated areas and areas around airports. In
general, the cost map c(·) may be used to model cost of interference
with commercial airspaces, cost of accidents, cost of noise pollution,
risks incurred, etc., and can be flexibly specified by government regu-
lation bodies.
Let po denote an origin point and pd denote a destination point.
Consider a sequence of highways SN = {H1, H2, . . . , HN} that satis-
fies the following:
H1(0) = po
Hi(1) = Hi+1(0), i = 0, 1, . . . , N − 1
HN (1) = pd
(1)
The interpretation of the above conditions is that the start point
of first highway is the origin, the end point of a highway is the start
point of the next highway, and the end point of last highway is the
destination. The highways H1, . . . , HN form a sequence of waypoints
for a UAV starting at the origin po to reach its destination pd.
Given only the origin point po and destination point pd, there are
an infinite number of choices for a sequence of highways that satisfy
(1). However, if one takes into account the cost of flying over any po-
sition p using the cost map c(·), we arrive at the air highway placement
problem:
(cid:33)
(cid:41)
(cid:40)(cid:32) N(cid:88)
(cid:90) 1
minSN ,N
i=1
subject to (1)
0
c(Hi(s))ds
+ R(N )
(2)
where R(·) is a regularizer, such as R(N ) = N 2.
The interpretation of (2) is that we consider air highways to be line
segments of constant altitude over a region, and UAV platoons travel on
these air highways to get from some origin to some destination. Any
UAV flying on a highway over some position p incurs a cost of c(p), so
that the total cost of flying from the origin to the destination is given by
the summation in (2). The air highway placement problem minimizes
the cumulative cost of flying from some origin po to some destination
pd along the sequence of highways SN = {H1, H2, . . . , HN}. The
regularization term R(N ) is used to prevent N from being arbitrarily
large.
3
from an origin po to any destination pd. Once V is found, the optimal
path P between po and pd can be obtained via gradient descent.
c(p)∇V (p) = 1
V (po) = 0
(5)
The Eikonal equation (5) can be efficiently computed numerically
using the fast marching method [19]; each computation takes on the
order of merely a second.
Note that (4) can be viewed as a relaxation of the air highway
placement problem defined in (2). Unlike (2), the relaxation (4) can be
quickly solved using currently available numerical tools. Thus, we first
solve the approximate air highway placement problem (4) by solving
(5), and then post-process the solution to (4) to obtain an approximation
to (2).
Given a single origin point po, the optimal cumulative cost func-
tion V (pd) can be computed. Suppose M different destination points
i , i = 1, . . . , M are chosen. Then, M different optimal paths
pd
Pi, i = 1, . . . , M are obtained from V .
C. From Paths to Waypoints
Each of the cost-minimizing paths Pi computed from the solution to
the Eikonal equation consists of a closely-spaced set of points. Each
path Pi is an approximation to the sequence of highways Si
Ni =
j}i=M,j=Ni
{Hi
defined in (2), but now indexed by the corresponding
i=1,j=1
path index i.
For each path Pi, we would like to sparsify the points on the path
to obtain a collection of waypoints, Wi,j, j = 1, . . . , Ni + 1, which
are the end points of the highways:
j(0) = Wi,j,
Hi
j(1) = Wi,j+1,
Hi
j = 1, . . . , Ni
(6)
There are many different ways to do this, and this process will not
be our focus. However, for illustrative purposes, we show how this
process may be started. We begin by noting the path's heading at the
destination point. We add to the collection of waypoints the first point
on the path at which the heading changes by some threshold θC, and
repeat this process along the entire path.
If there is a large change in heading within a small section of the
cost-minimizing path, then the collection of points may contain many
points which are close together.
In addition, there may be multiple
paths that are very close to each other (in fact, this behavior is desir-
able), which may contribute to cluttering the airspace with too many
waypoints. To reduce clutter, one could cluster the points. Afterwards,
each cluster of points can be replaced by a single point located at the
centroid of the cluster.
To the collection of points resulting from the above process, we
add the origin and destination points. Repeating the entire process for
every path, we obtain waypoints for all the cost-minimizing paths under
consideration. Figure 1 summarizes the entire air highway placement
process, including our example of how the closely-spaced set of points
on a path can be sparsified.
D. Results
To illustrate our air highway placement proposal, we used the San Fran-
cisco Bay Area as an example, and classified each point on the map
into four different regions: "regions around airports", "highly popu-
lated cities", "water", and "other". Each region has an associated cost,
reflecting the desirability of flying a vehicle over an area in the region.
In general, these costs can be arbitrary and determined by government
regulation agencies. For illustration purposes, we assumed the follow-
ing categories and costs:
• Region around airports: cairports = b,
Figure 1. Illustration of the air highway placement procedure.
B. The Eikonal Equation – Cost-Minimizing Path
Let s0, s1 ∈ R, and let P : [s0, s1] → R2 be a path starting from an
origin point po = P(s0) and ending at a destination point pd = P(s1).
Note that the sequence SN in (2) is a piecewise affine example of a path
P(s), s ∈ [s0, s1]; however, a path P that is not piecewise affine cannot
be written as a sequence of highways SN .
More concretely, suppose a UAV flies from an origin point po to
a destination point pd along some path P(s) parametrized by s. Then,
P(s0) = po would denote the origin, and P(s1) = pd would denote the
destination. All intermediate s values denote the intermediate positions
of the path, i.e. P(s) = p(s) = (px(s), py(s)).
Consider the cost map c(px, py) which captures the cost incurred
for UAVs flying over the position p = (px, py). Along the entire path
P(s), the cumulative cost C(P) is incurred. Define C as follows:
C(P) =
c(P(s))ds
(3)
(cid:90) s1
s0
For an origin-destination pair, we would like to find the path such
that the above cost is minimized. More generally, given an origin point
po, we would like to compute the function V representing the optimal
cumulative cost for any destination point pd:
C(P)
V (pd) =
min
P(·),P(s1)=pd
=
min
P(·),P(s1)=pd
(cid:90) s1
s0
c(P(s))ds
(4)
It is well known that the viscosity solution [26] to the Eikonal
equation (5) precisely computes the function V (pd) given the cost map
c [19, 27]. Note that a single function characterizes the minimum cost
• Cities: ccities = 1,
• Water: cwater = b−2,
Origin𝑝𝑜Destination𝑝𝑑Low cost regionHigh cost regionCost-minimizing path ℙOrigin𝑝𝑜Destination𝑝𝑑Low cost regionHigh cost regionLocations with large change in path headingOrigin𝑝𝑜Destination𝑝𝑑Low cost regionHigh cost regionSequence of highways ℍ1,ℍ2,…,ℍ5 • Other: cother = b−1.
This assumption assigns costs in descending order to the categories
"regions around airports", "cities", "other", and "water". Flying a UAV
in each category is more costly by a factor of b compared to the next
most important category. The factor b > 1 is a tuning parameter that
we adjusted to vary the relative importance of the different categories,
and we used b = 4 in the figures below.
Figure 2 shows the San Francisco Bay Area geographic map, cost
map, cost-minimizing paths, and contours of the value function V . The
region enclosed by the black boundary represents "region around air-
ports", which have the highest cost. The dark blue, yellow, and light
blue regions represent the "cities", the "water", and the "other"' cate-
gories, respectively. We assumed that the origin corresponds to the city
"Concord", and chose a number of other major cities as destinations.
A couple of important observations can be made here. First, the
cost-minimizing paths to the various destinations in general overlap,
and only split up when they are very close to entering their destina-
tion cities. This is intuitively desirable because having overlapping
cost-minimizing paths keeps the number of distinct air highways low.
Secondly, the contours, which correspond to level curves of the value
function, have a spacings corresponding to the cost map: the spacings
are large in areas of low cost, and vice versa. This provides insight into
the placement of air highways to destinations that were not shown in
this example.
Figure 3 shows the result of converting the cost-minimizing paths
to a small number of waypoints. The left plot shows the waypoints,
interpreted as the start and end points of air highways, over a white
background for clarity. The right plot shows these air highways over
the map of the Bay Area. Note that we could have gone further to
merge some of the overlapping highways. However, the purpose of
this section is to illustrate the natural occurrence of air highways from
cost-minimizing paths; post-processing of the cost-minimizing paths,
which serve as a guide for defining air highways, is not our focus.
E. Real-Time Highway Location Updates
Since (5) can be solved in approximately 1 second, the air highway
placement process can be redone in real-time if the cost map changes
at a particular time. This flexibility can be useful in any situation in
which unforeseen circumstances could cause a change in the cost of
a particular region of the airspace. For example, accidents or disaster
response vehicles may result in an area temporarily having a high cost.
On the other hand, depending on for instance the time of day, it may
be most desirable to fly in different regions of the airspace, resulting in
those regions temporarily having a low cost.
III. Unmanned Aerial Vehicle Platooning
Air highways exhibiting trunk routes that separate near destina-
tions motivate the use of platoons which fly on these highways. The
air highway structure along with the UAV platooning concept together
enable the use of reachability to analyze safety and goal satisfaction
properties. The structure reduces the likelihood of multiple-way con-
flicts, and makes pairwise analysis more indicative of the joint safety
of all UAVs. In addition to reducing complexity, the proposed struc-
ture is intuitive, and allows human participation in the monitoring and
management of the unmanned airspace.
Organizing UAVs into platoons implies that the UAVs cannot fly
in an unstructured way, and must have a restricted set of controllers
or maneuvers depending on the UAV's role in the airspace. To model
UAVs flying in platoons on air highways, we propose a hybrid sys-
tem whose modes of operations describe a UAV's role in the highway
structure. For the hybrid system model, reachability analysis is used to
enable successful and safe operation and mode transitions.
A. UAVs in Platoons
1. Vehicle Dynamics
Consider a UAV whose dynamics are given by
x = f (x, u)
(7)
4
where x represents the state, and u represents the control action. The
techniques we present in this paper do not depend on the dynamics of
the vehicles, as long as their dynamics are known. However, for con-
creteness, we assume that the UAVs are quadrotors that fly at a constant
altitude under non-faulty circumstances. For the quadrotor, we use a
simple model in which the x and y dynamics are double integrators:
px = vx
py = vy
vx = ux
vy = uy
ux,uy ≤ umax
(8)
where the state x = (px, vx, py, vy) ∈ R4 represents the quadrotor's
position in the x-direction, its velocity in the x-direction, and its po-
sition and velocity in the y-direction, respectively. The control input
u = (ux, uy) ∈ R2 consists of the acceleration in the x- and y- di-
rections. For convenience, we will denote the position and velocity
p = (px, py), v = (vx, vy), respectively.
In general, the problem of collision avoidance among N vehi-
cles cannot be tractably solved using traditional dynamic programming
approaches because the computation complexity of these approaches
scales exponentially with the number of vehicles. Thus, in our present
work, we will consider the situation where UAVs travel on air highways
in platoons, defined in the following sections. The structure imposed
by air highways and platooning enables us to analyze the safety and
goal satisfaction properties of the vehicles in a tractable manner.
2. Vehicles as Hybrid Systems
We model each vehicle as a hybrid system [23, 28] consisting of the
modes "Free", "Leader", "Follower", and "Faulty". Within each mode,
a vehicle has a set of restricted maneuvers, including one that allows
the vehicle to change modes if desired. The modes and maneuvers are
as follows:
• Free:
A Free vehicle is not in a platoon or on a highway, and its pos-
sible maneuvers or mode transitions are
– remain a Free vehicle by staying away from highways
– become a Leader by entering a highway to create a new
platoon
– become a Follower by joining a platoon that is currently
on a highway
• Leader:
A Leader vehicle is the vehicle at the front of a platoon (which
could consist of only the vehicle itself). The available maneu-
vers and mode transitions are
– remain a Leader by traveling along the highway at a pre-
specified speed vH
– become a Follower by merging the current platoon with a
platoon in front
– become a Free vehicle by leaving the highway
• Follower:
A Follower vehicle is a vehicle that is following a platoon leader.
The available maneuvers and mode transitions are
– remain a Follower by staying a distance of dsep behind the
vehicle in front in the current platoon
– remain a Follower by joining a different platoon on an-
other highway
– become a Leader by splitting from the current platoon
while remaining on the highway
– become a Free vehicle by leaving the highway
5
• Faulty:
If a vehicle from any of the other modes becomes unable to op-
erate within the allowed set of maneuvers, it transitions into the
Faulty mode. Reasons for transitioning to the Faulty mode in-
clude vehicle malfunctions, performing collision avoidance with
respect to another Faulty vehicle, etc. A Faulty vehicle is as-
sumed to descend via a fail-safe mechanism after some pre-
specified duration tfaulty to a different altitude level where it no
longer poses a threat to vehicles on the air highway system.
Such a fail-safe mechanism could be an emergency landing pro-
cedure such as those analyzed in [29–31]. Typically, emergency
landing involves identifying the type of fault and finding feasi-
ble landing locations given the dynamics during the fault. We
will omit these details and summarize them into tfaulty, the time
required to exit the current altitude level.
The available maneuvers and associated mode transitions are sum-
marized in Figure 4.
Figure 5. Notation for vehicles in platoons.
• join a new platoon
• leave a platoon to create a new one
• react to malfunctioning or intruder vehicles
We also propose more basic controllers to perform other simpler
actions such as
• follow the highway at constant altitude at a specified speed
• maintain a constant relative position and velocity with respect
to the leader of a platoon
In general, the control strategy of each vehicle has a safety com-
ponent, which specifies a set of states that it must avoid, and a goal
satisfaction component, which specifies a set of states that the vehi-
cle aims to reach. Together, the safety and goal satisfaction controllers
guarantee the safety and success of a vehicle in the airspace making any
desired mode transition. In this paper, these guarantees are provided us-
ing reachability analysis, and allow the multi-UAV system to perform
joint maneuvers essential to maintaining structure in the airspace.
B. Hamilton-Jacobi Reachability
1. General Framework
Consider a differential game between two players described by the sys-
tem
x = f (x, u1, u2), for almost every t ∈ [−T, 0]
(9)
where x ∈ Rn is the system state, u1 ∈ U1 is the control of Player 1,
and u2 ∈ U2 is the control of Player 2. We assume f : Rn×U1×U2 →
Rn is uniformly continuous, bounded, and Lipschitz continuous in x
for fixed u1, u2, and the control functions u1(·) ∈ U1, u2(·) ∈ U2 are
drawn from the set of measurable functionsa. Player 2 is allowed to use
nonanticipative strategies [32, 33] γ, defined by
γ ∈ Γ := {N : U1 → U2 u1(r) = u1(r)
for almost every r ∈ [t, s] ⇒ N [u1](r)
= N [u1](r) for almost every r ∈ [t, s]}
(10)
In our differential game, the goal of Player 2 is to drive the system
into some target set L, and the goal of Player 1 is to drive the system
away from it. The set L is represented as the zero sublevel set of a
bounded, Lipschitz continuous function l : Rn → R. We call l(·) the
implicit surface function representing the set L : L = {x ∈ Rn
l(x) ≤ 0}.
Given the dynamics (8) and the target set L, we would like to com-
pute the BRS, V(t):
a A function f : X → Y between two measurable spaces (X, ΣX ) and
(Y, ΣY ) is said to be measurable if the preimage of a measurable set in Y is
a measurable set in X, that is: ∀V ∈ ΣY , f−1(V ) ∈ ΣX, with ΣX , ΣY
σ-algebras on X,Y .
Figure 4. Hybrid modes for vehicles in platoons. Vehicles begin in the
"Free" mode before they enter the highway.
j=1 ⊆ {i}N
Suppose that there are N vehicles in total in the airspace contain-
ing the highway system. We will denote the N vehicles as Qi, i =
1 . . . , N. We consider a platoon of vehicles to be a group of M ve-
hicles (M ≤ N), denoted QP1 , . . . , QPM ,{Pj}M
i=1, in a
single-file formation. When necessary, we will use superscripts to de-
represents the ith vehicle in
note vehicles of different platoons: QP j
the jth platoon.
For convenience, let Qi denote the set of indices of vehicles with
If vehicle Qi is a free vehi-
respect to which Qi checks safety.
cle, then it must check for safety with respect to all other vehicles,
Qi = {j : j (cid:54)= i}.
If the vehicle is part of a platoon, then it
checks safety with respect to the platoon member in front and behind,
Qi = {Pj+1, Pj−1}. Figure 5 summarizes the indexing system of the
vehicles.
i
We will organize the vehicles into platoons travel along air high-
ways. The vehicles maintain a separation distance of dsep with their
neighbors inside the platoon.
In order to allow for close proximity
of the vehicles and the ability to resolve multiple simultaneous safety
breaches, we assume that when a vehicle exhibits unpredictable behav-
ior, it will be able to exit the altitude range of the highway within a
duration of tfaulty. Such a requirement may be implemented practically
as an fail-safe mechanism to which the vehicles revert when needed.
3. Objectives
Given the above modeling assumptions, our goal is to provide control
strategies to guarantee the success and safety of all the mode transi-
tions. The theoretical tool used to provide the safety and goal satisfac-
tion guarantees is reachability. The BRSs we compute will allow each
vehicle to perform complex actions such as
• merge onto a highway to form a platoon
Free:•Vehicle not in a platoon or on a highwayLeader•Leader of platoonFollower•Member of platoonFaulty•Descends after a duration of 𝑡faultyLeave highwayMerge onto highwayCreate new platoonMerge with platoon in frontJoin platoonFollow highwayFollow platoonLeave highway𝑄"𝑄#𝑄$𝑄%𝑄&𝑄'𝑄(Platoon 1𝑄""𝑄)𝑄*𝑄"+Platoon 2𝑄"#𝑃""=4𝑃#"=5𝑃$"=6𝑃%"=7𝑃"#=8𝑃##=9𝑃$#=10𝒬)=9𝒬$=1,2,4,5,…,12𝒬"#=1,2,3,…,11𝒬&=4,6Free vehiclesLeader vehiclesFollower vehicles V(t) := {x ∈ Rn ∃γ ∈ Γ such that ∀u1(·) ∈ U1,
∃s ∈ [t, 0], ξf (s; t, x, u1(·), γ[u1](·)) ∈ L}
(11)
where ξf is the trajectory of the system satisfying initial conditions
ξf (t; x, t, u1(·), u2(·)) = x and the following differential equation al-
most everywhere on [−t, 0]
d
ds
ξf (s; x, t, u1(·), u2(·))
= f (ξf (s; x, t, u1(·), u2(·)), u1(s), u2(s))
(12)
Many methods involving solving HJ PDEs [11] and HJ variational
inequalities (VI) [10, 12, 13] have been developed for computing the
BRS. These HJ PDEs and HJ VIs can be solved using well-established
numerical methods. For this paper, we use the formulation in [11],
which shows that the BRS V(t) can be obtained as the zero sublevel
set of the viscosity solution [26] V (t, x) of the following terminal value
HJ PDE:
∗
u
(t, x) = arg min
u∈U DxV (t, x) · f (x, u)
6
(17)
For our application, we will use several decoupled system models
and utilize decomposition techniques [34–36], which enables real-time
4D BRS computations and tractable 6D BRS computations.
2. Relative Dynamics and Augmented Relative Dynamics
Besides Equation (8), we will also consider the relative dynamics be-
tween two quadrotors Qi, Qj. These dynamics can be obtained by
defining the relative variables
px,r = px,i − px,j
py,r = py,i − py,j
vx,r = vx,i − vx,j
vy,r = vy,i − vy,j
(18)
DtV (t, x)+
min{0, max
u1∈U1
V (0, x) = l(x)
min
u2∈U2
DxV (t, x) · f (x, u1, u2)} = 0,
(13)
We treat Qi as Player 1, the evader who wishes to avoid collision,
and we treat Qj as Player 2, the pursuer, or disturbance, that wishes to
cause a collision. In terms of the relative variables given in (18), we
have
px,r = vx,r
py,r = vy,r
vx,r = ux,i − ux,j
vy,r = uy,i − uy,j
(19)
We also augment (18) with the velocity of Qi, given in (20), to
impose a velocity limit on the quadrotor.
px,r = vx,r
py,r = vy,r
vx,r = ux,i − ux,j
vy,r = uy,i − uy,j
vx,i = ux,i
vy,i = uy,i
(20)
C. Reachability-Based Controllers
Reachability analysis is useful for constructing controllers in a large
variety of situations. In order to construct different controllers, an ap-
propriate target set needs to be defined depending on the goal of the
controller. If one defines the target set to be a set of desired states, the
BRS would represent the states that a system needs to first arrive at in
order to reach the desired states. On the other hand, if the target set
represents a set of undesirable states, then the BRS would indicate the
region of the state space that the system needs to avoid. In addition, the
system dynamics with which the BRS is computed provide additional
flexibility when using reachability to construct controllers.
Using a number of different target sets and dynamics, we now
propose different reachability-based controllers used for vehicle mode
transitions in our platooning concept.
Figure 6. Illustration of a target set, a BRS, and their implicit surface func-
tions.
from which we obtain V(t) = {x ∈ Rn V (t, x) ≤ 0}. The target
set, reachable set, and implicit surface functions l(x) and V (t, x) rep-
resenting them are shown in Fig. 6. From the solution V (t, x), we can
also obtain the optimal controls for both players via the following:
u
u
∗
1(t, x) = arg max
u1∈U1
∗
2(t, x) = arg min
u2∈U2
DxV (t, x) · f (x, u1, u2)
min
u2∈U2
DxV (t, x) · f (x, u
∗
1, u2)
In the special case where there is only one player, we obtain an
(14)
1. Getting to a Target State
optimal control problem for a system with dynamics
x = f (x, u), t ∈ [−T, 0], u ∈ U.
The BRS in this case would be given by the HJ PDE
DtV (t, x) + min{0, min
u∈U DxV (t, x) · f (x, u)} = 0
V (0, x) = l(x)
where the optimal control is given by
(15)
(16)
The controller used by a vehicle to reach a target state is important in
two situations in the platooning context. First, a vehicle in the "Free"
mode can use the controller to merge onto a highway, forming a platoon
and changing modes to a "Leader" vehicle. Second, a vehicle in either
the "Leader" mode or the "Follower" mode can use this controller to
change to a different highway, becoming a "Leader" vehicle.
In both of the above cases, we use the dynamics of a single ve-
hicle specified in (8). The target state would be a position (¯px, ¯py)
representing the desired merging point on the highway, along with a
velocity (¯vx, ¯vy) = vH that corresponds to the speed and direction of
travel specified by the highway. For the reachability computation, we
define the target set to be a small range of states around the target state
¯xH = (¯px, ¯py, ¯vx, ¯vy):
7
LH = {x : px − ¯px ≤ rpx ,vx − ¯vx ≤ rvx ,
py − ¯py ≤ rpy ,vy − ¯vy ≤ rvy}.
(21)
Here, we represent the target set LH as the zero sublevel set of
the function lH (x), which specifies the terminal condition of the HJ
PDE that we need to solve. Once the HJ PDE is solved, we obtain the
BRS VH (t) from the subzero level set of the solution VH (t, x). More
concretely, VH (T ) = {x : VH (−T, x) ≤ 0} is the set of states from
which the system can be driven to the target LH within a duration of
T .
Depending on the time horizon T , the size of the BRS VH (T )
varies. In general, a vehicle may not initially be inside the BRS VH (T ),
yet it needs to be in order to get to its desired target state. Determining
a control strategy to reach VH (T ) is itself a reachability problem (with
VH (T ) as the target set), and it would seem like this reachability prob-
lem needs to be solved in order for us to use the results from our first
reachability problem. However, practically, one could choose T to be
large enough to cover a sufficiently large area to include any practically
conceivable initial state. From our simulations, a suitable algorithm for
getting to a desired target state is as follows:
1. Move towards ¯xH in pure pursuit with some velocity, until
VH (−T, x) ≤ 0. In practice, this step consistently drives the
system into the BRS.
2. Apply the optimal control extracted from VH (−T, x) according
to (17) until LH is reached.
2. Getting to a State Relative to Another Vehicle
In the platooning context, being able to go to a state relative to another
moving vehicle is important for the purpose of forming and joining
platoons. For example, a "Free" vehicle may join an existing platoon
that is on a highway and change modes to become a "Follower". Also,
a "Leader" or "Follower" may join another platoon and afterwards go
into the "Follower" mode.
To construct a controller for getting to a state relative to another
vehicle, we use the relative dynamics of two vehicles, given in (19).
In general, the target state is specified to be some position (¯px,r, ¯py,r)
and velocity (¯vx,r, ¯vy,r) relative to a reference vehicle. In the case of
a vehicle joining a platoon that maintains a single file, the reference
vehicle would be the platoon leader, the desired relative position would
be a certain distance behind the leader, depending on how many other
vehicles are already in the platoon; the desired relative velocity would
be (0, 0) so that the formation can be kept.
For the reachability problem, we define the target set to be a small
range of states around the target state ¯xP = (¯px,r, ¯py,r, ¯vx,r, ¯vy,r):
LP = {x : px,r − ¯px,r ≤ rpx ,vx,r − ¯vx,r ≤ rvx ,
py,r − ¯py,r ≤ rpy ,vy,r − ¯vy,r ≤ rvy}
(22)
The target set LP is represented by the zero sublevel set of the
implicit surface function lP (x), which specifies the terminal condi-
tion of the HJ PDE (13). The zero sublevel set of the solution to (13),
VP (−T, x), gives us the set of relative states from which a quadrotor
can reach the target in the relative coordinates within a duration of T .
In the BRS computation, we assume that the reference vehicle moves
along the highway at constant speed, so that uj(t) = 0. The follow-
ing is a suitable algorithm for a vehicle joining a platoon to follow the
platoon leader:
1. Move towards ¯xP in a straight line, with some velocity, until
VP (−T, x) ≤ 0.
2. Apply the optimal control extracted from VP (−T, x) according
to (14) until LP is reached.
3. Avoiding Collisions
A vehicle can use a goal satisfaction controller described in the previ-
ous sections when it is not in any danger of collision with other vehi-
cles. If the vehicle could potentially be involved in a collision within
the next short period of time, it must switch to a safety controller. The
safety controller is available in every mode, and executing the safety
controller to perform an avoidance maneuver does not change a vehi-
cle's mode.
In the context of our platooning concept, we define an unsafe con-
figuration as follows: a vehicle is either within a minimum separation
distance d to a reference vehicle in both the x and y directions, or is
traveling with a speed above the speed limit vmax in either of the x and
y directions. To take this specification into account, we use the aug-
mented relative dynamics given by (20) for the reachability problem,
and define the target set as follows:
LS = {x :px,r,py,r ≤ d ∨ vx,i ≥ vmax ∨ vy,i ≥ vmax}
(23)
We can now define the implicit surface function lS(x) correspond-
ing to LS, and solve the HJ PDE (13) using lS(x) as the terminal condi-
tion. As before, the zero sublevel set of the solution VS(t, x) specifies
the BRS VS(t), which characterizes the states in the augmented rela-
tive coordinates, as defined in (20), from which Qi cannot avoid LS
for a time period of t, if Qj uses the worst-case control. To avoid col-
lisions, Qi must apply the safety controller according to (14) on the
boundary of the BRS in order to avoid going into the BRS. The fol-
lowing algorithm wraps our safety controller around goal satisfaction
controllers:
1. For a specified time horizon t, evaluate VS(−t, xi − xj) for all
j ∈ Q(i).
Q(i) is the set of quadrotors with which vehicle Qi checks
safety.
2. Use the safety or goal satisfaction controller depending on the
values VS(−t, xi − xj), j ∈ Q(i):
If ∃j ∈ Q(i), VS(−t, xi−xj) ≤ 0, then Qi, Qj are in potential
conflict, and Qi must use a safety controller; otherwise Qi may
use a goal satisfaction controller.
D. Other Controllers
Reachability was used in Section C for the relatively complex maneu-
vers that require safety and goal satisfaction guarantees. For the sim-
pler maneuvers of traveling along a highway and following a platoon,
many well-known classical controllers suffice. For illustration, we use
the simple controllers described below.
1. Traveling along a highway
We use a model-predictive controller (MPC) for traveling along a high-
way; this controller allows the leader to travel along a highway at a
pre-specified speed. Here, the goal is for a leader vehicle to track an
air highway H(s), s ∈ [0, 1] while maintaining some constant velocity
vH specified by the highway. The highway and the specified velocity
can be written as a desired position and velocity over time, ¯p(t), ¯v(t).
Assuming that the initial position on the highway, s0 = s(t0) is speci-
fied, such a controller can be obtained from the following optimization
problem over the time horizon [t0, t1]:
minimize
(cid:90) t1
t0
(cid:8)(cid:107)p(t) − ¯p(t)(cid:107)2+
(cid:107)v(t) − ¯v(t)(cid:107)2 + 1 − s(cid:9)dt
subject to vehicle dynamics (7)
ux,uy ≤ umax,vx,vy ≤ vmax
s(t0) = s0, s ≥ 0
(24)
If we discretize time, the above optimization is becomes convex
optimization over a small number of decision variables, and can be
quickly solved.
2. Following a Platoon
Follower vehicles use a feedback control law tracking a nominal po-
sition and velocity in the platoon, with an additional feedforward term
given by the leader's acceleration input; here, for simplicity, we assume
perfect communication between the leader and the follower vehicles.
This following law enables smooth vehicle trajectories in the relative
platoon frame, while allowing the platoon as a whole to perform agile
maneuvers by transmitting the leader's acceleration command uP1 (t)
to all vehicles.
The i-th member of the platoon, QPi, is expected to track a rela-
tive position in the platoon ri = (ri
y) with respect to the leader's
position pP1, and the leader's velocity vP1 at all times. The resulting
control law has the form:
ui(t) = kp
(cid:2)vP1 (t)−vi(t)(cid:3)+uP1 (t) (25)
(cid:2)pP1 (t)+ri(t)−pi(t)(cid:3)+kv
x, ri
for some kp, kv > 0. In particular, a simple rule for determining ri(t)
in a single-file platoon is given for QPi as:
ri(t) = −(i − 1)dsep d
(26)
where dsep is the spacing between vehicles along the platoon and d is
the highway's direction of travel.
E. Summary of Controllers
We have introduced several reachability-based controllers, as well as
some simple controllers. Pairwise collision avoidance is guaranteed
using the safety controller, described in Section III-C-3. As long as a
vehicle is not in potential danger according to the safety BRSs, it is
free to use any other controller. All of these other controllers are goal
satisfaction controllers, and their corresponding mode transitions are
shown in Figure 7.
The controller for getting to an absolute target state, described in
Section III-C-1, is used whenever a vehicle needs to move onto a high-
way to become a platoon leader. This controller guarantees the success
of the mode transitions shown in blue in Figure 7.
The controller for getting to a relative target state, described in
Section III-C-2, is used whenever a vehicle needs to join a platoon to
become a follower. This controller guarantees the success of the mode
transitions shown in green in Figure 7.
For the simple maneuvers of traveling along a highway or follow-
ing a platoon, many simple controllers such as the ones suggested in
Section III-D can be used. These controllers keep the vehicles in either
the Leader or the Follower mode. Alternatively, additional controllers
can be designed for exiting the highway, although these are not consid-
ered in this paper. All of these non-reachability-based controllers are
shown in gray in Figure 7.
8
QP2, and the platoon trailer QPN checks safety against QPN−1. So
QP1 = {P2},QPN = {PN−1}. When all vehicles are using goal sat-
isfaction controllers to perform their allowed maneuvers, in most situa-
tions no pair of vehicles should be in an unsafe configuration. However,
occasionally a vehicle Qk may behave unexpectedly due to faults or
malfunctions, in which case it may come into an unsafe configuration
with another vehicle.
With our choice of Qi and the assumption that the platoon is in a
single-file formation, some vehicle Qi would get near the safety BRS
with respect to Qk, where Qk is likely to be the vehicle in front or
behind of Qi. In this case, a "safety breach" occurs. Our synthesis of
the safety controller guarantees the following: between every pair of
vehicles Qi, Qk, if VS(−t, xi − xk) > 0, then ∃ui to keep Qi from
colliding with Qk for a desired time horizon t, despite the worst case
(an adversarial) control from Qk. Therefore, as long as the number of
"safety breaches" is at most one for Qi, Qi can simply use the optimal
control to avoid Qk and avoid collision for the time horizon of t. Under
the assumption that vehicles are able to exit the current altitude range
within a duration of tfaulty, if we choose t = tfaulty, the safety breach
would always end before any collision can occur.
Within a duration of tfaulty, there is a small chance that additional
safety breaches may occur. However, as long as the total number of
safety breaches does not exceed the number of affected vehicles, col-
lision avoidance of all the vehicles can be guaranteed for the duration
tfaulty. However, as our simulation results show, placing vehicles in
single-file platoons makes the likelihood of multiple safety breaches
low during the presence of one intruder vehicle.
In the event that multiple safety breaches occur for some of the
vehicles due to a malfunctioning vehicle within the platoon or intrud-
ing vehicles outside of the platoon, vehicles that are causing safety
breaches must exit the highway altitude range in order to avoid col-
lisions. Every extra altitude range reduces the number of simultaneous
safety breaches by 1, so K simultaneous safety breaches can be re-
solved using K − 1 different altitude ranges. The general process and
details of the complete picture of multi-altitude collision avoidance is
part of our future work.
The concept of platooning can be coupled with any collision avoid-
ance algorithm that provides safety guarantees.
In this paper, we
have only proposed the simplest reachability-based collision avoidance
scheme. Existing collision avoidance algorithms such as [37] and [38]
have the potential to provide safety guarantees for many vehicles in
order to resolve multiple safety breaches at once. Coupling the pla-
tooning concept with the more advanced collision avoidance methods
that provide guarantees for a larger number of vehicles would reduce
the risk of multiple safety breaches.
Given that vehicles within a platoon are safe with respect to each
other, each platoon can be treated as a single vehicle, and perform colli-
sion avoidance with other platoons when needed. The option of treating
each platoon as a single unit can reduce the number of individual vehi-
cles that need to check for safety against each other, reducing overall
computation burden.
Figure 7. Summary of mode switching controllers. Reachability-based con-
trollers are shown as the blue and green arrows.
F. Safety Analysis
Under normal operations in a single platoon, each follower vehicle
Qi, i = P2, . . . , PM−1 in a platoon checks whether it is in the safety
BRS with respect to QPi−1 and QPi+1. So Qi = {Pi+1, Pi−1}
for i = P2, . . . , PN−1. Assuming there are no nearby vehicles out-
side of the platoon, the platoon leader QP1 checks safety against
G. Numerical Simulations
In this section, we consider several situations that vehicles in a platoon
on an air highway may commonly encounter, and show via simulations
the behaviors that emerge from the controllers we defined in Sections
III-C and III-D.
1. Forming a Platoon
We first consider the scenario in which Free vehicles merge onto an ini-
tially unoccupied highway. In order to do this, each vehicle first checks
safety with respect to all other vehicles, and uses the safety controller
if necessary, according to Section III-C-3. Otherwise, the vehicle uses
the goal satisfaction controller for getting to an absolute target set de-
scribed in Section III-C-1 in order to merge onto the highway, create a
platoon, and become a Leader vehicle if there are no platoons on the
highway. If there is already a platoon on the highway, then the vehi-
cle would use the goal satisfaction controller for getting to a target set
relative to the platoon leader as described in Section III-C-2 to join the
platoon and become a Follower.
Free:•Vehicle not in platoon or on highwayLeader•Leader of platoonFollower•Member of platoonLeave highwayMerge onto highway(get to absolute state)Create new platoon(get to absolute state)Merge with platoon in front(get to relative state)Join platoon(get to relative state)Follow highway(model predictive controller)Follow platoon(PD controller)Leave highway 9
For the simulation example, shown in Figure 8, the highway is
specified by a line segment beginning at the origin. The five vehicles,
Q1, Q2, . . . , Q5 are colored orange, purple, light blue, dark blue, and
yellow, respectively.
The first two plots in Figure 8 illustrate the use of safety and goal
satisfaction BRS for the first two vehicles. Since the goal satisfaction
BRSs are in 4D and the safety BRSs are in 6D, we compute and plot
their 2D slices based on the vehicles' velocities and relative velocities.
All vehicles begin as Free vehicles, so they each need to take into ac-
count five different BRSs: four safety BRSs and one goal satisfaction
BRS. For clarity, we only show the goal satisfaction BRS and the four
safety BRSs for one of the vehicles.
For Q1 (orange), an arbitrary point of entry on the highway is cho-
sen as the target absolute position, and the velocity corresponding to a
speed of 10 m/s in the direction of the highway is chosen as the target
absolute velocity. This forms the target state ¯xH = (¯px, ¯vx, ¯py, ¯vy),
from which we define the target set LH as in Section III-C-1.
At t = 4.2, Q1 (orange) is inside the goal satisfaction BRS for get-
ting to an absolute state, shown as the dotted orange boundary. There-
fore, it is "locked-in" to the target state ¯xH, and follows the optimal
control in (17) to ¯xH. During the entire time, Q1 checks whether it
may collide with any of the other vehicles within a time horizon of
tfaulty. To do this, it simply checks whether its state relative to each of
the other vehicles is within the corresponding safety BRS. As an exam-
ple, the safety BRS boundary with respect to Q2 (purple) is shown as
the orange dashed boundary around Q2 (purple); Q1 (orange) is safe
with respect to Q2 (purple) since Q1 (orange) is outside of the bound-
ary. In fact, Q1 is safe with respect to all vehicles.
After completing merging onto the empty highway, Q1 (orange)
creates a platoon and becomes its leader, while subsequent vehicles
begin to form a platoon behind the leader in the order of ascending
distance to Q1 (orange) according to the process described in Section
III-C-2. Here, we choose the target relative position (¯px,r, ¯py,r) to be a
distance dsep behind the last reserved slot in the platoon, and the target
relative velocity (¯vx,r, ¯vy,r) = (0, 0) with respect to the leader in order
to maintain the platoon formation. This gives us the target set LP that
we need.
At t = 8.0, Q2 (purple) is in the process of joining the platoon
behind Q1 (orange) by moving towards the target ¯xP relative to the
position of Q1 (orange). Note that ¯xP moves with Q1 (orange) as ¯xP
is defined in terms of the relative states of the two vehicles. Since Q2
is inside the goal satisfaction BRS boundary for joining the platoon
(purple dotted boundary), it is "locked-in" to the target relative state
¯xP , and begins following the optimal control in (14) towards the target
as long as it stays out of all safety BRSs. For example, at t = 5.9,
Q2 (purple) is outside of the safety BRS with respect to Q1 (orange),
shown as the purple dashed boundary around Q1 (orange). Again, from
the other safety BRS boundaries, we can see that Q2 is in fact safe with
respect to all vehicles.
In the bottom plots of Figure 8, Q1 (orange) and Q2 (purple) have
already become the platoon leader and follower, respectively. The rest
of the vehicles follow the same process to join the platoon. All 5 ve-
hicles eventually form a single platoon and travel along the highway
together. As with the first two vehicles, the goal satisfaction controllers
allow the remaining vehicles to optimally and smoothly join the pla-
toon, while the safety controllers prevent collisions from occurring.
2.
Intruder Vehicle
We now consider the scenario in which a platoon of vehicles encounters
an intruder vehicle. To avoid collision, each vehicle checks for safety
with respect to the intruder and any vehicles in front and behind of it in
the platoon. If necessary, the vehicle uses the reachability-based safety
controller to avoid collision, otherwise it uses the appropriate controller
to travel on the highway if it is a leader, or follow the leader if it is a
follower. After danger has passed, the vehicles in the platoon resume
normal operation.
Figure 9 shows the simulation result. At t = 9.9, a platoon of four
vehicles, Qi, i = 1, . . . , 4 (with Pi = i), travels along the highway
shown. An intruder vehicle Q5 (yellow) heads left, disregarding the
presence of the platoon. At t = 11.9, the platoon leader Q1 (red)
detects that it has gone near the boundary of the safety BRS (not shown)
with respect to the intruder Q5 (yellow). In response, Q1 (red) starts
using the safety controller to optimally avoid the intruder according to
(14); in doing so, it steers slightly off the highway.
Note that although in this particular simulation, the intruder travels
in a straight line, a straight line motion of the intruder was not assumed.
Rather, the safety BRSs are computed assuming the worst case control
of the intruder, according to (14).
As the intruder Q5 (yellow) continues to disregard other vehicles,
the followers of the platoon also get near the respective boundaries of
their safety BRSs with respect to the intruder. This occurs at t = 13.9,
where the platoon "makes room" for the intruder to pass by to avoid
collisions; all vehicles deviate from their intended path, which is to
follow the platoon leader or the highway. Note that in this case, we
have assumed that the platoon does not move as a unit in response to
an intruder to show more interesting behavior.
After the intruder has passed, eventually all vehicles become far
away from any safety BRSs. When this occurs, the leader resumes
following the highway, and the followers resume following the leader.
At t = 19.9, the platoon successfully gets back onto the highway.
3. Changing highways
In order to travel from origin to destination, a vehicle may need to
change highways several times before exiting an air highway system.
In this simulation, shown in Figure 10, two platoons are traveling on
two different highways that intersect. When the platoons are near the
highway intersection, two of the vehicles in the four-vehicle platoon
change highways and join the other platoon.
The t = 8.2 plot shows the two platoons of vehicles traveling on
the two air highways. One platoon has three vehicles, and the other
has four vehicles. At t = 12.3, the yellow vehicle begins steering off
its original highway in order to join the other platoon. In terms of the
hybrid systems modes, the yellow vehicle transitions from the Leader
mode to the Follower mode. At the same time, the green vehicle tran-
sitions from the Follower mode to the Leader mode, since the previous
platoon leader, the yellow vehicle, has left the platoon. By t = 16.9,
the yellow vehicle successfully changes highways and is now a fol-
lower in its new platoon.
At t = 16.9, the dark red vehicle is in the process of changing
highways. In this case, it remains in the Follower mode, since it is a
follower in both its old and new platoons. While the dark red vehicle
changes highways, the orange vehicle moves forward to catch up to
its new platoon leader, the green vehicle. By t = 23, all the vehicles
have finished performing their desired maneuvers, resulting in a two-
vehicle platoon and a five-vehicle platoon traveling on their respective
highways.
IV. Conclusions
To address the important and urgent problem of the traffic man-
agement of unmanned aerial vehicles (UAVs), we proposed to have
platoons of UAVs traveling on air highways. We showed how such
an airspace structure leads to much easier safety and goal satisfaction
analysis. We provided simulations which show that by organizing ve-
hicles into platoons, many complex maneuvers can be performed using
just a few different backward reachable sets.
For the placement of air highways over a region, we utilize the very
intuitive and efficient fast marching algorithm for solving the Eikonal
equation. Our algorithm allows us to take as input any arbitrary cost
map representing the desirability of flying over any position in space,
and produce a set of paths from any destination to a particular origin.
Simple heuristic clustering methods can then be used to convert the sets
of paths into a set of air highways.
On the air highways, we considered platoons of UAVs modeled
by hybrid systems. We show how various required platoon functions
(merging onto an air highway, changing platoons, etc.) can be imple-
mented using only the Free, Leader, and Follower modes of operation.
Using HJ reachability, we proposed goal satisfaction controllers that
guarantee the success of all mode transitions, and wrapped a safety
controller around goal satisfaction controllers to ensure no collision be-
tween the UAVs can occur. Under the assumption that faulty vehicles
can descend after a pre-specified duration, our safety controller guar-
antees that no collisions will occur in a single altitude level as long as
at most one safety breach occurs for each vehicle in the platoon. Addi-
tional safety breaches can be handled by multiple altitude ranges in the
airspace.
Acknowledgments
This work is supported in part by NSF under CPS:ActionWebs
(CNS-0931843) and CPS:FORCES (CNS1239166), by NASA under
grants NNX12AR18A and UCSCMCA-14-022 (UARC), by ONR un-
der grants N00014-12-1-0609, N000141310341 (Embedded Humans
MURI), and MIT 5710002646 (SMARTS MURI), and by AFOSR un-
der grants UPenn-FA9550-10-1-0567 (CHASE MURI) and the SURE
project.
References
[1] Tice, B. P., "Unmanned Aerial Vehicles – The Force Multiplier of the
1990s," Airpower Journal, 1991, pp. 41–55.
[2] Haulman, D. L., "U.S. Unmanned Aerial Vehicles in Combat, 1991-2003,"
Tech. rep., Air Force Historical Research Agency, Maxwell Air Force
Base, Alabama, 2003.
Inc.,
[3] Amazon.com,
Prime Air
"Amazon
http://www.amazon.com/b?node=8037720011
2016].
[online],"
[retrieved Oct.
2014,
17,
[4] Stewart, J., "Google tests drone deliveries in Project Wing trials," 2014.
[5] Prevot, T., Rios, J., Kopardekar, P., Robinson III, J. E., Johnson, M., and
Jung, J., "UAS Traffic Management (UTM) Concept of Operations to
Safely Enable Low Altitude Flight Operations," Proceedings of the AIAA
Aviation Technology, Integration, and Operations Conference, Jun. 2016,
AIAA paper number: 2016-3292, doi: 10.2514/6.2016-3292.
[6] DeBusk, W., "Unmanned Aerial Vehicle Systems for Disaster Relief: Tor-
nado Alley," AIAA Infotech@Aerospace 2010, Apr. 2010, AIAA paper
number: 2010-3506, 10.2514/6.2010-3506.
[7] AUVSI News, "UAS Aid in South Carolina Tornado Investigation
[online]," http://www.auvsi.org/blogs/auvsi-news/2016/01/29/tornado [re-
trieved Feb. 16, 2016].
[8] Jointed Planning and Development Office (JPDO), "Unmanned Aircraft
Systems (UAS) Comprehensive Plan – A Report on the Nation's UAS
Path Forward," Tech. rep., Federal Aviation Administration, 2013.
[9] National Aeronautics and Space Administration, "Challenge is On to De-
sign Sky for All," 2016, http://www.nasa.gov/feature/challenge-is-on-to-
design-sky-for-all, [retrieved Feb. 12, 2016].
[10] Bokanowski, O., Forcadel, N., and Zidani, H., "Reachability and Minimal
Times for State Constrained Nonlinear Problems without Any Controlla-
bility Assumption," SIAM Journal on Control and Optimization, Vol. 48,
No. 7, Jan. 2010, pp. 4292–4316, doi: 10.1137/090762075.
[11] Mitchell, I. M., Bayen, A. M., and Tomlin, C. J., "A time-dependent
Hamilton-Jacobi formulation of reachable sets for continuous dynamic
games," IEEE Transactions on Automatic Control, Vol. 50, No. 7, Jul.
2005, pp. 947–957.
[12] Barron, E. and Ishii, H., "The Bellman equation for minimizing the maxi-
mum cost," Nonlinear Analysis: Theory, Methods & Applications, Vol. 13,
No. 9, Sept. 1989, pp. 1067–1090, doi: 10.1016/0362-546X(89)90096-5.
[13] Fisac, J. F., Chen, M., Tomlin, C. J., and Sastry, S. S., "Reach-avoid prob-
lems with time-varying dynamics, targets and constraints," Proceedings of
the International Conference on Hybrid Systems: Computation and Con-
trol, 2015, pp. 11–20, doi: 10.1145/2728606.2728612.
[14] Chen, M., Zhou, Z., and Tomlin, C. J., "Multiplayer reach-avoid
games via low dimensional solutions and maximum matching," Proceed-
ings of the American Control Conference, 2014, pp. 1444–1449, doi:
10.1109/ACC.2014.6859219.
[15] Chen, M., Zhou, Z., and Tomlin, C. J., "Multiplayer Reach-Avoid Games
via Pairwise Outcomes," IEEE Transactions on Automatic Control, 2016,
pp. 1–1, doi: 10.1109/TAC.2016.2577619.
[16] Ding, J., Sprinkle, J., Sastry, S. S., and Tomlin, C. J., "Reacha-
bility calculations for automated aerial refueling," Proceedings of the
IEEE Conference on Decision and Control, 2008, pp. 3706–3712, doi:
10.1109/CDC.2008.4738998.
10
[17] Mitchell, I. M., "The Flexible, Extensible and Efficient Toolbox of Level
Set Methods," Journal of Scientific Computing, Vol. 35, No. 2-3, Jun.
2008, pp. 300–329, doi: 10.1007/s10915-007-9174-4.
[18] Osher, S. and Fedkiw, R., Level Set Methods and Dynamic Implicit Sur-
faces, Springer-Verlag, 2002.
[19] Sethian, J. A., "A fast marching level set method for monotonically
advancing fronts." Proceedings of the National Academy of Sciences,
Vol. 93, No. 4, Feb. 1996, pp. 1591–1595, doi: 10.1073/pnas.93.4.1591.
[20] Kavathekar, P. and Chen, Y., "Vehicle Platooning: A Brief Survey and Cat-
egorization," Proceedings of the ASME/IEEE International Conference on
Mechatronic and Embedded Systems and Applications, Parts A and B,
Vol. 3, 2011, pp. 829–845, doi: 10.1115/DETC2011-47861.
[21] McMahon, H. D., Hedrick, K. J., and Shladover, E. S., "Vehicle Mod-
elling and Control for Automated Highway Systems," American Control
Conference, 1990, May 1990, pp. 297–303.
[22] Hedrick, K. J., Zhang, G., Narendran, K. V., s. Chang, K., for Ad-
vanced Transit, P., (Calif.), H., and University of California, B. I. o. T. S.,
Transitional Platoon Maneuvers in an Automated Highway System, Cal-
ifornia PATH Program, Institute of Transportation Studies, University of
California at Berkeley, 1992.
[23] Lygeros, J., Godbole, D. N., and Sastry, S. S., "Verified hybrid con-
trollers for automated vehicles," IEEE Transactions on Automatic Control,
Vol. 43, No. 4, Apr. 1998, pp. 522–539.
[24] Gattami, A., Al Alam, A., Johansson, K. H., and Tomlin, C. J., "Estab-
lishing Safety for Heavy Duty Vehicle Platooning: A Game Theoretical
Approach," Vol. 44, Jan. 2011, pp. 3818–3823, doi: 10.3182/20110828-6-
IT-1002.02071.
[25] Sabau, S., Oara, C., Warnick, S., and Jadbabaie, A., "Optimal
Distributed Control
for Platooning via Sparse Coprime Factoriza-
tions," IEEE Transactions on Automatic Control, 2016, pp. 1–1, doi:
10.1109/TAC.2016.2572002.
[26] Crandall, M. G., Evans, L. C., and Lions, P.-L., "Some properties of vis-
cosity solutions of Hamilton-Jacobi equations," Transactions of the Amer-
ican Mathematical Society, Vol. 282, No. 2, Apr. 1984, pp. 487–502, doi:
S0002-9947-1984-0732102-X.
[27] Alton, K. and Mitchell, I., "Optimal path planning under defferent
the IEEE Inter-
norms in continuous state spaces," Proceedings of
national Conference on Robotics and Automation, pp. 866–872, doi:
10.1109/ROBOT.2006.1641818.
[28] Lygeros, J., Sastry, S., and Tomlin, C., Hybrid Systems: Foundations, ad-
vanced topics and applications, Springer Verlag, 2012.
[29] Adler, A., Bar-Gill, A., and Shimkin, N., "Optimal flight paths
for engine-out emergency landing," Proceedings of
the IEEE Chi-
nese Control and Decision Conference, 2012, pp. 2908–2915, doi:
10.1109/CCDC.2012.6244461.
[30] Coombes, M., Chen, W.-H., and Render, P., "Reachability Analysis of
Landing Sites for Forced Landing of a UAS," Journal of Intelligent
& Robotic Systems, Vol. 73, No. 1-4, Jan. 2014, pp. 635–653, doi:
10.1007/s10846-013-9920-9.
[31] Idicula, J., Akametalu, K., Chen, M., Tomlin, C., Ding, J., and Hook, L.,
"Where to Land: A Reachability Based Forced Landing Algorithm for
Aircraft Engine Out Scenarios," Nov. 2015, Seedling Technical Seminar,
NASA Dryden Flight Research Center, Report number DFRC-E-DAA-
TN28194.
[32] Evans, L. C. and Souganidis, P. E., "Differential games and representa-
tion formulas for solutions of Hamilton-Jacobi-Isaacs equations," Indiana
Univ. Math. J., Vol. 33, No. 5, 1984, pp. 773–797.
[33] Varaiya, P. P., "On the Existence of Solutions to a Differential Game,"
SIAM Journal on Control, Vol. 5, No. 1, Feb. 1967, pp. 153–162, doi:
10.1137/0305009.
[34] Chen, M. and Tomlin, C.
"Exact and efficient Hamilton-
Jacobi reachability for decoupled systems," Proceedings of the IEEE
Conference on Decision and Control, 2015, pp. 1297–1303, doi:
10.1109/CDC.2015.7402390.
J.,
[35] Chen, M., Herbert, S., and Tomlin, C. J., "Exact and Efficient Hamilton-
Jacobi-based Guaranteed Safety Analysis via System Decomposition,"
IEEE International Conference on Robotics and Automation 2017 (to ap-
pear), 2017.
11
[36] Chen, M., Herbert, S., Vashishtha, M., Bansal, S., and Tomlin, C. J., "A
General System Decomposition Method for Computing Reachable Sets
and Tubes," IEEE Transactions on Automatic Control (submitted), 2016.
[37] Bansal*, S., Chen*, M., and Tomlin, C. J., "Safe Sequential Path Planning
of Multi-Vehicle Systems Under Presence of Disturbances and Measure-
ment Noise," American Control Conference (to appear), 2017.
[38] Chen*, M., Shih*, J., and Tomlin, C. J., "Multi-Vehicle Collision Avoid-
ance via Reachability and Mixed Integer Programming," Proceedings of
the IEEE Conference on Decision and Control, 2016.
12
Figure 2. Cost-minimizing paths computed by the Fast Marching Method based on the assumed cost map of the San Francisco Bay Area.
Figure 3. Results of conversion from cost-minimizing paths to highway way points.
Bay Area Map, Cost-Minimizing Paths-1-0.500.51-1-0.8-0.6-0.4-0.200.20.40.60.81Cost Map, Cost-Minimizing Paths, Value Function-1-0.500.51-1-0.8-0.6-0.4-0.200.20.40.60.81-1-0.500.51-1-0.8-0.6-0.4-0.200.20.40.60.81-1-0.500.51-1-0.8-0.6-0.4-0.200.20.40.60.81 13
(a) The red vehicle is merging onto the highway while avoiding colli-
sions.
(b) The purple vehicle is joining the platoon while avoiding collisions.
(c) The last three vehicles follow the same process to join the platoon. (d) All five vehicles have successfully joined the platoon and now travel
on the highway together.
Figure 8. A simulation showing how five vehicles initially in the Free mode can form a platoon.
-150-100-50050100150200-300-250-200-150-100-50050t=4.2-150-100-50050100150-300-250-200-150-100-50050t=8-50050100-300-250-200-150-100-500t=10.9-50050100-300-250-200-150-100-500t=20.9 14
(a) A four-vehicle platoon travels along the highway. The yellow vehi-
cle disregards the others.
(b) Vehicles begin avoidance maneuvers as they get near safety BRS
boundaries.
(c) Safety controllers cause vehicles to spread out to "make room" for
the intruder to pass.
(d) After danger has passed, the platoon resumes normal operation.
Figure 9. A simulation showing how a platoon of four vehicles reacts to an intruder.
050100150200-180-160-140-120-100-80-60-40-20020t=9.9050100150200-180-160-140-120-100-80-60-40-20020t=11.9050100150200-180-160-140-120-100-80-60-40-20020t=13.9050100150200-180-160-140-120-100-80-60-40-20020t=19.9 15
(a) A three-vehicle platoon and a four-vehicle platoon travel on their
respective air highways.
(b) The yellow vehicle begins to join the new platoon; The green vehicle
becomes a leader.
(c) The dark red vehicle joins a new platoon; the orange vehicle catches
up to new platoon leader.
(d) New platoons now travel on their respective air highways.
Figure 10. A simulation showing two vehicles changing highways and joining a new platoon.
-200-1000100200-200-150-100-50050100150200t=8.2-200-1000100200-200-150-100-50050100150200t=12.3-200-1000100200-200-150-100-50050100150200t=16.9-200-1000100200-200-150-100-50050100150200t=23 |
1905.01358 | 1 | 1905 | 2019-04-29T11:16:34 | Agent based decision making for Integrated Air Defense system | [
"cs.MA"
] | This paper presents algorithms of decision making agents for an integrated air defense (IAD) system. The advantage of using agent based over conventional decision making system is its ability to automatically detect and track targets and if required allocate weapons to neutralize threat in an integrated mode. Such approach is particularly useful for futuristic network centric warfare. Two agents are presented here that perform the basic decisions making tasks of command and control (C2) like detection and action against jamming, threat assessment and weapons allocation, etc. The belief-desire-intension (BDI) architectures stay behind the building blocks of these agents. These agents decide their actions by meta level plan reasoning process. The proposed agent based IAD system runs without any manual inputs, and represents a state of art model for C2 autonomy. | cs.MA | cs | JOURNAL OF BATTLEFIELD TECHNOLOGY VOL 14, NO 1, MARCH 2011
1
AGENT-BASED DECISION MAKING FOR INTEGRATED AIR DEFENSE SYSTEM
Sumanta K. Das and Sumant Mukherjee1
Abstract. This paper presents algorithms of decision making agents for an integrated air defense (IAD) system. The
advantage of using agent based over conventional decision making system is its ability to automatically detect and track
targets and if required allocate weapons to neutralize threat in an integrated mode. Such approach is particularly
useful for futuristic network centric warfare. Two agents are presented here that perform the basic decisions making
tasks of command and control (C2) like detection and action against jamming, threat assessment and weapons
allocation, etc. The belief-desire-intension (BDI) architectures stay behind the building blocks of these agents. These
agents decide their actions by meta level plan reasoning process. The proposed agent based IAD system runs without
any manual inputs, and represents a state of art model for C2 autonomy.
INTRODUCTION
Conventional ways of decision making for command and
control (C2) for an Integrated Air Defense (IAD) system are
performed by human decision makers. The term IAD means
that different tactical air defense services like searching,
detecting, tracking, identifying and engaging targets using air
defense sensors (radars) & weapons (aircrafts & missiles) are
performed in an integrated fashion. Network Centric Warfare
(NCW) is a concept that makes IAD operations successful.
The C2 of NCW is viewed as a collaborative decision making
process. With the advent of synchronous or asynchronous
NCW in terms of both time and space [1], the conventional
methods and modes of implementing the decision making
processes of C2 has become obsolete [2]. The modern
networked-laid IAD systems demand advanced method of
decision making that should be enriched with artificial
intelligence (AI) techniques. At each level of service
execution decisions need to be taken autonomously by
intelligent computational entities or agents. These agents
should be capable
localized decision and
communicate with each other to achieve a collective goal.
take
to
the economics
The Belief-Desire-Intension (BDI) architectures [3] are based
on the philosophical tradition of understanding practical
reasoning. Recently these architectures are extended to
develop autonomous agents on the basis of a number of
disciplines ranging from
to cognitive
psychology to mathematics. The BDI architectures are
applied for developing agents that behave deliberatively and
reactively in a complex environment. In these architectures,
the mental attitudes of the agent are represented by the
attributes like beliefs, desires and intentions. The belief is the
knowledge of the agent about its world or environment.
Agent‟s desire or goal is the condition that agent wants to
satisfy. After satisfying the conditions, the agent has to
perform certain action to achieve the goal that is known as
intentions. Agents have different course of actions to achieve
different desires or goals i.e. known as plan repository or plan
library. JACK is the most widely used programming
language for developing the BDI agents [4].
In the recent times several studies have been performed to
understand and improve the agent based modeling in
different application domains. The agent technologies have
been successfully deployed for wireless battery powered
sensor network in [5] for graph colouring problem. Agent
based modeling and simulation tools are used for making
automated car driver [6]. Software agents can be embedded
on the web as a replacement of the human user. These agents
can do the work which human user is supposed to do. In such
situation dynamic service composition is essential. In [7] a
work
is presented where agents are evolving service
semantics cooperatively in a consumer driven approach. An
application of distributed computation by multi agent system
for traffic control is presented in [8]. Introducing learning
capability in BDI architectures is studied by [9]. A new
architecture is presented in that study as an extension of the
BDI architectures in which learning process is described as
plans. The manipulative abduction
reasons by
experiences and exhibitions of behaviour to find some pattern
in the environment is used for the learning process.
that
Search is an essential part of the agent‟s model. It is a
sequence of actions that takes any agent from the initial state
to the goal state. Search could be uninformed or informed
(heuristic). Heuristic search is an essential action for agents
that work in the real time. Two classes of heuristic search
methods are common, namely real time heuristic search and
incremental heuristic search. A detail comparison of these
two methods along with their advantages and disadvantages
is presented in [10].
The agent models are difficult to verify because there is
always a gap of understanding between agent logic and agent
programming. To overcome this problem an operational
semantics of agent programming language is presented in
[11]. In that study agent logic is first grounded by state based
semantics then denotational semantics is used to connect the
agent logic with agent programming.
An agent may pursue multiple goals at same time. In such
situation it may happens that pursuing multiple goals at the
same time simultaneously is not possible. This is known as
conflicting goals situation. The semantic representation of
conflicting goals is presented in [12]. Monitoring many
agents in a multi-agent architecture is a viable problem of
agent development work. Usual disagreements between
different agents arise in such situation.
are
identified, designed
In the present study C2 agents that are capable of taking
autonomous decisions
and
implemented for an IAD system. The OODA approach [13]
(i.e. Observe-orient-decide-act) is assumed for modeling the
tactical behavior of these agents. This control loop has since
long been used for understanding the human participation in
the complex C2 problem. Major roles of C2 of air defense
systems are threat assessment (TA) and weapon allocation
1 Institute for Systems Studies and Analyses, Defence Research and Development Organization, Metcalfe House, Delhi-110054, India.
2
DAS & MUKHERJEE: AGENT BASED DECISION MAKING FOR AIR DEFENSE SYSTEM
GCCC
LCCC1
LCCC2
…
LCCCn
SRdr11
…
SRdr13
AB11
AB13
TRdr111
SAM111
AP111
…
TRdr131
SAM131
AP131
Figure 1.
Hierarchical structure of a standard IAD system.
(WA). The objective of this study is to apply the practical
reasoning process of human decision makers to develop
autonomous agents responsible for TA and WA. The BDI
architectures are most suitable for
the
philosophical tradition of understanding practical reasoning.
These architectures are also suitable for developing team of
agents as similar to the hierarchical structure of air defense
system.
implementing
This paper is intended to contribute the application of BDI
architectures, as an extension of the goal based agent
architecture, for developing the decision making agents for an
IAD system. Main concern is to formulate the mental
attributes of a human decision maker in terms of belief, desire
and intension. This is a novel methodological application of
agent based modelling for IAD system and a technology
integration between agent oriented programming and Java
based combat simulation model. Two decision making agents
are proposed. The first one is related with identification of
jamming by surveillance radar in battlefield and the second
one is related with TA and WA. A brief discussion is
presented about the deployment status of these agents in a
simulated air combat scenario along with the lessons learned
from this study.
IAD SYSTEM
Air defense system has progressed steadily over recent years
to include highly sophisticated mission planning tools and
artificially intelligent capabilities [14, 15]. The Air force
mission support system (AFMSS, [14]), Power Scene and
Top Scene [14] all represent major advances in this field. Up
to now, most IAD operations consist of large teams of human
operators that control the IAD‟s actions.
The idea of using multi agent system (MAS) for weapons and
targets management in IAD is appropriate for distributed
architectures. In cooperative MAS, agents work together to
achieve one or more desired common goals. The overall
system goal is achieved through interactions and coordination
of the individual agents [3]. A distributed agent team has
advantages over a single, complex agent in many applications
[16]. For example, for search and rescue operations, multiple
robots can forage far more effectively than a single, complex
robot [17].
COMMAND AND CONTROL OF IAD SYSTEM
This section is intended to identify the possible information
processing agents for performing the task of C2 in an IAD
system. C2 of air defense system of most of the countries
follow certain hierarchical structure. Information is being
exchanged between different levels of this structure. The
Global Command and Control Centre (GCCC), Local
Command and Control Centre (LCCC), Surveillance Radar
(SRdr), Airbase Cadre (AB), tracking radar (TRdr), surface to
air missile (SAM), aircraft pilot (AP) are the different
components of an IAD system. The main roles of these
components are target detection and classification, threat
assessment (TA) and weapons allocation (WA).
Figure 1 shows the hierarchical structure of a standard IAD
system. The directions of flow of information between
different levels are shown by arrows. The arrow with dashed
line is purposely used to represent that command is passing
from higher to lower level. At the top of the tree is the higher
command unit which is known as the Global Command And
Control Centre (GCCC). Therefore in a MAS set-up these are
identified as one agent; namely the GCCC agent.
The GCCC agent first analyses the decisions given by the
different LCCCs located at diverse locations and takes its
own decision then passes it to the next level of the command
units i.e. to AB. The LCCC unit is identified as the second
agent. This agent analyses the information given by the
different SRdr(s) at diverse locations and takes their own
decision based on its perception and passes it to the GCCC
agent. The SRdr is the third type of agent. The LCCC agents
decide which target to engage and which weapon to allocate
to that target. The AB is identified as the fourth type of
agents. Based on the decisions given by the LCCC agent, the
AB agent decides which TRdr to track which target and
which SAM to engage which target. The TRdr and the SAM
systems are considered to be the fifth type of agents. Based
on the decisions given by the AB agent these agents engage
targets. In this study only two agents (namely, SRdr and
LCCC agent) are designed and implemented to show the
paradigm shift of agent based decision making for an IAD
system.
JOURNAL OF BATTLEFIELD TECHNOLOGY VOL 14, NO 1, MARCH 2011
3
Figure 2. Working principal of surveillance radar agent, nt stands for number of targets detected at time t. X
denotes target detected and X denotes the hidden target.
Figure 3. LCCC agent (4), prioritizing the clusters (1) and performing autonomously the target-interceptor
pairing (action) (3). Different targets (2) are grouped in different clusters (1) by MSDF module. For
prioritizing the clusters the LCCC agent uses the VAVP values (5).
BDI ARCHITECTURES OF C2 AGENTS
Two main questions are answered while constructing the BDI
architectures of the C2 agents. First one is what goals
(options or desires) the agent decides to achieve with its
current beliefs about the environment and second is how it is
going to achieve these chosen goals (intensions) by means of
some actions. These issues are resolved from the practical
reasoning applied by the human experts to the air defense
domain.
A. BDI architectures of the surveillance radar agent
Surveillance radars are required to detect aircrafts or missiles
flying towards them and often misdirected or confused by the
enemy targets that uses noise jamming. The experienced
radar operators can detect jamming and they generally decide
to keep the radar switch off in such situation. The goal of the
SRdr agent is to protect the radar from the noise jamming.
From the intensity of jamming this agent can decide which
action will be suitable for the radar.
The SRdr agent measures the intensity of jamming from the
difference between the numbers of targets detected at time
t+1 and t. If the difference is significant, the radar is jammed.
The working principal of SRdr agent is shown in the Figure
2. The SRdr agent is assumed to be deployed in the simulated
environment. It receives the information like "numbers of
targets (nt)" detected at time t from its environment. On the
basis of nt+1 and nt it identifies the occurrence of noise
jamming. An index is defined for this purpose namely
Normalized Target Difference (NTD = ((nt-nt+1)/nt )). On
the basis of the NTD values it decides what action it should
perform based on its belief. The main action of the SRdr
agent is to perform target detection in a jamming free
environment. Depending on the NTD values, the SRdr agent
can stay either in any two of the four states namely "Sense
Mode", "Sleep Mode", "Switch Off" and "Frequency
Hopping". If the "Jamming" is found then it can go either in
"Switch Off" or in "Frequency Hopping" mode.
B. BDI architectures of the LCCC agent
X X X X X X X X X X X X X X X X X X time = t time = t + t time = t1 time = t2 time = t3 time = t4 nt= 5 n t + t = 2 NTD = 0.6 Jammed Frequency Switch Off (if NTD>0.5) Hopping (if NTD > 0.75) (if 0.5< NTD <0.75) Z Z X X X V11 Y Y Y V-2,-2 V-1,-2 V0,-2 V1,-2 V2,-2 V-2,-1 V-1,-1 V0,-1 V1,-1 V2,-1 V-2,0 V-1,0 V0,0 V1,0 V2,0 V-2,1 V-1,1 V0,1 V11 V2,1 V-2,2 V-1,2 V0,2 V12 V2,2 1 2 3 4 5 Cluster 1 Cluster 2 Cluster 3 4
DAS & MUKHERJEE: AGENT BASED DECISION MAKING FOR AIR DEFENSE SYSTEM
Initialize clock = 0, simulation_time = 60, LCCC agent A;
Input: cluster id, package size, mission type, cluster locations, number or attacking aircraft. VAVP locations, interceptor locations.
Output: target priority list, target-interceptor pairing.
1.
2. While ( clock < simulation_time)
3.
4.
5. Endwhile
A.distance () ;
Clock ++ ;
3. A.distance ()
Start:
3.1. Compute distances (d1, d2) between clusters and VAVP and clusters and interceptors;
3.2. Add the distances in a beliefset-1 and beliefset-2 respectively.
3.3. Add attacking aircraft ranking in a beliefset-3.
3.4. Add attacking aircraft and interceptor availability separately in beliefsets-4 and 5 respectively;
3.5. Post an event (ev1) confirming that all belief updating is complete;
3.6. Meta-level plan reasoning using beliefset-1 to find closest clusters;
3.7. Post the closet cluster information by an event ev2.
3.8. Meta level plan reasoning using beliefsets-2 and 5 to find closet interceptor.
3.9. Post the cluster-interceptor pairing with an event ev3.
3.10. Meta level plan reasoning using beliefsets-3 and 4 to find closet attacking aircraft in the cluster.
3.11. Update the beliefsets-4 and 5.
End
Figure 4. Algorithm for implementing the LCCC agent.
Initialize clock = 0, simulation_time = 60, Surveillance Radar Agent S;
Input: number of targets (n) at t and t+1.
Output: identifying jamming, and action against jamming.
1.
2. While ( clock < simulation_time)
3.
4.
5. Endwhile
S.update (clock) ;
Clock ++ ;
3. S.update (clock)
Start :
3.1. Calculate NTD as ratio of nt / nt-1.
3.2. Add the NTD and clock time in beliefset-1.
3.3. Automatically post an event (ev1) containing NTD value and clock time.
3.4. Meta level plan reasoning (plan-1) for no action if NTD value is less than 0.5.
3.5. Meta level plan reasoning for action against jamming (plan-2) if NTD value is greater than 0.5.
3.6. Posting an event ev2 from plan-2 containing NTD value and clock time.
3.7. Meta level plan reasoning for frequency hopping action (plan-3) if NTD value lies between 0.5 to 0.75.
3.8. Meta level plan reasoning to switch off the radar (plan-4) if NTD value is greater than 0.75.
End
Figure 5. Algorithm for implementing the surveillance radar agent.
The LCCC agent is responsible for TA and WA. This agent
gets inputs from the multisensory data fusion (MSDF)
module. The MSDF groups the detected targets into different
clusters and sends the cluster information (cluster identity,
cluster location) along with the situational assessed inputs
about the enemy‟s intent (like mission type i.e. strike or
escort, package size i.e. small or large). Based on this
information and the LCCC agent‟s own beliefs (like VAVP
(vulnerable area and vulnerable points) value), LCCC agent
prioritizes the clusters and allocate interceptor to the
attacking aircraft. Figure 3 shows the LCCC agent residing in
a grid environment, evaluating the threat and prioritizing
targets along with target-interceptor pairing. For finding the
closest cluster this agent uses the meta-level plan reasoning
(MLPR) process based on the distance measure. The goal of
the LCCC agent is to optimally engage the detected targets
with its available interceptors subject to the restriction that no
target gets engaged by more than one interceptor.
META LEVEL PLAN REASONING
In this study, the concept of MLPR [3, 4, 9] is used
extensively by the C2 agents for taking optimal decisions.
MLPR is a method of selecting the appropriate plan from the
plan library to satisfy the agent‟s goal. This method is
generally used for BDI agent implementation. The actions in
MLPR are supposed to be optimal in some respect.
Sometimes, MLPR can also be used to enable the agent to
learn from the changing environment.
MLPR is implemented by using the getInstanceInfo() library
function provided by the JACK [17]. The getInstanceInfo()
method calculates the ranking of a plan by a PlanInstanceInfo
object. The ranking is done by calculating one index which is
a function of distance, mission type and package size.
Mission types and package sizes are assumed to be fuzzy set.
The membership values of these variables are obtained by
JOURNAL OF BATTLEFIELD TECHNOLOGY VOL 14, NO 1, MARCH 2011
5
(a)
(b)
Figure. 6. Architecture of (a) surveillance radar agent and (b) LCCC agents developed through JACK agent
programming language. Abbreviations: b: beleifset, ev: event, p: plan, H: handles, P: posts, M:modifies,
A: add, NTD: Normalized target difference.
example,
using a trapezoidal membership function. Each distance,
package size and mission type generates a distinct plan
(Figure 7). The plan with maximum ranking get selected by
the getInstanceInfo() function. In this way MLPR capability
is incorporated in the LCCC agent‟s architectures. The events
are posted either by the LCCC agent itself or by other plans.
(namely
For
the
NewClusterPriorityEvents)
the plan
"NewClusterPlan" when this plan is selected by the agent. In
the similar way the interceptor aircraft of the defender force
is allocated to the nearest aircraft. While allocating an
interceptor to the aircraft the agent also checks its availability
status so that multiple allocations do not take place. The
agent‟s algorithms are shown in the Figures 4 and 5 and
implemented through the JACK agent programming language
as shown in the Figure 6 (a and b).
ev2
from
event
is posted
The agent can decide which of the plans are applicable for a
particular event using either the three functions separately or
together. First one is the relevant() function. The agent uses
this function to select the plans that can handle such
particular event. Second is the context() function. This
function is used to select plan instances which are consistent
with the agent's current beliefs. If there are still multiple
plans left in the applicable plan set, the JACK provides the
getInstanceInfo() function to return a PlanInstanceInfo
object. This class has the def() method which can return the
rank of the plan. The MLPRs of the "NewClusterPlan" of the
LCCC agent and "Radar is Jammed" plan of the SRdr agent
are shown in the Figure 7. The rank of the "NewClusterPlan"
plan instance is the function of normalized distance between
the cluster mean and the VAVP value, package size and
mission type. The SRdr agent uses the relevant() function for
MLPR where as the LCCC agent uses both the context() and
the getInstanceInfo() functions together.
EVALUATION
Two approaches are used for evaluating the C2 agents. First
approach is the logical evaluation and second approach is the
statistical evaluation. Although logical evaluation is the most
widely used method for agent research, it can not quantify the
performance of the agents. A solution could be to use logical
evaluation for
the deadlock situations and
quantifying the performance by statistical measures. The
logical verification rectifies the conflicting/multiple goal
situations in the system. The statistical hypothesis testing
measures the performance of the agents.
identifying
A. Logical Evaluation
First approach is based on the logical verification of the
agent‟s model. In logical verification, the concept of goal
inference rule (gir) is used for detecting the conflicting goals
in the system ([12]). The girs‟ for SRdr agents is defined as
in the Figure 8. For example the gir,
{Jammed} ,{Frequency Hopping}k- Switch Off
represents that if the SRdr agent is Jammed (belief state
denoted by ), it may derive the goal to go for the "Switch
Off" plan (denoted by 1), but the goal to go for the
"Frequency Hopping" (denoted by 2) is in conflict (denoted
by k- ,where as k+ denotes non-conflicting goals)with the
goal to go for the "Switch off" plan (see Figure 8).
The girs are extended to default logic ([12]). The gir helps to
find out any sort of conflicting goals present in the model.
Consider that one wants to express that if a SRdr agent is
found "Jammed", it may go either for the "Switch Off" or
"Frequency Hopping" mode, but should not simultaneously
pursue these goals simultaneously, i.e., the goals "Switch Off"
6
DAS & MUKHERJEE: AGENT BASED DECISION MAKING FOR AIR DEFENSE SYSTEM
context()
{
ClusterPriority.get ($cls_id, $pck_size, $msn_type, $vavp_id, $distance);
}
static boolean relevant(Antijam ev)
{
return (ev.NTD >0.75);
}
public PlanInstanceInfo getInstanceInfo()
{
try {
if ($pck_size.equals("Big")) value1 = 2; else value1 = 1;
if ($msn_type. Equals("Strike")) value2 = 2; else value2 = 1;
rank = (int) ($distance.as_double() / 100.0 + value1 / 2 + value2 / 2);
return PlanInstanceInfo.def [9-rank]; }
catch (LogicException ex) { return PlanInstanceInfo.def [0]; }
}
Figure 7 Meta level plan reasoning by context(), relevant(), and getInstanceInfo() functions. The relevant() method in
this diagram is part of the "Radar is Jammed" plan of the surveillance radar agent. The context() and
getInstanceInfo() functions are part of the "NewClusterPlan" plan of the LCCC agent.
and "Frequency Hopping" are conflicting. Moreover, if a
SRdr agent has "Switch Off" mode, it wants to go "Sleep
Mode" with it, if it is in the "Frequency Hopping" mode, it
wants to remain in the "Sense Mode". This could be modeled
using the girs, as shown in the Figure 8.
as given in the Figure 9. Probability of false alarm plays a
vital role in the correct detection of jamming. Although a
single value of false alarm (i.e. 5%) is taken in this study, the
performance measure of the agent can be simulated for other
values of Pf.
Another form of logical evaluation adopted in the study is
representing the entire mechanism in the form of operational
semantics [11]. In this approach first agent model is grounded
with state-based semantic, then denotational semantics are
used to define the mathematical relation connecting agent
logic and agent programming. The operational semantic, state
based semantic, model semantic of LCCC agent use a
propositional language (L0) to represent their environment
with the operators like (conjunction), (disjunction) and
(negation). The L0 is infinite set of atomic proposition that
uses entailment ( = ) relation. The operational semantic
defines the input-output relation as a compositional function
mapping from initial states to the final state reached upon
termination. The state based semantics provides
the
ingredients for defining the operational semantics. The
denotational semantics provides the semantics for a modal
logic of agent programs [11].
B. Statistical Evaluation
This evaluation is based on the classical statistical approach
of hypothesis testing. The assumed hypothesis (also called
Null hypothesis, H0) is that the output data of the agent model
follows certain perfect statistical distribution.
The SRdr agent acts accordingly to the distribution pattern of
the number of targets detected (nt). The cut off value of the
NTD has a role in the performance measure of the SRdr
agent. If the number of targets detected (nt) follows the
Gaussian random distribution then the NTD follows the
Student's- t distribution. Some other form of transformations
is given in the Table 1.
The decision of correct detection of jamming is a classical
statistical problem of finding signal in the background of
random noise. The Kolmogorov-Smirnov (KS) statistic is
used to determine the underlying distribution pattern of NTD
To test the SRdr agent performance, 500 random numbers
were generated using the Gaussian random distribution with
mean and standard deviation equal to 20 and 10 respectively.
On the generated data, the NTD index is calculated. It is
found that after NTD transformation, the Gaussian random
number
statistical
distribution which is very similar to the Student's- t
distribution with parameter v = 2. If it is assumed that the nt
follows some other statistical distribution then the resulting
distributions of NTD would be of the form given in the Table
1.
into another
transforms
form of
The SRdr agent‟s algorithm is applied to the generated data.
The result of the simulation after applying the SRdr agent‟s
logic is shown in the Table 2. This table shows that out of
total 500 samples, 231 times (i.e. 41 %) the radar is found to
be jammed. This statistic is close to the Jamming Factor
(which was introduced in the simulation as random noise). It
was found that 91 times the radar is found "Switch Off". So it
has saved around 19% of the energy. Although the simulation
was started with Gaussian random data, after applying the
agent‟s logic the data transformed into statistical distribution
which is very similar to the Binomial distribution (because
the transformed data, the decisions, were either "Switch On"
or "Switch Off"). The closeness of this distribution is
measured by the KS statistic as shown in the Figure 9. This
statistic is used for performance measure of this agent. This
performance measure could be used to add the learning
capability in the agents.
For the LCCC agent, experiments have been performed with
three clusters, three types of cluster size (i.e. large, medium
and small) and two types of mission objectives (i.e. strike and
escort). So, total eighteen (3 3 2) possibilities of plans
instances are generated. Therefore, search space consists of
eighteen combinations. Hence, it is obvious with the increase
JOURNAL OF BATTLEFIELD TECHNOLOGY VOL 14, NO 1, MARCH 2011
7
Goal : K : if to achieve goal k for given belief then perform the plan .
Goal Inference Rule (gir):
{Jammed} ,{Frequency Hopping}k- Switch Off
{Jammed} ,{Switch Off}k- Frequency Hopping
{Frequency Hopping}k+ Sense Mode
{Switch Off}k+ Sleep Mode
Tautology : Frequency Hopping, Switch Off / Switch Off,
Tautology : Switch Off, Frequency Hopping / Frequency Hopping,
Frequency Hopping: Sense Mode / Sense Mode,
Switch Off: Sleep Mode / Sleep Mode.
Extension : {Switch Off, Sleep Mode}, {Frequency Hopping, Sense Mode}
Goal : G{Switch Off Sleep Mode}, G{Frequency Hopping Sense Mode}
Conflicting Goal :
G{Switch Off Frequency Hopping},
G{Sense Mode Sleep Mode},
G{Switch Off Sense Mode},
G{Frequency Hopping Sleep Mode}
Figure 8. Goal inference rule (gir) [12] for logical verification of surveillance radar agent.
G stands for goal.
of search space computation time taken by the agent
decreases. Similarly, the number of VAVP points also
influences the agent‟s performance. Although in this study
only the VAVP points are considered in the agent‟s beliefs,
the number of interceptor also influences the agent‟s
performance, therefore, can be included in the beliefset. In
general, the number of domains of the input parameters
determines the performance of the agents. Search space will
increase multiplicatively with the increase of the domain size.
Computation time to take an optimal decision using MLPR is
effected by these factors.
from
input
influences
the MSDF module
The
the
computation time required by the LCCC agent. It is found
that the number of cluster has direct influences on the
performance of the LCCC agents. In this study, only three
clusters are considered. How the agent will perform with
many clusters have not been studied. Similarly when the
number of mission type and cluster size changes it directly
influences the ranking of the plan instance.
DISCUSSION AND CONCLUSIONS
like
entities
aircrafts,
attacking
The SRdr and LCCC agents are deployed in a simulated
environment of air combat. The simulation is designed with
several
defending
interceptors, surveillance radars, air to air and surface to air
missiles, tracking radars. The simulated environment is
created by Java Netbeans IDE [18]. The agents are
programmed by the JACK agent programming language [17].
JACK supports BDI architectures and MLPR. The data
generated by this simulation are stored in the Oracle [19]
database. The output actually contains the information about
the states conditions of the environment and agents. The
agents analyse the environment and write their decisions
again in this database. The initial belief of the SRdr agent is
that no jamming has occurred. Over the time this agent keeps
on reading data from the database and adds it to its belief set
which automatically posts an event if it is greater than 0.5 (a
threshold value decided by the experts) as an indication of
noise jamming. The MSDF module collaboratively assess the
data obtain by different sensors and writes in the database.
The LCCC agent receives inputs from the MSDF output
through this database and decides accordingly. Based on
these decisions, resources are allocated to the attacking
aircrafts.
On each run of the simulation state situation of both the
environment and the agents are observed. The agents are
programmed such a way that these can automatically detect
any conflicting goal situation. For example, the SRdr agent
checks its present states and if it finds any conflicting states
situation as shown in the Figure 8 it throws an exception. In
this way the agent model is validated logically. For statistical
performance evaluation, the KS statistic is used. The KS
statistic with a lower value is always preferable.
is
in
this work
The main point emphasized
the
implementation details. Also, the MLPR is introduced so
agents can choose the right plan for the plan-repository using
a prioritizing mechanism. Such reasoning and the way it is
implemented can be used for many different application
domains. Main contribution of the paper is combining agent-
oriented programming with a Java based simulation
environment and implementing this for an IAD domain. Main
focus of the work is on how it is done and a significant effort
has been put in implementing these ideas. Given the level of
details it is certain that an advanced system may be
developed for further research is this field.
The present approaches of design, implementation and testing
of agent based system are found to be more suitable for
hierarchical structure of C2 that works on the principle of
practical reasoning. The way BDI architectures are used for
developing the C2 agents can be extended to build higher
order team agents. This could be a general frame work for
implementing decision making processes in an integrated
mode. The traditional optimization technique used for TA for
air defense system can be brought into this framework very
easily. This is an integrated approach of decision making for
selecting the optimal plan satisfying the agent‟s beliefs to
8
DAS & MUKHERJEE: AGENT BASED DECISION MAKING FOR AIR DEFENSE SYSTEM
No. of Targets Detected
Normalized Target
with parameters
Difference
Normal (20,10)
Student‟s t (2)
Triangular (20,10,30)
Gamma (=12.06, =0.08)
Uniform (10,30)
Gamma (=4.90,=0.22)
Exponential (10,20)
Laplace (=185.71,=1.0)
Table 1. Distribution pattern of Normalized Target
the
of
Difference
as
transformation
numbers of targets detected (nt).
from
distribution
obtained
of
Events
Number/ Percentage
Total Samples
500
Jamming
Frequency Hopping
Switch Off
231(46.2%)
134(26.8%)
97(19.4%)
Table 2. Simulation result of the surveillance radar
agent.
achieve desired goals. Usual methods of decision making do
not integrate the decision maker‟s beliefs and desires in direct
way although these components are essential attributes. This
approach is more suitable for futuristic network centric
warfare. The approach
terms of both
implementation (MLPR) and validation (logical as well as
statistical).
is novel
in
REFERENCES
[1] http://www.mindef.gov.sg/content/dam/imindef_media_l
ibrary/photos/scme/0015.res.
[2] http://www.dodccrp.org/events/12th_ICCRTS/Papers/01
0.pdf.
[3] G. Weiss, "Multiagent systems, a modern approach to
distributed modern approach to artificial intelligence,"
The MIT press, Cambridge, Massachusetts, 1999, pages-
585.
[4] R. H. Bordini, M. Dastani, J. Dix, A. E.F. Seghrouchni,
Multi-agent programming, languages, platforms and
applications, Springer, pages 313, 2005.
[5] A. Rogers et al., "Agent technologies for sensor
networks," IEEE Intelligent Systems, pp. 13-16, April,
2009.
[6] A. Pellecchia, C. Igel, J. Edelbrunner and G. Schoner,
"Making driver modeling attractive," IEEE Intelligent
Systems, pp. 8-12, 2005.
[7] M. Sensoy and P. Yolum, "Evolving service semantics
cooperatively: a consumer -driven approach", Auton.
Agent Multi-Agent Syst. 18, pp. 417-470, 2009.
[8] L. C. Bazzan, "Opportunities for multiagent systems and
multiagent reinforcement learning in traffic control",
Auton. Agent Multi-Agent Syst. 18, pp. 342-375, 2009.
[9] B. Subagdja, l. Sonenberg and I. Rahwan, "Intentional
learning agent architecture," Auton. Agent Multi-Agent
Syst. 18, pp. 417-470, 2009.
[10] S. Koenig, X. Sun, "Comparing
real-time and
incremental heuristic search for real time situated
agents," Auton. Agent Multi-Agent Syst., 18, pp. 313-341,
2009.
Figure 9. Kolmogorov Smirnov
statistics
agent's performance measurement.
for
[11] K. V. Hindriks, J.J. C. Meyer, "Toward a programming
theory for rational agents," Auton. Agent Multi-Agent
Syst. 19, pp. 4-29, 2009.
[12] M. Birna van Riemsdijk and M. Dastani, "Goals in
conflict: semantic foundations of goals
in agent
programming," Auton. Agent Multi-Agent Syst., 18, pp.
471-500, 2009.
[13] S. Karim, C. Heinze and S. Dunn, "Agent-based mission
management for a UAV", IEEE-ISSNIP 2004, pp. 481-
486, 2004.
[14] D. S. Alberts, J.J. Garstka and F. P. Stein, "Network
centric warfare, developing and leveraging information
superiority," 2nd Edn. (Revised), CCRP publication
series, 287 pages, (1999).
[15] L. Adelman, M. Christian, J. Gualtieri & T.A. Bresnick,
„Examining the effects of communication training and
team composition on the decision making of patriot air
defense teams‟, IEEE Trans Sys. Man Cyb. A28, 729 -- 41,
(1998).
[16] T. Balch. The impact of diversity on performance in
multi-robot foraging. In Proceedings of the third annual
conference on Autonomous Agents, pages 92 -- 99. ACM
Press, 1999.
[17] JACK®, Development Environment. Agent Oriented
Software Pty. Ltd., P.O. Box 639, Carlton South,
Victoria 3053, Australia..
[18] http://netbeans.org/
[19] http://www.oracle.com/in/index.html
[20] Q. Changwen and H. You, "A method of threat
assessment using multiple attribute decision making,",
IEEE, 2002 6th International Conference on Signal
Processing, vol.2, pp. 1091- 1095, Aug. 2002.
Dr Sumanta Kumar Das and Mr. Sumant Mukherjee are scientists in
the Institute for Systems Studies and Analyses, Defence Research
and Development Organization, Ministry of Defence, Delhi, India.
Both of them are working together for developing the agent based
models for integrated air defense system especially for network
centric warfare. E-mail: [email protected].
|
1412.1468 | 2 | 1412 | 2015-06-18T07:20:45 | Information-Sharing over Adaptive Networks with Self-interested Agents | [
"cs.MA"
] | We examine the behavior of multi-agent networks where information-sharing is subject to a positive communications cost over the edges linking the agents. We consider a general mean-square-error formulation where all agents are interested in estimating the same target vector. We first show that, in the absence of any incentives to cooperate, the optimal strategy for the agents is to behave in a selfish manner with each agent seeking the optimal solution independently of the other agents. Pareto inefficiency arises as a result of the fact that agents are not using historical data to predict the behavior of their neighbors and to know whether they will reciprocate and participate in sharing information. Motivated by this observation, we develop a reputation protocol to summarize the opponent's past actions into a reputation score, which can then be used to form a belief about the opponent's subsequent actions. The reputation protocol entices agents to cooperate and turns their optimal strategy into an action-choosing strategy that enhances the overall social benefit of the network. In particular, we show that when the communications cost becomes large, the expected social benefit of the proposed protocol outperforms the social benefit that is obtained by cooperative agents that always share data. We perform a detailed mean-square-error analysis of the evolution of the network over three domains: far field, near-field, and middle-field, and show that the network behavior is stable for sufficiently small step-sizes. The various theoretical results are illustrated by numerical simulations. | cs.MA | cs | Information-Sharing over Adaptive Networks with
Self-interested Agents
Chung-Kai Yu, Student Member, IEEE, Mihaela van der Schaar, Fellow, IEEE, and Ali H. Sayed, Fellow, IEEE
1
5
1
0
2
n
u
J
8
1
]
A
M
.
s
c
[
2
v
8
6
4
1
.
2
1
4
1
:
v
i
X
r
a
Abstract -- We examine the behavior of multi-agent networks
where information-sharing is subject to a positive communi-
cations cost over the edges linking the agents. We consider
a general mean-square-error formulation where all agents are
interested in estimating the same target vector. We first show
that, in the absence of any incentives to cooperate, the optimal
strategy for the agents is to behave in a selfish manner with
each agent seeking the optimal solution independently of the
other agents. Pareto inefficiency arises as a result of the fact
that agents are not using historical data to predict the behavior
of their neighbors and to know whether they will reciprocate and
participate in sharing information. Motivated by this observation,
we develop a reputation protocol to summarize the opponent's
past actions into a reputation score, which can then be used
to form a belief about the opponent's subsequent actions. The
reputation protocol entices agents to cooperate and turns their
optimal strategy into an action-choosing strategy that enhances
the overall social benefit of the network. In particular, we show
that when the communications cost becomes large, the expected
social benefit of the proposed protocol outperforms the social
benefit that is obtained by cooperative agents that always share
data. We perform a detailed mean-square-error analysis of the
evolution of the network over three domains: far field, near-field,
and middle-field, and show that the network behavior is stable
for sufficiently small step-sizes. The various theoretical results
are illustrated by numerical simulations.
Index Terms -- Adaptive networks, self-interested agents, repu-
tation design, diffusion strategy, Pareto efficiency, mean-square-
error analysis.
I.
INTRODUCTION
A DAPTIVE networks enable agents to share information
and to solve distributed optimization and inference tasks
in an efficient and decentralized manner. In most prior works,
agents are assumed to be cooperative and designed to fol-
low certain distributed rules such as the consensus strategy
(e.g., [2] -- [10]) or the diffusion strategy (e.g., [11] -- [19]). These
rules generally include a self-learning step to update the
agents' estimates using their local data, and a social-learning
step to fuse and combine the estimates shared by neighboring
agents. However, when agents are selfish, they would not obey
the preset rules unless these strategies conform to their own
interests, such as minimizing their own costs. In this work,
Copyright (c) 2015 IEEE. Personal use of this material
is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to [email protected].
This work was supported in part by NSF grants CCF-1011918, CSR-
1016081, and ECCS-1407712. An early short version of this work appeared
in the conference publication [1]. The authors are with the Department of
Electrical Engineering, University of California, Los Angeles, CA 90095-1594
USA (e-mail: [email protected], {mihaela,sayed}@ee.ucla.edu).
we assume that the agents can behave selfishly and that they,
therefore, have the freedom to decide whether or not they want
to participate in sharing information with their neighbors at
every point in time. Under these conditions, the global social
benefit for the network can be degraded unless a policy is
introduced to entice agents to participate in the collaborative
process despite their individual interests. In this article, we
will address this difficulty in the context of adaptive networks
where agents are continually subjected to streaming data, and
where they can predict in real-time, from their successive
interactions, how reliable their neighbors are and whether they
can be trusted to share information based on their past history.
This formulation is different from the useful work in [20],
which considered one particular form of selfish behavior in
the context of a game-theoretic formulation. In that work, the
focus is on activating the self-learning and social learning steps
simultaneously, and agents simply decide whether to enter into
a sleep mode (to save energy) or to continue acquiring and
processing data. In the framework considered in our work,
agents always remain active and are continually acquiring data;
the main question instead is to entice agents to participate
in the collaborative information-sharing process regardless of
their self-centered evaluations.
More specifically, we study the behavior of multi-agent
networks where information-sharing is subject to a positive
communication cost over the edges linking the agents. This
situation is common in applications, such as information
sharing over cognitive networks [21], online learning under
communication bandwidth and/or latency constraints
[22],
[23, Ch. 14], and over social learning networks when the
delivery of opinions involves some costs such as messaging
fees [24] -- [26]. In our network model, each agent is self-
interested and seeks to minimize its own sharing cost and
estimation error. Motivated by the practical scenario studied
in [21], we formulate a general mean-square error estimation
problem where all agents are interested in estimating the same
target parameter vector. Agents are assumed to be foresighted
and to have bounded rationality [27] in the manner defined
further ahead in the article. Then, we show that if left unat-
tended, the dominant strategy for all agents is for them not
to participate in the sharing of information, which leads to
networks operating under an inefficient Pareto condition. This
situation arises because agents do not have enough information
to tell beforehand if their paired neighbors will reciprocate
their actions (i.e., if an agent shares data with a second agent,
will the second agent reciprocate and share data back?) This
prediction-deficiency problem follows from the fact that agents
are not using historical data to predict other agents' actions.
One method to deal with this inefficient scenario is to
assume that agents adapt to their opponents' strategies and
improve returns by forming some regret measures. In [28], a
decision maker determines its action using a regret measure to
evaluate the utility loss from the chosen action to the optimal
action in the previous stage game. For multi-agent networks,
a regret-based algorithm was proposed in [20] and [29] for
agents to update their actions based on a weighted loss of the
utility functions from the previous stage games. However, these
works assume myopic agents and formulate repeated games
with fixed utility functions over each stage game, which is
different from the scenario considered in this article where
the benefit of sharing information over adaptive networks con-
tinually evolves over time. This is because, as the estimation
accuracy improves and/or the communication cost becomes
expensive, the return to continue cooperating for estimation
purposes falls and thus the act of cooperating with other agents
becomes unattractive and inefficient. In this case, the regret
measures computed from the previous stage games may not
provide an accurate reference to the current stage game.
In the performance evaluation, we are interested in ensuring
A second useful method to deal with Pareto inefficient and
non-cooperative scenarios is to employ reputation schemes
(e.g., [30] -- [33]). In this method, foresighted agents use rep-
utation scores to assess the willingness of other agents to
cooperate; the scores are also used to punish non-cooperative
behavior. For example, the works [31], [32] rely on discrete-
value reputation scores, say, on a scale 1-10, and these scores
are regularly updated according to the agents' actions. Similar
to the regret
in our
problem the utilities or cost functions of stage games change
over time and evolve based on agents' estimates. Conventional
reputation designs do not address this time variation within
the payoff of agents, which will be examined more closely in
our work. Motivated by these considerations, in Sec. IV, we
propose a dynamic/adaptive reputation protocol that is based
on the belief measure of future actions with real-time benefit
predictions.
learning references mentioned before,
In our formulation, we assume a general random-pairing
model similar to [10], where agents are randomly paired at the
beginning of each time interval. This situation could occur, for
example, due to an exogenous matcher or the mobility of the
agents. The paired agents are assumed to follow a diffusion
strategy [12] -- [15], which includes an adaptation step and a
consultation step, to iteratively update their estimates. Different
from conventional diffusion strategies, the consultation step
here is influenced by the random-pairing environment and by
cooperation uncertainty. The interactions among self-interested
agents are formulated as successive stage games of two players
using pure strategies. To motivate agents to cooperate with
each other, we formulate an adaptive reputation protocol to
help agents jointly assess the instantaneous benefit of de-
preciating information and the transmission cost of sharing
information. The reputation score helps agents to form a belief
of their opponent's subsequent actions. Based on this belief,
we entice agents to cooperate and turn their best response
strategy into an action choosing strategy that conforms to
Pareto efficiency and enhances the overall social benefit of
the network.
2
the mean-square-error stability of the network instead of exam-
ining equilibria as is common in the game theoretical literature
since our emphasis is on adaptation under successive time-
variant stage games. The performance analysis is challenging
due to the adaptive behavior by the agents. For this reason,
we pursue the mean-square-error analysis of the evolution
of the network over three domains: far-field, near-field, and
middle-field, and show that the network behavior is stable
for sufficiently small step-sizes. We also show that when
information sharing becomes costly, the expected social benefit
of the proposed reputation protocol outperforms the social
benefit that is obtained by cooperative agents that always share
data.
Notation: We use lowercase letters to denote vectors and
scalars, uppercase letters for matrices, plain letters for deter-
ministic variables, and boldface letters for random variables.
All vectors in our treatment are column vectors, with the
exception of the regression vectors, uk,i.
II. SYSTEM MODEL
A. Distributed Optimization and Communication Cost
Consider a connected network consisting of N agents.
When agents act independently of each other, each agent k
would seek to estimate the M × 1 vector wo that minimizes
an individual estimation cost function denoted by J est
k (w) :
CM → R. We assume each of the costs {J est
k (w)} is strongly
convex for k = 1, 2, . . . , N, and that all agents have the
same objective so that all costs are minimized at the common
location wo ∈ CM×1.
In this work, we are interested in scenarios where agents can
be motivated to cooperate among themselves as permitted by
the network topology. We associate an extended cost function
with each agent k, and denote it by Jk(w, ak). In this new
cost, the scalar ak is a binary variable that is used to model
whether agent k is willing to cooperate and share information
with its neighbors. The value ak = 1 means that agent k is
willing to share information (e.g., its estimate of wo) with
its neighbors, while the value ak = 0 means that agent k is
not willing to share information. The reason why agents may
or may not share information is because this decision will
generally entail some cost. We consider the scenario where a
positive transmission cost, ck > 0, is required for each act by
agent k involving sharing an estimate with any of its neighbors.
By taking ck into consideration, the extended cost Jk(w, a)
that is now associated with agent k will consist of the sum of
two components: the estimation cost and the communication
cost1:
Jk(w, ak) (cid:44) J est
k (w) + J com
where the latter component is modeled as
k
J com
k
(ak) (cid:44) akck
(ak)
(1)
(2)
1We focus on the sum of the estimation cost and the communication cost
due to its simplicity and meaningfulness in applications. Note that a possible
generalization is to consider a penalty-based objective function J est
k (w) +
p(J com
k (ak)) for some penalty function p(·).
We express the communication expense in the form (2) be-
cause, as described further ahead, when an agent k decides
to share information, it will be sharing the information with
one neighbor at a time; the cost for this communication will be
akck. With regards to the estimation cost, J est
k (w), this measure
can be selected in many ways. One common choice is the
mean-square-error (MSE) cost, which we adopt in this work.
At each time instant i ≥ 0, each agent k is assumed to
have access to a scalar measurement dk(i) ∈ C and a 1 ×
M regression vector uk,i ∈ C1×M with covariance matrix
Ru,k (cid:44) Eu∗
k,iuk,i > 0. The regressors {uk,i} are assumed
to have zero-mean and to be temporally white and spatially
independent, i.e.,
Eu∗
k,iu(cid:96),j = Ru,kδk(cid:96)δij
the Kronecker delta
(3)
in terms of
function. The data
{dk(i), uk,i} are assumed to be related via the linear regres-
sion model:
(4)
where wo is the common target vector to be estimated by the
agents. In (4), the variable vk(i) ∈ C is a zero-mean white-
v,k that is assumed to be spatially
noise process with power σ2
independent, i.e.,
dk(i) = uk,iwo + vk(i)
Ev∗
v,kδk(cid:96)δij
k(i)v(cid:96)(j) = σ2
(5)
We further assume that the random processes uk,i and v(cid:96)(i)
are spatially and temporally independent for any k, (cid:96), i, and
j. Models of the form (4) are common in many applications,
e.g., channel estimation, model fitting, target tracking, etc (see,
e.g., [15]).
Let wk,i−1 denote the estimator for wo that will be available
to agent k at time i − 1. We will describe in the sequel how
agents evaluate these estimates. The corresponding a-priori
estimation error is defined by
in terms of the estimation error vector
(6)
and it measures how close the weight estimate matches the
measurements {dk(i), uk,i} to each other. In view of model
(4), we can also write
(7)
ea,k(i) (cid:44) dk(i) − uk,iwk,i−1
ea,k(i) = uk,i(cid:101)wk,i−1 + vk(i)
(cid:101)wk,i−1 (cid:44) wo − wk,i−1
(8)
Motivated by these expressions and model (4), the instanta-
neous MSE cost that is associated with agent k based on the
estimate from time i − 1 is given by
k (wk,i−1) (cid:44) Eea,k(i)2
J est
= E(cid:107)(cid:101)wk,i−1(cid:107)2
= Edk(i) − uk,iwk,i−12
+ σ2
(9)
Note that this MSE cost conforms to the strong convexity of
J est
as we mentioned before. Combined with the action by
k
agent k, the extended instantaneous cost at agent k that is
based on the prior estimate, wk,i−1, is then given by:
Ru,k
v,k
Jk(wk,i−1, ak) = Eea,k(i)2 + akck
We will not be assuming fully-connected networks or fully-
paired protocols and will deal more generally with networks
that can be sparsely connected. Later in Sec. IV we will
(10)
B. Random-Pairing Model
3
We denote by Nk the neighborhood of each agent k, includ-
ing itself. We consider a random pairing protocol for agents
to share information at the beginning of every iteration cycle.
The pairing procedure can be executed either in a centralized
or distributed manner. Centralized pairing schemes can be used
when an online server randomly assigns its clients into pairs
as in crowdsourcing applications [31], [32], or when a base-
station makes pairing decisions for its mobile nodes for packet
relaying [34]. Distributed paring schemes arise more naturally
in the context of economic and market transactions [35]. In our
formulation, we adopt a distributed pairing structure that takes
neighborhoods into account when selecting pairs, as explained
next.
We assume each agent k has bi-directional links to other
agents in Nk and that agent k has a positive probability to be
paired with any of its neighbors. Once two agents are paired,
they can decide on whether to share or not their instantaneous
estimates for wo. We therefore model the result of the random-
pairing process between each pair of agents k and (cid:96) ∈ Nk\{k}
as temporally-independent Bernoulli random processes defined
as:
(cid:26)1, with probability pk(cid:96) = p(cid:96)k
1k(cid:96)(i) = 1(cid:96)k(i) =
(11)
0, otherwise
where 1k(cid:96)(i) = 1 indicates that agents k and (cid:96) are paired at
time i and 1k(cid:96)(i) = 0 indicates that they are not paired. We are
setting 1k(cid:96)(i) = 1(cid:96)k(i) because these variables represent the
same event: whether agents k and (cid:96) are paired, which results
in pk(cid:96) = p(cid:96)k. For (cid:96) /∈ Nk, we have 1k(cid:96)(i) = 0 since such pairs
will never occur. For convenience, we use 1kk(i) to indicate
the event that agent k is not paired with any agent (cid:96) ∈ Nk \
{k} at time i, which happens with probability pkk. Since each
agent will pair itself with at most one agent at a time from its
neighborhood, the following properties are directly obtained
from the random-pairing procedure:
pk(cid:96) = 1
(12)
(cid:88)
1k(cid:96)(i) = 1,
(cid:96)∈Nk
1k(cid:96)(i)1kq(i) = 0,
(cid:88)
(cid:96)∈Nk
for (cid:96) (cid:54)= q
(13)
We assume that the random pairing indicators {1k(cid:96)(i)} for all
k and (cid:96) are independent of the random variables {uk,t} and
{vk(t)} for any time i and t. For example, a widely used
setting in the literature is the fully-pairing network, which
assumes a fully-connected network topology [32], [36], i.e.,
Nk = N for every agent k, where N denotes the set of
all agents. The size N = N is assumed to be even and
every agent is uniformly paired with exactly one agent in the
network. Therefore, we have N/2 pairs at each time instant
and the random-pairing probability becomes
for (cid:96) (cid:54)= k
for (cid:96) = k
(cid:40) 1
N−1 ,
0,
pk(cid:96) =
(14)
demonstrate a simple random-pairing protocol which can be
implemented in a fully distributed manner.
agent k is paired with some neighboring agent (cid:96), which means
1k(cid:96)(i) = 1, we get
4
k,i[dk(i) − uk,iwk,i−1]
(15)
(16)
wk,i =
αkψk,i + (1 − αk)ψ(cid:96),i,
if a(cid:96)k(i) = 0
if a(cid:96)k(i) = 1
(21)
C. Diffusion Strategy
Conventional diffusion strategies assume that the agents are
cooperative (or obedient) and continuously share information
with their neighbors as necessary [12], [13], [15]. In the adapt-
then-combine (ATC) version of diffusion adaptation, each
agent k updates its estimate, wk,i, according to the following
relations:
(cid:88)
ψk,i = wk,i−1 + µu∗
α(cid:96)kψ(cid:96),i
wk,i =
(cid:96)∈Nk
where µ > 0 is the step-size parameter of agent k, and the
{α(cid:96)k, (cid:96) ∈ Nk} are nonnegative combination coefficients that
add up to one. In implementation (15) -- (16), each agent k
computes an intermediate estimate ψk,i using its local data,
and subsequently fuses the intermediate estimates from its
neighbors. For the combination step (16), since agent k is
allowed to interact with only one of its neighbors, then we
rewrite (16) in terms of a single coefficient 0 ≤ αk ≤ 1 as
follows:
if 1k(cid:96)(i) = 1 for some (cid:96) (cid:54)= k
otherwise
(17)
We can capture both situations in (17) in a single equation as
follows:
(cid:26)αkψk,i + (1 − αk)ψ(cid:96),i,
wk,i =
ψk,i,
wk,i = αkψk,i + (1 − αk)
1k(cid:96)(i)ψ(cid:96),i
(18)
(cid:88)
(cid:96)∈Nk
In formulation (15) and (18), it is assumed that once agents k
and (cid:96) are paired, they share information according to (18).
Let us now incorporate an additional layer into the algorithm
in order to model instances of selfish behavior. When agents
behave in a selfish (strategic) manner, even when agents k and
(cid:96) are paired, each one of them may still decide (independently)
to refuse to share information with the other agent for selfish
reasons (for example, agent k may decide that this cooperation
will cost more than the benefit it will reap for the estimation
task). To capture this behavior, we use the specific notation
ak(cid:96)(i),
the action taken by
agent k on agent (cid:96) at time i, and similarly for a(cid:96)k(i). Both
agents will end up sharing information with each other only if
ak(cid:96)(i) = a(cid:96)k(i) = 1, i.e., only when both agents are in favor
of cooperating once they have been paired. We set akk(i) = 1
for every time i. We can now rewrite the combination step
(18) more generally as:
instead of ak(i),
to represent
wk,i = αkψk,i + (1 − αk)
1k(cid:96)(i)[a(cid:96)k(i)ψ(cid:96),i+
(19)
From (19), when agent k is not paired with any agent at time
i (1kk(i) = 1), we get wk,i = ψk,i. On the other hand, when
(1 − a(cid:96)k(i)) ψk,i]
(cid:88)
(cid:96)∈Nk
wk,i = αkψk,i + (1 − αk) [a(cid:96)k(i)ψ(cid:96),i + (1 − a(cid:96)k(i)) ψk,i]
(20)
It is then clear that a(cid:96)k(i) = 0 results in wk,i = ψk,i, while
a(cid:96)k(i) = 1 results in a combination of the estimates of agents
k and (cid:96). In other words, when 1k(cid:96)(i) = 1:
(cid:26)ψk,i,
In the sequel, we assume that agents update and combine
their estimates using (15) and (19). One important question
to address is how the agents determine their actions {ak(cid:96)(i)}.
III. AGENT INTERACTIONS
When an arbitrary agent k needs to decide on whether to
set its action to ak(cid:96)(i) = 1 (i.e., to cooperate) or ak(cid:96)(i) =
0 (i.e., not to cooperate), it generally cannot tell beforehand
whether agent (cid:96) will reciprocate. In this section, we first show
that when self-interested agents are boundedly rational and
incapable of transforming the past actions of neighbors into a
prediction of their future actions, then the dominant strategy for
each agent will be to choose noncooperation. Consequently, the
entire network becomes noncooperative. Later, in Sec. IV, we
explain how to address this inefficient scenario by proposing
a protocol that will encourage cooperation.
A. Long-Term Discounted Cost Function
To begin with, let us examine the interaction between a
pair of agents, such as k and (cid:96), at some time instant i
(1k(cid:96)(i) = 1). We assume that agents k and (cid:96) simultaneously
select their actions ak(cid:96)(i) and a(cid:96)k(i) by using some pure
strategies (i.e., agents set their action variables by using data
or realizations that are available to them, such as the estimates
{wk,i−1, w(cid:96),i−1}, rather than select their actions according
to some probability distributions)2. The criterion for setting
ak(cid:96)(i) by agent k is to optimize agent k's payoff, which
incorporates both the estimation cost, affected by agent (cid:96)'s
own action a(cid:96)k(i), and the communication cost, determined
by agent k's action ak(cid:96)(i). Therefore, the instantaneous cost
incurred by agent k is a mapping function from the action
space (ak(cid:96)(i), a(cid:96)k(i)) to a real value. In order to account for
selfish behavior, we need to modify the notation used in (1)
to incorporate the actions of both agents k and (cid:96). In this way,
we need to denote the value of the cost incurred by agent k
at time i, after wk,i−1 is updated to wk,i, more explicitly by
2In our scenario, the discrete action set ak(cid:96)(i) ∈ {0, 1} will be shown to
lead to threshold-based pure strategies -- see Sec. IV-B.
=
Jk(ak(cid:96)(i), a(cid:96)k(i)) and it is given by:
Jk(ak(cid:96)(i), a(cid:96)k(i))
J est
k (wk,i = ψk,i),
J est
k (wk,i = αkψk,i + (1 − αk)ψ(cid:96),i),
J est
k (wk,i = ψk,i) + ck,
J est
k (wk,i = αkψk,i + (1 − αk)ψ(cid:96),i) + ck,
if (0, 0)
if (0, 1)
if (1, 0)
if (1, 1)
(22)
For example, the first line on the right-hand side of (22)
corresponds to the situation in which none of the agents
decides to cooperate. In that case, agent k can only rely on its
intermediate estimate, ψk,i, to improve its estimation accuracy.
In comparison, the second line in (22) corresponds to the
situation in which agent (cid:96) is willing to share its estimate but
not agent k. In this case, agent k is able to perform the second
combination step in (21) and enhance its estimation accuracy.
In the third line in (22), agent (cid:96) does not cooperate while agent
k does. In this case, agent k incurs a communication cost, ck.
Similarly, for the last line in (22), both agents cooperate. In
this case, agent k is able to perform the second step in (21)
while incurring a cost ck.
We can write (22) more compactly as follows:
Jk(ak(cid:96)(i), a(cid:96)k(i)) = J act
k (a(cid:96)k(i)) + ak(cid:96)(i)ck
(23)
where we introduced
J act
k (a(cid:96)k(i))
(cid:26)J est
(cid:44)
k (wk,i = ψk,i),
J est
k (wk,i = αkψk,i + (1 − αk)ψ(cid:96),i),
if a(cid:96)k(i) = 0
if a(cid:96)k(i) = 1
(24)
The function J act
k (a(cid:96)k(i)) helps make explicit the influence
of the action by agent (cid:96) on the estimation accuracy that is
ultimately attained by agent k.
Now, the random-pairing process occurs repeatedly over
time and, moreover, agents may leave the network. For this
reason, rather than rely on the instantaneous cost function in
(22), agent k will determine its action at time i by instead
minimizing an expected long-term discounted cost function of
the form defined by (25) where δk ∈ (0, 1) is a discount factor
to model future network uncertainties and the foresightedness
level of agent k. The expectation is taken over all randomness
for t ≥ i and is conditioned on the estimate wk,i−1 when
the actions ak(cid:96)(i) and a(cid:96)k(i) are selected. Formulation (25)
is meant to assess the influence of the action selected at time
i by agent k on its cumulative (but discounted) future costs.
More specifically, whenever 1k(cid:96)(i) = 1, agent k selects its
5
action ak(cid:96)(i) at time i to minimize the expected long-term
discounted cost given wk,i−1:
min
ak(cid:96)(i)∈{0,1} J
∞
k,i[ak(cid:96)(i), a(cid:96)k(i)wk,i−1]
(26)
Based on the payoff function in (25), we can formally regard
the interaction between agents as consisting of stage games
with recurrent random pairing. The stage information-sharing
game for 1k(cid:96)(i) = 1 is a tuple (N, A, J), where N (cid:44) {k, (cid:96)} is
the set of players, and A (cid:44) Ak × A(cid:96) is the Cartesian product
of binary sets Ak = A(cid:96) (cid:44) {1, 0} representing available
actions for agents k and (cid:96), respectively. The action profile is
a(i) (cid:44) (ak(cid:96)(i), a(cid:96)k(i)) ∈ A. Moreover, J = {J∞
(cid:96),i} is
the set of real-valued long-term costs defined over A → R
for agents k and (cid:96), respectively. We remark that since J∞
k,i
depends on wk,i−1, its value generally varies from stage to
stage. As a result, each agent k faces a dynamic game structure
with repeated interactions in contrast to conventional repeated
games as in [37], [38] where the game structure is fixed over
time. Time variation is an essential feature that arises when
we examine selfish behavior over adaptive networks.
k,i, J∞
Therefore, solving problem (26) involves the forecast of
future game structures and future actions chosen by the op-
ponent. These two factors are actually coupled and influence
each other; this fact makes prediction under such conditions
rather challenging. To continue with the analysis, we adopt
a common assumption from the literature that agents have
computational constraints. In particular, we assume the agents
have bounded rationality [27], [39], [40]. In our context, this
means that the agents have limited capability to forecast future
game structures and are therefore obliged to assume that future
parameters remain unchanged at current values. We will show
how this assumption enables each agent k to evaluate J∞
k,i in
later discussions.
Assumption 1 (Bounded rationality): Every agent k solves
the optimization problem (26) under the assumptions:
for t ≥ i
wk,t = wk,i−1, 1k(cid:96)(t) = 1k(cid:96)(i),
(27)
(cid:4)
We note that the above assumption is only made by the agent
at time i while solving problem (26); the actual estimates wk,t
and pairing choices 1k(cid:96)(t) will continue to evolve over time.
We further assume that the bounded rationality assumption is
common knowledge to all agents in the network3.
3Common knowledge of p means that each agent knows p, each agent
knows that all other agents know p, each agent knows that all other agents
know that all the agents know p, and so on [41].
J
∞
k,i [ak(cid:96)(i), a(cid:96)k(i)wk,i−1] (cid:44)
=
E(cid:104)
E(cid:104)
δt−i
k
δt−i
k
∞(cid:88)
∞(cid:88)
t=i
t=i
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = ak(cid:96)(i), a(cid:96)k(i) = a(cid:96)k(i)
(cid:105)
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = ak(cid:96)(i), a(cid:96)k(i) = a(cid:96)k(i)
(cid:105)
Jk(ak(cid:96)(t), a(cid:96)k(t))
J act
k (a(cid:96)k(t)) + ak(cid:96)(t)ck
(25)
TABLE I.
`````````
Agent k
Agent (cid:96)
a(cid:96)k(i) = 0
a(cid:96)k(i) = 1
THE EXPECTED LONG-TERM COST FUNCTIONS J 1
k,i AND J 1
(cid:96),i.
ak(cid:96)(i) = 0
ak(cid:96)(i) = 1
(cid:96) (ak(cid:96)(i) = 0)w(cid:96),i−1]
E[J act
k (a(cid:96)k(i) = 0)wk,i−1]
E[J act
(cid:96) (ak(cid:96)(i) = 0)w(cid:96),i−1] + c(cid:96)
k (a(cid:96)k(i) = 1)wk,i−1]
E[J act
E[J act
E[J act
E[J act
E[J act
E[J act
(cid:96) (ak(cid:96)(i) = 1)w(cid:96),i−1]
k (a(cid:96)k(i) = 0)wk,i−1] + ck
(cid:96) (ak(cid:96)(i) = 1)w(cid:96),i−1] + c(cid:96)
k (a(cid:96)k(i) = 1)wk,i−1] + ck
6
B. Pareto Inefficiency
In this section, we show that if no further measures are
taken, then Pareto inefficiency may occur. Thus, assume that
the agents are unable to store the history of their actions and
the actions of their neighbors. Each agent k only has access
to its immediate estimate wk,i−1, which can be interpreted as
a state variable at time i − 1 for agent k. In this case, each
agent k will need to solve (26) under Assumption 1. It then
follows that agent k will predict the same action for future
time instants:
ak(cid:96)(t) = ak(cid:96)(i),
for t > i
(28)
Furthermore, since the bounded rationality condition is com-
mon knowledge, agent k knows that the same future actions
are used by agent (cid:96), i.e.,
a(cid:96)k(t) = a(cid:96)k(i),
for t > i
(29)
However, this resulting action profile will be Pareto ineffi-
cient for both agents if it can be verified that the alternative
action profile (1, 1), where both agents cooperate, can lead to
improved payoff values for both agents in comparison to the
strategy (0, 0). To characterize when this is possible, let us
denote the expected payoff for agent k when agent (cid:96) selects
a(cid:96)k(i) = 0 by
k,i(ak(cid:96)(i)) (cid:44) E[J act
s0
k (a(cid:96)k(i) = 0)wk,i−1] + ak(cid:96)(i)ck
(33)
Likewise, when a(cid:96)k(i) = 1, we denote the expected payoff for
agent k by
k,i(ak(cid:96)(i)) (cid:44) E[J act
s1
k (a(cid:96)k(i) = 1)wk,i−1] + ak(cid:96)(i)ck
(34)
The benefit for agent k from agent (cid:96)'s sharing action, defined
k,i(ak(cid:96)(i)), is seen
as the improvement from s0
to be independent of ak(cid:96)(i):
k,i(ak(cid:96)(i)) to s1
(cid:1)
(30)
Using (28) and (29), agent k obtains
∞
k,i [ak(cid:96)(i), a(cid:96)k(i)wk,i−1]
J
(cid:12)(cid:12)(cid:12)wk,i−1
(cid:105)
(cid:12)(cid:12)(cid:12)wk,i−1
(cid:105)
Jk(ak(cid:96)(i), a(cid:96)k(i))
Jk(ak(cid:96)(i), a(cid:96)k(i))
k
δt−i
E(cid:104)
∞(cid:88)
1 − δk · E(cid:104)
(cid:0)E[J act
t=i
1
1
1 − δk
=
=
=
k (a(cid:96)k(i))wk,i−1] + ak(cid:96)(i)ck
Therefore, the optimization problem (26) reduces to the fol-
lowing minimization problem:
ak(cid:96)(i)∈{0,1} J 1
k,i (ak(cid:96)(i), a(cid:96)k(i))
(31)
min
where
k,i (ak(cid:96)(i), a(cid:96)k(i)) (cid:44) E[J act
k (a(cid:96)k(i))wk,i−1] + ak(cid:96)(i)ck (32)
J 1
is the expected cost of agent k given wk,i−1 -- compare
with (23). Table I summarizes the values of J 1
(cid:96),i for
both agents under their respective actions. From the entries
in the table, we conclude that choosing action ak(cid:96)(i) = 0
is the dominant strategy for agent k regardless of the action
chosen by agent (cid:96) because its cost will be the smallest it can
be in that situation. Likewise, the dominant strategy for agent
(cid:96) is a(cid:96)k(i) = 0 regardless of the action chosen by agent k.
Therefore, the action profile (ak(cid:96)(i), a(cid:96)k(i)) = (0, 0) is the
unique outcome as a Nash and dominant strategy equilibrium
for every stage game.
k,i and J 1
k (a(cid:96)k(i) = 1)wk,i−1]
(cid:3) (35)
bk(i)
(cid:44) s0
= E[J act
(cid:3)
so that
E[J act
k,i(ak(cid:96)(i))
k (wk,i)wk,i−1
Now, note from definition (6) that
k,i(ak(cid:96)(i)) − s1
k (a(cid:96)k(i) = 0)wk,i−1] − E[J act
k (wk,i = ψk,i)wk,i−1
k (wk,i = αkψk,i + (1 − αk)ψ(cid:96),i)wk,i−1
(cid:3)
(cid:105)
= E(cid:2)J est
− E(cid:2)J est
E(cid:2)J est
dk(i + 1) − uk,i+1wk,i2(cid:12)(cid:12)(cid:12)wk,i−1
= E(cid:104)
= E(cid:2)J est
(cid:3)
dk(i + 1) − uk,i+1ψk,i2(cid:12)(cid:12)(cid:12)wk,i−1
= E(cid:104)
uk,i+1(cid:101)ψk,i + vk(i + 1)2(cid:12)(cid:12)(cid:12)wk,i−1
= E(cid:104)
(cid:12)(cid:12)(cid:12)wk,i−1
= E(cid:104)
(cid:105)
(cid:107)(cid:101)ψk,i(cid:107)2
where (cid:101)ψk,i (cid:44) wo − ψk,i and, similarly,
= E(cid:2)J est
= E(cid:104)
(cid:107)αk(cid:101)ψk,i + (1 − αk)(cid:101)ψ(cid:96),i(cid:107)2
k (wk,i = αkψk,i + (1 − αk)ψ(cid:96),i)wk,i−1
+ σ2
k (wk,i = ψk,i)wk,i−1
k (a(cid:96)k(i) = 0)wk,i−1]
k (a(cid:96)k(i) = 1)wk,i−1]
(cid:12)(cid:12)(cid:12)wk,i−1
E[J act
+ σ2
v,k
(cid:105)
(cid:105)
(cid:105)
Ru,k
Ru,k
(36)
(37)
(cid:3)
v,k (38)
7
(a) γk,i > 1 and γ(cid:96),i > 1
(b) γk,i < 1 and γ(cid:96),i < 1
Fig. 1.
respectively).
Illustration of the behavior of the payoffs in terms of the size of the benefit-cost ratios ("S" and "NS" refer to the actions "share" and "do not share",
Then, the benefit bk(i) becomes
bk(i) = E(cid:104)
(cid:107)(cid:101)ψk,i(cid:107)2
− E(cid:104)
(cid:107)αk(cid:101)ψk,i + (1 − αk)(cid:101)ψ(cid:96),i(cid:107)2
(cid:12)(cid:12)(cid:12)wk,i−1
(cid:105)
Ru,k
(cid:105)
(cid:12)(cid:12)(cid:12)wk,i−1
(39)
Ru,k
Note that bk(i) is determined by the variable wk,i−1 and does
not depend on the actions a(cid:96)k(i) and ak(cid:96)(i). We will explain
how agents assess the information bk(i) to choose actions
further ahead in Sec. IV-C. Now, let us define the benefit-cost
ratio as the ratio of the estimation benefit to the communication
cost:
γk(i) (cid:44) bk(i)
ck
(40)
Then, the action profile (1, 1) in the game defined in Table I
is Pareto superior to the action profile (0, 0) when both of the
following two conditions hold
(cid:26)ck < bk(i)
c(cid:96) < b(cid:96)(i)
γk(i) > 1 and γ(cid:96)(i) > 1 ⇔
(41)
On the other hand, the action profile (0, 0) is Pareto superior
to the action profile (1, 1) if, and only if,
γk(i) < 1 and γ(cid:96)(i) < 1
(42)
In Fig. 1(a), we illustrate how the values of the payoffs com-
pare to each other when (41) holds for the four possibilities of
action profiles. It is seen from this figure that when γk(i) > 1
and γ(cid:96)(i) > 1, the action profile (S,S), i.e., (1, 1) in (32), is
Pareto optimal and that the dominant strategy (NS,NS), i.e.,
(0, 0) in (32), is inefficient and leads to worse performance
(which is a manifestation of the famous prisoner's dilemma
problem [42]). On the other hand, if γk(i) < 1 and γ(cid:96)(i) < 1,
then we are led to Fig. 1(b), where the action profile (NS,NS)
becomes Pareto optimal and superior to (S,S). We remark that
(NS,S) and (S,NS) are also Pareto optimal in both cases but
not preferred in this work because they are only beneficial for
one single agent.
IV. ADAPTIVE REPUTATION PROTOCOL DESIGN
As shown above, when both γk(i) > 1 and γ(cid:96)(i) > 1,
the Pareto optimal strategies for agents k and (cid:96) correspond
to cooperation; when both γk(i) < 1 and γ(cid:96)(i) < 1, the
Pareto optimal strategies for agents k and (cid:96) reduce to non-
cooperation. Since agents are self-interested and boundedly
rational, we showed earlier that if left without incentives, their
dominant strategy is to avoid sharing information because they
cannot tell beforehand if their paired neighbor will reciprocate.
This Pareto inefficiency therefore arises from the fact that
agents are not using historical data to predict other agents'
actions. We now propose a reputation protocol to summarize
the opponent's past actions into a reputation score. The score
will help agents to form a belief of their opponent's subsequent
actions. Based on this belief, we will be able to provide
agents with a measure that entices them to cooperate. We
will show, for example, that the best response rule for agents
will be to cooperate whenever γk(i) is large and not
to
cooperate whenever γk(i) is small, in conformity with the
Pareto-efficient design.
A. Reputation Protocol
Reputation scores have been used before in the literature as
a mechanism to encourage cooperation [32], [43], [44]. Agents
that cooperate are rewarded with higher scores; agents that do
not cooperate are penalized with lower scores. For example,
eBay uses a cumulative score mechanism, which simply sums
the sellers feedback scores from all previous periods to pro-
vide buyers and sellers with trust evaluation [45]. Likewise,
Amazon.com implements a reputation system by using an
average score mechanism that averages the feedback scores
from the previous periods [46]. However, as already explained
c(i)k(i)cbb(NS,NS)(NS,S)(S,NS)(S,S)kJ1 J1k iJ1 J1k c(i)k(i)cbb(NS,NS)(NS,S)(S,NS)(S,S)ki in [44], cheating can occur over time in both cumulative and
average score mechanisms because past scores carry a large
weight in determining the current reputation. To overcome this
problem, and in a manner similar to exponential weighting in
adaptive filter designs [47], an exponentially-weighted moving
average mechanism that gives higher weights to more recent
actions is discussed in [44]. We follow a similar weighting
formulation, with the main difference being that the reputation
scores now need to be adapted in response to the evolution of
the estimation task over the network. The construction can be
described as follows.
When 1k(cid:96)(i) = 1, meaning that agent k is paired with agent
(cid:96), the reputation score θ(cid:96)k(i) ∈ [0, 1] that is maintained by
agent k for its neighbor (cid:96) is updated as:
(43)
where rk ∈ (0, 1) is a smoothing factor for agent k to control
the dynamics of the reputation updates. On the other hand, if
1k(cid:96)(i) = 0, the reputation score θ(cid:96)k(i + 1) remains as θ(cid:96)k(i).
We can compactly describe the reputation rule as
θ(cid:96)k(i + 1) = rkθ(cid:96)k(i) + (1 − rk)a(cid:96)k(i)
θ(cid:96)k(i + 1) = 1k(cid:96)(i) [rkθ(cid:96)k(i) + (1 − rk)a(cid:96)k(i)]
+ (1 − 1k(cid:96)(i))θ(cid:96)k(i)
(44)
Directly applying the above reputation formulation, however,
can cause a loss in adaptation ability over the network.
For example, the network would become permanently non-
cooperative when agent (cid:96) chooses a(cid:96)k(i) = 0 for long con-
secutive iterations. That is because, in that case, the reputation
score θ(cid:96)k(i) will decay exponentially to zero, which keeps
agent k from choosing ak(cid:96)(i) = 1 in the future. In order to
avoid this situation, we set a lowest value for the reputation
score to a small positive threshold 0 < ε (cid:28) 1, i.e.,
θ(cid:96)k(i + 1) = 1k(cid:96)(i) · max{rkθ(cid:96)k(i) + (1 − rk)a(cid:96)k(i), ε}
(45)
+ (1 − 1k(cid:96)(i))θ(cid:96)k(i)
and thus θ(cid:96)k(i) ∈ [ε, 1].
The reputation scores can now be utilized to evaluate the
belief by agent k of subsequent actions by agent (cid:96). To explain
how this can be done, we argue that agent k would expect
the probability of a(cid:96)k(t) = 1, i.e., the probability that agent
(cid:96) is willing to cooperate, to be an increasing function of both
θ(cid:96)k(t) and θk(cid:96)(t) for t ≥ i. Specifically, if we denote this
belief probability by B(a(cid:96)k(t) = 1), then it is expected to
satisfy:
∂B(a(cid:96)k(t) = 1)
∂θ(cid:96)k(t)
≥ 0,
∂B(a(cid:96)k(t) = 1)
∂θk(cid:96)(t)
≥ 0
J
∞(cid:48)
k,i [ak(cid:96)(i)wk,i−1] =
δt−i
k
E(cid:104)
∞(cid:88)
=E(cid:104)
t=i
J act
k (a(cid:96)k(t)) + ak(cid:96)(t)ck
(cid:12)(cid:12)(cid:12)wk,i−1, Ki
(cid:105)
+ak(cid:96)(i)ck +
E(cid:104)
δt−i
k
∞(cid:88)
t=i+1
J act
k (a(cid:96)k(i))
J act
k (a(cid:96)k(t)) + ak(cid:96)(t)ck
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = ak(cid:96)(i), Ki
(cid:105)
(49)
8
The first property is motivated by the fact that according to
the history of actions, a higher value for θ(cid:96)k(t) indicates that
agent (cid:96) has higher willingness to share estimates. The second
property is motivated by the fact that lower values for θk(cid:96)(t)
mean that agent k has rarely shared estimates with agent (cid:96)
in the recent past. Therefore, it can be expected that agent
(cid:96) will have lower willingness to share information for lower
values of θk(cid:96)(t). Based on this argument, we suggest a first-
order construction for measuring belief with respect to both
θ(cid:96)k(t) and θk(cid:96)(t) as follows (other constructions are of course
possible; our intent is to keep the complexity of the solution
low while meeting the desired objectives):
B(a(cid:96)k(t) = 1) = θk(cid:96)(t) · θ(cid:96)k(t),
(47)
which satisfies both properties in (46) and where B(a(cid:96)k(t) =
1) ∈ [ε2, 1]. Therefore, the reputation protocol implements
(45) and (47) repeatedly. Each agent k will then employ the
reference knowledge Ki (cid:44) {θk(cid:96)(i), θ(cid:96)k(i), B(a(cid:96)k(i) = 1)} to
select its action ak(cid:96)(i) as described next.
t ≥ i
B. Best Response Rule
The belief measure (47) provides agent k with additional
information about agent (cid:96)'s actions. That is, with (47), agent
k can treat a(cid:96)k(t) as a random variable with distribution
B(a(cid:96)k(t) = 1) for t ≥ i. Then, the best response of agent
k is obtained by solving the following optimization problem:
(48)
min
∞(cid:48)
k,i [ak(cid:96)(i)wk,i−1]
ak(cid:96)(i)∈{0,1} J
where J∞(cid:48)
k,i [ak(cid:96)(i)wk,i−1] is defined by (49) and involves
an additional expectation over the distribution of a(cid:96)k(t) --
compare with (25). Similarly to Assumption 1, we assume
the bounded rationality of the agents extends to the reputation
scores θ(cid:96)k(t) for t ≥ i.
extend
the assumption of bounded rationality from (27) to also in-
clude:
Assumption 2 (Extended bounded rationality): We
θ(cid:96)k(t) = θ(cid:96)k(i),
(50)
(cid:4)
Now, using pure strategies, the best response of agent k is to
select the action ak(cid:96)(i) such that
for t ≥ i
if J∞(cid:48)
k,i [ak(cid:96)(i) = 1wk,i−1]
< J∞(cid:48)
k,i [ak(cid:96)(i) = 0wk,i−1]
0, otherwise
(51)
(46)
1,
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = ak(cid:96)(i), Ki
ak(cid:96)(i) =
(cid:105)
The following lemma shows how the best response rule
depends on the benefit-cost ratio γk(i) and the communication
cost ck:
Lemma 1: With Assumptions 1 and 2, the best response rule
(cid:40)
fk(·) becomes
ak(cid:96)(i) =
where
if γk(i) (cid:44) bk(i)
1,
0, otherwise
ck
> χk
θ(cid:96)k(i)
(52)
(53)
χk (cid:44) 1 − δkrk
δk(1 − rk)
Proof: See Appendix A.
We note that the resulting rule aligns the agents to achieve
the Pareto optimal strategy: to share information when γk(i)
is sufficiently large and not to share information when γk(i)
is small.
C. Benefit Prediction
To compute the benefit-cost ratio γk(i) = bk(i)/ck, the
agent still needs to know bk(i) defined by (35), which
depends on the quantities E[J act
k (a(cid:96)k(i) = 0)wk,i−1] and
E[J act
k (a(cid:96)k(i) = 1)wk,i−1]. It is common in the literature, as
in [29], [48], to assume that agents have complete information
about the payoff functions like the ones shown in Table I.
However, in the context of adaptive networks where agents
have only access to data realizations and not to their statistical
distributions, the payoffs are unknown and need to be estimated
or predicted. For example, in our case, the convex combination
αkψk,i + (1 − αk)ψ(cid:96),i is unknown for agent k before agent (cid:96)
shares ψ(cid:96),i with it. We now describe one way by which agent
k can predict bk(i); other ways are possible depending on how
much information is available to the agent. Let us assume a
special type of agents, which are called risk-taking [49]: agent
k chooses ak(cid:96)(i) = 1 as long as the largest achievable benefit,
denoted by ¯bk(i), exceeds the threshold:
(cid:40)
ak(cid:96)(i) =
if ¯bk(i)
ck
1,
0, otherwise
> χk
θ(cid:96)k(i)
(54)
(cid:105)
bk(i)
Using (39), the largest achievable benefit ¯bk(i) can be found
by solving the following optimization problem:
¯bk(i) (cid:44) max(cid:101)ψ(cid:96),i
= max(cid:101)ψ(cid:96),i
= E(cid:104)
(cid:110)E(cid:104)
(cid:107)(cid:101)ψk,i(cid:107)2
− E(cid:104)
(cid:107)αk(cid:101)ψk,i + (1 − αk)(cid:101)ψ(cid:96),i(cid:107)2
(cid:12)(cid:12)(cid:12)wk,i−1
(cid:107)(cid:101)ψk,i(cid:107)2
(cid:101)ψ(cid:96),i = −
(cid:12)(cid:12)(cid:12)wk,i−1
(cid:105)
since the maximum occurs when
(cid:12)(cid:12)(cid:12)wk,i−1
(cid:101)ψk,i
(cid:105)(cid:111)
(55)
Ru,k
Ru,k
Ru,k
αk
1 − αk
9
k,ivk(i)
(57)
Let us express the adaptation step (15) in terms of the
estimation error as
To continue, we assume that the step-size µ is sufficiently
small. Then,
k,iuk,i)(cid:101)wk,i−1 − µu∗
(cid:101)ψk,i = (I − µu∗
(cid:12)(cid:12)(cid:12)wk,i−1
(cid:105)
E(cid:104)
(cid:107)(cid:101)ψk,i(cid:107)2
= E(cid:104)
k,iuk,i)(cid:101)wk,i−1 − µu∗
= E(cid:104)(cid:101)w∗
(cid:107)(I − µu∗
(cid:12)(cid:12)(cid:12)wk,i−1
k,i−1(I − µu∗
× (cid:101)wk,i−1
= (cid:101)w∗
k,i−1Ωk(cid:101)wk,i−1 + O(µ2)
Ru,k
(cid:105)
k,iuk,i)Ru,k(I − µu∗
u,k)σ2
v,k
+ µ2Tr(R2
k,ivk(i)(cid:107)2
Ru,k
k,iuk,i)
(cid:105)
(cid:12)(cid:12)(cid:12)wk,i−1
(58)
k
where we are collecting terms that are second-order in the step-
size into the factor O(µ2),4 and where we introduced Ωk (cid:44)
Ru,k(I − 2µRu,k). We note that for sufficiently small step-
sizes:
Ω(cid:48)
Therefore, each agent k can approximate ¯bk(i) as
(cid:44) Ru,k(I − µRu,k)2 ≈ Ru,k(I − 2µRu,k) = Ωk
¯bk(i) = (cid:101)w∗
= (cid:101)w∗
k,i−1Ru,k(I − µRu,k)2(cid:101)wk,i−1
= (cid:107)(I − µRu,k)(cid:101)wk,i−1(cid:107)2
k(cid:101)wk,i−1
k,i−1Ω(cid:48)
(59)
(60)
Ru,k
D. Real-Time Implementation
Expression (60) is still not useful for agents because it
requires knowledge of both Ru,k and (cid:101)wk,i−1. With regards
to Ru,k, we can use the instantaneous approximation Ru,k ≈
u∗
k,iuk,i to get
k,iuk,i)2(cid:101)wk,i−1
k,iuk,i(cid:101)wk,i−1
k,i−1u∗
= (1 − µ(cid:107)uk,i(cid:107)2)2(cid:101)w∗
k,iuk,i(I − µu∗
k,i−1u∗
With regards to (cid:101)wk,i−1, we assume that agents use a moving-
average filter as in [50] to approximate wo iteratively as
follows:
(61)
¯bk(i) ≈ (cid:101)w∗
(cid:98)wo
k,i = (1 − ν)(cid:98)wo
(cid:101)wk,i−1 ≈ (cid:98)wo
k,i − wk,i−1
k,i−1 + νψk,i
(62)
(63)
where ν ∈ (0, 1) is a positive forgetting factor close to 0 to give
higher weights on recent results. We summarize the operation
of the resulting algorithm in the following listing.
(56)
4This approximation simplifies the algorithm construction. However, when
we study the network performance later in (74) we shall keep the second-order
terms.
Algorithm 1 Diffusion Strategy with an Adaptive Reputation
Scheme.
Let {wk,−1 = 0} and {θk(cid:96)(−1) = 1} for all k and (cid:96). Define χk (cid:44)
1−δkrk
δk(1−rk) .
loop
k,i[dk(i) − uk,iwk,i−1]
k,iuk,i(cid:101)wk,i−1
k,i−1u∗
Generate {1k(cid:96)(i)} for all k and (cid:96).
% Stage 1 (Adaptation): For each k:
ψk,i = wk,i−1 + µu∗
¯bk(i) ≈ (1 − µ(cid:107)uk,i(cid:107)2)2(cid:101)w∗
(cid:98)wo
k,i = (1 − ν)(cid:98)wo
(cid:101)wk,i−1 ≈ (cid:98)wo
(cid:40)
% Stage 2 (Action Selection): For all k and (cid:96),
if 1k(cid:96)(i) = 1 then
1,
k,i − wk,i−1
k,i−1 + νψk,i
> χk
θ(cid:96)k(i)
ak(cid:96)(i) =
if ¯bk(i)
ck
otherwise
0,
else ak(cid:96)(i) = 0.
end if
% Stage 3 (Reputation Update): For all k and (cid:96),
θ(cid:96)k(i + 1) =1k(cid:96)(i) · max{rkθ(cid:96)k(i) + (1 − rk)a(cid:96)k(i), ε}
+ (1 − 1k(cid:96)(i))θ(cid:96)k(i)
% Stage 4 (Combination): For all k,
1k(cid:96)(i)(cid:2)a(cid:96)k(i)ψ(cid:96),i +(1 − a(cid:96)k(i)) ψk,i
(cid:3)
(cid:88)
(cid:96)∈Nk
wk,i = αkψk,i +(1 − αk)
end loop
V. STABILITY ANALYSIS AND LIMITING BEHAVIOR
In this section, we study the stability of Algorithm 1 and its
limiting performance after sufficiently long iterations. In order
to pursue a mathematically tractable analysis, we assume that
the maximum benefit ¯bk(i) is estimated rather accurately by
each agent k. That is, instead of the real-time implementation
(61) -- (63), we consider (60) that
¯bk(i) = (cid:107)(I − µRu,k)(cid:101)wk,i−1(cid:107)2
(cid:12)(cid:12)(cid:12)(cid:101)wk,i−1
k,iuk,i(cid:101)wk,i−1
k,iuk,iu∗
expression (61) given (cid:101)wk,i−1:
E(cid:104)
(1 − µ(cid:107)uk,i(cid:107)2)2(cid:101)w∗
k,i−1u∗
= (cid:101)w∗
Ru,k − 2µE[u∗
k,i−1
(cid:104)
k,iuk,i] + O(µ2)
(cid:105)
Ru,k
This consideration is motivated by taking the expectation of
(65)
By subtracting (64) from (65), we can see that the difference
between (64) and (65) is at least in the order of µ:
(64)
k,i−1
2µ(R2
u,k − E[u∗
k,iuk,iu∗
k,iuk,i]) + O(µ2)
(66)
Therefore, for small enough µ the expected value of the
realization given by (61) approaches the true value (64). The
performance degradation from the real-time implementation
error will be illustrated by numerical simulations in Sec. VI.
Under this condition, and for µ (cid:28) 1, we shall argue that
the operation of each agent is stable in terms of both the
estimation cost and the communication cost. Specifically, for
the estimation cost, we will provide a condition on the step-
size to ensure that the mean-square estimation error of each
(cid:105)(cid:101)wk,i−1
(cid:105)(cid:101)wk,i−1
(cid:101)w∗
(cid:104)
10
k
agent is asymptotically bounded. Using this result, we will
further show that the communication cost for each agent k,
and which is denoted by J com
, is upper bounded by a constant
value that is unrelated to the transmission cost ck. This result
will be in contrast to the case of always cooperative agents
where J com
k will be seen to increase proportionally with ck.
This is because in our case, the probability of cooperation,
Prob{ak(cid:96)(i) = 1}, will be shown to be upper bounded by
the ratio co/ck for some constant co independent of ck. It will
then follow that when the communication becomes expensive
(large ck), self-interested agents using the adaptive reputation
scheme will become unwilling to cooperate.
A. Estimation Performance
In conventional stability analysis for diffusion strategies, the
combination coefficients are either assumed to be fixed, as
done in [16], [19], [51] -- [54], or their expectations conditioned
on the estimates wk,i−1 are assumed to be constant, as in [55].
These conditions are not applicable to our scenario. When
self-interested agents employ the nonlinear threshold-based
action policy (54), the ATC diffusion algorithm (15) and (19)
ends up involving a combination matrix whose entries are
dependent on the estimates wk,i−1 (or the errors (cid:101)wk,i−1).
This fact introduces a new challenging aspect into the analysis
of the distributed strategy. In the sequel, and in order to
facilitate the stability and mean-square analysis of the learning
process, we shall examine the performance of the agents in
the network in three operating regions: the far-field region,
the near-field region, and a region in between. We will show
that the evolution of the estimation errors in these operating
regions can be described by the same network error recursion
given further ahead in (99). Following this conclusion, we will
then be able to use (99) to provide general statements about
stability and performance in the three regions.
To begin with, referring to the listing in Algorithm 1,
we start by noting that we write down the following error
recursions for each agent k:
(cid:101)ψk,i = (I − µu∗
(cid:88)
(cid:101)wk,i =
k,iuk,i)(cid:101)wk,i−1 − µu∗
g(cid:96)k(i)(cid:101)ψ(cid:96),i
(cid:96)∈Nk
k,ivk(i)
where the combination coefficients {g(cid:96)k(i), (cid:96) ∈ Nk} used in
(68) are defined as follows and add up to one:
(cid:88)
g(cid:96)k(i) (cid:44) (1 − αk)1(cid:96)k(i)a(cid:96)k(i) ≥ 0
gkk(i) (cid:44) 1 −
g(cid:96)k(i) ≥ 0
(cid:96)∈Nk\{k}
(67)
(68)
(69)
(70)
Note that, in view of the pairing process, at most two of
evaluating the squared weighted norm of (cid:101)ψk,i in (67), which
the coefficients {g(cid:96)k(i)} in (68) are nonzero in each time
instant. The subsequent performance analysis will depend on
is seen to be:
(cid:107)(cid:101)ψk,i(cid:107)2
Ru,k
Ru,k
k,iuk,iRu,k)
k,iuk,i)Ru,ku∗
k,iuk,i)(cid:101)wk,i−1(cid:107)2
= (cid:107)(I − µu∗
− µ(cid:101)w∗
+ µ2(cid:107)vk(i)(cid:107)2Tr(u∗
k,i−1(I − µu∗
− µv∗
k(i)uk,iRu,k(I − µu∗
(cid:88)
g(cid:96)k(i)(cid:107)(cid:101)ψ(cid:96),i(cid:107)2
(cid:107)(cid:101)wk,i(cid:107)2
E(cid:104)
(cid:88)
g(cid:96)k(i)(cid:107)(cid:101)ψ(cid:96),i(cid:107)2
E(cid:107)(cid:101)wk,i(cid:107)2
Ru,k ≤
(cid:96)∈Nk
Ru,(cid:96)
(cid:96)∈Nk
Ru,(cid:96)
Ru,k ≤
so that, under expectation,
Now, from (68) we can use Jensen's inequality and the
convexity of the squared norm to write
k,iuk,i)(cid:101)wk,i−1 (71)
k,ivk(i)
(cid:105)
(72)
(73)
(cid:105)
E(cid:104)
g(cid:96)k(i)(cid:107)(cid:101)ψ(cid:96),i(cid:107)2
= E(cid:104)
(cid:16)
We note that g(cid:96)k(i) is a function of the random variables
{1(cid:96)k(i), a(cid:96)k(i)}. The random pairing indicator 1(cid:96)k(i) is inde-
pendent of u(cid:96),i and v(cid:96)(i). As for a(cid:96)k(i), which is determined
by ¯b(cid:96)(i) and θk(cid:96)(i), we can see from expressions (64) and
(45) that both ¯b(cid:96)(i) and θk(cid:96)(i) only depend on the past data
prior to time i and therefore are independent of u(cid:96),i and v(cid:96)(i).
Consequently, g(cid:96)k(i) is independent of u(cid:96),i and v(cid:96)(i), and we
get
Ru,(cid:96)
Ru,(cid:96)
Ru,(cid:96)
(74)
g(cid:96)k(i)
we get
(cid:96),iv(cid:96)(i)
(cid:17)(cid:105)
u,(cid:96))σ2
v,(cid:96)
(cid:96),iu(cid:96),iRu,(cid:96))
(cid:96),iu(cid:96),i)Ru,(cid:96)u∗
(cid:96),iu(cid:96),i)(cid:101)w(cid:96),i−1(cid:107)2
(cid:107)(I − µu∗
− µ(cid:101)w∗
+ µ2(cid:107)v(cid:96)(i)(cid:107)2Tr(u∗
(cid:96),i−1(I − µu∗
(cid:96),iu(cid:96),i)(cid:101)w(cid:96),i−1
= E(cid:104)
(cid:105)
(cid:96) (i)u(cid:96),iRu,(cid:96)(I − µu∗
− µv∗
(cid:96),iu(cid:96),i)(cid:101)w(cid:96),i−1(cid:107)2
g(cid:96)k(i)(cid:107)(I − µu∗
+ µ2Eg(cid:96)k(i)Tr(R2
Using the fact that u(cid:96),i is independent of g(cid:96)k(i) and (cid:101)w(cid:96),i−1,
E(cid:104)
(cid:105)
(cid:96),iu(cid:96),i)(cid:101)w(cid:96),i−1(cid:107)2
(cid:12)(cid:12)(cid:12)g(cid:96)k(i), (cid:101)w(cid:96),i−1
= E(cid:104)E(cid:104)
(cid:105)(cid:105)
g(cid:96)k(i)(cid:107)(I − µu∗
(cid:96),iu(cid:96),i)(cid:101)w(cid:96),i−1(cid:107)2
(cid:105)
= E(cid:104)
g(cid:96)k(i)(cid:107)(I − µu∗
(cid:96),i−1Σ(cid:96)(cid:101)w(cid:96),i−1
g(cid:96)k(i)(cid:101)w∗
Σ(cid:96) (cid:44) E(I − µu∗
= Ru,(cid:96) − 2µR2
(cid:96),iu(cid:96),i)Ru,(cid:96)(I − µu∗
u,(cid:96) + µ2E(u∗
(cid:96),iu(cid:96),i)
(cid:96),iu(cid:96),iRu,(cid:96)u∗
(cid:96),iu(cid:96),i)
If the regression data happens to be circular Gaussian, then a
closed-form expression exists for the last fourth-order moment
term in (76) in terms of Ru,(cid:96) [47]. We will not assume
Gaussian data. Instead, we will assume that the fourth-order
moment is bounded and that the network is operating in the
slow adaptation regime with a sufficiently small step-size so
that terms that depend on higher-order powers of µ can be
where
(76)
(75)
Ru,(cid:96)
Ru,(cid:96)
(cid:105)
(cid:105)
E(cid:104)
ignored in comparison to other terms. Under this assumption,
we replace (75) by:
Ru,(cid:96)
= E(cid:104)
g(cid:96)k(i)(cid:107)(I − µu∗
= E(cid:104)
(cid:96),iu(cid:96),i)(cid:101)w(cid:96),i−1(cid:107)2
g(cid:96)k(i)(cid:101)w∗
(cid:96)(cid:101)w(cid:96),i−1
(cid:105)
(cid:96),i−1Ω(cid:48)
(77)
(cid:96)(cid:101)w(cid:96),i−1 from (64). Note that
(cid:96) = Ru,(cid:96)(I − µRu,(cid:96))2 from (59) and ¯b(cid:96)(i) =
(cid:105)
g(cid:96)k(i)¯b(cid:96)(i)
(cid:96) = O(µ2)
Σ(cid:96) − Ω(cid:48)
=E(cid:104)
(78)
(cid:105)
(cid:101)w∗
where Ω(cid:48)
(cid:96),i−1Ω(cid:48)
E(cid:104)
g(cid:96)k(i)(cid:107)(cid:101)ψ(cid:96),i(cid:107)2
Ru,(cid:96)
g(cid:96)k(i)¯b(cid:96)(i)
Eg(cid:96)k(i)
(79)
To continue, we introduce the following lemma which provides
useful bounds for ¯bk(i).
+µ2Tr(R2
u,(cid:96))σ2
v,(cid:96)
Lemma 2 (Bounds on ¯bk(i)): The values of ¯bk(i) defined
by (64) are lower and upper bounded by:
Therefore, expression (74) becomes
min(cid:107)(cid:101)wk,i−1(cid:107)2
ρ2
Ru,k ≤ ¯bk(i) ≤ ρ2
max(cid:107)(cid:101)wk,i−1(cid:107)2
Ru,k
where
ρmax (cid:44) max
1≤k≤N
ρmin (cid:44) min
1≤k≤N
λmax(I − µRu,k)
λmin(I − µRu,k)
11
(80)
(81)
(82)
Proof: We introduce the eigendecomposition of the co-
variance matrix, Ru,k (cid:44) UkΛkU∗
k , where Uk is a unitary
matrix and Λk (cid:44) diag{λ1,k, ..., λM,k} is a diagonal matrix
with positive entries. Then, Ru,k can be factored as
where
1
2
u,k
R
(cid:44) UkΛ
1
2
k U
∗
1
k , Λ
2
k
Ru,k = R
1
2
u,kR
1
2
u,k
(83)
(cid:112)
(cid:112)
(cid:44) diag{
u,k and I − µRu,k are commutable.
λM,k}
λ1,k, ...,
(84)
1
2
It is easy to verify that R
Using this property, we obtain the following inequality:
u,kR
u,k(I − µRu,k)2R
1
2
1
2
k(cid:101)wk,i−1
k,i−1(I − µRu,k)R
k,i−1R
(cid:101)w∗
k,i−1Ω(cid:48)
= (cid:101)w∗
= (cid:101)w∗
≤ λmax((I − µRu,k)2)(cid:101)w∗
max(cid:107)(cid:101)wk,i−1(cid:107)2
≤ ρ2
(cid:105)
Ru,k
maxE(cid:104)
g(cid:96)k(i)¯b(cid:96)(i)
E(cid:104)
≤ ρ2
1
2
u,k(I − µRu,k)(cid:101)wk,i−1
u,k(cid:101)wk,i−1
k,i−1Ru,k(cid:101)wk,i−1
1
2
(85)
g(cid:96)k(i)(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
Ru,(cid:96)
(cid:105)
(86)
We can obtain the lower bound for ¯bk(i) by similar arguments.
Using the upper bound from Lemma 2, we have
µ2 (cid:88)
(cid:96)∈Nk
Tr(R2
u,(cid:96))σ2
v,(cid:96)
Eg(cid:96)k(i) ≤ µ2κ
max
(cid:105)
Ru,k ≤ ρ2
E(cid:107)(cid:101)wk,i(cid:107)2
Then, from (79) and (86) we deduce from (73) that
E(cid:104)
(cid:88)
g(cid:96)k(i)(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
+ µ2 (cid:88)
is ready to check that {Eg(cid:96)k(i)} are
From (69) -- (70),
nonnegative and add up to 1. Therefore, the second term on
the right-hand side of (87) is a convex combination and has
the following upper bound:
Eg(cid:96)k(i)
u,(cid:96))σ2
v,(cid:96)
Tr(R2
(cid:96)∈Nk
(cid:96)∈Nk
(87)
Ru,(cid:96)
it
(88)
(89)
(cid:105)
+ µ2κ
where
κ (cid:44) max
1≤k≤N
Tr(R2
u,k)σ2
v,k
(cid:88)
E(cid:104)
(cid:96)∈Nk
g(cid:96)k(i)(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
Ru,(cid:96)
Ru,k ≤ ρ2
max
Therefore, we have
E(cid:107)(cid:101)wk,i(cid:107)2
errors {(cid:107)(cid:101)wk,i−1(cid:107)2
(90)
Since the combination coefficients {g(cid:96)k(i)} and the estimation
Ru,k} are related in a nonlinear manner (as re-
vealed by (54, (64), and (69) -- (70)), the analysis of Algorithm 1
becomes challenging. To continue, we examine the behavior of
the agents in the three regions of operation mentioned before.
During the initial stage of adaptation, agents are generally
away from the target vector wo and therefore have large
estimation errors. We refer to this domain as the far-field region
of operation, and we will characterize it by the condition:
(cid:107)(cid:101)wk,i−1(cid:107)2
norm of estimation error (cid:107)(cid:101)wk,i−1(cid:107)2
where χk and ρmin are defined in (53) and (82), respectively,
and the parameter φ is close to 1 and in the range of 1 ≥
φ (cid:29) 0. That is, in the far-field regime, the weighted squared
exceeds a threshold
with high probability. We note that the far-field condition (91)
can be more easily achieved when ck is small. We also note
that when the event in (91) holds with high-probability, then
agent k is enticed to cooperate since:
Far-Field: Prob
> φ
(91)
ckχk
ρ2
minε
Ru,k >
(cid:27)
(cid:26)
Ru,k
(cid:107)(cid:101)wk,i−1(cid:107)2
Ru,k >
ckχk
ρ2
minε
(a)
⇒ ¯bk(i)θ(cid:96)k(i) > ckχk
⇒ ak(cid:96)(i) = 1
(b)
(92)
where step (a) is by (80) and the fact θ(cid:96)k(i) ∈ [ε, 1], and step
(b) is by (54). Consequently, in the far-field region it holds
with high likelihood that
Prob{ak(cid:96)(i) = 1} ≈ 1
(93)
This approximation explains the phenomenon that with large
estimation errors, agents are willing to cooperate and share
12
estimates. We then say that under the far-field condition (91),
the combination coefficients in (69) -- (70) can be expressed as
(94)
(cid:88)
g(cid:96)k(i) = (1 − αk)1(cid:96)k(i)
gkk(i) = 1 −
(cid:96)∈Nk\{k}
g(cid:96)k(i)
(95)
(96)
(97)
E(cid:107)(cid:101)wk,i(cid:107)2
(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
col{(cid:107)(cid:101)w1,i(cid:107)2
with expectation values as follows:
= αk + (1 − αk)1kk(i)
In this case, expression (90) becomes
g(cid:96)k (cid:44) Eg(cid:96)k(i) = (1 − αk)p(cid:96)k
gkk (cid:44) Egkk(i) = αk + (1 − αk)pkk
g(cid:96)kE(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
(cid:88)
Ru,k ≤ ρ2
Ru,(cid:96)
max
(cid:96)∈Nk
+ µ2κ
Ru,(cid:96)
. If we stack {(cid:107)(cid:101)wk,i(cid:107)2
Ru,1 , ...,(cid:107)(cid:101)wN,i(cid:107)2
EX i (cid:22) ρ2
where we used the independence between 1(cid:96)k(i) and
Ru,k} into a vector X i (cid:44)
Ru,N} and collect the combination
coefficients {g(cid:96)k} into a left-stochastic matrix G, then we
obtain the network error recursion as:
(98)
where e ∈ 1N is the vector of all ones and the notation x (cid:22)
y denotes that the components of vector x are less than or
equal to the corresponding components of vector y. Taking the
maximum norm from both sides and using the left-stochastic
matrix property (cid:107)GT(cid:107)∞ = 1, we obtain:
(cid:44) (cid:107)EX i(cid:107)∞ < ρ2
max
1≤k≤N
max(cid:107)EX i−1(cid:107)∞ + µ2κ
maxGT (EX i−1) + µ2κe
E(cid:107)(cid:101)wk,i(cid:107)2
(99)
Let us consider the long-term scenario i (cid:29) 1. The far-field
regime (91) is more likely to occur when ck is small. Let us
now examine the situation in which the communication cost
is expensive (ck (cid:29) 0). In this case, the agents will operate in
a near-field regime, which we characterize by the condition:
Ru,k
Near-Field: Prob
ckχk
ρ2
max
where ρmax is defined in (81). We note that
Ru,k <
> φ
(100)
(cid:26)
(cid:107)(cid:101)wk,i−1(cid:107)2
(cid:27)
(cid:107)(cid:101)wk,i−1(cid:107)2
Ru,k <
ckχk
ρ2
max ⇒ ¯bk(i)θ(cid:96)k(i) < ckχk
In this regime, it then holds with high likelihood that
⇒ ak(cid:96)(i) = 0
Prob{ak(cid:96)(i) = 0} ≈ 1
and the combination coefficients in (69) -- (70) then become:
that
gkk(i) = 1, g(cid:96)k(i) = 0
(103)
This means
agents will now be operating non-
cooperatively since the benefit of sharing estimates is small
relative to the expensive communication cost ck (cid:29) 0. Using
similar arguments to (97) -- (99), we arrive at the same network
recursion (99) for this regime.
(101)
(102)
Following similar arguments to (97) -- (99), we again arrive at
the same network recursion (99).
We therefore conclude that after long iterations, the esti-
mation performance can be approximated by recursion (99)
for general values of ck. Consequently, sufficiently small
step-sizes that satisfy the following condition guarantees the
stability of the network error (cid:107)EX i(cid:107)∞:
ρ2
max < 1 ⇔ µ <
max
1≤k≤N
which leads to
2
λmax(Ru,k)
lim sup
i→∞ (cid:107)EX i(cid:107)∞ ≤
µ2κ
1 − ρ2
max
13
(112)
(113)
However, there exists a third possibility that for moderate
values of ck, agents operate at a region that does not belong to
neither the far-field nor the near-field regimes. In this region,
agents will be choosing to cooperate or not depending on their
local conditions. To facilitate the presentation, let us introduce
the notation IE for an indicator function over event E where
IE = 1 if event E occurs and IE = 0 otherwise. Then, the
action policy (54) can be rewritten as:
ak(cid:96)(i) = I¯bk(i)θ(cid:96)k(i)>ckχk
min)(cid:107)(cid:101)wk,i−1(cid:107)2
(104)
From Lemma 2, we know that ¯bk(i) is sandwiched within
(cid:107)(cid:101)wk,i−1(cid:107)2
an interval of width (ρ2
. As the error
becomes smaller after sufficient iterations, the
feasible region of ¯bk(i) shrinks. Therefore, it is reasonable
to assume that ¯bk(i) becomes concentrated around its mean
¯bk(i) (cid:44) E¯bk(i) for i (cid:29) 0. It follows that we can approximate
ak(cid:96)(i) by replacing ¯bk(i) by ¯bk(i):
max − ρ2
Ru,k
Ru,k
ak(cid:96)(i) ≈ I¯bk(i)θ(cid:96)k(i)>ckχk , as i (cid:29) 1
(105)
To continue, we further assume that after long iterations, which
means that repeated interactions between agents have occurred
for many times, the reputation scores gradually become sta-
tionary, i.e., we can model {θ(cid:96)k(i)} as
θ(cid:96)k(i) = θ(cid:96)k + n(cid:96)k(i), as i (cid:29) 1
(106)
where θ(cid:96)k (cid:44) Eθ(cid:96)k(i) and n(cid:96)k(i) is a random disturbance
which is independent and identically distributed (i.i.d.) over
time i and assumed to be independent of all other random
variables. Under this modeling, we can obtain the indepen-
(cid:105)
I¯b(cid:96)(i)(θ(cid:96)k+n(cid:96)k(i))>ckχk · (cid:107)(cid:101)w(cid:96),i−1(cid:107)2
dence between n(cid:96)k(i) and (cid:101)w(cid:96),i−1 and write that for k (cid:54)= (cid:96)
E(cid:104)
g(cid:96)k(i)(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
= E(cid:104)
(1 − αk)1(cid:96)k(i)a(cid:96)k(i)(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
(cid:105)
≈ (1 − αk)p(cid:96)kE(cid:104)
E(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
(cid:44) (1 − αk)p(cid:96)kE(cid:2)I¯bk(i)(θ(cid:96)k+n(cid:96)k(i))>ckχk
E(cid:104)
gkk(i)(cid:107)(cid:101)wk,i−1(cid:107)2
(cid:17)
= E(cid:104)(cid:16)
(cid:88)
(cid:107)(cid:101)wk,i−1(cid:107)2
E(cid:107)(cid:101)wk,i−1(cid:107)2
For (cid:96) = k, we can use similar arguments to write
(1 − αk)1(cid:96)k(i)a(cid:96)k(i)
= g
where
(cid:48)
g
(cid:96)k
(cid:96)∈Nk\{k}
≥ 0
1 −
(109)
(107)
(108)
(cid:105)
(cid:105)
(cid:105)
(cid:3)
(cid:48)
(cid:96)k
Ru,k
Ru,k
Ru,(cid:96)
Ru,(cid:96)
Ru,(cid:96)
Ru,(cid:96)
Ru,(cid:96)
(cid:48)
(cid:96)k
≈ g
where
(cid:96)∈Nk\{k}
Therefore, expression (90) becomes
E(cid:107)(cid:101)wk,i(cid:107)2
Ru,k ≤ ρ2
max
(cid:48)
g
kk
(cid:44) 1 −
(cid:88)
(cid:96)∈Nk
(cid:48)
(cid:96)k ≥ 0
g
(cid:88)
E(cid:107)(cid:101)w(cid:96),i−1(cid:107)2
(cid:48)
g
(cid:96)k
Ru,(cid:96)
Recalling that we are assuming sufficiently small µ, we have
max
1≤m≤M
max
1≤m≤M
ρ2
max = max
1≤k≤N
≈ max
1≤k≤N
(cid:44) 1 − 2µβ
(1 − µλm(Ru,k))2
1 − 2µλm(Ru,k)
β (cid:44) min
1≤k≤N
λmin(Ru,k)
(114)
(115)
where
It is straightforward to verify that the bound on the right-hand
side of (113) is upper-bounded by µκ/2β, which is O(µ).
Consequently, we conclude that
E(cid:107)(cid:101)wk,i(cid:107)2
lim sup
i→∞
= O(µ)
(116)
Ru,k
which establishes the mean-square stability of the network
under the assumed conditions and for small step-sizes.
B. Expected Individual and Public Cost
In this section we assess the probability that agents will
opt for cooperation after sufficient time has elapsed and the
network has become stable. From (116), we know that after
sufficient iterations, the MSE cost at each agent k is bounded,
say, as
for
Ru,k ≤ ηµ,
(117)
for some constant η. Based on this result, we can upper bound
the cooperation rate of every agent k, defined as the probability
that agent k would select ak(cid:96)(i) = 1 for every pairing agent
(cid:96).
i (cid:29) 1
Theorem 1: (Upper bound on cooperation rate) After suffi-
cient iterations, the cooperation rate for each agent k is upper
bounded by:
E(cid:107)(cid:101)wk,i(cid:107)2
(cid:27)
, 1
,
(cid:26) co
ck
(110)
Prob{ak(cid:96)(i) = 1} ≤ min
for any ck < ∞
(118)
where co is independent of ck and defined as
+ µ2κ (111)
co (cid:44) ηµρ2
χmin
max
, χmin (cid:44) min
1≤k≤N
χk
(119)
Fig. 2. The feasible region of the probability of cooperation Prob{ak(cid:96)(i) =
1} for agent k.
Proof: From (54), the cooperation rate of agent k is
bounded by:
Prob{ak(cid:96)(i) = 1} = Prob(cid:8)¯bk(i)θ(cid:96)k(i) > ckχk
(cid:9)
(cid:9)
≤ Prob(cid:8)¯bk(i) > ckχk
(cid:107)(cid:101)wk,i−1(cid:107)2
≤ Prob
(cid:40)
Ru,k >
ckχk
ρ2
max
(cid:41)
then
(120)
14
On the other hand, the expected estimation cost for each agent
k for i (cid:29) 0 is:
EJ est
= E(cid:107)(cid:101)wk,i−1(cid:107)2
k = Edk(i) − uk,iwk,i−12
+ σ2
Ru,k
v,k
v,k
≤ ηµ + σ2
where we use (9) and the fact that (cid:101)wk,i−1 is independent of
{dk(i), uk,i}. It follows that for i (cid:29) 1 the expected extended
cost at each agent k is bounded by
E [Jk(wk,i−1, ak(cid:96)(i))] = EJ est
k + EJ com
(124)
k
≤ ηµ + σ2
v,k + (1 − pkk)co
(125)
If we now define the public cost as the accumulated expected
extended cost over the network:
J pub (cid:44) N(cid:88)
N(cid:88)
k=1
E [Jk(wk,i−1, ak(cid:96)(i))]
N(cid:88)
σ2
v,k + co
k=1
k=1
(1 − pkk)
J pub ≤ N ηµ +
(126)
(127)
which shows that J pub is uniformly bounded by a constant
value independent of ck.
For comparison purposes, let us consider a uniform trans-
mission cost ck = c and consider the case in (18) where the
agents are controlled so that they always share estimates with
their paired agents, i.e., ak(cid:96)(i) = 1 for all k, (cid:96), and i whenever
1k(cid:96)(i) = 1. Then, the random combination coefficients are the
same as (94) -- (95) defined in the far-field since agents always
choose to cooperate. Let us denote the network mean-square-
deviation (MSD) of cooperative agents by
MSDcoop = lim
i→∞
1
N
(128)
N(cid:88)
k=1
E(cid:107)(cid:101)wk,i(cid:107)2
which has a closed-form expression provided in [55]. There-
fore, we can characterize the performance of cooperative
agents by MSDcoop and note that for i (cid:29) 0, we have
Ru,k ≥ N β · MSDcoop
(129)
where β is defined in (115). Consequently, after sufficient
iterations,
the expected public cost for cooperative agents
becomes
(cid:88)
E1k(cid:96)(i)
(cid:96)∈Nk\{k}
J pub
coop =
+ σ2
v,k + c
Ru,k
N(cid:88)
k=1
N(cid:88)
k=1
≥ N β · MSDcoop +
σ2
v,k +c
(1 − pkk)
(130)
N(cid:88)
k=1
E(cid:107)(cid:101)wk,i(cid:107)2
E(cid:107)(cid:101)wk,i(cid:107)2
N(cid:88)
k=1
where we used the fact that θk(cid:96)(i) ≤ 1 and the upper bound on
is a nonnegative random
Ru,k ≤ ηµ, we can use Markov's
Ru,k
¯bk(i) from (80). Since (cid:107)(cid:101)wk,i−1(cid:107)2
variable with E(cid:107)(cid:101)wk,i−1(cid:107)2
(cid:107)(cid:101)wk,i−1(cid:107)2
inequality [56] to write
Ru,k >
ckχk
ρ2
max
(cid:26)
Prob
max
ηµρ2
ckχk ≤
co
ck
≤
(121)
(cid:27)
Combining (120) and (121), we obtain that the cooperation
rate for i (cid:29) 1 is upper bounded by co/ck. Using the fact that
Prob{ak(cid:96)(i) = 1} ≤ 1, we get (118).
As illustrated in Fig. 2, the feasible region of Prob{ak(cid:96)(i) =
1} is the intersection area of (118) and 0 ≤ Prob{ak(cid:96)(i) =
1} ≤ 1. We note that co has an order of µ. It describes the fact
that when µ is small, the long term estimation errors reduce
and agents have less willingness to cooperate and thus the
cooperation rate Prob{ak(cid:96)(i) = 1} becomes low. Now, the
expected communication cost for each agent k is
(cid:88)
(cid:88)
(cid:96)∈Nk\{k}
(cid:96)∈Nk\{k}
EJ com
k =
=
E[1k(cid:96)(i)ak(cid:96)(i)]ck
pk(cid:96) · Prob{ak(cid:96)(i) = 1}ck
(122)
where {k} is excluded from the summation since there is
no communication cost required for using own data. From
Theorem 1, we know that when ck is large, the expected
communication cost has an upper bound which is independent
(cid:88)
of ck, i.e., for ck ≥ co,
EJ com
k ≤
(cid:96)∈Nk\{k}
pk(cid:96)
co
ck
ck = (1 − pkk)co
(123)
1=coccocProb(i)=1} a{Prob(i)=1} a{=Prob(i)=1} a{1 15
Comparing (127) with (130), we get J
pub
coop ≥ J pub whenever
N(cid:88)
k=1
N β · MSDcoop + c
N(cid:80)
⇔ c ≥ co +
N [ηµ − β · MSDcoop]
(1 − pkk)
k=1
(1 − pkk) ≥ N ηµ + co
(1 − pkk)
N(cid:88)
k=1
(131)
In other words, when the transmission cost c exceeds the above
threshold, self-interested agents using the reputation protocol
obtain a lower expected public cost than cooperative agents
running the cooperative ATC strategy.
VI. NUMERICAL RESULTS
Fig. 3. A network topology with N = 20 agents.
We consider a network with N = 20 agents. The network
topology is shown in Fig. 3. The noise variance profile at the
agents is shown in Fig. 4. In the simulations, we consider
that the transmission cost ck = c is uniform and the matrix
Ru,k = Ru is uniform and diagonal. Figure 5 shows the entries
of the target vector wo of size M = 10 and the diagonal entries
of Ru. We set the step-size at µ = 0.01 and the combination
weight at αk = 0.5 for all k. The parameters used in the
reputation update rule are ε = 0.1 and the initial reputation
scores θk,(cid:96)(0) = 1 for all agents k and (cid:96). The discount factor
δk and the smoothing factor rk for all k are set to 0.99 and
0.95, respectively, and the forgetting factor ν is set to 0.01.
In this simulation, we consider a distributed random-pairing
mechanism as follows. At each time instant, each agent inde-
pendently and uniformly generates a random continuous value
from [0, 1]. Then, the agent holding the smallest value in the
network, say agent k, is paired with the neighboring agent in
Nk who has not been paired and holds the smallest value in
Nk \ {k}. Then, similar steps are followed by the agent who
has not been paired and holds the second smallest value in
the network. The random-pairing procedure continues until all
agents complete these steps.
In Fig. 6, we simulate the learning curves of instantaneous
public costs for small and large communication costs. It is
seen that in both cases, using the proposed reputation protocol,
the network of self-interested agents reaches the lowest public
cost. Therefore, the network is efficient in terms of public cost.
Furthermore, we note that in these two cases, there is only
small difference between the performance of the reputation
protocol using Algorithm 1 and the real-time implementation.
To see the general effect of c, in Fig. 7 we simulate the
public cost in steady-state versus the communication cost c.
We observe that for large and small c, the reputation protocol
performs as well as the non-sharing and the cooperative
network, respectively. The only imperfection occurs around
the switching region. Without real-time implementations, the
reputation protocol has a small degradation in the range of
c ∈ [10−2, 10−1]. While using real-time implementations (61) --
(63), we can see that the degradation happens in a wider range
of c ∈ [10−4, 10−1]. In Fig. 8, we simulate the network benefit
defined as the largest achievable ¯bk(i) averaged over all agents
Fig. 4. The noise variance profile used in the simulations.
in steady state, i.e.,
bnet (cid:44) lim
i→∞
1
N
N(cid:88)
k=1
E[¯bk(i)]
(132)
where ¯bk(i) follows the real-time implementation in (61) -- (63).
We note that the network benefits, bnet, for the non-sharing and
cooperative cases are invariant for different communication
costs since the behavior of agents is independent of c. More-
over, as expected in the form of (61), cooperative networks
generally give smaller steady-state estimation errors and thus
result in lower bnet. When the communication cost c increases,
self-interested agents following the proposed reputation pro-
tocol have less willingness to cooperate, and therefore have
larger estimation errors and predict higher values of the benefit
¯bk(i) in general.
VII. CONCLUSION
In this work, we studied the distributed information-sharing
network in the presence of self-interested agents. We showed
that without using any historical information to predict fu-
ture actions, self-interested agents with bounded rationality
become non-cooperative and refuse to share information. To
entice them to cooperate, we developed an adaptive reputation
protocol which turns the best response of self-interested agents
into an action-choosing rule.
17192061112810181659413151147321234567891011121314151617181920−10−8−6−4−20 (dB)σv,k2Agent number k 16
Fig. 5. Entries of wo and Ru used in the simulations.
Fig. 7. Simulations of steady-state public costs J pub.
(a) Small communication cost c = 0.0001.
Fig. 8. Simulations of average largest achievable benefit ¯bk(i) over agents.
for
t=i+1
t=i+1
δt−i
k (cid:52)J 2
k (t) < 0 (135)
δt−i
k (cid:52)J 1
k (t) +
∞(cid:48)
k,i [ak(cid:96)(i) = 1wk,i−1] < J
J
∞(cid:48)
k,i [ak(cid:96)(i) = 1wk,i−1] − J
J
= ck +
∞(cid:48)
k,i [ak(cid:96)(i) = 0wk,i−1]
Using (49), this condition translates into requiring
∞(cid:88)
∞(cid:48)
k,i [ak(cid:96)(i) = 0wk,i−1]
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = 1, Ki
(cid:105)
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = 0, Ki
(cid:105)
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = 1, Ki
(cid:105)
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = 0, Ki
∞(cid:88)
k (t) (cid:44) E(cid:104)
− E(cid:104)
k (t) (cid:44) E(cid:104)
− E(cid:104)
where, for simplicity, we introduced
(cid:52)J 1
ak(cid:96)(t)ck
J act
k (a(cid:96)k(t))
J act
k (a(cid:96)k(t))
ak(cid:96)(t)ck
(cid:52)J 2
(134)
(cid:105)
(136)
(137)
(b) Large communication cost c = 0.5.
Fig. 6. Learning curve of public cost J pub for small and large communication
costs.
APPENDIX A
We can represent
the best response rule (51) by some
mapping function, fk(·), that maps the available realizations
of θ(cid:96)k(i) and wk,i−1 to ak(cid:96)(i), i.e.,
ak(cid:96)(i) = fk(θ(cid:96)k(i), wk,i−1)
(133)
We show the form of the resulting fk(·) later in (52). For now,
we note that construction (51) requires us to find the condition
so that
ak(cid:96)(t) = ak(cid:96)(i),
for t ≥ i
(cid:52)J 1
k (t) = ck
(138)
(139)
Following similar argument to (28), we combine Assumptions
1 and 2 to conclude that
2468100.10.150.20.25Index number m(m)Entries of w2468102468Index number m(m,m)Diagonal entries of RouRuwo050100150200250−4−3−2−101234TimePublic cost (dB)c = 0.0001Non−sharingCooperative (18)Reputation (19), w/o real-time implement.Reputation (19), w/ real−time implement.pubJ050100150200250−4−3−2−101234TimePublic cost (dB)c = 0.5Non−sharingCooperative (18)Reputation (19), w/o real-time implement.Reputation (19), w/ real−time implement.pubJ10−610−410−2100−4−3−2−1012Communication costPublic cost (dB)Non−sharingCooperative (18)Reputation (19), w/o real-time implement.Reputation (19), w/ real−time implement.pubJCommunication cost10-610-410-2100Largest achievable benefit (dB)-22-20-18-16-14-12Non-sharingCooperative (15)Reputation (16), w/ real-time implement. (cid:105)
(cid:105)
(cid:12)(cid:12)(cid:12)wk,t−1
(cid:12)(cid:12)(cid:12)wk,t−1
since these two conditional expectations depend only on
wk,t−1. Therefore, expression (140) becomes:
(cid:105)
E(cid:104)
Similarly, using the assumption wk,t−1 = wk,i−1 for t ≥ i
from (27), we have
J act
k (a(cid:96)k(t))
J act
k (a(cid:96)k(t) = 1)
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = j, Ki
= E(cid:104)
+ E(cid:104)
× B(a(cid:96)k(t) = 1)
(cid:12)(cid:12)(cid:12)wk,t−1, ak(cid:96)(i) = j, Ki
(cid:105)
(cid:12)(cid:12)(cid:12)wk,t−1, ak(cid:96)(i) = j, Ki
(cid:105)
J act
k (a(cid:96)k(t) = 0)
× (1 − B(a(cid:96)k(t) = 1))
(140)
for j = 0 or 1. From (37) and (38), we can write
E(cid:104)
E(cid:104)
J act
k (a(cid:96)k(t) = 0)
J act
k (a(cid:96)k(t) = 1)
J act
k (a(cid:96)k(t) = 1)
J act
k (a(cid:96)k(t) = 0)
(cid:105)
(cid:105)
(cid:105)
(cid:12)(cid:12)(cid:12)wk,t−1, ak(cid:96)(i) = j, Ki
= E(cid:104)
(cid:12)(cid:12)(cid:12)wk,t−1, ak(cid:96)(i) = j, Ki
= E(cid:104)
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = j, Ki
(cid:12)(cid:12)(cid:12)wk,t−1
(cid:105)
(cid:16)E(cid:104)
− E(cid:104)
(cid:12)(cid:12)(cid:12)wk,t−1
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = j, Ki
(cid:12)(cid:12)(cid:12)wk,t−1
(cid:105)
(cid:105)
(cid:105)
J act
k (a(cid:96)k(t) = 1)
J act
k (a(cid:96)k(t))
J act
k (a(cid:96)k(t) = 0)
+ B(a(cid:96)k(t) = 1) ·
J act
k (a(cid:96)k(t) = 0)
E(cid:104)
= E(cid:104)
= E(cid:104)
E(cid:104)
where we used the definition for bk(t) from (35). We note
that using (39) we have bk(t) = bk(i) due to the assumption
wk,t−1 = wk,i−1. As a result, it follows that
J act
k (a(cid:96)k(t))
= E(cid:104)
J act
k (a(cid:96)k(t) = 0)
− B(a(cid:96)k(t) = 1)bk(i)
Let us denote by θj
k(cid:96)(t) the value of θk(cid:96)(t) at time t if ak(cid:96)(i) =
j is selected at time i. We utilize the assumption θ(cid:96)k(t) =
θ(cid:96)k(i) to rewrite B(a(cid:96)k(t) = 1) as
B(a(cid:96)k(t) = 1) = θj
(141)
k(cid:96)(t)θ(cid:96)k(i)
It then follows that
E(cid:104)
(cid:12)(cid:12)(cid:12)wk,i−1, ak(cid:96)(i) = j, Ki
(cid:12)(cid:12)(cid:12)wk,t−1
(cid:105)
(cid:105)
− θj
J act
k (a(cid:96)k(t) = 0)
J act
k (a(cid:96)k(t))
= E(cid:104)
k(cid:96)(t)θ(cid:96)k(i)bk(i)
Therefore, using (137) we get
k (t) = [θ0
(cid:52)J 2
k(cid:96)(t) − θ1
k(cid:96)(t)]θ(cid:96)k(i)bk(i)
(142)
Now, we recall that following Assumption 1 and the considered
scenario 1k(cid:96)(i) = 1, we have 1k(cid:96)(t) = 1k(cid:96)(i) = 1 for t ≥ i.
As a result, the reputation update in (45) can be approximated
(cid:12)(cid:12)(cid:12)wk,t−1
(cid:105)
(cid:12)(cid:12)(cid:12)wk,t−1
(cid:105)(cid:17)
J act
k (a(cid:96)k(t) = 0)
− B(a(cid:96)k(t) = 1)bk(t)
where we introduced
χk (cid:44) 1 − δkrk
δk(1 − rk)
(cid:40)
The best response rule fk(·) therefore becomes
> χk
if γk(i) (cid:44) bk(i)
ck
θ(cid:96)k(i)
ak(cid:96)(i) =
1,
0, otherwise
17
k(cid:96)(t) and θ1
k(cid:96)(t) are given by:
by expression (43) for sufficiently small ε. Then, under (138),
the future reputation scores θ0
k(cid:96)(t) = θk(cid:96)(i)rt−i
θ0
k(cid:96)(t) = θk(cid:96)(i)rt−i
θ1
= θk(cid:96)(i)rt−i
k + (1 − rk)
k + (1 − rt−i
t−i−1(cid:88)
q=0
)
(143)
(144)
rq
k
k
k
Therefore, expression (142) becomes
)θ(cid:96)k(i)bk(i),
for t > i
(145)
Using (139) and (145), agent k then chooses ak(cid:96)(i) = 1 if
(cid:52)J 2
k
k (t) = −(1 − rt−i
∞(cid:88)
∞(cid:88)
δt−i
k
ck +
∞(cid:88)
t=i+1
δt−i
k
ck <
ck −
∞(cid:88)
t=i+1
t=i+1
δt−i
k
⇔
t=i
ck
1 − δk
⇔
⇔ γk(i) (cid:44) bk(i)
ck
< θ(cid:96)k(i)bk(i)δk ·
>
χk
θ(cid:96)k(i)
)θ(cid:96)k(i)bk(i) < 0
k
k
δt−i
(1 − rt−i
(1 − rt−i
(cid:18) 1
1 − δk −
k
)θ(cid:96)k(i)bk(i)
rk
1 − δkrk
(cid:19)
(146)
(147)
(148)
REFERENCES
[1] C.-K. Yu, M. van der Schaar, and A. H. Sayed, "Reputation design
for adaptive networks with selfish agents," in Proc. IEEE SPAWC,
Darmstadt, Germany, Jun. 2013, pp. 160 -- 164.
[2] P. Braca, S. Marano, and V. Matta, "Enforcing consensus while moni-
toring the environment in wireless sensor networks," IEEE Trans. Signal
Process., vol. 56, no. 7, pp. 3375 -- 3380, Jul. 2008.
[3] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione,
"Gossip algorithms for distributed signal processing," Proceedings of
the IEEE, vol. 98, no. 11, pp. 1847 -- 1864, Nov. 2010.
[4] S. Sardellitti, M. Giona, and S. Barbarossa, "Fast distributed average
consensus algorithms based on advection-diffusion processes," IEEE
Trans. Signal Process., vol. 58, no. 2, pp. 826 -- 842, Feb. 2010.
[5] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and coop-
eration in networked multi-agent systems," Proceedings of the IEEE,
vol. 95, no. 1, pp. 215 -- 233, Jan. 2007.
[6] S. Kar and J. M. F. Moura, "Distributed consensus algorithms in
sensor networks: Link failures and channel noise," IEEE Trans. Signal
Process., vol. 57, no. 1, pp. 355 -- 369, Jan. 2009.
[7] -- -- , "Convergence rate analysis of distributed gossip (linear parame-
ter) estimation: Fundamental limits and tradeoffs," IEEE J. Sel. Topics
in Signal Process., vol. 5, no. 5, pp. 674 -- 690, Aug. 2011.
[8] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multi-
agent optimization," IEEE Trans. Autom. Control, vol. 54, no. 1, pp.
48 -- 61, Jan. 2009.
[9] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging,"
Systems and Control Letters, vol. 53, no. 1, pp. 65 -- 78, Sep. 2004.
[10] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Randomized gossip
algorithms," IEEE Trans. Inf. Theory, vol. 52, no. 6, pp. 2508 -- 2530,
Jun. 2006.
[11] C. G. Lopes and A. H. Sayed, "Diffusion least-mean squares over
adaptive networks: Formulation and performance analysis," IEEE Trans.
Signal Process., vol. 56, no. 7, pp. 3122 -- 3136, Jul. 2008.
[12] A. H. Sayed, "Adaptive networks," Proceedings of the IEEE, vol. 102,
no. 4, pp. 460 -- 497, Apr. 2014.
[13] A. H. Sayed, S.-Y. Tu, J. Chen, X. Zhao, and Z. Towfic, "Diffusion
strategies for adaptation and learning over networks," IEEE Signal
Process. Mag., vol. 30, no. 3, pp. 155 -- 171, May 2013.
[14] A. H. Sayed, Adaptation, Learning, and Optimization over Networks.
Foundations and Trends in Machine Learning, vol. 7, no. 4-5, pp. 311-
801, NOW Publishers, Jul. 2014.
[15] -- -- , "Diffusion adaptation over networks," in Academic Press Library
in Signal Processing, R. Chellapa and S. Theodoridis, Eds., vol. 4, pp.
323 -- 454, Academic Press, Elsevier, 2014.
[16] F. S. Cattivelli and A. H. Sayed, "Diffusion LMS strategies for dis-
tributed estimation," IEEE Trans. Signal Process., vol. 58, no. 3, pp.
1035 -- 1048, Mar. 2010.
[17] S. Chouvardas, K. Slavakis, and S. Theodoridis, "Adaptive robust
distributed learning in diffusion sensor networks," IEEE Trans. Signal
Process., vol. 59, no. 10, pp. 4692 -- 4707, Oct. 2011.
[18] N. Takahashi and I. Yamada, "Link probability control for probabilistic
diffusion least-mean squares over resource-constrained networks," in
Proc. IEEE ICASSP, Dallas, TX, Mar. 2010, pp. 3518 -- 3521.
J. Chen and A. H. Sayed, "Diffusion adaptation strategies for distributed
optimization and learning over networks," IEEE Trans. Signal Process.,
vol. 60, no. 8, pp. 4289 -- 4305, Aug. 2012.
[19]
[20] O. N. Gharehshiran, V. Krishnamurthy, and G. Yin, "Distributed energy-
aware diffusion least mean squares: Game-theoretic learning," IEEE J.
Sel. Topics in Signal Process., vol. 7, no. 5, pp. 821 -- 836, Oct. 2013.
[21] C.-K. Yu, M. van der Schaar, and A. H. Sayed, "Distributed spectrum
sensing in the presence of selfish users," in Proc. IEEE CAMSAP, Saint
Martin, Dec. 2013, pp. 392 -- 395.
[22] A. Filali, A. Hafid, and M. Gendreau, "Bandwidth and computing
resources provisioning for grid applications and services," in Proc. IEEE
ICC, Dresden, Germany, Jun. 2009, pp. 1 -- 6.
[23] R. Bekkerman, M. Bilenko, and J. Langford, Scaling up Machine
Learning, Parallel and Distributed Approaches. Cambridge University
Press, 2011.
[24] D. Acemoglu and A. Ozdaglar, "Opinion dynamics and learning in
social networks," Dyn. Games Appl., vol. 1, no. 1, pp. 3 -- 49, Mar. 2011.
[25] A. Jadbabaie, P. Molavi, A. Sandroni, and A. Tahbaz-Salehi, "Non-
bayesian social learning," Game. Econ. Behav., vol. 76, no. 1, pp. 210 --
225, Sep. 2012.
[26] X. Zhao and A. H. Sayed, "Learning over social networks via diffusion
adaptation," in Proc. Asilomar Conference on Signals, Systems, and
Computers, Pacific Grove, CA, USA, Nov. 2012, pp. 709 -- 713.
[27] R. S. G Gigerenzer, Bounded Rationality: The Adaptative Toolbox.
MIT Press, 2002.
[28] D. Foster and R. Vohra, "Regret in the on-line decision problem," Games
and Economic Behavior, vol. 29, no. 1-2, pp. 7 -- 35, Oct. 1999.
[29] O. N. Gharehshiran, V. Krishnamurthy, and G. Yin, "Distributed track-
ing of correlated equilibria in regime switching noncooperative games,"
IEEE Trans. Autom. Control, vol. 58, no. 10, pp. 2435 -- 2450, Oct. 2013.
[30] G. J. Mailath and L. Samuelson, Repeated Games and Reputations:
[31]
Long-Run Relationships. Oxford University Press, 2006.
J. Xu and M. van der Schaar, "Social norm design for information
exchange systems with limited observations," IEEE J. Sel. Areas Com-
mun., vol. 30, no. 11, pp. 2126 -- 2135, Dec. 2012.
18
[32] Y. Zhang and M. van der Schaar, "Reputation-based incentive protocols
in crowdsourcing applications," in Proc. IEEE INFOCOM, Orlando,
Florida, USA, Mar. 2012, pp. 2140 -- 2148.
J. Carter, E. Bitting, and A. A. Ghorbani, "Reputation formalization for
an information-sharing multi-agent system," in Computational Intelli-
gence, vol. 18, no. 4, 2002, pp. 515 -- 534.
[33]
[34] Y. Yang, H. Hu, J. Xu, and G. Mao, "Relay technologies for wimax and
lte-advanced mobile systems," IEEE Commun. Mag., vol. 47, no. 10,
pp. 100 -- 105, Oct. 2009.
[35] C. Aliprantis, G. Camera, and D. Puzzello, "A random matching theory,"
Games Econ. Behav., vol. 59, pp. 1 -- 16, 2007.
[36] G. Ellison, "Cooperation in the prisoner's dilemma with anonymous
random matching," Rev. Econ. Stud., vol. 61, no. 3, pp. 567 -- 588, 1994.
[37] Y. Shoham and K. Leyton-Brown, Multiagent Systems: Algorithmic,
Game Theoretic and Logical Foundations. Cambridge University Press,
2008.
J. Xu, Y. Song, and M. van der Schaar, "Sharing in networks of strategic
agents," IEEE J. Sel. Topics in Signal Process., vol. 8, no. 4, pp. 717 --
731, Aug. 2014.
[38]
[39] H. A. Simon, "A behavioral model of rational choice," Quart. J. Econ.,
vol. 69, no. 1, pp. 99 -- 118, Feb. 1955.
[40] H. Park and M. van der Schaar, "On the impact of bounded rationality
in peer-to-peer networks," IEEE Signal Process. Lett., vol. 16, no. 8,
pp. 675 -- 678, Aug. 2009.
[41] M. J. Osborne and A. Rubinstein, A Course in Game Theory. Cam-
bridge, MA, 1994.
[42] A. Rapaport and A. Chammah, The Prisoners Dilemma. Univ. of
Michigan Press, 1965.
[43] R. Jurca and B. Faltings, "Eliciting truthful feedback for binary reputa-
tion mechanisms," in Proc. IEEE/WIC/ACM International Conference
on Web Intelligence, Beijing, China, Sep. 2004, pp. 214 -- 220.
[44] M. Fan, Y. Tan, and A. B. Whinston, "Evaluation and design of online
cooperative feedback mechanisms for reputation management," IEEE
Trans. Knowl. Data Eng., vol. 17, no. 2, pp. 244 -- 254, Feb. 2005.
[45] D. Houser and J. Wooders, "Reputation in auctions: Theory, and
evidence from ebay," J. Econom. Manage. Strat., vol. 15, no. 2, pp.
353 -- 369, Jun. 2006.
[46] A. Jøsang, R. Ismail, and C. Boyd, "A survey of trust and reputation
systems for online service provision," Decis. Support Syst., vol. 43,
no. 2, pp. 618 -- 644, Mar. 2007.
[47] A. H. Sayed, Adaptive Filters. Wiley, NJ, 2008.
[48] A. Nayyar, A. Gupta, C. Langbort, and T. Basar, "Common information
based Markov perfect equilibria for stochastic games with asymmetric
information: Finite games," IEEE Trans. Autom. Control, vol. 59, no. 3,
pp. 555 -- 570, Mar. 2014.
[49] G. Wyatt, "Risk-taking and risk-avoiding behavior -- The impact of
some dispositional and situational variables," Journal of Psychology,
vol. 124, pp. 437 -- 447, 1990.
[50] Ø. L. Rørtveit, J. H. Husøy, and A. H. Sayed, "Diffusion LMS with
communication constraints," in Proc. 44th Asilomar Conference on
Signals, Nov. 2010, pp. 1645 -- 1649.
J. Chen and A. H. Sayed, "Distributed Pareto optimization via diffusion
strategies," IEEE J. Sel. Topics in Signal Process., vol. 7, no. 2, pp.
205 -- 220, Apr. 2013.
[51]
[52] X. Zhao and A. H. Sayed, "Performance limits for distributed estimation
over LMS adaptive networks," IEEE Trans. Signal Process., vol. 60,
no. 10, pp. 5107 -- 5124, Oct. 2012.
J. Chen and A. H. Sayed, "On the learning behavior of adaptive
networks -- Part I: Transient analysis," IEEE Trans. Inf. Theory, vol. 61,
no. 6, pp. 3487 -- 3517, Jun. 2015.
[53]
[54] -- -- , "On the learning behavior of adaptive networks -- Part II:
Performance analysis," IEEE Trans. Inf. Theory, vol. 61, no. 6, pp.
3518 -- 3548, Jun. 2015.
19
[55] X. Zhao and A. H. Sayed, "Asynchronous adaptation and learning over
networks -- Part I: Modeling and stability analysis," IEEE Trans. Signal
Process., vol. 63, no. 4, pp. 811 -- 826, Feb. 2015.
[56] A. Papoulis and S. U. Pillai, Probability, Random Variables and
Stochastic Processes. McGraw-Hill, 2002.
Chung-Kai Yu (S'08) received the B.S. and MS
degrees in electrical engineering from the National
Taiwan University (NTU), Taiwan, in 2006 and 2008,
respectively.
From 2008 to 2011, he was a Research Assistant
in the Wireless Broadband Communication System
Laboratory at NTU. He is currently working toward
the Ph.D. degree in electrical engineering at
the
University of California, Los Angeles. His research
interests include game-theoretic learning, adaptive
networks, and distributed optimization.
Mihaela van der Schaar (F'10) is Chancellor's
Professor of Electrical Engineering at University of
California, Los Angeles. She is an IEEE Fellow since
2010, a Distinguished Lecturer of the Communica-
tions Society for 2011-2012, the Editor in Chief of
IEEE Transactions on Multimedia (2011 - 2013). She
received an NSF CAREER Award (2004), the Best
Paper Award from IEEE Transactions on Circuits and
Systems for Video Technology (2005), the Okawa
Foundation Award (2006), the IBM Faculty Award
(2005, 2007, 2008), the Most Cited Paper Award
from EURASIP: Image Communications Journal (2006), theGamenets Con-
ference Best Paper Award (2011) and the 2011 IEEE Circuits and Systems
Society Darlington Award Best Paper Award. She holds 33 granted US patents.
She is also the founding and managing director of the UCLA Center for Engi-
neering Economics, Learning, and Networks (see http://netecon.ee.ucla.edu).
For more information about her research visit: http://medianetlab.ee.ucla.edu/
Ali H. Sayed (S'90-M'92-SM'99-F'01) is professor
and former chairman of electrical engineering at the
University of California, Los Angeles, USA, where
he directs the UCLA Adaptive Systems Laboratory.
An author of more than 440 scholarly publications
and six books, his research involves several areas
including adaptation and learning, statistical signal
processing, distributed processing, network science,
and biologically inspired designs.
Dr. Sayed has received several awards including
the 2014 Athanasios Papoulis Award from the Euro-
pean Association for Signal Processing, the 2013 Meritorious Service Award,
and the 2012 Technical Achievement Award from the IEEE Signal Processing
Society. Also, the 2005 Terman Award from the American Society for Engi-
neering Education, the 2003 Kuwait Prize, and the 1996 IEEE Donald G. Fink
Prize. He served as Distinguished Lecturer for the IEEE Signal Processing
Society in 2005 and as Editor-in Chief of the IEEE TRANSACTIONS ON
SIGNAL PROCESSING (20032005). His articles received several Best Paper
Awards from the IEEE Signal Processing Society (2002, 2005, 2012, 2014).
He is a Fellow of the American Association for the Advancement of Science
(AAAS). He is recognized as a Highly Cited Researcher by Thomson Reuters.
|
1605.08883 | 1 | 1605 | 2016-05-28T12:24:27 | User-based solutions for increasing level of service in bike-sharing transportation systems | [
"cs.MA",
"physics.soc-ph"
] | Bike-sharing transportation systems have been well studied from a top-down viewpoint, either for an optimal conception of the system, or for a better statistical understanding of their working mechanisms in the aim of the optimization of the management strategy. Yet bottom-up approaches that could include behavior of users have not been well studied so far. We propose an agent-based model for the short time evolution of a bike-sharing system, with a focus on two strategical parameters that are the role of the quantity of information users have on the all system and the propensity of user to walk after having dropped their bike. We implement the model in a general way so it is applicable to every system as soon as data are available in a certain format. The model of simulation is parametrized and calibrated on processed real time-series of bike movements for the system of Paris. After showing the robustness of the simulations by validating internally and externally the model, we are able to test different user-based strategies for an increase of the level of service. In particular, we show that an increase of user information can have significant impact on the homogeneity of repartition of bikes in docking stations, and, what is important for a future implementation of the strategy, that an action on only 30% of regular users is enough to obtain most of the possible amelioration. | cs.MA | cs | User-based solutions for increasing level of
service in bike-sharing transportation systems
Juste Raimbault
Graduate School, Ecole Polytechnique, Palaiseau, France
and LVMT, Ecole Nationale des Ponts et Chauss´ees,
Champs-sur-Marne, France
[email protected]
Abstract. Bike-sharing transportation systems have been well studied
from a top-down viewpoint, either for an optimal conception of the sys-
tem, or for a better statistical understanding of their working mecha-
nisms in the aim of the optimization of the management strategy. Yet
bottom-up approaches that could include behavior of users have not been
well studied so far. We propose an agent-based model for the short time
evolution of a bike-sharing system, with a focus on two strategical pa-
rameters that are the role of the quantity of information users have on
the all system and the propensity of user to walk after having dropped
their bike. We implement the model in a general way so it is applicable to
every system as soon as data are available in a certain format. The model
of simulation is parametrized and calibrated on processed real time-series
of bike movements for the system of Paris. After showing the robustness
of the simulations by validating internally and externally the model, we
are able to test different user-based strategies for an increase of the level
of service. In particular, we show that an increase of user information
can have significant impact on the homogeneity of repartition of bikes
in docking stations, and, what is important for a future implementation
of the strategy, that an action on only 30% of regular users is enough to
obtain most of the possible amelioration.
Keywords: bike-sharing transportation system, agent-based modeling,
bottom-up complex system management
1
Introduction
Bike-sharing transportation systems have been presented as an ecological and
user-friendly transportation mode, which appears to be well complementary to
classic public transportation systems ([1]). The quick propagation of many im-
plementations of such systems across the world confirms the interesting poten-
tialities that bike-sharing can offer [2]. O'Brien & al. propose in [3] a review
on the current state of bike-sharing across the world. Inspired by the relatively
good success of such systems in Europe, possible key factors for their quality
have been questioned and transposed to different potential countries such as
China ([4,5]) or the United States ([6]).
6
1
0
2
y
a
M
8
2
]
A
M
.
s
c
[
1
v
3
8
8
8
0
.
5
0
6
1
:
v
i
X
r
a
2
Bike-sharing Agent-based Modeling
The understanding of system mechanisms is essential for its optimal exploita-
tion. That can be done through statistical analysis with predictive statistical
models ([7,8,9,10]) or data-mining techniques ([3,11]), and can give broader re-
sults such as structure of urban mobility patterns. Concerning the implementa-
tion, a crucial point in the design of the system is an optimal location of stations.
That problem have been extensively studied from an Operational Research point
of view ([12,13] for example). The next step is a good exploitation of the system.
By nature, strong asymmetries appear in the distribution of bikes: docking sta-
tions in residential areas are emptied during the day contrary to working areas.
That causes in most cases a strong decrease in the level of service (no parking
places or no available bikes for example). To counter such phenomena, operators
have redistribution strategies that have also been well studied and for which
optimal plans have been proposed ([14,15,16]).
However, all these studies always approach the problem from a top-down
point of view, in the sense of a centralized and global approach of the issues,
whereas bottom-up strategies (i. e. local actions that would allow the emergence
of desired patterns) have been to our knowledge not much considered in the
literature. User-based methods have been considered in [17,18] in the case of a
car-sharing system, but the problem stays quite far from a behavioral model of
the agents using the system, since it explores the possibility of implication of
users in the redistribution process, or of shared travels what is not relevant in
the case of bikes. Indeed the question of a precise determination of the influ-
ence of users behaviors and parameters on the level of service of a bike-sharing
systems remains open. We propose an agent-based model of simulation in order
to represent and simulate the system from a bottom-up approach, considering
bikers and parking as stations as agents and representing their interactions and
evolutions in time. That allows to explore user-targeted strategies for an increase
of the level of service, as the incitation to use online information media or to
be more flexible on the destination point. Note that our work aims to explore
effects of user-based policies, but does not pretend to give recommendations to
system managers, since our approach stays technical and eludes crucial political
and human aspects that one should take into account in a broader system design
or management context.
The rest of the paper is organized as follows. The model and indicator used
to quantify its behavior are described in Section 2. Next, Section 3 presents
the implementation and results, including internal and external validations of
the model by sensitivity analysis and simplified calibration on real data, and
also exploration of possible bottom-up strategies for system management. We
conclude by a discussion on the applicability of results and on possible develop-
ments.
2 Presentation of the model
Introduction The granularity of the model is the scale of the individual biker
and of the stations where bikes are parked. A more integrated view such as
Bike-sharing Agent-based Modeling
3
flows would not be useful to our purpose since we want to study the impact
of the behavior of individuals on the overall performance of the system. The
global working scheme consists in agents embedded in the street infrastructure,
interacting with particular elements, what is inspired from the core structure
of the Miro model ([19]). Spatial scale is roughly the scale of the district; we
don't consider the whole system for calculation power purposes (around 1300
stations on all the system of Paris, whereas an interesting district have around
100 stations), what should not be a problem as soon as in- and outflows allow to
reconstruct travels entering and getting out of the area. Tests on larger spatial
zones showed that generated travel were quite the same, justifying this choice of
scale. Focusing on some particular districts is important since issues with level
of service occur only in narrow areas. Time scale of a run is logically one full
day because of the cyclic nature of the process ([20]).
Formalisation The street network of the area
is an euclidian network (V ⊂ R2, E ⊂ V × V )
in a closed bounded part of R2. The time is
discretized on a day, so all temporal evolu-
tion are defined on T = [0, 24] ∩ τN with τ
time step (in hours). Docking stations S are
particular vertices of the network for which
constant capacities c(s ∈ S) are defined, and
that can contain a variable number of bikes
pb(s) ∈ {0, . . . , c}T . We suppose that tem-
poral fields O(x, y, t) and D(x, y, t) are de-
fined, corresponding respectively to probabili-
ties that a given point at a given time becomes
the expected departure (resp. the expected ar-
rival) of a new bike trip, knowing that a trip
starting (resp. arriving) at that time exists. Boundaries conditions are repre-
sented as a set of random variables (NI (i, t)). For each possible entry point i ∈ I
(I ⊂ V is a given set of boundaries points) and each time, NI (i, t) gives the
number of bikes trips entering the zone at point i and time t. For departures,
a random time-serie ND(t) represents the number of departures in the zone at
time t. Note that these random variables and probabilities fields are sufficient to
built the complete process of travel initiation at each time step. Parametrization
of the model will consist in proposing a consistent way to construct them from
real data.
Fig. 1: Flowchart of the decision
process of bikers, from the start of
their travel to the drop of the bike.
Docking stations are fixed agents, only their functions pb will vary through
time. The other core agents are the bikers, for which the set B(t) is variable. A
biker b ∈ B(t) is represented by its mean speed ¯v(b), a distance r(b) correspond-
ing to its "propensity to walk" and a boolean i(b) expressing the capacity of
having access to information on the whole system at any time (through a mobile
device and the dedicated application for example). The initial set of bikers B(0)
is taken empty, as t = 0 corresponds to 3a.m. when there is approximately no
travels on standard days.
Des$na$on?
Info
?
YES
NO
Free
sta$on
closest
To
objec$ve
Objec$ve
Realize
Travel
At
sta$on
?
NO
Find
sta$on
Within
radius
YES
Free
parking?
YES
Drop
bike
NO
Info
?
NO
Find
sta$on
randomly
YES
Go
to
closest
with
free
parking
and
drop
bike
4
Bike-sharing Agent-based Modeling
We define then the workflow of the model for one time step. The following
scheme is sequentially executed for each t ∈ T , representing the evolution of the
system on a day.
For each time step the evolution of the system follows this process :
– Starting new travels. For a travel within the area, if biker has information,
he will adapt his destination to the closest station of its destination with free
parking places, if not his destination is not changed.
• For each entry point, draw number of new traveler, associate to each a
destination according to D and characteristics (information drawn uni-
formly from proportion of information, speed according to fixed mean
speed, radius also).
• Draw new departures within the area according to O, associate either
destination within (in proportion to a fixed parameter pit, proportion of
internal travels) the area, or a boundary point (travel out of the area).
If the departure is empty, biker walks to an other station (with bikes if
has information, a random one if not) and will start his travel after a
time determined by mean walking speed and distance of the station.
• Make bikers waiting for start for which it is time begin their journey
(correspond to walkers for which a departure station was empty at a
given time step before)
– Make bikers advance of the distance corresponding to their speed. Travel
path is taken as the shortest path between origin and destination, as effective
paths are expected to have small deviation from the shortest one in urban
bike travels [8].
– Finish travels or redirect bikers
is finished (gets out of the area)
go to a random station within r(b)
• if the biker was doing an out travel and is on a boundary point, travel
• if has no information, has reached destination and is not on a station,
• if is on a station with free places, drop the bike
• if is on a station with no places, choose as new destination either the clos-
est station with free places if he has information, or a random one within
r(b) (excluding already visited ones, implying the memory of agents).
Fig. 1 shows the decision process for starting and arriving bikers. Note that
walking radius r(b) and information i(b) have implicitly great influence on the
output of the model, since dropping station is totally determined (through a
random process) by these two parameters when the destination is given.
Evaluation criteria In order to quantify the performance of the system, to com-
pare different realizations for different points in the parameter space or to evalu-
ate the fitness of a realization towards real data, we need to define some functions
of evaluation, proxies of what are considered as "qualities" of the system.
Bike-sharing Agent-based Modeling
5
Temporal evaluation functions These are criteria evaluated at each time step
and for which the output on the all shape of the time-series will be compared.
– Mean load factor ¯l(t) = 1S
– Heterogeneity of bike distribution: we aggregate spatial heterogeneity of load
factors on each station through a standard normalized heterogeneity indica-
(cid:80)
pb(s)
c(s)
s∈S
tor, defined by h(t) =
2
s(cid:54)=s(cid:48)∈S
1
d(s,s(cid:48) )
s(cid:54)=s(cid:48)∈S
(cid:80)
·(cid:80)
(cid:12)(cid:12)(cid:12)(cid:12) pb (s,t)
c(s) − pb (s(cid:48) ,t)
c(s(cid:48) )
d(s,s(cid:48))
(cid:12)(cid:12)(cid:12)(cid:12)
Aggregated evaluation functions These are criteria aggregated on a all day quan-
tifying the level of service integrated on all travels. We note T the set of travels
for a realization of the system and A the set of travel for which an "adverse
event" occured, i. e. for which a potential dropping station was full or a starting
station was empty. For any travel v ∈ T , we denote by dth(v) the theoretical
distance (defined by the network distance between origin and initial destination)
and dr(v) the effective realized distance.
– Proportion of adverse events: proportion of users for which the quality of
service was doubtful. A =
service Dtot = 1T ·(cid:80)
v∈T dr(v)
dth(v)
A
T
– Total quantity of detours: quantification of the deviation regarding an ideal
We also define a fitness function used for calibration of the model on real data.
If we note (lf (s, t))s∈S,t∈T the real time-series extracted for a standard day by
a statistical analysis on real data, we calibrate on the mean-square error on all
time-series, defined for a realization of the model by
M SE =
1
ST
(
pb(s, t)
c(s)
− lf (s, t))2
(cid:88)
(cid:88)
t∈T
s∈S
3 Results
3.1 Implementation and parametrization
Implementation The model was implemented in NetLogo ([21]) including GIS
data through the GIS extension. Preliminary treatment of GIS data was done
with QGIS ([22]). Statistical pre-treatment of real temporal data was done
in R ([23]), using the NL-R extension ([24]) to import directly the data. For
complete reproducibility, source code (including data collection scripts, sta-
tistical R code and NetLogo agent-based modeling code) and data (raw and
processed) are available on the open git repository of the project at http:
//github.com/JusteRaimbault/CityBikes.
Concerning the choice of the level of representation in the graphical inter-
face, we followed Banos in [25] when he argues that such exploratory models
can really be exploited only if a feedback through the interface is possible. It is
necessary to find a good compromise for the quantity of information displayed in
6
Bike-sharing Agent-based Modeling
Fig. 2: Example of the graphical output of the model for a particular district (Chatelet).
The map shows docking stations, for each the color gradient from green to red gives
the current loading factor (green : empty, red : full).
the graphical interface. In our case, we represent a map of the district, on which
link width is proportional to current flows, stations display their load-factor by
a color code (color gradient from green, lf (s) = 0, to red, lf (s) = 1). Bikes are
also represented in real time, what is interesting thanks to an option that allow
to follow some individuals and visualize their decision process through arrows
representing original destination, provenance and new destination (should be im-
plemented in further work). This feature could be seen as superficial at this state
of the work but it appears as essential regarding possible further developments
of the project (see discussion section). Fig. 2 shows an example of the graphical
interface of the implementation of the model of simulation.
Data collection All used data are open data, in order to have good reproducibility
of the work. Road network vector layer was extracted from OpenStreetMap
([26]). Time-series of real stations statuts for Paris were collected automatically1
all 5 minutes during 6 month and were imported into R for treatment with [27]
and the point dataset of stations was created from the geographical coordinates
with [28].
Parametrization The model was designed in order to have real proxies for most of
parameters. Mean travel speed is taken as ¯v =14km.h−1 from [29], where data of
trips where studied for the bike system of the city of Lyon, France. To simplify, we
1 from the dedicated website api.jcdecaux.com
Bike-sharing Agent-based Modeling
7
take same speed for all bikers : v(b) = ¯v. A possible extension with tiny gaussian
distribution around mean speed showed in experiments to bring nothing more. It
has been shown in [3] that profiles of use of bike systems stays approximatively
the same for european cities (but can be significantly different for cities as Rio or
Taipei), what justify the use of these inferred data in our case. We also use the
determined mean length of travel from [16] (here that parameter should be more
sensible to the topology so we prefer extract it from this second paper although
it seems to have subsequent methodological bias compared to the first rigorous
work on the system of Lyon), which is 2.3km, in order to determine the diameter
of the area on which our approach stays consistent. Indeed the model is built in
order to have emphasis on travels coming from the outside and on travels going
out, internal travels have to stay a small proportion of all travels. In our case,
a district of diameter 2km gives a proportion of internal travels pit ≈ 20%. We
will take districts of this size with this fixed proportion in the following.
The crucial part of the parametriza-
tion is the construction of O, D fields
and random variables NI , ND from
real data. Daily data were reduced
through sampling of time-series of
load-factors of all stations and dimen-
sion of the representation of a day
was significantly reduced through a
k-means clustering procedures (classi-
cally used in time-series clustering as
it is described in [30]). These reduced
points were then clustered again in or-
der to isolate typical weekdays from
week-ends, where the use profiles are
typically different and from special
days such as ones with very bad cli-
mate or public transportation strikes.
That allowed to create the profile of a
"standard day" that was used to infer
O, D fields through a spatial Gaussian
multi-kernel estimation (see [31]). The
characteristic size of kernels 1/σ is an
essential parameter for which we have
no direct proxy, and that will have to
be fixed through a calibration proce-
dure. The laws for NI , ND were taken
as binomial: for an actual arrival, we
consider each possible travel and in-
crease the number of drawing of each binomial law of entries by 1 at the time
corresponding to mean travel time (depending on the travel distance) before
arrival time. Probabilities of binomial laws are 1/Card(I) since we assume in-
Fig. 3: Statistical analysis of outputs.
For some aggregated outputs (here the
overall quantity of detours and the pro-
portion of adverse events), we plotted his-
tograms of the statistical distribution of
the functions on many realizations of the
model for a point in the parameter space.
Two points of the parameter space, corre-
sponding to (r = 300, pinf o = 50, σ = 80)
(green histogram) and (r = 700, pinf o =
50, σ = 80) (red) are plotted here as exam-
ples. Gaussian fits are also drawn. The rela-
tive good fit shows the internal consistence
of the model and we are able to quantify the
typical number of repetitions needed when
applying the model : supposing normal dis-
tributions for the indicator and its mean, a
95% confidence interval of size σ/2 is ob-
tained with n = (2 · 2σ·1.96/σ)2 ≈ 60
DetoursFrequency1.01.21.41.61.82.002468Adverse eventsFrequency0.000.050.100.150.2002468 8
Bike-sharing Agent-based Modeling
dependence of travels. For departure, we just increase by one drawings of the
binomial law at current time for an actual departure.
What we call parameter space in
the following consists in the 3 dimen-
sional space of parameters that have
not been fixed by this parametriza-
tion, i. e. the walking radius r (taken
as constant on all bikers, as for the
speed), the information proportion
pinf o what is the probability for a
new biker to have information and the
"size" of the Gaussian kernels σ (note
that the spread of distributions is de-
creasing with σ).
3.2 Robustness assessment,
exploration and calibration
Fig. 4: Simplified calibration procedure.
We plot the surface of the mean-square er-
ror on time-series of load-factors as a func-
tion of the two parameters on which we
want to calibrate. For visibility purpose,
only one surface was represented out of
the different obtained for different values
of walking radius. The absolute minimum
obtained for very large kernel has no sense
since such value give quasi-uniform proba-
bilities because of total recovering of Gaus-
sian kernels. We take as best realization the
second minimum, which is located around a
kernel size of 50 and a quantity of informa-
tion of 30%, which seem to be reasonable
values afterwards.
Internal consistence of the model Be-
fore using simulations of the model to
explore possible strategies, it is nec-
essary to assess that the results pro-
duced are internally consistent, i. e.
that the randomness introduced in the
parametrization and in the internal
rules do not lead to divergences in re-
sults. Simulations were launched on a
large number of repetitions for differ-
ent points in the parameter space and
statistical distribution of aggregated
outputs were plotted. Fig. 3 shows ex-
ample of these results. The relative
good gaussian fits and the small devi-
ation of distributions confirm the in-
ternal consistence of the model. We obtain the typical number of repetitions
needed to have a 95% confidence interval of length half of the standard devia-
tion, what is around 60, and we take that number in all following experiments
and applications. These experiments allowed a grid exploration of the parame-
ter space, confirming expected behavior of indicators. In particular, the shape
of M SE suggested to use the simplified calibration procedure presented in the
following.
Robustness regarding the study area The sensitivity of the model regarding ge-
ometry of the area was also tested. Experiments described afterwards were run
on comparable districts (Chatelet, Saint-Lazare and Montparnasse), leading to
the same conclusions, what confirms the external robustness of the model.
kernel.size20406080100info.proportion020406080100mse.lf.timeseries10500110001150012000MSE on lf−time−series Bike-sharing Agent-based Modeling
9
Reduced calibration procedure Using experiments launched during the grid ex-
ploration of the parameter space, we are able to assess or the regularity of some
aggregated criteria, especially of the mean-square error on loads factors of sta-
tions. We calibrate on kernel size and quantity of information. For different
values of the walking radius, the obtained area for the location of the minimum
of the mean-square error stays quite the same for reasonable values of the ra-
dius (300-600m). Fig. 4 shows an example of the surface used for the simplified
calibration. We extract from that the values of around 50 for kernel size and 30
for information proportion. The most important is kernel size since we cannot
have real proxy for that parameter. We use these values for the explorations of
strategies in the following.
3.3 Investigation of user-based strategies
Influence of walking radius Taking for kernel-size and quantity of information
the values given by the calibration, we can test the influence of walking radius
on the performance of the system. Note that we make a strong assumption, that
is that the calibration stay valid for different values of the radius. As we stand
previously, this stays true as soon as we stay in a reasonable range of values (we
obtained 300m to 600m) for the radius. The influence of variations of walking
radius on indicators were tested. Most interesting results are shown in figure
5. Concerning the indicators evaluated on time-series (h and ¯l(t)), it is hard
to have a significant conclusion since the small difference that one can observe
between curves lies inside errors bars of all curves. For A, we see a decreasing
of the indicator after a certain value (300m), what is significant if we consider
that radius under that value are not realistic, since a random place in the city
should be at least in mean over 300m from a bike station. However, the results
concerning the radius are not so concluding, what could be due to the presence
of periodic negative feedbacks: when the mean distance between two stations is
reached, repartitions concerns neighbor stations as expected, but the relation is
not necessarily positive, depending on the current status of the other station.
A deeper understanding and exploration of the behavior of the model regarding
radius should be the object of further work.
Influence of information For the quantity of information, we are on the contrary
able to draw significant conclusions. Again, behavior of indicators were studied
according to variations of pinf o. Most significant are shown on figure 6. Results
from time-series are also not concluding, but concerning aggregated indicators,
we have a constant and regular decrease for each and for different values of the
radius. We are able to quantify a critical value of the information for which
most of the progress concerning indicator A (adverse events) is done, that is
around 35%. We observe for this value an amelioration of 4% in the quantity of
adverse events, that is interesting when compared to the total number of bikers.
Regarding the management strategy for an increase in the level of service, that
implies an increase of the penetration rate of online information tools (mobile
10
Bike-sharing Agent-based Modeling
(a) Time series of heterogeneity indica-
tor h(t) for different values of walking
radius. Small differences between means
could mislead to a positive effect of radius
on heterogeneity, but the error bars of all
curves recover themselves, what makes any
conclusion non-significant.
(b) Influence of walking radius on the
quantity of adverse events A. After 400m,
we observe a relative decrease of the pro-
portion. However, values under 300-400m
should be ignored since these are smaller
than the mean distance of a random point
to a station.
Fig. 5: Results on the influence of walking radius.
application e. g.) if that rate is below 50%. If it is over that value, we have shown
that efforts for an increase of penetration rate would be pointless.
4 Discussion
4.1 Applicability of the results
We have shown that increases of both walking radius and information quantity
could have positive consequences on the level of service of the system, by reducing
the overall number of adverse events and the quantity of detours especially in the
case of the information. However, we can question the possible applicability of the
results. Concerning walking radius, first a deeper investigation would be needed
for confirmation of the weak tendency we observed, and secondly it appears that
in reality, it should be infeasible to play on that parameter. The only way to
approach that would be to inform users of the potential increase in the level of
service if they are ready to make a little effort, but that is quite optimistic to
think that they will apply systematically the changes, either because they are
egoistic, because they won't think about it, or because they will have no time.
Concerning the information proportion, we cannot also force users to have
information device (although a majority of population owns such a device, they
won't necessarily install the needed software, especially if that one is not user-
friendly). We should proceed indirectly, for example by increasing the ergonomics
of the application. An other possibility would to improve information displayed
at docking stations that is currently difficult to use.
0.30.40.50.60.70100200timeheterogeneity300400500600700800900radius1002003004005006007008000.0800.090radiusadverse Bike-sharing Agent-based Modeling
11
(a) Influence of proportion of information
on adverse events A for two different val-
ues of walking radius. We can conclude sig-
nificantly that the information has a pos-
itive influence. Quantitatively, we extract
the threshold of 35% that corresponds to
the majority of decrease, that means that
more is not necessarily needed.
(b) Influence of information on quantity of
detours Dtot. Curves for r = 300m and
r = 700m are shown (scale color). Here
also, the influence is positive. The effect
is more significant for high values of walk-
ing radius. The inflection is around 50%
of users informed, what is more than for
adverse events.
Fig. 6: Results on the influence of proportion of information.
4.2 Possible developments
Other possible management strategies Concerning user parameters, other choices
could have been made, as including waiting time at a fixed place, either for a
parking or a bike. The parameters chosen are both possible to influence and quite
adapted to the behavioral heuristic used in the model, and therefore were consid-
ered. Including other parameters, or changing the behavioral model such as using
discrete choice models may be possible developments of our work. Furthermore,
only the role of user was so far explored. The object of further investigation
could be the role of the "behavior" of docking stations. For example, one could
fix rules to them, as close all parkings over a certain threshold of load-factor, or
allow only departures or parkings in given configurations, etc. Such intelligent
agents would surely bring new ways to influence the overall system, but will also
increase the level of complexity (in the sense of model complexity, see [32]), and
therefore that extension should be considered very carefully (that is the reason
why we did not integrate it in this first work).
Towards an online bottom-up pilotage of the bike-sharing system Making the
stations intelligent can imply making them communicate and behave as a self-
adapting system. If they give information to the user, the heterogeneity of the
nature and quantity of information provided could have strong impact on the
overall system. That raises of course ethical issues since we are lead to ask if
it is fair to give different quantities of information to different users. However,
the perspective of a bottom-up piloted system could be of great interest from
a theoretical and practical point of view. One could think of online adaptive
algorithms for ruling the local behavior of the self-adapting system, such as ant
0.060.080.100.120255075100infoadverse300400500600700radius1.21.51.82.10255075100infodetours300400500600700radius 12
Bike-sharing Agent-based Modeling
algorithms ([33]), in which bikers would depose virtual pheromones when they
visit a docking station (corresponding to their information on travel that is easy
to obtain), that would allow the system to take some local decisions of redirecting
bikers or closing stations for a short time in order to obtain an overall better level
of service. Such methods have already been studied to improve level of service
in other public transportation systems like buses [34].
Conclusion
This work is a first step of a new bottom-up approach of bike-sharing systems.
We have implemented, parametrized and calibrated a basic behavioral model
and obtained interesting results for user-based strategies for an increase of the
level of service. Further work will consist in a deeper validation of the model,
its application on other data. We suggest also to explore developments such as
extension to other types of agents (docking stations), or the study of possible
bottom-up online algorithm for an optimal pilotage of the system.
References
1. Peter Midgley. The role of smart bike-sharing systems in urban mobility. JOUR-
NEYS, 2:23–31, 2009.
2. Paul DeMaio. Bike-sharing: History, impacts, models of provision, and future.
Journal of Public Transportation, 12(4):41–56, 2009.
3. Oliver O'Brien, James Cheshire, and Michael Batty. Mining bicycle sharing data
for generating insights into sustainable transport systems. Journal of Transport
Geography, 2013.
4. Zhili Liu, Xudong Jia, and Wen Cheng. Solving the last mile problem: Ensure
the success of public bicycle system in beijing. Procedia-Social and Behavioral
Sciences, 43:73–78, 2012.
5. Xue Geng, Kai TIAN, Yu ZHANG, and Qing LI. Bike rental station planning and
design in paris [j]. Urban Transport of China, 4:008, 2009.
6. Jonathan Gifford and Arlington Campus. Will smart bikes succeed as public trans-
portation in the united states? Center for Urban Transportation Research, 7(2):1,
2004.
7. Pierre Borgnat, Patrice Abry, and Patrick Flandrin. Mod´elisation statistique cy-
clique des locations de v´elo'v `a lyon. In XXIIe colloque GRETSI (traitement du sig-
nal et des images), Dijon (FRA), 8-11 septembre 2009. GRETSI, Groupe d'Etudes
du Traitement du Signal et des Images, 2009.
8. Pierre Borgnat, Eric Fleury, C´eline Robardet, Antoine Scherrer, et al. Spatial anal-
ysis of dynamic movements of v´elo'v, lyon's shared bicycle program. In European
Conference on Complex Systems 2009, 2009.
9. Pierre Borgnat, Patrice Abry, Patrick Flandrin, Jean-Baptiste Rouquier, et al.
Studying lyon's v´elo'v: a statistical cyclic model. In European Conference on Com-
plex Systems 2009, 2009.
10. Pierre Borgnat, Patrice Abry, Patrick Flandrin, C´eline Robardet, Jean-Baptiste
Rouquier, and Eric Fleury. Shared bicycles in a city: A signal processing and data
analysis perspective. Advances in Complex Systems, 14(03):415–438, 2011.
Bike-sharing Agent-based Modeling
13
11. Andreas Kaltenbrunner, Rodrigo Meza, Jens Grivolla, Joan Codina, and Rafael
Banchs. Urban cycles and mobility patterns: Exploring and predicting trends in a
bicycle-based public transport system. Pervasive and Mobile Computing, 6(4):455–
466, 2010.
12. Jenn-Rong Lin, Ta-Hui Yang, and Yu-Chung Chang. A hub location inventory
model for bicycle sharing system design: Formulation and solution. Computers &
Industrial Engineering, 2011.
13. Jenn-Rong Lin and Ta-Hui Yang. Strategic design of public bicycle sharing sys-
tems with service level constraints. Transportation research part E: logistics and
transportation review, 47(2):284–294, 2011.
14. Alvina GH Kek, Ruey Long Cheu, and Miaw Ling Chor. Relocation simulation
model for multiple-station shared-use vehicle systems. Transportation Research
Record: Journal of the Transportation Research Board, 1986(1):81–88, 2006.
15. Rahul Nair and Elise Miller-Hooks. Fleet management for vehicle sharing opera-
tions. Transportation Science, 45(4):524–540, 2011.
16. Rahul Nair, Elise Miller-Hooks, Robert C Hampshire, and Ana Busi´c. Large-scale
vehicle sharing systems: Analysis of v´elib'. International Journal of Sustainable
Transportation, 7(1):85–106, 2013.
17. Matthew Barth and Michael Todd. Simulation model performance analysis of a
multiple station shared vehicle system. Transportation Research Part C: Emerging
Technologies, 7(4):237–259, 1999.
18. Matthew Barth, Michael Todd, and Lei Xue. User-based vehicle relocation tech-
niques for multiple-station shared-use vehicle systems. TRB Paper No. 04-4161,
2004.
19. Arnaud Banos, Annabelle Boffet-Mas, Sonia Chardonnel, Christophe Lang, Nicolas
Marilleau, Thomas Th´evenin, et al. Simuler la mobilit´e urbaine quotidienne: le
projet miro. Mobilit´es urbaines et risques des transports, 2011.
20. Patrick Vogel, Torsten Greiser, and Dirk Christian Mattfeld. Understanding bike-
sharing systems using data mining: Exploring activity patterns. Procedia-Social
and Behavioral Sciences, 20:514–523, 2011.
21. U. Wilensky. Netlogo. Center for Connected Learning and Computer-Based Mod-
eling, Northwestern University, Evanston, IL., 1999.
22. QGIS Development Team. QGIS Geographic Information System. Open Source
Geospatial Foundation, 2009.
23. R Core Team. R: A Language and Environment for Statistical Computing. R
Foundation for Statistical Computing, Vienna, Austria, 2013.
24. Jan C Thiele, Winfried Kurth, and Volker Grimm. Agent-based modelling: Tools
for linking netlogo and r. Journal of Artificial Societies and Social Simulation,
15(3):8, 2012.
25. Arnaud Banos. Pour des pratiques de mod´elisation et de simulation lib´er´ees en
G´eographie et SHS. PhD thesis, UMR CNRS G´eographie-Cit´es, ISCPIF, D´ecembre
2013.
26. Jonathan Bennett. OpenStreetMap. Packt Publishing, 2010.
27. Alex Couture-Beil. rjson: Json for r. R package version 0.2, 13, 2013.
28. Timothy H Keitt, Roger Bivand, Edzer Pebesma, and Barry Rowlingson. rgdal:
bindings for the geospatial data abstraction library. R package version 0.7-1, URL
http://CRAN. R-project. org/package= rgdal, 2011.
29. Pablo Jensen, Jean-Baptiste Rouquier, Nicolas Ovtracht, and C´eline Robardet.
Characterizing the speed and paths of shared bicycle use in lyon. Transportation
research part D: transport and environment, 15(8):522–524, 2010.
14
Bike-sharing Agent-based Modeling
30. T Warren Liao. Clustering of time series data-a survey. Pattern Recognition,
38(11):1857–1874, 2005.
31. Alexandre B Tsybakov. Introduction to nonparametric estimation. (introduction
`a l'estimation non-param´etrique.). 2004.
32. Franck Varenne, Marc Silberstein, et al. Mod´eliser & simuler. Epist´emologies et
pratiques de la mod´elisation et de la simulation, tome 1. 2013.
33. Nicolas Monmarch´e. Algorithmes de fourmis artificielles: applicationsa la classifi-
cation eta l'optimisation. PhD thesis, ´Ecole Polytechnique, 2004.
34. Carlos Gershenson. Self-organization leads to supraoptimal performance in public
transportation systems. PLoS ONE, 6(6):e21469, 06 2011.
|
1603.08961 | 3 | 1603 | 2016-07-11T22:16:35 | Betting and Belief: Prediction Markets and Attribution of Climate Change | [
"cs.MA",
"physics.soc-ph"
] | Despite much scientific evidence, a large fraction of the American public doubts that greenhouse gases are causing global warming. We present a simulation model as a computational test-bed for climate prediction markets. Traders adapt their beliefs about future temperatures based on the profits of other traders in their social network. We simulate two alternative climate futures, in which global temperatures are primarily driven either by carbon dioxide or by solar irradiance. These represent, respectively, the scientific consensus and a hypothesis advanced by prominent skeptics. We conduct sensitivity analyses to determine how a variety of factors describing both the market and the physical climate may affect traders' beliefs about the cause of global climate change. Market participation causes most traders to converge quickly toward believing the "true" climate model, suggesting that a climate market could be useful for building public consensus. | cs.MA | cs |
Betting and Belief: Prediction Markets and Attribution of Climate
Change∗
John J. Nay†, Martin Van der Linden, Jonathan M. Gilligan
October 3, 2018
ABSTRACT
Despite much scientific evidence, a large fraction of the American public doubts that greenhouse gases are causing
global warming. We present a simulation model as a computational test-bed for climate prediction markets. Traders
adapt their beliefs about future temperatures based on the profits of other traders in their social network. We simulate
two alternative climate futures, in which global temperatures are primarily driven either by carbon dioxide or by solar
irradiance. These represent, respectively, the scientific consensus and a hypothesis advanced by prominent skeptics.
We conduct sensitivity analyses to determine how a variety of factors describing both the market and the physical
climate may affect traders' beliefs about the cause of global climate change. Market participation causes most traders
to converge quickly toward believing the "true" climate model, suggesting that a climate market could be useful for
building public consensus.
1 INTRODUCTION
The climate change debate has become strongly polarized over the past two decades. Although the scientific consensus
on the anthropogenic nature of climate change strongly increased, beliefs about climate change did not evolve much
within the public (Vandenbergh et al. 2014). In addition, the divide on anthropogenic climate change between liberals
and conservatives has grown steadily as the question is becoming increasingly politicized and potentially disconnected
from scientific evidence (Kahan et al. 2011). The costs of misinformed climate policies are high. If climate change is
not human-induced but is believed to be so, public resources will be spent on unnecessary efforts. On the other hand,
if climate change is human-induced but not recognized to be so, the costs of inaction could be devastating. Effective
climate policies require acting quickly, so it would be valuable to bring the public to a prompt and accurate consensus
on the issue.
Attempts to foster such consensus face many social and psychological challenges, some of which could be ad-
dressed by creating climate prediction markets where participants can "put their money where their mouths are" (Hsu
2011, Vandenbergh et al. 2014). The idea of using prediction markets to efficiently aggregate information about un-
certain event outcomes has been widely discussed (Horn et al. 2014). Prediction markets have interesting theoretical
properties (Set and Selten 1998, Hanson 2012), and perform well in terms of prediction accuracy and information
aggregation in experiments (Hanson et al. 2006, Healy et al. 2010) simulation models (Klingert and Meyer 2012,
Jumadinova and Dasgupta 2011) and the real-world (Wolfers and Zitzewitz 2006, Pathak et al. 2015, Dreber et al.
2015). However, to the best of our knowledge, the idea that prediction markets can generate consensus on the factors
affecting uncertain events has never been explored quantitatively.
Bloch, Annan, and Bowles (2010) have proposed using derivatives markets to reduce the scientific uncertainty in
estimating the impact of future climate change. Existing prediction markets (e.g. hypermind, betfair, and PredictIt)
focus on near term events such as elections months away, so it is difficult to extrapolate empirical findings to the
climate case. Furthermore, we are interested in investigating the unobservable beliefs of traders. Therefore, we turn
∗Copyright © 2016, IEEE.
†Corresponding Author: [email protected] and johnjnay.com.
1
to simulation modeling informed by climate and economic theory. We simulate a prediction market where traders
exchange securities related to climate outcomes to explore whether, and under what social and climate conditions,
prediction markets may be useful for increasing convergence of climate beliefs. Our work can be extended as part of
a computational design process for effective climate prediction markets and we release all our code along with this
paper.
From a public policy perspective, changing the explanatory models of market participants is one of the most
important roles that prediction markets might play. This is perhaps the most important social benefit of prediction
markets given the predictive power of statistical supervised learning models, which could possibly provide information
at a much lower cost than creating and maintaining a market (Goel et al. 2010). Effective climate policy not only
requires an accurate consensus on future climate outcomes.
It also requires an accurate consensus on the causal
mechanisms influencing such outcomes. If people agree that temperature will rise, but some believe it will be due to
greenhouse gases, while others believe that it will be caused by increased solar activity, inconsistent and ineffective
policies may be implemented.
2 RELATED WORK
Agent-based simulations of prediction markets have been studied (Klingert and Meyer 2012, Tseng et al. 2010, Ju-
madinova and Dasgupta 2011), including some that feature communication between agents. In these models, however,
beliefs about the uncertain outcomes are constructed in rather abstract ways. In particular, beliefs are not based on
structural models from which agents could derive causal implications, so these models are not suitable for investigating
the convergence of the underlying explanatory models agents employ for prediction.
Tseng et al. (2010) created an agent-based model (ABM) of a continuous double auction market with multiple
market strategies, including two variants of a zero-intelligence agent. They compared the behavior of their simulation
with data from a prediction market for the outcome of political elections. They found that despite their simplicity,
zero-intelligence agents capture some salient features of real market data. Klingert and Meyer (2012) compared
the predictive accuracy of different kinds of simulated markets (continuous double auctions and logarithmic market
scoring rules) and reached similar conclusions to experimental work by Hanson et al. (2006). None of these models
featured agents learning and updating their beliefs.
Jumadinova and Dasgupta (2011) created a continuous double auction model in which agents update their beliefs
about uncertain events based on newly-acquired information. The better the information set, the more likely the agents
were to put higher weight on last period's prices when revising their beliefs. However, little structure was imposed
on the information set and the way it was used to generate the weights for updating beliefs, so that model does not
permit studying the convergence of trader beliefs about predictive models. Ontan´on and Plaza (2009) studied the
effect of deliberation on a simulated prediction market, in which agents used case-based reasoning (Aamodt and Plaza
1994) to debate uncertain outcomes with their neighbors in a social network. This model could in principle be used
to assess convergence of predictive models, but this was not done and the process of forming and revising beliefs
through argumentation is not well suited to studying stochastic time series, such as those relevant to climate prediction
markets.
3 MODEL DESIGN
In our model, traders bet on global temperature anomalies six years in the future. During the six-year period, traders
buy and sell futures. Every year, traders update their models and forecast of future temperatures based on newly
available data. At the end of each six-year period, winners collect gains and traders revise beliefs about climate
models, based on their ideology and the beliefs of top earners in their social network. In this section, we describe
the models used to generate future temperature data, agents beliefs about those models, the market procedures, and
the social network connecting agents. We conclude with details on model dynamics and an overview of the model
parameters we experimentally vary. A full Overview, Design concepts and Details (ODD) specification can be found
on the project website.
2
3.1 Temperature Models
For climate time-series, we use the annual anomaly of global mean temperature. For years from 1880 -- 2014 we use
the GISTEMP global mean land-sea annual temperature anomalies (GISTEMP Team 2016, Hansen et al. 2010) and
for years from 2015 onward, we project future climates under two alternative theories: In both theories, changes in
global temperature are proportional to changes in a deterministic forcing plus a stochastic noise term. For simplicity,
we choose two alternative expressions for the deterministic climate forcing: one, which corresponds to conventional
climate science, takes the natural logarithm of the atmospheric carbon dioxide concentration in parts per million
(Archer 2012, p. 37), and the other, which corresponds to an alternative theory advocated by many who doubt or reject
conventional climate science, takes the total solar irradiance (the brightness of sunlight, in Watts per square meter, at
the top of the atmosphere) averaged over the 11-year sunspot cycle (Soon 2005). Most scientific models of the earth's
climate include many forcings, including carbon dioxide, other greenhouse gases, aerosols, total solar irradiance, and
more. In these models, the changing CO2 concentration is, by a large margin, the strongest single forcing (IPCC 2013,
p. 14). Choosing only one forcing term for each of the competing models simplifies comparison because each model
has the same number of adjustable parameters, therefore we consider only CO2.
For CO2 we used historical emissions through 2005, harmonized with RCP 8.5 representative concentration path-
way from 2005 onward (Kolp and Riahi 2009, Riahi et al. 2011) and for TSI we used the harmonized historical values,
with a projection through 2100 from Velasco Herrera et al. (2015), which is, to our knowledge, the only prediction of
TSI for the entire 21st century.
Warming coefficients for each model were determined by linear regression of historical temperatures from 1880 --
2014 against the historical values for each model's forcing term. The noise model was determined by fitting an
ARMA(p,q) model to the residuals from the regression, using the Stan probabilistic modeling language and the
rstan package (Carpenter et al. 2016). Stan proved more numerically stable than the R nlme package for fitting
ARMA noise models. We identified the optimal model for the auto-correlated noise term by performing the regression
analysis for all combinations of p,q ∈ {0,1,2} and using the Widely Applicable Information Criterion to select the
optimal noise model (Watanabe 2013, Gelman et al. 2014). In both cases (TSI and logCO2), the optimal noise model
was AR(1).
Future climates were generated by applying the future climate forcings to Eq. 1:
Tmodel(t) = βmodelFmodel(t) + ε,
(1)
where Tmodel(t) is the temperature at time t, under a given model of what causes warming, Fmodel(t) is the forcing (TSI
or logCO2) at time t, βmodel is the regression coefficient, and ε is a noise term. The coefficient β and the parameters
of the ARMA noise model were fit to the historical data (1880 -- 2014). An example realization of future climates for
both the ln(CO2) and TSI models is shown in Fig. 1.
3.2 Climate Beliefs
Traders use one of the two models (temperature depends on CO2 or TSI) to forecast future temperature. These models
are interpreted as the trader's beliefs about the true climate process, the driving factor of long-term global temperatures.
They represent pervasive positions on climate change in the public debate. In order to approximately match the current
configuration of beliefs in climate change in the United States, during model initialization, the CO2 model is randomly
assigned to half of the traders, while the TSI model is assigned to the other half of the traders. These random model
assignments are made before the market initially opens.
Both when the true data generating process is CO2 and TSI, at model initialization approximately half of the
traders use the true data-generating model to make predictions. Traders using the true model do not necessarily make
perfectly accurate predictions however. Although these traders believe in the correct functional form of the model,
they still need to calibrate their model based on limited noisy data. Therefore, the values these traders assign to the
parameters of the model will typically be different from the exact parameters in the data-generating process.
3.3 Traders and Markets
Traders are initially endowed with a single experimental currency unit (ECU). Traders use their model to forecast the
distribution of future temperature and determine their reservation price for different securities. Each security pays 1
ECU at the end of the trading sequence if the temperature at the end of the sequence falls into a certain range.
3
Figure 1: Historical measurements of temperature and a realization of possible future temperatures under two alternate
models of climate physics.
Traders are risk-neutral expected utility maximizers. Therefore, their reservation price for a security is simply
their assessment of the probability that the temperature will fall in the range covered by the security at the end of
the sequence. At each time-step (year), the agents use the new year's temperature data to re-estimate the coefficients
for the model they believe explains climate change, using Bayesian linear regression with an AR(1) noise model.
Traders use the joint posterior probability distribution of regression and noise coefficients to estimate the probability
distribution for the temperature at the end of the current trading sequence. Stan was sufficiently fast that we could
perform a full Bayesian analysis at each time step.
Traders then use this posterior probability distribution to assign reservation prices for buying and selling securities.
Based on their reservation price, agents behave as "zero-intelligence" traders (Gode and Sunder 1993). They attempt to
sell securities at a random price above their reservation, and to buy securities at a random price below their reservation.
These trading strategies are simple but provide accurate approximations of behavior in prediction markets (Klingert
and Meyer 2012), and in financial markets more broadly.
Based on traders' sell and buy orders, traders exchange securities following a continuous-double auction (CDA)
procedure (see Model Dynamics below for more details). CDAs or some close variants are common procedures to
match buy and sell orders. CDA are notably used on large stock markets (Tseng et al. 2010).
3.4 Social Network
Traders are part of a social network where each agent forms two links at random, and then forms links randomly,
ensuring that each agent is connected to at least two other agents. Every time securities are realized, each trader
looks at the performance of her richest neighbor in the network. Traders start with the same initial amount of ECU
and differences in ECU can only come from market interactions. Therefore, if some trader is poorer than her richest
neighbor, the trader interprets it as a signal that her richest neighbor may have a better model of the climate. Then,
the trader considers adopting the model of her richest neighbor. For each trader, the willingness to revise her belief
is determined by how ideologically loaded her belief is (Kahan et al. 2011), which is a parameter we vary in our
experiments.
An example of a snapshot of a social network is depicted in Fig. 2. A segmentation parameter controls the ho-
mophily of the network: the extent to which traders are preferentially linked with other traders who share their initial
belief in the cause of climate change (CO2 or TSI), as opposed to traders with the opposite intial belief. This parameter
varies from 0 (no preference for like-minded traders) to 1 (traders are only connected to like-minded traders). Although
traders can change their beliefs over time, the connections between traders do not change as the market unfolds, i.e the
4
0123419001950200020502100YearTemperature anomaly ((cid:176)C)True driver of future climateCO2TSIHistorical Figure 2: An example of the social network among traders. Yellow and blue circles represent traders who believe in
the two different models of climate change. Lines indicate social network connections. In a highly segmented network
(seg close to 1), most links connect like-minded traders and few connect traders with opposing beliefs.
edges are fixed. We vary the segmentation parameter in our experiments.
3.5 Model Dynamics
The time periods t are grouped into trading sequences. In a given sequence, the potential payments associated with
traded securities are all based on the temperature at the end of the sequence. For instance, the third trading sequence
might start in period t = 1964 and end in period t = 1970. In this case, a security traded in the third sequence pays 1
ECU if the temperature at t = 1970 falls into the range of temperatures covered by the security.
At each time t, traders are assumed to know the past value of the temperature T0:t, carbon dioxide CO20:t, and
In a sequence finishing at time t∗, traders also have common knowledge of CO2t:t∗
total solar irradiance TSI0:t.
and TSIt:t∗, the future values of carbon dioxide and total solar irradiance up to t∗. However, at any t, traders do
not know the value of any future temperatures. Thus, the traders know the forcing terms for the rest of the trading
sequence, but do not know the value of Tt∗. Traders can only predict Tt∗ using their approximate model and their
knowledge of T0:t, CO20:t∗, and TSI0:t∗. Notice that because T0:t, CO20:t∗, and TSI0:t∗ are common knowledge, in
each period t, any two traders with the same approximate model m form the same stochastic beliefs about future
temperatures Pt,m(Tt∗ T0:t ,CO20:t∗,TSI0:t∗). The probability distribution Pt,m incorporates both epistemic uncertainty
(the trader does not know the true values of the coefficients of the climate model) and aleatory uncertainty (in addition
to deterministic warming, the temperature exhibits stochastic noise that is directly modeled).
At each time t, traders:
1. re-calibrate m, their approximate model at time t, based on the new set of temperature data available at t,
2. use the posterior probability distribution from (1) to assign beliefs about the probability distribution of future
temperatures at time t∗: Pt,m(Tt∗ T0:t ,CO20:t∗,TSI0:t∗) and use it to determine the expected value they attach to
each security, and
3. trade on the CDA market as follows:
• Every trader i chooses at random a security sB
i she will try to buy.
5
• Every trader i also chooses at random a security sS
i she will try to sell among the securities she owns a
positive amount of (if any).
• Traders then decide of their selling price pS
expected values E(sB
i ) for securities sB
approximate model at time t). Then traders set pS
(see Model Parameters below for more details).
i ) and E(sS
i and buying price pB
i . To do so, traders first compute their
i (where expected values are with respect to i's
i at random below E(sB
i )
i and sS
i at random above E(sS
i ) and pB
• Traders go to the market one at the time, in an order drawn randomly for each t.
• When trader i comes to the market, she places limit orders in the order book. These orders specify that i is
willing to buy sB
i at any price below pB
i , and to sell sS
i at any price above pS
i .
• The market maker attempts to match i's orders with some order which was put in the book before i came
to the market.
• If there are outstanding sell offers for sB
i at a price below pB
i from the seller who sells at the lowest price below pB
i , a trade is concluded. Trader i buys one unit
i , and the sell and buy offers are removed from
of sB
the order book.
• If there are outstanding buy offers for sS
the buyer who buys at the highest price above pS
book.
i at price above pS
i , a trade is concluded. Trader i sells one unit to
i , and the sell and buy offers are removed from the order
• When all traders have come to the market, any remaining outstanding offer is removed from the order
book, and the trading period is concluded.
At t∗, when the sequence ends, there is only one security s∗ associated with a range of temperatures including the
actual temperature Tt∗. At t∗, traders:
1. receive 1 ECU per unit of s∗ they own, and
2. consider adopting their neighbors' approximate model as described in the behavioral parameters sub-section
below.
3.6 Model Parameters
The model depends on the following parameters, which we vary in simulation experiments to determine their effects
on the convergence of beliefs. We group the parameters into climate, network and individual behavioral factors.
• Climate parameters:
-- true.model: temperature data-generating process (CO2 or TSI).
• Network parameters:
-- n.traders: the number of traders.
-- n.edg: the number of edges in the social network.
-- seg: the segmentation parameter for the social network.
• Behavioral parameters:
-- risk.tak : determines the distribution of risk tolerance with respect to successfully buying or selling
securities. Higher risk tolerance corresponds to demanding more aggressive prices for buying and selling,
and hence, a higher risk of not consummating a trade that would be mutually advantageous to the buyer
and seller.
For each trader i, the level of risk.taki is drawn uniformly at random from [0, risk.tak]. The higher
risk.taki, the higher the price i will demand for selling securities, and the lower the price i will offer for
buying.
Formally, in each period, trader i picks her buying or selling price for s uniformly at random in the interval
[(1− risk.taki)reservi, reservi,t ] for buying and [reservi, (1 + risk.taki)reservi,t ] for selling.
6
Figure 3: Convergence over trading sequences for different degrees of social-network connection and segmentation.
-- ideo : determines the degree of "ideology" of traders. For each trader i, the level of ideoi is drawn
uniformly at random from [0, ideo]. If ideo is high, traders will not revise their approximate models
easily, even when faced with evidence that their richest neighbor is doing better than them. Formally, for
each trader i and each sequence, ideoi is the probability that i adopts the approximate model of her richest
neighbor if that neighbor is doing better than i at the end of the sequence (in monetary terms).
4 RESULTS
4.1 Historical Climate
As a validation that the market is operating correctly, we ran the market using actual historical temperatures from 1880
to 2014, with market betting from 1931 to 2014. Greenhouse gas concentrations did not become high enough to begin
dominating natural climate variations until the 1970s or so, and it took until the 1990s for the temperature record to
show clear signs of anthropogenic interference (IPCC 2013). Consistent with this, our simulated historical trading
sequence (Fig. 3) shows convergence to belief in the TSI model until the early 1970s, after which traders converge
toward believing the CO2 model.
7
2.4 edges per trader, segmentation = 0.052.4 edges per trader, segmentation = 0.959.6 edges per trader, segmentation = 0.059.6 edges per trader, segmentation = 0.95025507510002550751000255075100025507510019311950197520002014Year% traders believing CO2 model (n=2,688)Historical climate change1931−20142.4 edges per trader, segmentation = 0.052.4 edges per trader, segmentation = 0.959.6 edges per trader, segmentation = 0.059.6 edges per trader, segmentation = 0.95025507510002550751000255075100025507510020152025205020752100Year% traders believing true model (n = 5,376)True driver of future climateCO2TSI2015−2098 4.2 Future Scenarios
Our primary focus is on simulating the market in future scenarios. This is more relevant for policy design and more
interesting theoretically due to the increasing divergence of global temperature values under the two future scenarios.
The sensitivity analysis is based on a Latin hypercube sampling of 500 parameter sets from the following distributions
(Beachkofski and Grandhi 2002, Carnell 2012).
• ideo ∼ Uniform(0,1)
• n.edge ∼ Uniform(100,200) (mapped into integer)
• n.traders ∼ Uniform(50,250) (mapped into integer)
• risk.tak ∼ Uniform(0,1)
• seg ∼ Uniform(0,1)
• true.model ∼ Bernoulli(0.5)
The future CO2 forcing is taken from the RCP 8.5 scenario (Riahi et al. 2011) and the future TSI forcing is taken
from a prediction of 21st Total Solar Irradiance by Velasco Herrera et al. (2015). Because of the stochastic noise term,
the temperature time series were different in each simulation.
We used the model to perform 10 full simulations for each of the 500 input parameter sets and average the 10
convergence scores. We conducted multiple simulations for the same parameter set because there is stochasticity in the
tempertature time series, the social network structure, and the agent decision models. We then conducted a partial rank
correlation coefficient analysis on the relationship between the input matrix, X, and the resulting simulated outcome
vector of mean belief convergence scores, y (Marino et al. 2008, Pujol et al. 2014, Saltelli et al. 2009). Partial
correlation computes the linear relationship between the part of the variation of Xi and y that are linearly independent
of other Xj ( j (cid:54)= i). The difference between the partial correlation and the partial rank correlation that we use here
is that we first rank-transform that data in order to capture potentially non-linear relationships. We conduct 1,000
bootstrapped estimations of the partial rank correlation coefficients to obtain 95% confidence intervals.
The sensitivity analysis averages over time and thus masks time trends. We randomly drew from the above dis-
tributions for ideology, risk tolerance, and the number of traders, and conducted an experiment crossing 95th and 5th
percentile values for the number of edges per trader and the segmentation of the social network and both values for the
true model. We collected the time series of belief-convergence across these eight designs to visualize the distributions
of convergence over time (Fig. 3). Under most parameterizations, the median fraction of traders believing in the true
model reaches 75% in 10 -- 20 years.
5 DISCUSSION
Our sensitivity analysis, Fig. 4, shows that ideology, risk tolerance, the number of traders, the number of edges per
trader, the segmentation of the social network, and the true model all statistically significantly affect convergence.
Increasing the number of edges per trader increases the flow of information through the market, which causes traders
to converge toward believing the "true" climate model.
Segmented social networks reduce convergence by creating an "echo-chamber" effect, in which lack of interac-
tion between traders with different views reduces traders' access to information that could persuade them to change
their beliefs. This is apparent in Fig. 3, where the rate of convergence in highly segmented markets (seg = 0.95) is
considerably slower, especially in the first three trading sequences (18 years), than for markets with low segmentation
(seg = 0.05).
When the true model is CO2-driven there is more convergence toward the true model. We believe this is because
when models are fit to the historical data, the residuals from the CO2-driven model are considerably smaller than those
from the TSI-driven model. Thus, projections of future climate change using the TSI model will exhibit significantly
larger stochastic noise than the CO2 models, and this noise makes it more difficult for traders to identify the true model
when the true model is TSI.
Traders with higher risk tolerance will price buy and sell orders more aggressively, so they will earn more (or
lose less) on completed trades, but will have a greater risk of failing to complete trades. Among traders with the
8
Figure 4: Estimated effects of model parameters on convergence of beliefs in future scenario. Positive numbers
accelerate convergence and negative numbers retard it. True model refers to the true climate model being CO2, so
traders converge to the true model more quickly when carbon dioxide drives climate change than when total solar
irradiance does. Bars represent 95% confidence intervals.
correct model of climate, those with higher risk tolerance will earn more profit on completed trades (and their coun-
terparties will lose more). A trader's wealth is an important source of information, and we believe this is why greater
risk taking enables traders to identify the correct model more quickly. Additionally, the risk of failing to complete
trades adds volatility, which also puts information in play. We have observed a similar phenomenon in a very differ-
ent context, where adding stochastic noise to player decisions in iterated games improves the accuracy with which
machine-learning algorithms can identify the players' strategies (Nay and Gilligan 2015).
What is not clear to us is why increasing the number of traders slows convergence, even if the number of edges per
trader remains fixed. This is a topic for future research.
6 CONCLUSIONS
We simulate two alternative climate futures: one where CO2 is the primary driver of global temperature and one where
variations in solar intensity are the primary driver. These represent the two most plausible competing views in the
public discourse and our analysis is agnostic about which is "true".
Market participation causes traders to converge toward believing the "true" climate model under a variety of model
parameterizations in a relatively short time: In markets with low segmentation, the number of traders believing the
true model rises from 50% to 75% in roughly 12 years, and even in highly segmented markets, 75% convergence
is achieved in 18 -- 24 years.
Ideally, we would like to compare belief convergence with and without a prediction
market but because the only source of climate belief in our current model is market interactions we cannot make
this comparison. However, we do use actual temperatures for our model under historical 20th century conditions and
observe convergence to the CO2 model, whereas in the real world there is no convergence of beliefs.
Both the historical and future simulation results suggest that a climate prediction market could be useful for pro-
ducing broad agreement about the causes of climate change, and could have persuasive power for people who are
not persuaded by the overwhelming consensus among climate researchers that greenhouse gases (and especially CO2)
are responsible for the majority of observed climate change (IPCC 2013). We also find that market segmentation has
9
−1.0−0.50.00.51.0IdeologyRisk tolerance# Traders# Edges/TraderSegmentationTrue modelPartial Rank Correlation Coefficient a large effect on the speed of convergence, so transparency and effective communication about the performance of
traders with different beliefs will be important.
The fact that rapid convergence to the true model occurs regardless of which model is actually "true" may persuade
those who doubt the scientific consensus that the market is ideologically neutral, and that the deck is not stacked to
produce a pre-determined result.
All code and data for the model is available at github.com/JohnNay/predMarket. This project is a computational
test-bed for public policy design: our code can be extended to test the effects of trading strategies, cognitive models,
future climate scenarios, and market designs on the evolution of trader beliefs.
ACKNOWLEDGEMENTS
We thank Yevgeniy Vorobeychik for feedback on this project and Manuel Naumovich Velasco for sharing the data
from his Total Solar Irradiance prediction. U.S. National Science Foundation grants EAR-1416964, EAR-1204685,
and IIS-1526860 partially supported this research. All views expressed are the authors' and not the funder's.
References
Aamodt, A., and E. Plaza. 1994. "Case-Based Reasoning: Foundational Issues, Methodological Variations, and System
Approaches". AI Communications 7:39 -- 59.
Archer, D. 2012. Global Warming: Understanding the Forecast. 2nd ed. Wiley.
Beachkofski, B.,
and R. Grandhi.
43rd
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference. American Insti-
tute of Aeronautics and Astronautics.
"Improved Distributed Hypercube Sampling".
2002.
In
Bloch, D., J. Annan, and J. Bowles. 2010. "Cracking the Climate Change Conundrum with Derivatives". Wilmott
Journal 2:271 -- 287.
Carnell, R. 2012. "Latin Hypercube Samples". R package, Comprehensive R Archive Network.
Carpenter, B., A. Gelman, M. Hoffman, D. Lee, B. Goodrich, M. Betancourt, M. A. Brubaker, J. Guo, P. Li, and
A. Riddell. 2016. "Stan: A Probabilistic Programming Language". Journal of Statistical Software. (in press).
Dreber, A., T. Pfeiffer, J. Almenberg, S. Isaksson, B. Wilson, Y. Chen, B. A. Nosek, and M. Johannesson. 2015. "Using
Prediction Markets to Estimate the Reproducibility of Scientific Research". Proceedings of the National Academy
of Sciences 112:15343 -- 15347.
Gelman, A., J. Hwang, and A. Vehtari. 2014. "Understanding Predictive Information Criteria for Bayesian Models".
GISTEMP Team 2016. "GISS Surface Temperature Analysis (GISTEMP)". Technical report, NASA Goddard Institute
Statistics and Computing 24:997 -- 1016.
for Space Studies.
Gode, D. K., and S. Sunder. 1993. "Allocative Efficiency of Markets with Zero-Intelligence Traders: Market as a
Partial Substitute for Individual Rationality". Journal of Political Economy 101:119 -- 137.
Goel, S., D. M. Reeves, D. J. Watts, and D. M. Pennock. 2010. "Prediction Without Markets". In Proceedings of the
11th ACM Conference on Electronic Commerce, EC '10, 357 -- 366. New York, NY, USA: ACM.
Hansen, J., R. Ruedy, M. Sato, and K. Lo. 2010. "Global Surface Temperature Change". Reviews of Geo-
Hanson, R. 2012. "Logarithmic Market Scoring Rules for Modular Combinatorial Information Aggregation". The
physics 48:RG4004.
Journal of Prediction Markets:3 -- 15.
Hanson, R., R. Oprea, and D. Porter. 2006. "Information Aggregation and Manipulation in an Experimental Market".
Journal of Economic Behavior & Organization 60:449 -- 459.
Healy, P. J., S. Linardi, J. R. Lowery, and J. O. Ledyard. 2010. "Prediction Markets: Alternative Mechanisms for
Complex Environments with Few Traders". Management Science 56:1977 -- 1996.
Horn, C. F., B. S. Ivens, M. Ohneberg, and A. Brem. 2014. "Prediction markets -- A literature review 2014". The
Journal of Prediction Markets 8:89 -- 126.
Hsu, S.-L. 2011. "A Prediction Market for Climate Outcomes". University of Colorado Law Review 83:179 -- 256.
IPCC 2013. Climate Change 2013: The Physical Science Basis. Cambridge University Press.
Jumadinova, J., and P. Dasgupta. 2011. "A Multi-Agent System for Analyzing the Effect of Information on Prediction
Markets". International Journal of Intelligent Systems 26:383 -- 409.
10
Kahan, D. M., H. Jenkins-Smith, and D. Braman. 2011. "Cultural Cognition of Scientific Consensus". Journal of Risk
Research 14:147 -- 174.
Klingert, F. M., and M. Meyer. 2012. "Comparing Prediction Market Mechanisms Using An Experiment-Based Multi-
Agent Simulation.". In Proceedings 26th European Conference on Modelling and Simulation, edited by K. G.
Troitzch, Mohring, and U. Lotzmann, 654 -- 661.
Kolp, P., and K. Riahi. 2009. "RCP Database". Technical report.
Marino, S., I. B. Hogue, C. J. Ray, and D. E. Kirschner. 2008. "A Methodology for Performing Global Uncertainty
and Sensitivity Analysis in Systems Biology". Journal of Theoretical Biology 254:178 -- 196.
Nay, J. J., and J. M. Gilligan. 2015. "Data-Driven Dynamic Decision Models". In Proceedings of the 2015 Winter
Simulation Conference, edited by L. Yilmaz, W. Chan, I. Moon, T. Roeder, C. Macal, and M. Rosetti, 2752 -- 2763:
IEEE Press.
Ontan´on, S., and E. Plaza. 2009. "Argumentation-Based Information Exchange in Prediction Markets". In Argumen-
tation in Multi-Agent Systems, edited by I. Rahwan and P. Moraitis, Number 5384 in Lecture Notes in Computer
Science, 181 -- 196. Springer Berlin Heidelberg.
Pathak, D., D. Rothschild, and M. Dudik. 2015. "A Comparison of Forecasting Methods: Fundamentals, Polling,
Prediction Markets, and Experts". Journal of Prediction Markets 9:1 -- 31.
Pujol, G., B. Iooss, and A. J. Lemaitre. 2014. "Sensitivity: Sensitivity Analysis". R package, Comprehensive R Archive
Network.
nomics 1:43 -- 62.
Riahi, K., S. Rao, V. Krey, C. Cho, V. Chirkov, G. Fischer, G. Kindermann, N. Nakicenovic, and P. Rafaj. 2011. "RCP
8.5: A Scenario of Comparatively High Greenhouse Gas Emissions". Climatic Change 109:33 -- 57.
Saltelli, A., K. Chan, and E. M. Scott. 2009. Sensitivity Analysis. Chichester: Wiley.
Set, R., and R. Selten. 1998. "Axiomatic Characterization of the Quadratic Scoring Rule". Experimental Eco-
Soon, W. W.-H. 2005. "Variable Solar Irradiance as a Plausible Agent for Multidecadal Variations in the Arctic-wide
Surface Air Temperature Record of the Past 130 Years". Geophysical Research Letters 32:L16712.
Tseng, J.-J., C.-H. Lin, C.-T. Lin, S.-C. Wang, and S.-P. Li. 2010. "Statistical Properties of Agent-based Models
in Markets with Continuous Double Auction Mechanism". Physica A: Statistical Mechanics and its Applica-
tions 389:1699 -- 1707.
Vandenbergh, M. P., K. T. Raimi, and J. M. Gilligan. 2014. "Energy and Climate Change: A Climate Prediction
Market". UCLA Law Review 61:1962 -- 2017.
Velasco Herrera, V. M., B. Mendoza, and G. Velasco Herrera. 2015. "Reconstruction and Prediction of the Total Solar
Irradiance: From the Medieval Warm Period to the 21st century". New Astronomy 34:221 -- 233.
Watanabe, S. 2013. "A Widely Applicable Bayesian Information Criterion". Journal of Machine Learning 14:867 -- 897.
Wolfers, J., and E. Zitzewitz. 2006. "Prediction Markets in Theory and Practice". Working Paper 12083, NBER.
11
|
1911.04646 | 1 | 1911 | 2019-11-12T03:07:36 | LAC-Nav: Collision-Free Mutiagent Navigation Based on The Local Action Cells | [
"cs.MA",
"cs.AI"
] | Collision avoidance is one of the most primary requirement in the decentralized multiagent navigations: while the agents are moving towards their own targets, attentions should be paid to avoid the collisions with the others. In this paper, we introduce the concept of local action cell, which provides for each agent a set of velocities that are safe to perform. Based on the realtime updated local action cells, we propose the LAC-Nav approach to navigate the agent with the properly selected velocity; and furthermore, we coupled the local action cell with an adaptive learning framework, in which the effect of selections are evaluated and used as the references for making decisions in the following updates. Through the experiments for three commonly considered scenarios, we demonstrated the efficiency of the proposed approaches, with the comparison to several widely studied strategies. | cs.MA | cs |
LAC-Nav: Collision-Free Mutiagent Navigation
Based on The Local Action Cells
Li Ning and Yong Zhang
Shenzhen Institutes of Advanced Technology, Chinese Academy of Science
{li.ning, zhangyong}@siat.ac.cn
July 11, 2021
Abstract
Collision avoidance is one of the most primary problems in the decentralized
multiagent navigation: while the agents are moving towards their own targets, at-
tentions should be paid to avoid the collisions with the others. In this paper, we
introduced the concept of the local action cell, which provides for each agent a
set of velocities that are safe to perform. Consequently, as long as the local action
cells are updated on time and each agent selects its motion within the correspond-
ing cell, there should be no collision caused. Furthermore, we coupled the local
action cell with an adaptive learning framework, in which the performance of se-
lected motions are evaluated and used as the references for making decisions in
the following updates. The efficiency of the proposed approaches were demon-
strated through the experiments for three commonly considered scenarios, where
the comparisons have been made with several well studied strategies.
1
Introduction
Collision-free navigation is a fundamental and important problem in the design of the
multiagent systems, which are widely applied in the fields such as robots control and
traffic engineering. When moving the agents in an environment with static or dynamic
obstacles, it is usually a necessary requirement to well plan the trajectories such that
no collision is caused. As the number of agents increases and the environment area
becomes large, planning the realtime motions for all agents in the centralized manner
causes huge amount of the calculations, and is often restricted by the efficiency of the
communication between the agent and the planning monitor. Therefore, it is natural
(sometimes necessary) to consider the decentralized navigation approaches, by which
the individual agent is responsible for sensing the nearby obstacles and performing
the proper motion to progress towards its destination without causing any collisions.
On the other hand, as a consequence of the decentralized navigation, it is in general
difficult for the agents to fully coordinate before making the independent moves. Thus
they should also be considered and avoided as the obstacles to each other.
1
As noticed in the existing works, when avoiding collisions with the other agents,
it is important to take into account the fact that they are also intelligent to perform
the collision avoidance motions (otherwise, undesirable oscillations may be observed
during the navigation). Consequently, it is not necessary for any individual agent to
take all the responsibility of making sure that the performed motion is safe. ORCA
[9] is a well known decentralized approach that guarantees to generate the optimal
reciprocal collision-free velocities, except for some certain conditions with densely
packed agents. BVC [11] has been proposed to restrict the agents moving inside the
non-intersecting areas, and thus the collision avoidance is guaranteed. After the safe
field of the motions (i.e. the safe range of velocities or the safe area of positions) is
determined, both of the ORCA-based approaches and the BVC-based approaches usu-
ally select the motion that is closest to the preferred motion, within the safe field. Such
a greedy strategy is natural and widely used in the local-search-based optimizations.
However, it may cause the less efficient performance in the multiagent navigation, as
the agents may refuse to detour until there is no chance to approach the target. In the
worst case, with the greedy selection, agents may get stuck in a loop of two or more
situations (also known as the deadlocks). Although some tricks have been proposed to
fix such drawbacks (including the ideas described in [11]), they are not always valid in
the concrete implementations, and the improvements vary from case to case.
In this work, in order to improve the navigation efficiency, we extend the buffered
Voronoi cell [11] in the velocity space, and consider the relative velocities for their
effects on causing the potential conflicts. In the selection of the motion to perform, the
traveling progress has been also considered, and consequently the agents may detour
earlier, as long as approaching directly to the target leads to less progress in the moving
distance.
In this work, we consider a set A of the disk-shaped agents
Problem formation.
moving in the plane. For any time point, agent ai ∈ A of position pi ∈ R2 is free
to change its velocity vi ∈ R2, and after a short time δ > 0, it moves to pi + δ ·
vi, if there is no collisions between the agents (i.e. the distance between any pair of
agents is at least the sum of their radii). For a decentralized navigation approach, it
runs independently for each individual agent ai, and based on the observations of the
environment, it updates the velocity in order to guide agent ai to arrive at the given
and fixed destination/target di ∈ R2. On the measure of the approach's performance,
we want all the agents arriving at their destinations/targets as soon as possible, without
causing any collisions.
Our contributions. We introduced the concept of the local action cell to specify
the underlying choices for the selection of the motion to perform, and proposed two
approaches (LAC-Nav and LAC-Learn) that guarantee to provide the collision-free
navigations. While the LAC-Nav approach simply perform the action of the largest pe-
nalized length (among all choices in the local action cell), the LAC-Learn approach
evaluates the performed actions and adjust the selection based on an adaptive learning
framework. The experiment results have shown that the proposed approaches perform
more efficiently in the completion time (formally defined in the section of "Experi-
2
ments"), compared to several well studied approaches.
Related works. The velocity-based collision-free navigation have been extensively
studied in the last two decades. The idea of reciprocal velocity obstacles (RVO, [10])
was introduced to reduce the problem of calculating the collision-free motion to solv-
ing a low-dimensional linear program, based on the definition of velocity obstacles
[3], and it was further improved to derive the optimal reciprocal collision avoidance
(ORCA, [9]) framework, which guarantees the optimal reciprocal collision-free mo-
tions, except for some certain conditions with densely packed agents. While the safety
of the final motion is guaranteed by ORCA, the ALAN [5] online learning framework
has been proposed for adapting the preferred motions of multiple agents without the
need for offline training; and the CNav [6] is designed to allow the agents to take the
others' preferred motion into account and adjust accordingly to achieve the better co-
ordination in the crowd environments. Notice that although the efficiency of CNav has
been demonstrated through the experiments, it requires the the spreading of some pri-
vate information of the agents, such as their preferred motions or their targets, which
is often a controversial issue in the practical applications.
As the well known Voronoi diagram can be used to divide the working space into
non-intersecting areas, it has been also adopted for the collision-free path planning with
multiple robots [4, 2]. Inspired by the algorithms for the coverage control of the agents
[7], and a Voronoi-cell-based algorithm [1] which is introduced to avoid collisions
within a larger probabilistic swarm, the buffered Voronoi cell (BVC, [11]) approach
has been proposed to achieve the collision avoidance guarantee for the multiagent nav-
igation, based on only the information of the positions. With the up-to-date information
of the others' positions, the agents are restricted to move in the non-intersecting areas,
and thus there should be no collisions. In [8], a trajectory planning algorithm was pro-
posed to navigate the agents under the higher-order dynamic limits, in which BVC is
used as the low-level strategy to avoid collisions.
2 The Local Action Cells
In this paper, we assume that all the agents in A have the same radius r for the simplic-
ity of the argument (for the case when the agents have different radii, the arguments in
this paper can be directly extended by substituting the classical Voronoi diagram with
its weighted variant). Thus for any time and any pair of non-colliding agents ai and aj,
it always holds that (cid:107)pij(cid:107)2 ≥ 2r, where pij stands for pj − pi.
Recall that in [11], the buffered Voronoi cell of agent ai is defined as
+ r ≤ 0,∀j (cid:54)= i
p ∈ R2
¯Vi =
·
,
(cid:27)
(cid:26)
(cid:19)
p − pi + pj
(cid:18)
Di =(cid:8)v ∈ R2pi + δ · v ∈ ¯Vi
pij
(cid:107)pij(cid:107)2
2
(cid:9) ,
which implies a safe velocity domain
for agent ai to change and maintain its velocity in order to reach a point in ¯Vi, where δ ∈
R+ is the length of the time interval between two consecutive updates. Equivalently,
3
domain Di can be presented as
(cid:26)
Di =
v ∈ R2 v · uij ≤ (cid:107)pij(cid:107)2 − 2r
2δ
(cid:27)
,∀j (cid:54)= i
,
where uij is the unit vector along the same direction with pj−pi, i.e. uij = pij/(cid:107)pij(cid:107)2.
Obviously, domain Di is the intersection of the half-planes Pij's for each agent aj (cid:54)=
ai, with
(cid:27)
.
Assuming that agent ai is moving at velocity vi and agent aj is moving at velocity vj,
we estimate the colliding risk by calculating
(cid:26)
Pij =
v ∈ R2 v · uij ≤ (cid:107)pij(cid:107)2 − 2r
(cid:26)
(cid:107)pij(cid:107)2 − 2r
2δ
− vj · uij
(cid:27)
2δ
vij = max
0,
(cid:26)
(cid:27)
,
(cid:107)pij(cid:107)2 − 2r
vij · τ
θij = min
1,
and
and define the safe half-plane ¯Pij of agent i according to agent aj as a subset of Pij
¯Pij = {(1 − λ + θij · λ) · vv ∈ Pij} ,
where λ ∈ [0, 1] is the relax factor indicating how much the agent considers the long-
sighted decision, and it is set to 0.5 through this paper.
Now, we are ready to define the local action cell (LAC) of agent ai, denoted by Ci,
as a subset of velocities in the intersection of all the safe half-planes, i.e.
(cid:26)
Ci =
v ∈ ∩j ¯Pij (cid:107)v(cid:107)2 ≤ min{vmax,
(ρ(v) − ρ(di − pi)) mod 2π ∈ ∆
(cid:107)di − pi(cid:107)2
},
δ
(cid:27)
where vmax indicates the maximum moving speed, di is the destination/target of agent
ai, ρ(·) denotes the angle (in radians) of the clockwise rotation of the argument vector
to align with the positive direction of the x-axis, and ∆ is a set of candidate angles
which is defined by
(cid:110)
(cid:111)
∆ =
k · π
4
k ∈ Z, 0 ≤ k < 8
,
through this paper. (See Figure 1 for an illustration of the local action cell of an agent
moving through two neighbors.)
3 Collision-Free Navigation
In this section, we introduce a distributed approach, named LAC-Nav, for the collision-
free navigation with multiple agents. As shown in Algorithm 1, the approach is straight
4
Figure 1: The local action cell of an agent (the black one) moving through two neigh-
bors.
forward with the following steps executed in loops: for each agent ai, calculate the
current local action cell; and then select a proper velocity from the cell.
Algorithm 2 follows the definition of the local action cell and describes the calcu-
lation details; Algorithm 3 shows how the new velocity is selected: Given the current
local action cell Ci, each velocity v ∈ Ci is at first evaluated according to the penalized
length ζv · (cid:107)v(cid:107)2, where ζv is the factor that is initialized as 0 < ζ ≤ 1 and decreased
exponentially on the angle between v and the direction of di − pi. Finally, the velocity
of the maximum penalized length is returned as the result.
Algorithm 1: LAC-Nav(ai): The LAC-based navigation algorithm running on
agent ai.
1 while ai is not at the destination do
2
3
4
Ci := LAC(ai);
vnew
i
agent ai moves at velocity vnew
:= SelectVel(Ci);
;
i
;
·
i
(cid:107)di−pi(cid:107)2
};
δ
Algorithm 2: LAC(ai): Calculate the current local action cell of agent ai.
1 vmax
i
i := vmax
2 v0
3 for k = 1 to 7 do
4
i ) =(cid:0)ρ(v0
:= min{vmax,
di−pi
(cid:107)di−pi(cid:107)2
i such that (cid:107)vk
i k ∈ Z∗, 0 ≤ k ≤ 7};
calculate vk
2π;
5 Ci := {vk
6 for agent aj with j (cid:54)= i do
i (cid:107)2 = vmax
i ) + k · π/4(cid:1) mod
and ρ(vk
i
7
8
9
10
calculate the safe half-plane ¯Pij;
for v ∈ Ci do
Γv := {θ · v θ · v ∈ ¯Pij, 0 ≤ θ ≤ 1};
v := arg maxu∈Γv (cid:107)u(cid:107)2;
11 Return: Ci;
While calculating the local action cells, it is not necessary to consider all the agents
in the environment. When the distance between agent ai and agent aj is at least (cid:96) :=
5
Algorithm 3: SelectVel(Ci): Select a velocity inside cell Ci as the new velocity
to move at.
1 for v ∈ Ci do
2
4αv
αv := (ρ(v) − ρ(di − pi)) mod 2π;
ζv := ζ
wv := ζv · (cid:107)v(cid:107)2;
:= arg maxv∈Ci wv
π ;
3
4
5 vnew
6 Return: vnew
i
i
;
2 · vmax · τ + 2 · r, it holds directly that θij = 1 and θji = 1. Thus the corresponding
safe half-planes can be ignored in the calculation of the agents' local action cells, which
implies it is sufficient to consider only the neighbors within distance (cid:96).
Processing complexity. When considering only the agents within a distance (cid:96), the
number of an agent's neighbors is at most 3 · (cid:96)2/r2, since there is no overlap between
the neighbors and for each of them, at least 1/3 of the body is covered by the disk of
radius (cid:96). Consequently, the loop of Lines 6 − 10 is executed for a constant time within
one step of update of an individual agent. Thus, the processing complexity of LAC
is determined by the efficiency to detect the neighbors in the specified range. In the
simulations, the neighbors can be efficiently derived through querying in a KD-Tree
that maintains all the positions, and in more practical cases, the neighbor detection is
often executed in a parallel process, and it can be assume that the required information
is always ready when it is needed.
Learning with LAC.
In LAC-Nav, the new velocity is selected according to the pe-
nalized length, which can be roughly seen as an estimate of the traveling distance of
the next move. On the other hand, it is also common to evaluate the performed ac-
tions and record the results, which also provides the information that may be useful for
making decisions in the future. In the case when a specific behavior should perform
well for a period of time, selecting the action of the best known evaluation should be
more promising than trying based on the estimates only. Generally, the evaluations are
learned as the agent keeps running in the "sense-evaluate-act" cycles.
Following the ALAN learning framework [5], we propose the LAC-Learn approach,
in which the reward of the latest performed action is defined as the summation of the
penalized lengths of the velocities in the resulting local action cell. Notice that by this
definition, the reward naturally incorporates the considerations of the goal-oriented
performance and the politeness performance, which are treated as two separate com-
ponents in ALAN. In fact, the lengths of the velocities approaching to the destination
reflect how efficient the performed action is for getting the agent closer to the goal;
and the lengths of velocities in the local action cell as a whole reflects the efficiency
in avoiding the crowding situations. In spite of the definition of the action reward,
LAC-Learn selects the new velocity in a different way from the one used in ALAN.
With LAC-Learn, the selected new velocity is the one corresponding to the action that
6
maximizes a linear combination of the reward and the penalized velocity length.
Algorithm 4: LAC-Learn(ai): Navigation algorithm of agent ai while learning
with the local action cells.
1 while ai is not at the destination do
2
3
4
5
6
7
8
Ci := LAC(ai);
Wi := CalcWeights(Ci);
Ri := UpdateReward(αi, Wi, Ri);
Si := UpdateWUCB(αi, Ri, Si);
αi := SelectAct(αi, Wi, Ri, Si);
vnew
i
agent ai moves at velocity vnew
:= Ci[αi];
;
i
Algorithm 5: SelectAct(αi, Wi, Ri, Si): Select the action for agent ai to per-
form.
1 i := 0;
2 α := Null;
3 if αi = 0 then
4
5
6
if Wi[0] ≥ η · min{vmax,
} then
(cid:107)di−pi(cid:107)2
α := 0;
i := 0;
δ
7
8
else
i := min{1, i + β};
9 if α = Null then
10
11
12
take s from [0, 1] uniformly at random;
if s < 1 − i then
α := arg maxα∈∆ ((1 − γ) · Ri[α] + γ · Wi[α]);
13
14
else
α := arg maxα∈∆ Si[α];
15 Return: α;
Inside an execution cycle of some agent ai, after the local action cell is calculated
by LAC (Algorithm 2), the penalized length of each velocity in Ci is calculated as
what has been done in Line 2 − 4 of SelectVel (Algorithms 3), and saved in a set Wi.
In UpdateReward, the reward of the last performed action is updated to the sum of all
weights in Wi, as mentioned earlier.
Notice that although the velocities given by Ci may vary from step to step, in the
local view, they can always be interpreted as the actions corresponding to the angles
specified in ∆. For example, without considering the variation of the length, the veloc-
ity pointing to the destination can always be interpreted as the action corresponding to
7
angle 0 ∈ ∆. For an action/angle α ∈ ∆, we use Ci[α] to denote the velocity v ∈ Ci
such that (ρ(v) − ρ(di − pi)) mod 2π = α.
Following the ALAN learning framework, we calculate (by UpdateWUCB) and
maintain (in Si) the upper confidence bound within a moving time window (i.e. a se-
quence of consecutive time steps), which is used when the agent explores in the action
space. As defined in [5], the wUCB score of action α during the last T ∈ Z+ steps is
defined by
(cid:115)
wUCB(α) := ¯Ri(α) +
2 ln(ν)
να
,
where ¯Ri(α) is the average reward of action α, να denotes the number of times action α
has been chosen, and ν denotes the total number of performed actions, all with respect
to the moving time window.
Similar to the context-aware action selection approach proposed in [5], SelectAct
(Algorithm 5) decides with the "win-stay, lose-shift" strategy and the adaptive -greedy
strategy in which the wUCB suggested action is chosen for the exploration.
When the agent is in the winning state (i.e. the goal-oriented action αi = 0 is per-
formed in the last update and is still a good choice in the sense that the corresponding
velocity is little constrained), it is natural to keep forwarding to the goal. Otherwise, if
the agent is in the losing state, it performs the -greedy strategy to exploit on the action
that maximizes a linear combination of the action reward and the penalized length of
the corresponding velocity. With a small and adaptively adjusted probability, the agent
explores and performs the action that maximizes the wUCB score.
the hyper-parameters η ∈ [0, 1] (Line 4 in Algorithm 5) and γ ∈ [0, 1] (Line 12 in
Algorithm 5) are determined depends on the scenarios.
4 Experiments
In this section, we present the results of running experiments with LAC-Nav and LAC-
Learn, on a computer of 7 Intel Core i7-6700 CPU (3.40 GHz) processors. The simu-
lations are implemented in Python 3.5, while the update processes of individual agents
have been speeded up by applying the multitasking scheme. For one second, there are
100 updates performed for each agent, and therefore we set δ := 0.01 in the imple-
mentation of LAC-Nav and LAC-Learn.
Scenarios. For the experiments, we considered three scenarios (Figure 2): the reflec-
tion scenario, the circle scenario and the crowd scenario, where
• in the reflection scenario, two groups of agents start from the left side and right
side of the area, respectively (Figure 2(a)). For each agent, the target is the posi-
tion on the other side that is symmetric to its start position (Figure 2(d)). Through
navigating the agents to the target positions, the picture of the starting configu-
ration is reflected.
8
• in the circle scenario, the agents start in layers of circles (Figure 2(b)), and each
agent targets the antipodal position (Figure 2(e)). That is, the picture of starting
configuration is going to be "rotated" by half of a circle, around the origin/center.
• in the crowd scenario, the start positions (Figure 2(c)) and target positions (Fig-
ure 2(f)) are randomly picked from a small area.
(a) Reflection: start
(b) Circle: start
(c) Crowd: start
(d) Reflection: target
(e) Circle: target
(f) Crowd: target
Figure 2: Experiment scenarios.
For each of these three scenarios, we compare the performances of LAC-Nav and
LAC-Learn, with the performances of approaches including BVC [11], CNav [6],
ALAN [5] and ORCA [9].
In this work, we consider two measurements: the completion time and the average
detour-distance ratio, as the evaluation of the algorithm's performance for the multia-
gent navigation tasks.
• the completion time of running a navigation algorithm is defined as the time (in
seconds) when the last agent arrives at its target, assuming all the agents start
from time 0;
• the average detour-distance ratio is defined as the average of the ratios between
the actual travel distance and the optimal travel distance (i.e. the length of the
straight line from the start position to the target position), over all the agents.
• the average detour-time ratio is defined as the average of the ratios between the
actual travel time and the optimal travel time (i.e. the time of moving in a straight
line from the start position to the target position, at the maximum speed), over
all the agents.
9
While the completion time justifies the algorithm's global performance on finishing
the navigation tasks, by investigating the detour-distance/time ratio, it provides a view
on the variance of the individual agent's behavior with different algorithms.
In the experiments for all scenarios, the agent's radius is uniformly set as r = 10,
and the maximum moving speed is set as vmax = 50. In addition, as mentioned in the
beginning of this section, within each second there are 100 updates performed for each
of the agents, which implies the that the time interval between two consecutive updates
is 0.01, i.e. δ = 0.01 in all the experiments.
Recall that when calculating the local action cells, the hyper-parameter τ > 0
is needed to locate the safe half-planes. Through all experiments involving the local
action cells, we set τ = 0.05. In addition, the penalty factor ζ is also needed in the cal-
culation of the local action cells (thus it is required when running LAC-Nav and LAC-
Learn). Through the experiments, we set ζ = 0.95, with which the goal-orthogonal
action (with angle π/2 from the direction pointing to the goal) is penalized by 0.9025,
and the goal-opposite action (along the direction leaving the goal) is penalized by about
0.8145.
For LAC-Learn, we set the mixing factor (Line 12 of Algorithm 5) as γ = 0.75 for
the reflection scenario and the circle scenario, and γ = 0.95 for the crowd scenario.
and the length of the moving time window (for the calculation of wUCB) as T = 8,
which is the minimum choice as there are 8 actions in the used ∆. Furthermore, the
incremental step (for adjusting the exploration probability) is set as β = 0.1 (Line 8 of
Algorithm 5).
For ORCA, the collision-free time window is set as τ = 0.02, i.e. twice of the
update interval's length. Recall that with CNav and ALAN, ORCA is also called to
make sure the performed velocity is collision-free, where the time window are also set
as τ = 0.02. Notice that the time window τ > 0 in ORCA has different meaning from
the hyper-parameter using the same symbol in the calculation of the local action cells,
even though they are both related to the avoidance of the potential collisions.
For CNav, the hyper-parameter for mixing the goal-oriented reward and the constrained-
reduction reward is set as γ = 0.5 for the reflection scenario, and γ = 0.9 for the circle
scenario and the crowd scenario; the number of constrained neighbors of which the
action's effect is estimated is set as k = 3; the number of the neighbor-based actions is
set as s = 3.
For ALAN, the hyper-parameter for mixing the goal-oriented reward and the polite-
ness reward is set as γ = 0.5; the length of the moving time window for the calculation
of wUCB is set as T = 5; and the incremental step for adjusting the exploration prob-
ability is set as β = 0.1.
Results.
In Figure 3, it shows the experiment results for
• the reflection scenario of 100 agents (50 agents on each side);
• the circle scenario of 120 agents (in 5 circles around the same center point);
• the crowd scenario of 100 agents located in the area of size 600 × 600.
10
Overall, LAC-Nav and LAC-Learn outperform almost all the other approaches in the
completion time. The only exception is in the reflection scenario, BVC has shown the
advantages and it completes earlier than LAC-Learn.
Figure 3: Experiment results, where ctime (s) stands for the completion time (in sec-
onds); addr stands for the average detour-distance ratio; and adtr stands for the average
detour-time ratio.
In general, the efficiency of the LAC based approaches is due to the fact that it
considers both of the task to arrive at the target and the intension to move as much
as possible in every step. The later consideration prevents the agent from the non-
necessary halting before it arrives at the target. By maintaining and comparing the
penalized lengths of all the candidate actions (according to ∆), even though the agent
still has the change to move directly towards the target, it detours as long as there is
an other action that provides a better moving (penalized) velocity. As shown in the
reflection scenario, this kind of active detouring results in a more fluent navigation as
the agents (of the antagonistic moving directions) pass by each other (Figure 4).
Recall that according to the definition of the safe half-planes, the local action cell
is depressed if there are neighbors approaching. Therefore, with the same relative po-
sition, it is easier for an agent to "follow" a leaving-away neighbor, if they have the
similar preferred trajectories. As a consequence result, in the case when there are more
conflicts, such as the circle scenario, after gathered around the central area, instead of
squeezing through (as what happens with the other approaches), the agents with the
LAC based approaches spin as a whole to resolve the conflicts (Figure 5).
Although the local action cell can be seen as a variant or extension of the buffered
Voronoi cell, it should be noticed that the LAC based approaches perform distinguish-
ably from BVC, except for the simple situation such as the reflection scenario. For the
more crowding situations (like the circle scenario and the crowd scenario), the indi-
11
ReflectionCircleCrowd0102030ctime (s)18.2824.9619.5716.4525.8517.3016.5232.7822.8622.8528.4319.7125.5132.9926.8236.1134.9429.06ReflectionCircleCrowd012addr1.081.711.231.032.541.161.031.851.161.082.151.291.162.081.291.202.211.36ReflectionCircleCrowd024adtr1.112.882.811.032.992.641.043.751.861.092.362.781.212.571.451.262.551.55LAC-LearnLAC-NavBVCCNavALANORCA (a) LAC-Nav
(b) CNav
(c) ORCA
Figure 4: The antagonistic agents pass by each other in the reflection scenario, where
the points of black outline and colored (red/blue) inside are the current positions of the
agents, and the simply colored (red/blue) points are the target positions of the agents.
(a) LAC-Nav
(b) CNav
(c) ORCA
Figure 5: Agents resolve the conflicts in the circle scenario, where the points of black
outline and colored inside are the current positions of the agents, and the simply colored
points are the target positions of the agents.
vidual agents with LAC-Nav or LAC-Learn spend more time on average, while on the
other hand the global completion time is shorter. By investigating into the experiment
processes, it can be found that the LAC based approaches caused less stuck agents than
the other approaches did. This fact can also be revealed by checking the completion
time of the first 90% arrivals (Figure 6), in which the approaches' performances are
less distinguishable.
Figure 6: The completion time (in seconds) of the first 90% arrivals.
12
ReflectionCircleCrowd010203090% ctime (s)17.1524.0316.7416.4323.9312.7716.4731.4814.8719.3924.6315.9618.1527.8115.7719.7026.6415.92LAC-LearnLAC-NavBVCCNavALANORCA 5 Discussions
In this work, we introduced the definition of the local action cells, and proposed two
approaches LAC-Nav and LAC-Learn, of which the efficiency in the completion time
have been experimentally demonstrated. In order to improve the approaches' perfor-
mance, besides trying with different parameter values, there are some natural directions
that also extend the proposed approaches or make a variant.
Adaptive λ. Recall that in the definition of the safe half-plane, we have set the relax
factor as λ := 0.5. Intuitively, λ indicates how much the agents would like to compro-
mise in the next move, in order to avoid the collisions that may happen in a near future.
Although it is valid to select any value [0, 1] from the theoretical respect, it should be
noted that a very small λ may cause the local action cell being depressed too much,
and a very large λ may help little for the long-sighted consideration. The value 0.5 is a
balanced choice, and it also follows from an important idea in the reciprocal collision
avoidance: each agent take half of the responsibility to avoid the coming collisions.
However, it will be more interesting if λ can be dynamic adjusted as the agents learned
more information about the environment.
Non-uniform ∆.
In the candidate set ∆ used through this paper, the angles between
any consecutive actions are uniform. While it is natural to use another uniform candi-
date set of different size, say ∆ of size 4, 12 or 16, it is also valid to include the actions
between which the angles are arbitrary, such as the neighbor-based actions considered
in CNav. In order to prevent the actions being penalized too much, it should be better
to set ζ close to 1 or bound maximum penalty, when the size of ∆ becomes large.
Continuous LAC.
In this paper, we defined the local action cells as sets of finite
number of actions. However, it may be more natural to consider the continuous area
spanned by the velocities in a cell. There is a direct way to extend the definition of
the local action cell to include all the linear combinations between every pair of the
adjacent velocities. Formally speaking, we can define,
C∗
i = {λv · v + λu · u v and u are adjacent in Ci,
λv + λu ≤ 1, λv ≥ 0, λu ≥ 0},
which is a continuous area in the velocity space. With the continuous version the local
action cell, the agents are no longer restricted to select the actions from ∆, and they
can move in any angle as long as the corresponding velocity has a positive length.
References
[1] S. Bandyopadhyay, S. Chung, and F. Y. Hadaegh. Probabilistic swarm guid-
ance using optimal transport. In 2014 IEEE Conference on Control Applications
(CCA), pages 498 -- 505, Oct 2014.
13
[2] P. Bhattacharya and M. L. Gavrilova. Roadmap-based path planning - using the
voronoi diagram for a clearance-based shortest path. IEEE Robotics Automation
Magazine, 15(2):58 -- 66, June 2008.
[3] Paolo Fiorini and Zvi Shiller. Motion planning in dynamic environments using
velocity obstacles. The International Journal of Robotics Research, 17(7):760 --
772, 1998.
[4] Santiago Garrido, Luis Moreno, and Dolores Blanco. Voronoi diagram and fast
marching applied to path planning. volume 2006, pages 3049 -- 3054, 01 2006.
[5] Julio E. Godoy, Ioannis Karamouzas, Stephen J. Guy, and Maria Gini. Adaptive
In Proceedings of the 2015 International
learning for multi-agent navigation.
Conference on Autonomous Agents and Multiagent Systems, AAMAS'15, pages
1577 -- 1585, 2015.
[6] Julio Erasmo Godoy, Ioannis Karamouzas, Stephen J Guy, and Maria L Gini.
Implicit coordination in crowded multi-agent navigation. In Proceedings of 17th
AAAI Conference on Artificial Intelligence, AAAI'16, pages 2487 -- 2493, 2016.
[7] L. C. A. Pimenta, V. Kumar, R. C. Mesquita, and G. A. S. Pereira. Sensing and
coverage for a network of heterogeneous robots. In 2008 47th IEEE Conference
on Decision and Control, pages 3947 -- 3952, Dec 2008.
[8] Baskın S¸enbas¸lar, Wolfgang Honig, and Nora Ayanian. Robust trajectory ex-
ecution for multi-robot teams using distributed real-time replanning.
In Niko-
laus Correll, Mac Schwager, and Michael Otte, editors, Distributed Autonomous
Robotic Systems, pages 167 -- 181, Cham, 2019. Springer International Publishing.
[9] Jur Van Den Berg, Stephen Guy, Ming Lin, and Dinesh Manocha. Reciprocal
n-body collision avoidance. In Robotics Research, pages 3 -- 19. Springer, 2011.
[10] Jur P. van den Berg, M. Chiao Lin, and Dinesh Manocha. Reciprocal velocity ob-
stacles for real-time multi-agent navigation. 2008 IEEE International Conference
on Robotics and Automation, pages 1928 -- 1935, 2008.
[11] D. Zhou, Z. Wang, S. Bandyopadhyay, and M. Schwager. Fast, on-line collision
avoidance for dynamic vehicles using buffered voronoi cells. IEEE Robotics and
Automation Letters, 2(2):1047 -- 1054, April 2017.
14
|
1204.4427 | 1 | 1204 | 2012-04-19T18:24:11 | Coupling Clinical Decision Support System with Computerized Prescriber Order Entry and their Dynamic Plugging in the Medical Workflow System | [
"cs.MA"
] | This work deals with coupling Clinical Decision Support System (CDSS) with Computerized Prescriber Order Entry (CPOE) and their dynamic plugging in the medical Workflow Management System (WfMS). First, in this paper we argue some existing CDSS representative of the state of the art in order to emphasize their inability to deal with coupling with CPOE and medical WfMS. The multi-agent technology is at the basis of our proposition since (i) it provides natural abstractions to deal with distribution, heterogeneity and autonomy which are inherent to the previous systems (CDSS, CPOE and medical WfMS), and (ii) it introduces powerful concepts such as organizations, goals and roles useful to describe in details the coordination of the different components involved in these systems. In this paper, we also propose a Multi-Agent System (MAS) to support the coupling CDSS with CPOE. Finally, we show how we integrate the proposed MAS in the medical workflow management system which is also based on collaborating agents | cs.MA | cs | Coupling Clinical Decision Support System with
Computerized Prescriber Order Entry and their
Dynamic Plugging in the Medical Workflow
System
Lotfi BOUZGUENDA
Miracl laboratory/ University of Sfax/ISIM
Route de Tunis, km 10. BP 242, 3021 Sakeit Ezzit, Sfax-Tunisia
[email protected]
Abstract. This work deals with coupling Clinical Decision
Support System (CDSS) with Computerized Prescriber
Order Entry (CPOE) and their dynamic plugging in the
medical Workflow Management System (WfMS). First, in
this paper we argue some existing CDSS representative of
the state of the art in order to emphasize their inability to
deal with coupling with CPOE and medical WfMS. The
multi-agent technology is at the basis of our proposition
since (i) it provides natural abstractions to deal with
distribution, heterogeneity and autonomy which are
inherent to the previous systems (CDSS, CPOE and medical
WfMS), and (ii) it introduces powerful concepts such as
organizations, goals and roles useful to describe in details
the coordination of the different components involved in
these systems. In this paper, we also propose a Multi-Agent
System (MAS) to support the coupling CDSS with CPOE.
Finally, we show how we integrate the proposed MAS in the
medical workflow management system which is also based
on collaborating agents.
Keywords: Clinical Decision Support System , Computerized
Prescriber Order Entry, Multi-agent Technology and Medical
Workflow management System .
I.
INTRODUCTION
The purpose of Clinical Decision Support System (or
CDSS for short) is to assist health professionals with
decision making tasks, as determining diagnosis or
analysis of patient data [1].
In spite of the growing multiplicity of CDSS and their
effectiveness certified in the decision making tasks at the
time and the location of care, the state of the art of the
existing CDSS ([2-6]) emphasizes four main challenges
that require to be resolved. The first challenge is that the
clinical data that must be entered is already contained
elsewhere in a digital form in that hospital’s system, and
some CDSSs (alert system, drug-drug detection system,
medicinal errors prevention system, etc.) are not able to
access and exploit this information. The second challenge
concerns the appropriate decision making by the health
professionals to a patient. Indeed, the majority of existing
CDSS exploit only the list of the actives prescripts of the
Manel TURKI
University of Monastir/ Faculty of pharmacy
Rue Ibn Sina, 5000, Monastir-Tunisia
[email protected]
patient, which is compared with a frozen source of
knowledge fulfilled under form of mapping table. In others
terms, they do not take into account the clinical context of
the patient. A such CDSS must consider several
information’s such as patient’s symptoms, medical history,
family history and genetics. It needed to interact with
others clinical data sources such as patient administrative
data, physio-pathologic profile, biological analysis data,
etc. The third challenge is that the CDSS is often not
connected to Computerized Prescriber Order Entry
(CPOE) which assists the doctor in preparing the
prescripts.
The last challenge concerns the integration issue of the
CDSS and CPOE in the medical workflow management
system. In fact, often the clinical user must stop clinical
process on the current system, switch to the CDSS or
CPOE, and reenter data necessary into the CDSS that may
already exist in another healthcare system.
An obvious example of clinical workflow is the patient
care in an accident case. In this case, various health
professionals such as the doctors, the nurses, the
pharmacists,
must
cooperate,
synchronize
their
intervention processes, share the access to the clinical data
sources and use the CDSS, the CPOE without stop the
clinical process to act in a coherent way and in order to
give care to the patient.
Giving the previous observations, the problem being
addressed in this paper can be resumed according to the
following two questions: “how to design and develop a
CDSS which considers the previous challenges? How do
we integrate the CDSS and CPOE in the medical workflow
management system?
The contribution of the paper is to provide a Multi-
Agent System (MAS) to support the coupling CDSS with
CPOE. Then, it explains how we integrate the proposed
MAS in the medical Workflow management system.
Our approach is based on the following principles:
The use of multi-agent technology, which provides
relevant high-level features to design and implement
CDSS, CPOE and clinical workflow management
-
-
system. Indeed, Agent technology provides natural
abstractions to deal with autonomy, distribution,
heterogeneity and cooperation which are inherent to
the previous systems,
A coordination model based on organizational
dimension. A good coordination model among clinical
workflow components requires an organizational
model which attributes roles to each component and
constraints their interactions,
- An engineering perspective, which takes into account
the existing standards. It defines an architecture
which is compliant with the WfMC reference model
[7] and the involved agents communicate among them
using the FIPA-ACL standard [8]. This compliance
allows a very large usability and interoperability of the
solution.
The remainder of this paper is organized as follows.
Section 2 exposes some related works regarding the design
and the development of CDSS in order to underline their
inability to deal with the challenges as mentioned
previously. Section 3 presents our Multi-Agent System
(MAS) for supporting the coupling CDSS with CPOE.
This section first justifies the interest of use of multi-agent
technology to design and develop the CDSS, CPOE and
WfMS systems. It exposes then our proposed multi-agent
system. Section 4 shows how we integrate dynamically the
proposed MAS in the medical workflow management
system and presents an organizational model based on
message methodology [9] that structures the interactions
between agents and highlights the coordination at macro
level. Section 5 concludes the paper and gives some
perspectives.
II.
RELATED WORKS
Regarding the design and development of CDSS, the
use of agent technology is not new, and several works have
been proposed in the literature ([2-6]).
[2] defends the interest of using agents to extend the
medical expert systems. Indeed, according to the [2] the
agents can resolve some issues by checking several
conditions that could be ignored by humans and as a
consequence the elimination of some mistakes from the
physicians’ decisions. More precisely, [2] proposes a new
cooperative medical diagnosis system called “Contract Net
Based Medical Diagnosis System”. This latter owns two
specific features namely the autonomy and the flexibility
during the treatment of medical diagnosis problems. The
proposed system in [2] does not support the interaction
either with the CPOE or with the medical WfMS.
[3] proposes a CDSS called “SAPHIRE”. The main
purpose of this system is to support the definition,
deploying and performing clinical guidelines to a patient.
It is composed of a set of collaborating agents running in a
heterogeneous distributed environment. SAPHIRE has
two main advantages. Firstly, it proposes a specific agent
called “EHR” agent” that is responsible to access and
extract clinical data from the Electronic Healthcare.
Secondly, it supports the interaction with several modules
in the clinical workflow. The main limits of this work are
the following: (i) the non possibility to couple with the
CPOE and (ii) the use of agent only for the design of
CDSS.
[4] provides a framework based on multi-agent system
paradigm in order to build a comprehensive clinical
decision applications aimed at various medical conditions.
This work has not addressed the coupling issue with the
healthcare systems such as CPOE, clinical workflow
system, etc.
[5] proposes a multi-agent system called IMASC to
assist physicians and other health professionals with
decision making tasks. In spite of the system is powerful,
it does not support the previous challenges, as mentioned
in introduction, and namely the coupling with CPOE and
the dynamic integration in the medical workflow
management system.
[6] provides a prototype called MET3 which aims at
data collection, diagnosis formulation and treatment
planning. MET3 is based on multi-agent technology. It
supports the interoperability issue since it runs on different
platforms. It is also capable to interact which hospital
information systems and particularly with an electronic
patient record system via HL7 messages to provide
realistic integration with existing healthcare systems.
Our work differs from the previous works on three
points. First, it addresses the coupling CDSS with CPOE
aspect which has not been addressed in the previous
works. Second, the proposed architecture for medical
Workflow management system is compliant with the
Workflow Management Coalition reference model. Third,
to the best of our knowledge, the dynamic plugging of
CDSS and CPOE in the medical Workflow management
system through the agent technology has not been
addressed. Thus, we believe that our solution is currently
unique in trying to take into account the agents to deal
with coupling CDSS with CPOE and their dynamic
integration in the medical WfMS.
III.
A MULTI-AGENT SYSTEM FOR COUPLING CDSS WITH
CPOE
This section is devoted to the motivation for using
multi-agent technology and presents our multi-agent
system for coupling CDSS with CPOE.
A. Motivation for using multi-agent technology
The multi-agent technology can help the design and
the development of CDSS, CPOE and clinical workflow
system thanks to the following high- level properties [10]:
- Autonomy of agents eases the cooperation since it
avoids regular and direct interventions of the systems:
the medical information system or clinical workflow
management system can be agentified to support the
cooperation and provides needed information to each
health professional. Agents can also play the role of
interface between the actors and the system (filtering
and notification of events, providing relevant views of
the whole system and its evolution).
- Natural abstractions to deal with cooperation. A lot of
sophisticated protocols like Contract-Net Protocols
and Negotiation mechanisms are available and could
be used to coordinate processes ([11] [12]). Agent
technology also provides organizational concepts to
abstract and structure a system as a computational
community made of groups, roles and interaction.
Pro-active and reactive attitudes of agents ease the
control and enactment of clinical processes as well as
reactions to events, and hence the synchronization of
related activities. Being able to exhibit goal-directed
behaviour, agents can take the initiative to select and
engage cooperation with others actors.
Social abilities of agents also ease the cooperation
needed to enact complex clinical workflows and to
provide an abstraction to high- level concepts like
commitments, reputations and so on.
-
-
B. The proposed MAS for coupling CDSS with CPOE
In order to support the coupling CDSS with CPOE, we
propose a Multi-Agent System (MAS) where each agent
plays a specific role or function and exploits one or several
clinical data sources. More precisely, the architecture of
our MAS is organized around five clinical data sources
and ten specialized agent. Let us detail the role/function of
each agent.
The communication between agents is assured thanks
to the FIPA-ACL language [8].
The Information Collection Agent (ICA) is responsible
to extract information’s from two clinical data sources
(patient administrative data and physio-pathologic profile)
such as name, type, date of birth, service, weight, size,
diagnosis, allergy, against indications, situations at risk
and so on. It is to signal that these information’s are
managed by a medical information system and they can be
also exploited by the others systems like CDSS and CPOE
in our context. In this case, this agent answers to the two
first challenges as mentioned in introduction i.e. we
haven’t need to enter these information’s again.
Clinica l data so urces
Computer ized Presciber Order
E ntry , CPOE
Clinical Decision
Suppo rt System, CDSS
Drug Errors De te c t ion
Agent
Informa t ion Colle c tion
Age nt
Drug Se lec t ion Agent
The ra peut ic file Agent
informat ion De l ive ry
Agent
P re sc ription ana lyze r
Age nt
Pha rma cy S toc k-
Che cke r Agent
Adminis tra tion Mode
S ele c tion Agent
P rotocols Re cording
Agent
Drug-Drug Inte rac t ion
De tec t ion Agent
Pa tient
Administrative
Data
Physio
Pathologic
profile
Data bank of
drugs
Drug
Repository
Biologica l
Analysis
data
In terac tio n b etween agen ts
b elong in g to th e same system
In teractio n be tween
ag en ts belo ng ing to the
d ifferen t systems
I nteractio n b etween
agents an d clin ica l
data sources
Figure 1. A MAS for Coupling CDSS with CPOE
The Drug Selection Agent (DSA) cooperates with the
ICA and the Pharmacy Stock Checker Agent (PSCA)
when selecting the drug form data bank of drugs. The role
of the PSCA is to verify from the drug repository if the
selected drug is available or in rupture of supply. In
unavailability case, it recommends to substitute the drug
by another generic. The DSA also interacts with the
Administration Mode Selection Agent (AMSA) who
provides relevant information’s regarding administration
mode, for each selected drug, such as unit of catch, dosage
by default, and borders dosage by catch, a day and by
kilogram of weight. To do this, it enters in connection with
the Drug Errors Detection Agent.
The Prescription Analyzer Agent (PAA) triggers two
agents to analyze and validate the current prescription: the
Drug Errors Detection Agent (DEDA) and the Drug-Drug
Interaction Detection Agent (DDIDA). More precisely, the
Drug Errors Detection Agent (DEDA) interacts with the
AMSA and DDIDA in order to verify if there is an error
such as error of dose, error of choice (i.e. a drug not
conforms to the reference), drug-drug interaction or error
of administration mode and so on. In the literature, we
distinguish two types of medicinal errors. A medicinal
error is said potential if it was discerned before drug
arrives up to the patient. It is said proved to be if it was
discerned after the catch of drug by the patient. In our
work, we support the first type of medicinal error in order
to guarantee a well service to the patient.
The Drug-Drug Interaction Detection Agent (DDIDA)
consists in proving if the association of medicaments is
likely to cause undesirable effects, allergies, etc. To do
this, it interacts with the Drug Errors Detection Agent and
exploits the biological analysis data source.
The Protocols Recording Agent (PRA) stores all types
of prescripts as protocols which will be then
called by their title. As a consequence, the time of
prescript becomes considerably reduced.
The Information Delivery Agent (IDA) is responsible
to deliver the information to the concerned clinical user
like the chemist, the nurse, etc.
The Therapeutic File Agent (TFA) preserves the
validated prescription as archived files in order to be
exploited in the future.
IV.
DYNAMIC PLUGGING OF CDSS AND CPOE IN THE
MEDICAL WORKFLOW MANAGEMENT SYSTEM
The purpose of this section is (i) to rethink the
reference model of a WfMS with agent technology in order
to deal with medical processes and (ii) to show how we
integrate the MAS that supports the coupling CDSS with
COPE in the new architecture of WfMS.
A. The reference architecture of a WfMS
According to this architecture, a WfMS includes a
Workflow Enactment Service (WES) and supports the
following interfaces (see figure 2) [7]:
Interface 1 with Process Definition,
-
Interface 2 with Workflow Client Applications,
-
Interface 3 with Invoked Applications,
-
-
-
Interface 4 with others WESs,
Interface 5 with Administration and Monitoring.
P roce s s e s
de finitions Tools
1
M onitoring and
adm inis tra tion Tools
5
Workflow Enac tm ent
Se rvice (WES) (engine )
Othe rs WES
4
3
2
C lients
Applica tions
Invoked applica tions
Figure 2. The reference model of a WfMS [7]
Two main components of this architecture are the
WES and Interface 4. The aim of the WES, on which one
imposes no constraints upon its internal structure, is to
manage the execution of one or several instances of
processes while the aim of Interface 4 is to connect WfMS
together in order to share the execution of a workflow
process between different WESs of different organizations.
However, this reference architecture is inadequate in
the clinical context since the WES must not only execute
clinical processes but also must mix different concurrent
activities including clinical decision making, analyzing
and validation of prescription, etc. In others words, the
WES need to cooperate with CDSS and CPOE.
B. Revisiting the WES with agents and dynamic
plugging of CDSS and CPOE
Figure 3 explains how we have rethought the
Workflow Enactment Service using agents. This
architecture includes (i) as many Clinical Workflow agents
as running clinical process instances, (ii) an Agent
Manager responsible for these Clinical Workflow Agents,
(iii) a Connector Agent that interacts with CDSS and
CPOE specialized respectively in clinical decision making
and prescription elaborating, and, (iv) a new interface,
Interface 6, to support the communication between a
Connector Agent and the proposed MAS.
Regarding clinical Workflow Agents, the idea is to
implement each clinical process instance (stored in the
Clinical Processes database, CP) as a software process, and
to encapsulate this instance within an (pro-) active agent.
Such a clinical Workflow Agent includes a workflow
engine that, as and when the clinical process instance
progresses, reads the CP schema definition (specified in
XML schema), and triggers the action(s) to be done
according to its current state. This clinical Workflow
Agent supports interface 3 with the applications that are to
be used to perform pieces of work associated to process’
tasks.
W ork f low
Eng ine
C linic a l
W o r kf low
A ge nt
1 , 2 ,5
A ge nt M anage r
C linic a l
P r oce sses
D a taba se
4
C lin ic a l
W ork flow
A gen t
C linica l
W or kf low
A ge nt
3
3
C onne c tion
A ge n t
6
C o m p u te r iz e d
P r e sc r ib e r O r d e r E n t r y
( C PO E )
C l in i c a l D e c is io n
S u p p o r t S y s t e m
( C D S S )
Figure 3. Dynamic plugging of the proposed MAS in the medical Workflow
Management System
The Agent Manager controls and monitors the running
of clinical Workflow Agents:
Upon a request for a new instance of a clinical process,
the agent manager creates a new instance of the
corresponding clinical process agent schema, initializes its
parameters according to the context, and launches the
running of its workflow engine.
It ensures the persistency of clinical Workflow Agents
that execute long-term clinical processes extending to a
long time in which task performances are interleaved with
periods of inactivity.
It coordinates clinical Workflow Agents in their use of
shared resources.
It assumes interfaces 1, 2 and 5 of the WfMS with the
environment.
The role of the Connector Agent is to help clinical
Workflow Agents to find the clinical information they
need. More precisely, the Connector Agent interacts with
CDSS specialized in clinical decision making. This
requires defining a new interface, Interface 6 that supports
the communication between a Connector Agent, the CDSS
and the COPE. Interface 4 of the reference architecture
cannot be used for such a communication since it only
supports the execution of a clinical process between
different workflow engines.
In our proposition, we consider coordination as a
specific component when designing and implementing a
clinical Workflow, and consequently, we separate
coordination activities from clinical processes execution.
That is why we introduced mediation infrastructure,
gathering the CDSS and COPE. This infrastructure of
course independent of the WES, and is dynamically
plugged to it only when it is necessary.
The standard FIPA-ACL (Foundations of Intelligent
Physical Agents-Agent Communication Language [8]), is
used to support the interaction, through message
exchange, between the agents inside the architecture of
WfMS. FIPA-ACL offers a convenient set of performatives
for supporting the cooperation between agents (e.g.
inform, ask, propose, agree, cfp,...). Moreover, FPA-ACL
supports exchange messages between heterogeneous
agents since (i) the language used to specify the message is
free and (ii) a message can refer to ontology. This is very
interesting in the context of inter organization clinical
processes since ontology can be used to solve semantic
interoperability problems.
C. An organizational model based on message
methodology for ruling the interaction between the
clinical workflow components
Our organizational model (see figure 4) is organized
around the following components:
-
Three organizations represented by triangles:
-
Ten agents represented by circles :
- Nine tasks represented by polygons:
-
Three roles represented by moon:
-
and three goals represented by double circles:
For clarity reason of the organizational model we give
only the tasks, goals and roles for three agents (Agent
Manager and Drug-Drug Interaction Detection Agent and
Prescription Analyzer Agent.
Agent
Ro le
Organ iza t ion
T as k
Serv ic e
GoalGoal
«int erac t ion» : WAPI4, WAPI 6, F IPA-AC L
«execut ion»
Resource
«play» , «w ish»
C rea t ing C linica l
W ork f low Agent s
C linical
Proces ses
Agen t
M anager
Perform ing
c linic al
Goal
proces ses
Superv is ing
the execut ion
of process
isnt anc es
Manager
Phys io
Pa thologic
prof ile
D ata
bank of
drugs
Superv is ing
the
Presc ript ion
Ana lyz ing
C rea t ing
Pres c rip t ion
Goal
Pat ient
Adm in
Da ta
Therapeut i
c f ile Agent
Informa t ion
C ollec t ion
Agent
I nformat ion
De livery
Agent
Agen t s
Coordinator
D rug
Selec t ion
Agen t
Presc ript io
n Ana lyser
Agent
Pharmac y
Stock
Check er
Agen t
Adm in
Mode
Se lec t ion
Agent
C l inica l
Work f low
A gent
Connec tor
Agen t
D rug
Repos itory
Computer ized Presc r iber Order Ent ry
Cl inica l W ork f low Management
Pro toc ols
R ec ording
Agen t
D SA C reation
PRA C reation
T A C rea t ion
I DA C reat ion
Bio log ical
Analys is
da ta
Drug Errors
D etec t ion
Agent
D rug-D rug
Interac t ion
De tec t ion
Agen t
Aga ins t Ind icat ion
detec t ion
Dos age Ex ceeding
D etec t ion
Allergy Det ect ion
U ndes irable Ef fec t
D tec t ion
D rug-Drug
In terac t ion
D etec tion
Analy zer
Va lida t ion of
Goal
Pres c rip t ion
C linical D ec is ion Support
Figure 4. An organizational model for structuring the communication
between CDSS, CPOE and clinical WfMS
Let us detail how each organization operates.
The Clinical Workflow Management Organization
enables one or several clinical workflow agents to interact
with their connector to which make a clinical decision or
elaborate a validated prescription.
The Computerized Prescriber
Order
Entry
Organization enables the Prescription Analyzer Agent to
interact at the first step with the drug selecting Agent.
This latter interacts in its turn with the information
collection agent, pharmacy stock checker Agent and
Administration mode selection agent. At the second time,
the Prescription Analyzer Agent connects to the Drug-
Drug interaction Detection Agent and Drug Errors
Detection Agent to analyze the current prescription. After
the validation, it offers the following services thanks to the
three agents: Information delivery, protocols recording and
therapeutic file creation.
The Clinical Decision Support Organization enables
the drug errors detection and drug-drug interaction
detection agents to enter in connection with the Connector
Agent or the Prescription Analyzer Agent in order to give
respectively help to the clinical user or assist the doctor to
validate his prescription. This organization involves only
those two agents.
CONCLUSION
V.
This paper has dealt with coupling CDSS with CPOE
and their dynamic plugging in the medical Workflow
Management System. More precisely, it has presented in
the one hand, a multi-agent system allowing the coupling
CDSS with CPOE. In the other hand, it has presented an
architecture of clinical WfMS which revisits the reference
model defined by the WfMC in terms of collaborating
agents. Some of these agents implement clinical processes,
while others are dedicated to the CDSS and CPOE.
The idea defended in this paper is that the agent
technology is appropriate to face the coupling CDSS with
CPOE and their dynamic integration in the clinical
WfMS. Several reasons motivate this idea.
Firstly, the agent technology is appropriate to model
clinical processes since it provides natural abstractions to
deal with distribution, heterogeneity and autonomy that
are inherent to clinical processes.
Secondly, the organizational dimension which is
inherent to multi-agent systems [9], is fundamental to
highlight the coordination of the different components
involved in the three systems (CDSS, CPOE and clinical
WfMS) under consideration by clearly separating the
macro-level (coordination) from the micro-level (agent).
Thus, the internal architecture of the agents may be
thought and implemented independently from their
coordination.
As future work, we plan (i) to study in depth the
behavior of all agents involved in the CDSS, CPOE and
WfMS systems, (ii) to formalize and build the knowledge’s
used by the agents in order to resolve the semantic
conflicts, (iii) to implement these agents using the Jade
[13] (Java Agent Development Framework) platform that
offers the possibility to simulate the execution of these
agents in autonomic way and in distributed mode and (iv)
to evaluate the proposed tool according to the three key
criteria: scalability, openness and efficiency.
REFERENCES
[1] Web site of Wikipedia: http//www.wekipedia.org.
[2] B. Iantovics, “Cognitive Medical Multi-agent Systems”, in
Broad Research in Artificial Intelligence and Neuroscience,
Volume 1, Issue 1, January 2010.
[3] G. B. Laleci, A. Dogac, M. Olduz, I. Tasyurt, M. Yuksel, A.
Okcan., “SAPHIRE: A Multi-Agent System for Remote Healthcare
Monitoring through Computerized Clinical Guidelines”, in Agent
Technology and e-Health, pp. 25-44, 2008.
[4] S. Wilk, W. Michalowski, D. O’Sullivan., “Engineering of a
Clinical Decision Support Framework or the Point of Care Use, in
AMIA’08 Symposium Proceedings, page – 814, 2008.
[5] G. Czibula, I. G. Czibula, G. S. Cojocar, and A. M. Guran.,
“IMASC - An Intelligent MultiAgent System for Clinical Decision
Support, Proceedings of the First International Conference on
Complexity and Intelligence of the Artificial and Natural Complex
Systems, pp.185-190, 2009.
[6] J.S. Shirabad, S. Wilk, W. Michalowski, K. Farion.,
“Implementing an Integrative Multi-agent Clinical Decision Support
System with Open Source Software”. In international journal of
Medical Systems, 2010.
[7] The Workflow Management Coalition, The Workflow Reference
Model. In: Technical Report WfMC-TC-1003, November 1994.
[8] Web site of FIPA-ACL: http://www.fipa.org.
[9] G. Caire, W. Coulier, F. Garijo, J. Gómez-Sanz, J. Pavón, P.
Kearney and P. Massonet., “The Message Methodology” Book
Chapter 9 in Methodologies and Software Engineering for Agent
Systems, Volume 11, Part III, 177-194, 2004.
[10] N.R. Jennings, P. Faratin, T.J. Norman, P. O'Brien, M.E.
Wiegand, C.Voudouris, J.L. Alty, T. Miah, E.H. Mamdani,. ADEPT.-
Managing Business Processes using Intelligent Agents Proceedings of
the 16th Annual Conference of the British Computer Society
Specialist Group on Expert Systems (ISIP Track), 1996.
[11] L. Bouzguenda, R. Bouaziz, E. Andonoff.: Dynamic plugging of
business processes in cross organizational workflow. In international
journal of computer sciences and applications, Technomatimatics
Resarch foundation, Bombay India, volume V issue IIIB, 2008.
[12] E. Andonoff, W. Bouaziz, C. Hanachi, L. Bouzguenda., "An
Agent Based Model for Autonomic Coordination of Inter-
Organizational Business Process". In International Journal of
computing and Informatics, INFORMATICA'09, 2009.
[13] Web site of Jade: http//www.jade.org.
|
1506.02796 | 1 | 1506 | 2015-06-09T06:42:06 | Agent-Based Product Configuration: towards Generalized Consensus Seeking | [
"cs.MA",
"cs.SE"
] | This paper will present an evolution of a fuzzy agent based platform which performed products configuration. As a first step, we used the notion of consensus to establish robust results at the end of the configuration process. We implemented the concept of generalized consensus which implied the consideration of consensuses from the beginning, in this way robust data are treated during the entire process and the final result enables the designer to distinguish the robust components and flexible ones in a set of configurations. | cs.MA | cs | Agent-Based Product Configuration: towards Generalized
Consensus Seeking
Benoît Beroule1, Alain-Jérôme Fougères1,2 and Egon Ostrosi1
1 University of Technology of Belfort-Montbéliard, Laboratory IRTES-M3M
{benoit.beroule, alain-jerome.fougeres, egon.ostrosi}@utbm.fr
90010 Belfort – FRANCE
2 ESTA, School of Business & Engineering
90004 Belfort – FRANCE
[email protected]
Abstract
This paper will present an evolution of a fuzzy agent based
platform which performed products configuration. As a first step,
we used the notion of consensus to establish robust results at the
end of the configuration process. We implemented the concept of
“generalized consensus” which implied the consideration of
consensuses from the beginning, in this way robust data are treated
during the entire process and the final result enables the designer
to distinguish the robust components and flexible ones in a set of
configurations.
Keywords: Fuzzy Agents, Consensus, Consensus Seeking, Agent-
Based Design, Collaborative and Distributive Design.
1. Introduction
in
The collaborative and distributive design concept raises
several issues [1, 2]. Design software must take into account
numerous parameters which all have to play a more or less
important role
the final result, furthermore much
information will transit between the different entities of
software to communicate the intermediate outcomes and
ideally update parameters if the user wishes to change one of
them during the process.
To best achieve these objectives we have chosen an agent-
based approach which will efficiently process the abundant
information and will be easy to use [3, 4, 5].
In this article, we will focus on a product configuration
platform named FAPIC (Fuzzy Agent for Product Integrated
Configuration) which
implements a collaborative and
distributive approach based on agents [6, 7]. To be as close
as possible to the real situation, the FAPIC platform agents
deal permanently with fuzzy values [8, 9]. Such values allow
a precise parameterization and consequently optimal results.
To improve the results the platform provides, the latter
implements the concept of "consensuses" [10, 11]. In
concrete terms, the agents are grouped to form clusters of
similar agents and these clusters are considered as new
agents in the process. We expect this approach will create
groups of optimal similar product configurations which will
enable robust parts of the product configuration to be
identified and creates a range of products or just chosen
among several proposals.
This paper is organized as follows: In the second part, the
agent-based product configuration process will be detailed
step by step. Then in the third part, the concept of
"consensus" and its purpose will be presented. Moreover in
the fourth part, the new concept of "generalized consensus"
and its implementation in the platform will be introduced.
Furthermore
the
configuration of an aerial conveyor will be detailed by using
several approaches previously described in detail. Finally in
the sixth part a comparison between the different approaches
will be made. The last section, the conclusion shows interest
in the proposed approach.
the example of
fifth part,
the
in
2. Agent-Based Product Configuration
2.1 Configuration Process
the product separated
The FAPIC platform is an agent-based software design to
represents
make product configuration. Each agent
information about
into several
categories we call communities [8].
The product configuration process performed by the FAPIC
platform is divided into four steps (Figure 1) [9].
First, fuzzy relationships are built between agents from the
same
communities
themselves based on the data provided by the user. The
then between
communities
the
purpose of this step is to link the agents and create a network
which will be used to compute the coming result.
In the second step, the solution agents are evaluated. A
rating is assigned to each of them depending on their
consistency with the requirements imposed by the customer,
the constraints determined by the experts and the functions
the product must perform. This rating is a fuzzy value
determined by the relationships created during the first step;
the higher the value, the more compatible the solution.
During the third step, each solution agent determines the
optimal configuration for the product concerned based on its
local point of view. Then the configurations are evaluated
considering all the points viewed in the previous steps.
Finally,
the fourth step, consensuses of optimum
configurations are created by grouping them according to
their resemblance. The purpose of this step is to provide
ranges of products instead of just a sole configuration.
in
R, F,
C, S
Building fuzzy
relationships in
product
configuration
(1)
iℜ~ }
{
F~
R~
C~
Searching the
fuzzy set of
consensual
solutions
(2)
cS~
cS~
Domains
expertise
Customer
Designers
Generating fuzzy
optimal product
solutions
S~
(3)
S~
Seeking
G~
fuzzy
consensus
(4)
R: set of requirements
iℜ~ }: fuzzy relationships
{
C~
: fuzzy set of constraints
G~ : fuzzy consensual configurations
F: set of functions
R~ : fuzy set of requirements
cS~ : fuzzy consensual solutions
C: set of constraints
F~ : fuzzy set of functions
S~ : fuzzy optimal solution
Fig. 1 Fuzzy product configuration approach
2.2 Configuration Process
=<
~,~,~,~
~
ΟΡΙΑΜα
The FAPIC platform is a fuzzy agent-based system for the
distributed and collaborative configuration of products [12].
Thus we defined a fuzzy agent-based system αM~ by Eq. (1):
(1)
Where A~ is the fuzzy set of agents, I~ is the fuzzy set of
interactions defined in αM~ , P~ is the fuzzy set of roles that
agents of A~ can play, and O~ is
the fuzzy set of
organizations defined for fuzzy agent of A~ .
>
A fuzzy agent-based system αM~ being defined, we can now
define a fuzzy agent
~α of A~ by Eq. (2):
i
.
(2)
,
)~(~
αΓ
i
)~(~
α∆
i
,
)~(~
αΠ
i
~,
ΚΦΦΦ
~
α
i
are respectively: the
~α , the
i
~α , the function of fuzzy action
i
Where
function of fuzzy observation of fuzzy the agent
function of fuzzy decision of
~α , and the fuzzy knowledge of
of
i
In concrete
the platform consists of several
communities (requirements, functions, solution and one or
~α .
i
terms,
more constraints domains) and each community contains
agents which represent elements of the community (Figure
2.a).
The agents communicate by sending and receiving messages
to and from other agents from linked communities to
establish local results.
As shown in the figure 2 b), the process is basically divided
into four parts:
ir and
jf which informed
jf and
ks which informed
lc and
ks which informed
1) Interactions between each
the other agents of R and F.
2) Interactions between each
the other agents of F and S
3) Interactions between each
the other agents of C and S
Requirements
Customer
Solutions
Constraints
Domain experts
a)
4) Interaction between each
consensual results.
ks to establish the new
Moreover the use of agents allows the platform to be
dynamic because the agents are constantly running. When
they are not computing calculus, they are waiting for
messages from other agents. Consequently, any users
(customer and/or experts) may send new data at any time. If
data are sent before step 4, thus after the computation of the
final result, the latter is recalculated. If data are sent during
the process, the former data are updated and the final result
will be changed according to the new information sent,
directly without displaying the out of date result based on
the initial data [13].
Community of fuzzy
requirement agents:
R~
Community of fuzzy
function agents:
F~
Community of fuzzy
solution agents:
S~
Community of fuzzy
constraint agents:
C~
1.1
R – {ri}
1.2
2.1
F – {fj}
3.2
2.2
4
S – {sk}
3.1
C – {cl}
:ri
1
:fj
2
:sk
3
:cl
b)
Fig. 2 a) Agent-based architecture of FAPIC platform divided in four communities of agents, and b) Illustration of interactions between fuzzy agents during
Phase 2 of the product configuration (Searching the fuzzy set of consensual solution agents).
3. Configuration Consensus Seeking
3.1 Consensus Seeking
The fourth step of the configuration process consists of
creating consensuses of agents between the solution and
configuration communities. The purpose of this step is to
create not just a sole optimal configuration but a set of
configurations divided into groups named "consensuses".
Consensuses contain one or more agents which are quite
similar regarding the constraints imposed by the different
experts of domains and the customer. These solutions allow
the
several product
configurations which best fit the imposed requirements.
customer
choose
among
to
T(
)THEN
)
Card
←
)G(
r
ENDIF
rG
r
;
)
(
/g,s
µ
i
p
j
1k
p,i
2k
p,i
=
j
∑
Gg
∈
(
= ∑
1
Gg
∉
p
j
Card
);G(
p
(
g,s
µ
i
)
)
j
/
Card
)GG(
p
−
−
α
1k
p,i
(
1
−
+
)
2k
α
p,i
byDO
agent
each
'S ←
r
r
(
receiveMsg
IF
S
update
1=
)G(
r
rto
p
FOR
=
k
p,i
ENDFOR
k
*p,i
=
S
*p
(
SIF
←
'S!
*p
*p =
diffuse
diffuse
ENDIF
WHILE
(
(
(
kmax
+
S
)r,i
k
,...,
1,i
{ }i
s
)THEN
*p
i
r
S,s(,T,G,s(
i
S
r
*p
));
G,g(,T,C,g(
G*p
j
j
*p
))
receiveMsg
T(
rG
))
To establish consensuses, the agents are divided into groups
(initially one agent per group). Then an agent evaluates its
knowledge to establish a new set of groups by merging the
previous ones according to their fuzzy value before sending
a message containing its evaluation to another agent which
will do the same. Once each agent computes its own
evaluation, new groups are created and the algorithm restarts
until the formed groups are not modified anymore (This
algorithm is detailed in the Table 1).
Table 1: Consensuses seeking algorithm (algorithm detailled in [7])
Description
Algorithm
is
// agents of S search for partition of fuzzy config. agents
// save the current partition
rS
rG
// if reception of a new partition
// update the groups of fuzzy configurations agents
// iteration on r groups of fuzzy configuration agents
// compute k1
// compute k2
]1..0∈α
[
// compute k ,
// end of iteration on r groups of fuzzy configuration agents
// max k founded in the group *p of fuzzy config. agents
// assign
// inform only if modification
// diffuse assigning to fuzzy configuration agents
is to the part of fuzzy solutions
*pS
// diffuse assigning to the fuzzy consensus cluster agents
// end of diffusion of changes
// while information of modifications is sent by agents of G
*pC
3.2 Implementation on the FAPIC platform
In the FAPIC platform, the consensus seeking algorithm
is used to create consensuses between the solutions and
the configurations agents (Figure 3).
Fig. 3 Process of configuration consensus seeking on FAPIC platform
To best present the results, they are shown in a matrix
which represents the fuzzy compatibility between solutions
and configurations. The algorithm groups the highest
values on blocks spread on the matrix diagonal. These
blocks represent the consensuses; they are made up of
agents which share sufficient characteristics
to be
considered as a range of products.
It can be necessary to rearrange some consensuses (for
instance if they contain several low values), by excluding
one or more configurations and obtaining a robust set of
results which matches as closely as possible all the
constraints and requirements.
Figure 4 shows
the
configuration of an office chair. The solutions are divided
into four groups: seat (S1 to S5), back (S6 to S10), armrest
(S11 to S15) and stand (S16 to S20). Each blue rectangle
the consensuses created by
highlights a set of configurations made up of four solutions
(one from each groups) which create a consensus.
Fig. 4 Visualization of the consensuses on FAPIC platform (Note that the matrix representation of consensus is an artificial representation, since the
consensus knowledge is distributed over all agents).
4. Generalized Consensus Seeking
4.1 Presentation
The next evolution of the platform should provide more
robust results by introducing the concept of "generalized
consensus".
The principle is to use the consensus seeking algorithm to
create consensuses from each community before starting
the main process, and then consider them as new super-
agents containing elementary agents. The process can be
run considering these super-agents which will be treated as
standard agents.
In these conditions, the consensuses will be considered
from the beginning of the process to obtain more robust
results.
4.2 Implementation on the FAPIC platform
To apply the new concept of "generalized consensus" to
the FAPIC platform, the consensuses are created directly
inside
functions and constraints
requirements,
the
the
first step of
the product
communities before
configuration process. The characteristics of the new
created super-agents are calculated averaging
those
containing elementary agents (Figure 5).
Fig. 5 Process of generalized consensus seeking on FAPIC platform
These new Agents are identical to the previous ones from a
computer science point of view, so the platform may run
the process normally with these agents to obtain an optimal
result (step 3) composed of a set of consensuses.
To create consensuses inside a community, the relations
between the agents of this community must be defined by
the operators concerned (the customer or the experts) by
filling the relative matrix directly on the user interface.
If these relations are not defined properly, the consensuses
cannot be created and the platform will use the elementary
agents during the configuration process.
5. Example – A Conveyor
5.1 Presentation
To illustrate the concept of "generalized consensus", the
process will be detailed on a concrete example. The
product to configure is an aerial conveyor used to transport
heavy loads in a factory.
In this case, the platform will be parameterized with
function, requirement and solution communities (the
constraints are not taken into account).
The requirements are defined by the customer (see Table
2). Then the set of functions are established (see Table 5)
thanks to the functional diagram (see Figure 8), with the
following notations:
- CS : Control Signal
- EE : Electrical Energy
- ME : Mechanical Energy
Finally, several solutions are proposed and evaluated to
perform each function (see Table 3). The evaluation is a
fuzzy value which represents the effectiveness of the
solution concerned. Then the set of solutions may be
deduced (see Table 4).
In this case, the requirement and the function communities
are defined with internal relations which enable the
creation of consensuses
the generalized
consensus. No constraints are taken into account and the
solution community does not need to be divided into
consensuses.
to apply
5.2 Results without the generalized consensus
If we compute the optimal configuration without using
the concept of generalized consensus, we obtain the
following optimal result (Figure 6).
S2
S10
S1
S12 S15 S18 S20 S23 S25 S28
S3
S5
S7
S9
Table 2: Convoyor’s requirements
Fig. 6 Simple optimal result
This result is considered as optimal but is accompanied
by a set of configuration consensuses. Each consensus
contains a set of configurations which are optimal
according to the local point of view of at least one
solution agent, these configurations form a group of
products which share similar characteristics.
5.3 Results with the generalized consensus
If we compute the optimal configurations using the
concept of generalized consensus, the FAPIC platform
will determine consensuses from the function and the
requirement communities. These new agents will be used
in the process to obtain the following results finally
(Figure 7).
S2
S1
S10
S12 S15 S18 S20 S23 S25 S28
S3
S9
S5
S7
S2
S1
S10
S13 S15 S18 S20 S23 S25 S29
S3
S5
S7
S9
S2
S1
S10
S12 S15 S18 S20 S23 S25 S29
S3
S5
S7
S9
S2
S1
S10
S13 S15 S18 S20 S23 S25 S28
S3
S5
S7
S9
Fig. 7 Optimal results with generalized consensus
We may notice that the former configuration determined
without using the generalized consensus is present, but this
time, there are four optimal results instead of just one. In
this case only the optimal results are displayed and they
already highlight a configuration consensus.
The base of each configuration found is the same and the
difference is in the S12 and S13 solutions on the one hand
then the S28 and S29 on the other hand in concrete terms,
these results suggest that the function "import support +
load" may be realized by a pivot (S12) or a screw (S13),
and the function "transmit load + support on rail" may be
realized thanks to welding (S28) or bending (S29). The
other solutions form a robust part of the conveyor
configuration.
Easy to assemble
Easy accessible pieces
Quick to install
Easy to dismantle
Minimize work
Min. rails degradation
Adaptable to rails
Ad. different rails
Ad. different supports
Respect security norms
R1
R2
R3
R4
R5
R6
R7
R8
R9
R10
R11 Minimize noise
R12
Support heavy loads
R13
R14
R15
R16
R17
R18
R19
R20
R21
R22
R23
R24
Environment resistant
Easily washable
Long life time
Ad. Different speeds
Not too cumbersome
Allow CAD
Elect. motor admission
Impermeability
High temper. resistant
Use client pieces
Use easy manuf. pieces
Minimize price
Table 3: Convetor’s solutions
Defined by customer
Defined by customer
Electrical wires
Direct contact
Prog. Logic controller
Contactor
Prog. Logic controller
Potentiometer
Defined by customer
Reducer
Belt
Pivot
Screw
S1
S2
S3
S4
S5
S6
S7
S8
S9
S10
S11
S12
S13
S14 Mandrel
S15 Mechanical contact
S16
S17
S18
S19
S20
S21
S22
S23
S24
S25
S26
S27
S28
S29
S30
Hydrostatic contact
Mecha. contact by sides
4 tensioners
3 tensioners
Adhesion tensioner/rail
Pinion/rack
Caterpillar/rail
Cable and kook
Metallic cable
Pivot
Screw
Mandrel
Welded
Bend
Fig. 8 Conveyor’s functional diagram
Table 4: Conveyor’s solution evaluation
Table 5: Conveyor’s Functions
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
Import CS
Import EE
Transmit EE
Activate EE
Regulate EE
Convert EE/ME
Transmit ME
Import support + load
Guide support + load
Import support + load
Transport support + load
Secure support + load
Export support + load
Transmit support + load
7. Conclusions
The purpose of the implementation of the generalized
consensus concept was to obtain more robust results. With
this approach, the notion of consensuses is present at each
step of the configuration process and we may observe that
the change affects the process because the latter takes this
notion directly into account and provides a group of
similar results which will be used to create a range of
products without further processing.
The next step is to use the concepts and models developed
in
the "city
organization". We consider the implementation may be
duplicated with several arrangements
to
numerous problems which will imply considering the
contribution of a large amount of entities.
the FAPIC platform
to respond
to apply
to
it
References
[1] O. Shai and Y. Reich, "Infused design. I. Theory", Res Eng
Design, Vol. 15, 2004, pp. 93-107.
[2] O. Shai and Y. Reich, "Infused design. II. Practice", Res Eng
Design, Vol. 15, 2004, pp. 108-121.
[3] P. Leitão, "Agent-based distributed manufacturing control: A
survey", Engineering Applications of
state-of-the-art
Artificial Intelligence, Vol. 22, No. 7, 2009, pp. 979-991.
[4] L. Monostori., J. Vancza, and S.R.T. Kumara, "Agent-Based
Systems for Manufacturin,. Annals of the CIRP, Vol. 55, No.
2, 2006, pp. 697-720.
[5] H.V.D. Parunak, A.D. Baker, and S.J. Clark, "The AARIA
agent architecture: From manufacturing requirements to
agent-based system design", Integrated Computer-Aided
Engineering, Vol. 8 No. 1, 2001, pp. 45-58
[6] E. Ostrosi, A.-J. Fougères, and M. Ferney, "Fuzzy Agents for
Product Configuration in collaborative and distributed design
Process", Applied Soft Computing, Vol. 12, No. 8, 2012, pp.
2091-2105.
[7] A.-J. Fougères and E. Ostrosi, "Fuzzy Agent-based Approach
for Consensual Design Synthesis in Product Integrated
Configuration", Integrated Computer-Aided Engineering,
Vol. 20, No. 3, 2013, pp. 259-274.
[8] E.K. Antonsson and K.N. Otto, "Imprecision in Engineering
Design". ASME Journal of Mechanical Design, Vol. 117B,
1995, pp. 25-32.
[9] B. Agard and M. Barajas, "The use of fuzzy logic in product
family development: literature review and opportunities",
Journal of Intelligent Manufacturing, Vol. 23, No. 5, 2012,
pp. 1445-1462.
[10] W. Ren and R.W. Beard, "Consensus seeking in multi-agent
systems under dynamically changing interaction topologies",
IEEE Trans. Autom. Control, Vol. 50, No. 5, 2005, pp. 655-
661.
[11] R. Olfati-Saber, J.A. Fax, and R.M. Murray, "Consensus
and cooperation
in networked multi-agent systems",
Proceedings of the IEEE, Vol. 95, No. 1, 2007, pp. 215-233.
[12] A.-J. Fougères, "A Modelling Approach Based on Fuzzy
Agents", Int. J. of Computer Science Issues, Vol. 9, No. 6,
2013, pp. 19-28.
[13] B. Beroule, A.-J. Fougères, E. Ostrosi, "Engineering change
management through consensus seeking by fuzzy agents",
Proceedings of The 2nd World Conference on Complex
Systems (WCCS14), 2014, Agadir, Morocco.
representation, and
Benoît Beroule received the B.S. degree from the department of
Computer Science of University of Technology of Belfort-Montbéliard
(UTBM). He is currently pursuing the Ph.D. degree in the Laboratory
IRTES-SeT (Systems and Transports) of UTBM.
Alain-Jérôme Fougères holds a PhD in Artificial Intelligence from
the University of Technology of Compiègne - France (1997). He is
a Professor of Computer Science at ESTA and he conducts his
research at ESTA-Lab. He is also associated with the Laboratory
IRTES-M3M of University of Technology of Belfort-Montbéliard.
Initially, his areas of interests and his scientific contributions were:
1) assistance to the redaction of formal specifications, where
problems focused on the natural language processing, the
knowledge
formal
specifications; and 2) design of agent-based systems, in particular
architecture,
interactions, communication, and cooperation
problems. Over the last ten years, his research has focused on
agent-based mediation for cooperative work.
Egon Ostrosi holds a Ph.D. from the University “Louis Pasteur” of
Strasbourg (ULP), and is currently Associate Professor in the
Laboratory of Mecatronics3M (M3M) of IRTES at the University of
Technology of Belfort – Montbéliard. His current research
concerns integrated mechanical product design and modeling,
knowledge engineering and its application in mechanical design,
and concurrent engineering. He has been actively involved with
the development of methods for concurrent product design, CAD
product modeling, product configuration, and collaborative and
distributed design process.
techniques of
the
|
1909.06711 | 1 | 1909 | 2019-09-15T02:02:22 | Cognitive swarming in complex environments with attractor dynamics and oscillatory computing | [
"cs.MA",
"cs.NE",
"cs.RO",
"nlin.AO",
"q-bio.NC"
] | Neurobiological theories of spatial cognition developed with respect to recording data from relatively small and/or simplistic environments compared to animals' natural habitats. It has been unclear how to extend theoretical models to large or complex spaces. Complementarily, in autonomous systems technology, applications have been growing for distributed control methods that scale to large numbers of low-footprint mobile platforms. Animals and many-robot groups must solve common problems of navigating complex and uncertain environments. Here, we introduce the 'NeuroSwarms' control framework to investigate whether adaptive, autonomous swarm control of minimal artificial agents can be achieved by direct analogy to neural circuits of rodent spatial cognition. NeuroSwarms analogizes agents to neurons and swarming groups to recurrent networks. We implemented neuron-like agent interactions in which mutually visible agents operate as if they were reciprocally-connected place cells in an attractor network. We attributed a phase state to agents to enable patterns of oscillatory synchronization similar to hippocampal models of theta-rhythmic (5-12 Hz) sequence generation. We demonstrate that multi-agent swarming and reward-approach dynamics can be expressed as a mobile form of Hebbian learning and that NeuroSwarms supports a single-entity paradigm that directly informs theoretical models of animal cognition. We present emergent behaviors including phase-organized rings and trajectory sequences that interact with environmental cues and geometry in large, fragmented mazes. Thus, NeuroSwarms is a model artificial spatial system that integrates autonomous control and theoretical neuroscience to potentially uncover common principles to advance both domains. | cs.MA | cs | Noname manuscript No.
(will be inserted by the editor)
Cognitive swarming in complex environments with attractor
dynamics and oscillatory computing
Joseph D. Monaco1 · Grace M. Hwang2 · Kevin M. Schultz2 ·
Kechen Zhang1
9
1
0
2
p
e
S
5
1
]
A
M
.
s
c
[
1
v
1
1
7
6
0
.
9
0
9
1
:
v
i
X
r
a
Received: date / Accepted: date
Abstract Neurobiological theories of spatial cogni-
tion developed with respect to recording data from
relatively small and/or simplistic environments com-
pared to animals' natural habitats. It has been unclear
how to extend theoretical models to large or complex
spaces. Complementarily, in autonomous systems tech-
nology, applications have been growing for distributed
control methods that scale to large numbers of low-
footprint mobile platforms. Animals and many-robot
groups must solve common problems of navigating
complex and uncertain environments. Here, we intro-
duce the 'NeuroSwarms' control framework to inves-
tigate whether adaptive, autonomous swarm control
of minimal artificial agents can be achieved by direct
analogy to neural circuits of rodent spatial cognition.
NeuroSwarms analogizes agents to neurons and swarm-
ing groups to recurrent networks. We implemented
neuron-like agent interactions in which mutually visible
agents operate as if they were reciprocally-connected
place cells in an attractor network. We attributed a
phase state to agents to enable patterns of oscilla-
tory synchronization similar to hippocampal models
of theta-rhythmic (5 -- 12 Hz) sequence generation. We
demonstrate that multi-agent swarming and reward-
approach dynamics can be expressed as a mobile form
of Hebbian learning and that NeuroSwarms supports
a single-entity paradigm that directly informs theoret-
ical models of animal cognition. We present emergent
This work was supported by NSF award NCS/FO 1835279,
JHU/APL internal research and development awards, and the
JHU/Kavli Neuroscience Discovery Institute.
1 Biomedical Engineering Department, Johns Hopkins Uni-
versity School of Medicine, Baltimore, MD, 21205, USA.
E-mail: [email protected]
2 The Johns Hopkins University/Applied Physics Laboratory,
Laurel, MD, 20723, USA.
behaviors including phase-organized rings and trajec-
tory sequences that interact with environmental cues
and geometry in large, fragmented mazes. Thus, Neu-
roSwarms is a model artificial spatial system that inte-
grates autonomous control and theoretical neuroscience
to potentially uncover common principles to advance
both domains.
Keywords swarming · multi-robot groups · place
cells · oscillations · spatial navigation · emergence
1 Introduction
Spatial cognition in rodents has been extensively stud-
ied in non-naturalistic environments such as linear or
circular tracks, radial arm mazes, and T-mazes, or small
open-field arenas such as squares or cylinders of approx-
imately 1 -- 2 m2 area. Such experimental conditions have
allowed individual place fields of hippocampal pyrami-
dal neurons (O'Keefe and Dostrovsky, 1971) and the
activity of other spatial cells (Knierim, 2006; Moser
et al., 2008; Savelli et al., 2008; Poulter et al., 2018;
Wang et al., 2018) to be exquisitely controlled and an-
alyzed, leading to a detailed neural coding account of
distributed representations that subserve spatial learn-
ing, memory, and planning in mammals (O'Keefe and
Nadel, 1978; Moser and Paulsen, 2001; Knierim and
Hamilton, 2011; Monaco and Abbott, 2011; Pfeiffer and
Foster, 2013; Hartley et al., 2014; Burgess, 2014; Schiller
et al., 2015; Foster, 2017), potentially extending to gen-
eral cognitive computations in humans (Bellmund et al.,
2018; Kunz et al., 2019). However, the multiplicity of
Poisson-distributed hippocampal place fields exposed
in larger environments (Fenton et al., 2008; Rich et al.,
2014) and species differences in mapping 3-dimensional
contexts (Yartsev and Ulanovsky, 2013; Casali et al.,
2
Monaco et al.
2019) suggest large and/or complex environments as
the next frontier in understanding spatial navigation.
Computational models of rodent spatial networks
have typically emulated the restricted environments
of experimental studies (for computational efficiency,
ease of analysis, and compatibility with published
data). Despite these limitations, recent theoretical re-
sults have demonstrated the importance of sensory
and cortical feedback in stabilizing and shaping hip-
pocampal and entorhinal cortical spatial representa-
tions (Monaco et al., 2011; Poll et al., 2016; Renn´o-
Costa and Tort, 2017; Ocko et al., 2018); this relation-
ship has been supported by experimental approaches
to the animals' own active sensing behaviors such as
lateral head scanning (Monaco et al., 2014; Yadav and
Doreswamy, 2017) and closed-loop control of orient-
ing distal cues (Jayakumar et al., 2019). Addition-
ally, extending prior theoretical frameworks such as
the attractor map formalism (Zhang, 1996; Tsodyks,
1999; Samsonovich and McNaughton, 1997; Knierim
and Zhang, 2012) to large spatial contexts has re-
vealed substantial increases in the computational and
mnemonic capacities of these network models (Hedrick
and Zhang, 2016). Thus, a theoretical approach to
naturalistic and dynamical spatial coding in large or
complex environments may require closed-loop systems
that integrate sensory information with internal spatial
maps in continuously adapting loops. Complementary
to the animal studies, investigating the behaviors and
performance of completely specified but artificial spa-
tial systems including virtual agents and/or mobile
robotics platforms may help to elucidate the compu-
tational principles of spatial cognition in naturalistic
contexts (Hasselmo, 2018; Tomov et al., 2018; Savelli
and Knierim, 2019; Gaussier et al., 2019).
Hippocampal phenomena that have been theo-
rized to support biological spatial cognition include
(1) self-stabilizing activity patterns in attractor map
networks and (2) temporal-phase organization relative
to global oscillations. On the basis of these phenomena,
we introduce a brain-inspired dynamical controller for
self-organized swarms of autonomous agents. Our key
realization was that each virtual or robotic agent can
be represented as a spatial neuron (e.g., a place cell)
whose place preference follows the location of the agent
in its local environment. The analogy of a multi-agent
group (or swarm) to a space-coding neural network
follows immediately. If we further suppose that inter-
agent distances map to 'synaptic weights' and, conse-
quently, that relative agent movements map to changes
in those weights, then spatial configurations of the
swarm constitute an attractor map network (Zhang,
1996; Tsodyks, 1999; Samsonovich and McNaughton,
1997) and the swarm's internal motion dynamics con-
stitute learning based on synaptic modification (Hebb,
1949; Oja, 1982).
Additionally, spatial activity in the rodent hip-
pocampus is strongly modulated by continuous theta
oscillations (5 -- 12 Hz) during locomotion (Vanderwolf,
1969; Buzs´aki, 2005). The resulting 'phase precession,'
a monotonic advance in timing from late to early within
each theta cycle, may enhance the precision and tempo-
ral organization of spatial codes (O'Keefe and Recce,
1993; Jensen and Lisman, 2000; Foster and Wilson,
2007; Drieu et al., 2018) in ways that support decision-
making and/or deliberative planning during subsequent
sleep or quiescent states (Buzs´aki, 1989; Johnson and
Redish, 2007; Buzs´aki and Moser, 2013; Wikenheiser
and Redish, 2015; Papale et al., 2016; Muessig et al.,
2019). Recently, in contrast to phase precession, we dis-
covered a novel class of spatial phase-coding neurons
in open field environments termed 'phaser cells' that
were located predominantly in the rat lateral septum
and characterized by a strong coupling of theta-phase
timing to firing rates (Monaco et al., 2019a). This
coupling supports an intrinsic neuronal mechanism of
phase-coding that may theoretically transform spatial
information to synchronize downstream targets using
the global theta oscillation (Monaco et al., 2019a).
Accordingly, to examine the effects of temporal phase
organization, our dynamical swarm controller considers
each agent to have an internal phase variable (anal-
ogous to the theta-phase of a place cell or phaser
cell) that may interact via oscillatory coupling with
its neighbors' phases. Such oscillator-based swarm-
ing has been previously generalized as the 'swarmala-
tors' framework (O'Keeffe et al., 2017). Thus, together
with attractor dynamics, these phenomena may pro-
vide brain-like solutions to problems of decentralized
self-organization and distributed communication in au-
tonomous swarming.
We refer to our conception of this swarm controller
as NeuroSwarms (Monaco et al., 2019b; Fig. 1A). In
this paper, we derive an operational NeuroSwarms im-
plementation (Section 2), present emergent swarming
behaviors in simulations of a fragmented and het-
erogeneous environment (Section 3.1), demonstrate
NeuroSwarms as a dual system which can be ex-
pressed through single-entity simulations that help
inform biological theory (Section 3.2), evaluate adap-
tations of reward-approach behaviors in a large hairpin
maze (Section 3.3), and discuss implications for au-
tonomous systems design and biological spatial cogni-
tion in large, complex environments (Section 4).
Cognitive swarms with attractors and oscillators
3
Fig. 1 Conceptual schematic and theoretical neuroscientific inspiration for the NeuroSwarms controller. A. An artificial spa-
tial system of mobile virtual or robotic agents communicate over sparse recurrent channels (bottom) just as spatial neurons
in biological neural circuits produce reverberating activity patterns that reflect energy minima in the dynamical state-space
of the system (e.g., fixed-point attractors; top; adapted from Knierim and Zhang, 2012). B. Example simulation of the spatial
self-organization of an activity bump on an attractor map. In an attractor map network, the environment is represented by
a continuum of locations with overlapping place fields, leading to network connectivity that produces self-reinforcing spatial
activity patterns. Adapted from Zhang (1996). C. Schematic of a minimal model of temporal-phase coding in which an exci-
tatory external input (green) is rhthmically modulated by a continuous inhibitory oscillation (blue) such as the hippocampal
theta rhythm. Adapted from Monaco et al. (2019a) as permitted by the CC-BY 4.0 International License (creativecom-
mons.org/licenses/by/4.0/).
2 Methods
instance, a rate-based network following
(cid:88)
j
,
2.1 Hippocampal model mechanisms
= −ri + g
dri
dt
Jijrj + Ii
(1)
Self-stabilizing attractor maps. Hippocampal place
cells fire within a contiguous region of the animal's local
environment, or 'place field' (O'Keefe and Dostrovsky,
1971). Place fields are thought to collectively form
cognitive maps (O'Keefe and Nadel, 1978) that are
stabilized (at least in part) via attractor dynamics,
such as fixed points or continuous manifolds of the
network energy surface, that drive activity toward low-
dimensional spatial or task representations (Fig. 1B;
Knierim and Zhang, 2012). Attractor map models have
shown that recurrent connectivity between place cells
with nonlinear integration of inputs is nearly sufficient
to achieve stable spatial attractors (Zhang, 1996; Sam-
sonovich and McNaughton, 1997; Tsodyks, 1999). For
where ri is the rate of unit i, Ii is the unit's total in-
put, and g is a sigmoidal nonlinearity, only further re-
quires that the recurrent weights Jij encode the degree
of place-field overlap between units (i.e., the strength of
learned spatial associations). Such an encoding follows
from a kernel function of field-center distances, e.g.,
Jij := F (xi − xj) = A exp
− B ,
(2)
(cid:18)
−xi − xj2
(cid:19)
σ2
where xk is the field-center position of unit k, σ is the
Gaussian scale constant, and A and B determine the
strength of local excitation vs. long-range inhibition, re-
spectively. While this formulation violates Dale's law, it
illustrates the typical parsimony of attractor map mod-
els (Tsodyks, 1999). A network constructed from equa-
4
Monaco et al.
tions (1) and (2) supports self-organization of its activ-
ity into a singular, contiguous 'bump' that emerges as
the network relaxes (Fig. 1B; Zhang, 1996). The activ-
ity bump can then respond to input changes due to, e.g.,
movement through the environment or internal pro-
cessing. These conditions are encapsulated within Neu-
roSwarms by analogizing (1) inter-agent visibility (e.g.,
for a line-of-sight communication channel) to sparse, re-
current network connectivity, and (2) a local kernel of
inter-agent distances to spatially-weighted synapses.
Oscillatory organization. Hippocampal phase preces-
sion relative to the theta rhythm is clearest on linear
tracks (O'Keefe and Recce, 1993), on which place fields
typically arrange in an unambiguous sequential order
that may enable learning of temporally compressed
'theta sequences' (Feng et al., 2015). While the phe-
nomenon is less clear in open, 2-dimensional environ-
ments, phase precession has been observed in particular
traversals of the firing fields of place cells as well as grid
cells in entorhinal cortex (Climer et al., 2013; Jeewa-
jee et al., 2014). However, our analysis of phaser cells
in the lateral septum (Monaco et al., 2019a) revealed
a more direct phase code for 2-dimensional position
that was consistent with a minimal model of temporal
phase-coding (Fig. 1C). To study how the phaser cell
code may contribute to sequence formation in open
environments, the NeuroSwarms controller adds an in-
ternal phase variable to each agent and couples the
modulation frequency of that phase to each agent's
total input (see equation (12) below).
Internalized place fields for neural control. There are
two reasons why neural swarming control must decou-
ple each agent's physical location from its internal self-
localization. First, the multiplicity of agents is a qual-
itative difference with brain circuits; every hippocam-
pal neuron in a biological network corresponds to the
same single agent (e.g., a rat) and has particular inputs
from internal processing or sensory inputs (Wang et al.,
2016) that contribute to the appearance and location of
the cell's place field. Given the analogy of agents to neu-
rons (Fig. 1A), an individual rat has many place fields
but an individual swarm agent should have only one.
Further, the location of an agent's place field cannot be
identical to the agent's physical location, which depends
on the momentary vicissitudes of a entity operating in
the external world (or a simulation thereof). Second,
experimental studies have compellingly demonstrated
that spatial path planning in hippocampal networks
may rely on activating sequences of place cells repre-
senting remote locations (Gupta et al., 2010; Pfeiffer
and Foster, 2013; ´Olafsd´ottir et al., 2018; Momennejad
et al., 2018), indicating that internal representations
should be separable from an animal's (or agent's) cur-
rent physical location. Thus, NeuroSwarms assigns a
distinct cue-based place preference to each agent.
2.2 Mobile Hebbian learning with a global oscillation
Hebbian learning in neural network models typically
increments or decrements a synaptic weight according
to a learning rate and a measure of the activity corre-
lation between the pre-synaptic (input) and the post-
synaptic (output) neurons (Hebb, 1949; Levy and Stew-
ard, 1979; Oja, 1982; Eichenbaum, 2018). For the Neu-
roSwarms controller, the conceptual similarity of the
synaptic strength relation in a neural network and the
physical distance relation in a multi-agent group allows
us to construct a neural activation and learning model
for the motion of artificial mobile agents.
ij/σ2) ,
Following the Gaussian attractor-map kernel from
equation (2), we explicitly relate a recurrent synaptic
weight matrix W ∈ RNs×Ns, prior to learning-based
updates, to swarm state via
Wij = Vij exp(−D2
(3)
for inter-agent visibility V ∈ {0, 1}Ns×Ns , inter-agent
distances D, and spatial constant σ. To provide for
environmental
interactions, we consider a minimal
reward-approach mechanism for a set of reward coordi-
nates that serve as attractive locations. Thus, we like-
wise incorporate a feedforward matrix W r ∈ RNs×Nr
for reward learning,
ik exp(−Dr
W r
(4)
for agent-reward visibility V r ∈ {0, 1}Ns×Nr , agent-
reward distances Dr, and spatial constant κ. The re-
ward weights are based on an exponential kernel to al-
low for long-range approach behaviors. (We emphasize
that the NeuroSwarms framework encompasses the gen-
eral equivalence between synaptic weights and agent
distances, but the particular implementation that we
present here is one of many possible designs.)
ik = V r
ik/κ) ,
To define neuron-like inputs, we consider that each
agent's internal place-field location derives from the
conjunction of sensory cue inputs related to a preferred
location. We define time-continuous sensory cue inputs
c ∈ RNs×Nc following
ik − ci ,
τc cik = V c
(5)
for cue k, agent-cue visibility V c ∈ {0, 1}Ns×Nc , fixed
agent-cue preferences V c∗ ∈ {0, 1}Ns×Nc, and integra-
tion time-constant τc. Thus, the product in equation (5)
yields the integer number of preferred cues visible from
ikV c∗
Cognitive swarms with attractors and oscillators
5
each agent's position. This means that place-field size
is not independently controlled but determined by the
relative cue richness of the environment: more cues will
generally increase agent heterogeneity and spatial selec-
tivity. Similarly, we compute reward inputs r ∈ RNs×Nr
following
τr rik = V r
ik − ri ,
(6)
for reward k, and integration time-constant τr. Unlike
sensory cues, all agents respond equally to all (visible)
rewards. Lastly, we define recurrent swarming inputs
q ∈ RNs×Ns following
τq qij = Vij cos(θj − θi) − qij ,
(7)
for 'post-synaptic' agent i, 'pre-synaptic' agent j, inte-
gration time-constant τq, and internal oscillatory phase
state θ. The cosine term in equation (7) confers phase-
modulation of the input in which excitation/inhibition
depends on whether units i and j are in/out of sync.
Having defined the input signals, we consider 'net
currents' as gain-modulated, visibility-normalized quan-
tities for sensory cue inputs,
Iqi =
Wijqij ,
(10)
where the parameter gains gc, gr, and gs sum to 1. Be-
cause the net inputs are bounded in equations (8) -- (10),
we simply apply linear rectification rather than a sat-
urating nonlinearity (cf. equation (1)) to their sum to
calculate post-synaptic activation
p = [Ic + Ir + Iq]+ ,
(11)
which is the remaining component needed to com-
pute Hebbian (or any two-factor) learning. In terms of
swarming, however, the agents are phase-coupled via
equation (7). Thus, in the same way that a spiking
neuron can be reduced to a phase description of its
orbit on the phase plane, we consider that p drives the
agents' internal phase state, e.g.,
θ = ω0 + ωI p ,
fi =
(12)
cik ,
W r
ikrik ,
(8)
(9)
reward inputs,
Ici =
Iri =
gc(cid:80)
k V c
ik
gr(cid:80)
k V r
ik
gs(cid:80)
j Vij
Nc(cid:88)
k=1
Nr(cid:88)
k=1
Ns(cid:88)
j=1
and recurrent swarming inputs,
where ωI sets the maximum increase in input-modulated
angular frequency above the baseline frequency ω0.
The net effect of this mechanism is that agents have
place-cell-like spatial tuning with phaser-cell-like phase
coding and synchronization.
The core of the NeuroSwarms controller comprises
the learning-based updates to W and W r. A naıve
Hebbian rule, such as dWij = ηpiqj, would cause
weights to grow unbounded, leading to ictogenesis in
recurrent networks or spatial singularities in swarms.
Instead, after updating agent activations via equa-
tion (11), we compute updated weights W (cid:48) as
ij = Wij + ∆t ηVij pi(qij − piWij) ,
W (cid:48)
with simulation time-step ∆t and learning rate η, which
effectuates a pre-synaptic normalization according to
Oja's rule (Oja, 1982). Similarly, the updated feedfor-
ward weights W r(cid:48) are computed for reward k as
W r (cid:48)
ik pi(rik − piW r
ik + ∆t ηrV r
ik = W r
ik) .
(13)
(14)
The normalization effected by equations (13) and (14)
is due to a subtractive term, quadratic in the post-
synaptic activation p, that depresses the growth of
overly active synapses. In place-cell network models,
feedback inhibition typically serves to spread out place
fields to more efficiently map an environment (Savelli
and Knierim, 2010; Monaco and Abbott, 2011), but
the lack of explicit inhibition in NeuroSwarms allows
synaptic depression to provide a similar repulsive role
due to the distance -- weight equivalence of equation (3).
2.3 Neural swarm control: closing the loop
To integrate with swarming, the controller attempts
to drive the agents' kinematic states to the equiva-
lent desired inter-agent distances, in effect replacing
the typical attraction and repulsion fields of conven-
tional swarming models (e.g., Gazi and Passino, 2011).
The updated weights W (cid:48) and W r (cid:48) can be converted
directly to desired distances by inverting the Gaussian
swarming kernel in equation (3),
D(cid:48)
ij =
(cid:113)−2σ2 log W (cid:48)
(15)
ij ,
and the exponential reward kernel in equation (4),
ij = −κ log W r (cid:48)
Dr (cid:48)
ij .
To compute the resultant swarm motion, the desired
positional shift of agent i is averaged across its visible
neighbors, i.e.,
(16)
Vij (D(cid:48)
ij − Dij)
xj − xi
xj − xi ,
(17)
2(cid:80)
1
j Vij
Ns(cid:88)
j=1
6
Monaco et al.
∆xb
i = (1 − βsi)∆xi + βsi∆xi nsi ,
(20)
vs =
and the resultant reward-related motion is similarly
computed as the average
1(cid:80)
k V r
ik
f r
i =
Nr(cid:88)
k=1
ik (Dr(cid:48)
V r
ik − Dr
ik)
k − xi
xr
xr
k − xi .
(18)
The net positional shift is calculated as a linear combi-
nation of the swarm- and reward-related shifts,
∆x = αf + (1 − α)f r ,
(19)
where α = 0.5 for all simulations presented. The re-
maining processing of ∆x in our NeuroSwarms im-
plementation serves to embed the foregoing dynamics
within 'physical' simulations of irregular or complex
2-dimensional environments. First, our example evi-
ronments (Fig. 2) of ∼500-point height (for arbitrary
points units) were processed for wall proximity and nor-
mal vectors for all interior locations. Thus, a 'barrier
aware' positional shift ∆xb is calculated as
for an exponential kernel βs = exp(−d/λ) of distance d
to the nearest wall with a constant of λ = 20 points, and
the normal vectors ns of the nearest wall. These shifts
update the internal place-field locations xs ← xs + ∆xb
of each swarm agent. Second, 'physical' agent locations
are updated based on the instantaneous velocity needed
for each agent to approach their internal field locations,
vs = (xs−x)/∆t, which is processed through a momen-
tum filter,
vµ = µv + (1 − µ)vs ,
(21)
with the actual velocity (prior to updating) v and coeffi-
cient µ; a speed-limiting nonlinearity based on a kinetic-
energy maximum Emax,
vmax =(cid:112)2Emax/m ,
(cid:18) vµi
vki = vmaxi tanh
vmaxi
(cid:19) vµi
vµi ,
(22)
(23)
where m is agent mass; and barrier avoidance as in
equation (20),
vi = (1 − βi)vki + βivki ni ,
(24)
for proximity β and normal vectors n. Finally, agent
locations are updated by x ← x + v∆t.
2.3.1 Single-entity simulations
In keeping with the neuroscientific motivation for Neu-
roSwarms (Fig. 1), our implementation allows for sin-
gleton simulations analogous to conventional models of
neural networks in an animal such as a rat navigating
a maze. With only minor adjustments, NeuroSwarms
can operate with a single agent (i.e., N = 1) that
owns a collection of 'virtual' or 'mental' swarming par-
ticles (e.g., Ns = 300) that guide the agent's spatial
behavior. In this sense, the virtual swarm represents
a highly dynamic spatial field that provides the agent
with various options for constructing a path through
the environment. The dynamics of the virtual swarm
are as described above up to equation (20). An array
V δ ∈ {0, 1}Ns indicates which particles' positions are
visible to the agent and serves to additionally mask the
learning updates in equations (13) and (14). To produce
motion, single-agent velocity is instead calculated using
a cubic-activation-weighted average of the swarm,
∆t(cid:80)
1
j V δ
j p3
j
Ns(cid:88)
i=1
V δ
i p3
i (xsi − x) ,
(25)
prior to processing the 'physical' embedding of the
agent's motion in equations (21) -- (24). Thus, the agent
constructs a path toward the most highly activated and
visible swarm particles.
2.4 NeuroSwarms simulations
Simulated environments (Fig. 2) contain fixed-position
rewards and cues depicted as gold stars and purple
shapes, respectively. Environments are otherwise de-
fined by a set of linear barrier segments (e.g., walls) that
form a closed shape defining an interior space that be-
comes the set of allowable agent locations. Simulations
are initialized by setting all velocities, input signals, and
activations to zero, randomly choosing internal phase
states, and randomly assigning agent positions to al-
lowable locations within a set of 'spawn discs' defined
in the environment. Random number generator seeds
are reused for simulations meant to compare parame-
ter values, unless otherwise specified. Environments are
specified as vector image files in Tiny SVG format with
XML text nodes defining reward, cue, and spawn disc
locations. Unless noted, parameters were set to the de-
fault values displayed in Table 1. The python source
code will be made available upon reasonable request.
Cognitive swarms with attractors and oscillators
7
Fig. 2 Example post-initialization (t = 0.1 s) swarm states for NeuroSwarms simulations. (Left) A single-agent simulation in
the 'multireward' arena, which contains 3 rewards (gold stars; northwest, southwest, southeast), 7 cues (purple shapes), and
3 spawn discs. White enclosed areas constitute the set of allowable locations for swarm agents; black regions constitute barriers
and disallowed locations. Initial particle positions are sampled from the spawn discs and initial phases are random. Green
circle in southwest: the single agent; dots: 300 virtual swarm particles with internal phase state indicated by color. (Right)
A multi-agent simulation in the 'hairpin' maze, which contains 5 connected hallways, 3 rewards, 7 cues, and 4 spawn discs.
Circles: 300 swarm agents with internal phase state indicated by color; reward (gold star) size is for visual differentiation only
and has no effect in the model.
Table 1 Parameters, default values, and descriptions (with
units) for the NeuroSwarms controller implementation.
∆t
duration
N
N
Ns
0.01
180.0
300
1
300
†
Dmax
Emax
µ
mmulti
msingle
σ†
κ†
η
ηr
ω0
ωI
1.0
3e3
0.9
0.3
3.0
1.0
1.0
1.0
1.0
0.0
1.0
s, integration time-step of simulation
s, total simulation time
no. of 'physical' agents (multi-agent)
no. of 'physical' agents (single-entity)
no. of internal fields (multi-agent) or
virtual particles (single-entity)
max. inter-agent visibility range
kg·points2/s2, max. kinetic energy
momentum coefficient of agent motion
kg, mean agent mass (multi-agent)
kg, agent mass (single-entity)
spatial scale of swarm interaction
spatial scale of reward interaction
learning rate for swarm connections
learning rate for reward connections
cycles/s, baseline oscillatory frequency
cycles/s, max. increase in oscillatory
frequency due to neural activation
gain of sensory cue inputs
gain of reward inputs
gain of swarming inputs
s, time-constant of sensory cue inputs
s, time-constant of reward inputs
s, time-constant of swarming inputs
points, reward contact radius
0.4
0.2
0.4
0.5
0.5
0.1
0.0
gc
gr
gs
τc
τr
τq
drad
† These parameter values are multiplicatively scaled to the
notional environment size, defined in points as the radius of
a disc with the same area as the set of allowable locations in
the environment's interior.
3 Results
3.1 Emergent swarming behaviors
We designed the multireward arena (Fig. 2, left) to
characterize emergent swarming and reward approach
behaviors, and the hairpin maze (Fig. 2, right) to
assess behavioral adaptation in large, fragmented en-
vironments. We observed several emergent dynamical
behaviors in simulations of both multi-agent swarm-
ing and single-entity locomotion (Section 2, Methods).
The most notable and persistent behaviors included
the emergence of phase-sorted spatial formations such
as line segments, rings, or concentric loops (Fig. 3).
These behaviors were analogous in form to (1) the
'phase wave' states observed in certain swarmalator
regimes (O'Keeffe et al., 2017), and (2) the hippocam-
pal phenomena of theta sequences and theta-rhythmic
phase assemblies (Foster and Wilson, 2007; Drieu et al.,
2018). Further, by inspection of simulation movies, we
observed two dynamical
features. First, agent sub-
groups forming line segments and rings continuously
phase-synchronized in a shared oscillation that was in-
dependent from the absolute movement or rotation of
the formation in space. Second, line or ring formations
would often break apart and re-form new configura-
tions that typically involved other agents or formations
that were able to phase-synchronize with elements of
the subgroup. These alternating disintegrative and ag-
gregative dynamics may be consistent with analyses of
8
Monaco et al.
persistent homologies in place-cell networks with tran-
sient connectivity (Babichev and Dabaghian, 2017).
These spatiotemporal dynamics are evident across
frame captures of multi-agent (Fig. 3A) and single-
entity (Fig. 3B) simulations. While phase-ordered
groups could appear far from rewards (Fig. 3A, last
two frames, smaller red circles), swarm agents typically
approached a reward location and formed a rotat-
ing ring centered on the reward position (Fig. 3A,
southeast corner, last three frames). Such reward rings
appeared in single-entity simulations, but the virtual
swarm particles (Section 2.3.1) additionally exhibited
particularly extended line segments that often traced
out phase-ordered trajectory sequences; e.g., the agent
followed an extended sequence to the reward located in
the southeast corner (Fig. 3B, last two frames). Further,
we observed that the size of reward rings decreased over
time, reflecting a relaxation of phase and momentum
given the centrally organizing reward location.
When the reward kernel's spatial scale κ (equa-
tion (4); Table 1) was increased, streams of virtual
swarm particles formed around distal rewards as the
particles' motion was modulated by agent visibility
interacting with the geometry of the environment. As
shown in the first frame of Fig. 3C, a step-like pat-
tern formed near the northwest reward location while
a wavy pattern formed near the southeast reward lo-
cation. Both virtual swarm formations presented path
choices to the single agent located in the large central
compartment of the arena. As expected (Section 2.3.1),
virtual swarm particles that were not visible to the
agent remained fixed in place due to masking of the
weight updates in equations (13) and (14). In addi-
tion to single rings, double and even triple concentric
loops of nested, non-overlapping, phase-sorted rings
were observed in some simulations. An example of a
double loop forming is shown in the southeast cor-
ner at t = 16.74 s (Fig. 3C). Strikingly, we did not
predict, expect, or adapt the NeuroSwarms controller
design to observe these emergent behaviors; we simply
implemented abstractions for place-cell spatial tuning,
phaser-cell oscillatory synchronization, and a distance --
weight equivalence for Hebbian learning with notions
of visibility and environmental geometry that provided
spatial barriers to communication (Section 2). These
behaviors would be unexpected as well from conven-
tional swarming algorithms (Gazi and Passino, 2011).
3.2 Reward-based behavior in a compartmented arena
To assess the spatial performance of NeuroSwarms,
we examined the ability of single-entity behavior to
find all three rewards in the multireward arena. We fo-
cused on the parameter constants governing the spatial
scale of swarm (σ) and reward (κ) interactions (equa-
tions (3) and (4); Table 1) and found (σ, κ) values for
which the agent approached multiple rewards regard-
less of its initial spawn location. Due to the random
initialization of location, we selected 40 simulations for
analysis in which the agent was spawned in the south-
west corner (as in Fig. 2, left). The agent successfully
captured one, two, or all three rewards in 11, 28, and
1 simulation(s) at elapsed times ranging from 4 -- 108,
20 -- 179, and ∼160 s, respectively. Frame captures of
reward approaches are shown in Fig. 4A for the simula-
tion in which all three rewards were found. The ability
of the agent to approach multiple fixed rewards over
time was an unexpected and emergent behavior: based
on our NeuroSwarms implementation, we had predicted
that the rewards would serve as stable attractors in the
absence of additional mechanisms such as adaptation
or reward learning. However, in accordance with those
expectations, we observed simulations which failed to
explore much of the environment after approaching a
single reward location. For the same parameters but a
different random seed than shown in Fig. 4A, a failed
exploration occurred (Fig. 4B) when the virtual parti-
cles split into two fixed-point, out-of-phase attractors
that essentially 'trapped' the agent.
To counteract these unsuccessful equilibria, we
implemented a 'reward capture' mechanism in the
NeuroSwarms controller based on a minimum con-
tact radius, drad. This feature causes rewards to cease
being attractive locations to the virtual swarm par-
ticles upon contact by the agent, thus releasing the
agent from reward-related attractors before further
exploration is prevented. Indeed, having capturable
rewards with drad = 12 points enabled a simulation
that was otherwise identical to the failed case (Fig. 4B)
to successfully navigate the arena to capture all three
rewards (Fig. 4C). Thus, a notion of reward adaptation
or reward consumption may be crucial to achieving
continuous exploration.
For the 40 single-entity simulations with fixed re-
wards, the bottom panel of Fig. 5A reveals strong at-
tractors at the southeast and northwest corners of the
arena associated with reward locations. To demonstrate
the effect of the contact radius on exploration when re-
wards were capturable, the trajectories resulting from
contact radii of 1, 4, 10, and 15 points are shown in the
top row of Fig. 5A; these values produced 1, 3, 8, and
30 (out of 40) simulated trajectories, respectively, that
successfully contacted all three rewards (Fig. 5A, red
traces). In a few simulations, the single-entity agent
spawned in the southwest corner, found the southeast
Cognitive swarms with attractors and oscillators
9
Fig. 3 Temporal evolution of swarming and single-entity approaches to rewards. A. Three agent-clusters were initially popu-
lated in the multireward arena. The internal place-field location of each agent is indicated by a small black dot (e.g., t = 1.26 s,
inset, top right). Phase sorting is indicated by sequentially ordered colors of the circle markers representing agent positions.
A reward-centered phase ring was created (t = 9.71 s) with a decreasing diameter over time (southeast corner, t = 22.49 s
and t = 24.32 s). NeuroSwarms parameters: σ = 1.5, gc = 0.2, gr = 0.3, gs = 0.5; Table 1. B. A single-entity agent (green
circle) was guided by Ns = 300 virtual particles (phase-colored dots). Swarm particles formed phase sequences leading the
agent from the southwest corner to the reward location in the southeast corner of the arena by t = 20.3 s. NeuroSwarms
parameters: σ = 4, κ = 1.5, gc = 0.2, gr = 0.3, gs = 0.5; Table 1. C. Step-like patterns of particles appeared near rewards
that were occluded from the perspective of the single agent by corners in the environmental geometry. While the agent became
'indecisive' around t = 10.24 s, as it was pulled simultaneously in both directions, the agent ultimately found its way to the
southeast reward by t = 16.74 s. NeuroSwarms parameters: σ = 4, κ = 8, gc = 0.2, gr = 0.3, gs = 0.5; Table 1.
reward first, and then later returned to the southwest
corner in order to collect all three rewards; such a wan-
dering trajectory suggests that the model might qualify
as an ergodic system under these conditions, but that
hypothesis would be appropriately addressed by future
analytical studies. These results demonstrate that the
sensitivity of reward capture modulates exploratory
variability by mitigating the effect of reward-related
attractors. Histograms of the time-to-capture profile
across agent spawn sites and reward locations reflect
the structure of the environment as well as the different
possible sequences of reward contact (Fig. 5B). Thus,
the contact radius for capturable rewards exerted sub-
stantial control over the likelihood of the single-entity
agent finding all rewards in the environment.
3.3 Behavioral adaptation in large hairpin mazes
A key challenge facing current state-of-the-art swarm
controllers is the inability to rapidly adapt to dynamic
changes in complex environments. The hairpin maze
is well suited to study such adapation because swarm
agents spawned from certain hallways do not have line-
of-sight visibility to rewards that may be located in ad-
jacent hallways. A form of behavioral adapation can be
assessed based on whether agents spawned into reward-
free hallways can nonetheless navigate to rewards in
other parts of the maze.
We examined multi-agent swarming dynamics in the
hairpin maze under several conditions: pure swarming
(equation (13); Fig. 6A); swarming with sensory cue
inputs (equation (8); Fig. 6B); and swarming with sen-
sory cue inputs and reward approach (equation (14);
Fig. 6C). The sample frames shown in Fig. 6 demon-
strate the emergence of phase-ordered structures in
each of these conditions with the clear distinction that
tightly configured reward rings became prevalent when
reward learning was activated (Fig. 6C). In that condi-
tion, with the same NeuroSwarms features as studied
above in the multireward arena, it was also clear that
agents in the second and third hallways had difficulty
leaving to find another hallway with a reward. We ex-
pected this was due to (1) the parity of swarming and
reward spatial constants (σ, κ), which perhaps overem-
phasized swarming at the cost of reward-following in
highly-partitioned environments, and (2) the need for
more sensitive reward-capture. Thus, we simulated
a condition with fixed and capturable rewards using
drad = 10 points but also increased the spatial constants
10
Monaco et al.
Fig. 4 Single-entity reward-approach behavior with fixed or capturable rewards. The agent was initialized to the spawn disc
in the south west corner of the multireward arena. A. A rare example in which the single agent (green circle) captured all three
rewards when rewards were fixed (i.e., they remained attractive despite previous contact with the agent): southwest reward at
∼8.9 s, southeast reward at ∼33 s, and northwest reward at ∼160 s. Movie frames show the initial contacts with each reward
(gold stars). NeuroSwarms parameters: σ = 4, κ = 1.5, gc = 0.2, gr = 0.3, gs = 0.5; Table 1. B. With the same parameters as
(A) but initialized with a different random seed, this final frame of a simulation shows the converged state after the agent was
attracted to the southwest corner and remained there for the duration. The red ellipse highlights that the agent was 'stuck'
between two fixed-point attractors that formed through mutual phase-desynchronization. C. With the identical parameters
and random seed as (B), rewards were made to be 'capturable' at a minimum contact radius of drad = 12 points. Thus,
rewards ceased to be attractive locations once the agent made initial contact. The agent captured the southwest reward at
∼5 s, the southeast reward at ∼27 s, and the northwest reward at ∼60 s. Transparent/white stars indicate captured rewards.
NeuroSwarms parameters: σ = 4, κ = 1.5, gc = 0.2, gr = 0.3, gs = 0.5, drad = 12; Table 1.
with 3.3-fold bias for the reward value (σ, κ) = (2, 6.6)
(equations (3) and (4); Table 1). Multi-agent trajec-
tories for this enhanced reward-exploration regime are
shown in Fig. 6D: with fixed rewards (top panel), the
reward attractors dominate the dynamics and agents
generally stayed within their initial hallways; with
capturable rewards (bottom panel), there was sub-
stantially more path variability between agents, spatial
coverage increased (cf. the spiral patterns characteristic
of agents' exits from reward locations after contact),
and many more agents were able to traverse from one
hallway to the next.
To assess the converged state of multi-agent dynam-
ics in the hairpin maze, we simulated N = 300 agents
for 300 s using the same parameters and fixed rewards
as the top panel of Fig. 6D. The temporal progression of
swarm state across the simulation frames presented in
Fig. 7 shows distinct stages exhibited by the four initial
clusters of the swarm. The two clusters that spawned in
reward-free hallways eventually found their way around
the barriers to adjacent hallways after milling in vari-
ous line segment or ring formations for nearly a minute
(Fig. 7). All of the clusters successfully converged onto
the three reward locations in the maze, but the two that
traversed hallways left some agents behind. The pro-
gression of those swarm clusters from initial positions
to ring/arc formations to linear trajectory sequences to
fixed-point reward attractors illustrates a high degree
of spontaneous adaptation to the circumstances in the
hairpin maze. These dynamics were self-organized and
emergent, providing behaviors that resulted in nearly
complete convergence to reward locations. Thus, Neu-
roSwarms demonstrated autonomous spatial navigation
to unknown, occluded, and remote rewards in a large
and complex environment.
Cognitive swarms with attractors and oscillators
11
Fig. 5 Dispersion of exploratory trajectories with capturable rewards. A. Superimposed agent trajectories are shown from
40 single-entity simulations of 180 s duration in which the agent was initialized to the southwest corner (Section 2.3.1). With
fixed (non-capturable) rewards, only 1 simulation (bottom, red trace) contacted all three rewards in the arena (see Fig. 4A)
and there was minimal variance in the exploratory paths taken by the agent in the other simulations (black traces). The dense
sampling of the northwest and southeast reward location indicates these were strong attractors for the agent. With increasing
contact radii of 1, 4, 10, or 15 points (top), exploratory variance increased, the reward attractors became relatively weaker, and
higher proportions of agent trajectories successfully visited all three rewards (red traces). NeuroSwarms parameters: σ = 4,
κ = 1.5, gc = 0.2, gr = 0.3, gs = 0.5. Gold stars: reward locations. B. For 700 single-entity simulations with random initial
agent locations and drad = 15, histograms for each of the agent spawn locations (central, southeast, or southwest) display the
time-to-capture profile of each of the three rewards. NeuroSwarms parameters same as the top right panel of (A).
4 Discussion
We introduced the NeuroSwarms controller as a model
for studying neural control paradigms of artificial
swarming agents. We demonstrated that NeuroSwarms
also acts as a two-way bridge between artificial systems
and theoretical models of animal cognition. This reci-
procity arises due to a single-entity paradigm in which
NeuroSwarms controls a single agent in which an inter-
nal, virtual 'cognitive swarm' guides the agent's spatial
behavior. Both modes of operation, multi-agent and
single-entity, share the same underlying neural mech-
anisms (with differences described in Section 2.3.1).
This duality enables developments in artificial sys-
tems to also inform advances in neurobiological theo-
ries of spatial cognition. Additionally, this duality will
aid discovery of neural dynamics in large, uncertain,
and/or complex environments based on closed-system
approaches to distributed spatial coding. We presented
behaviors responding to environmental complexities
such as multiple reward sites (that optionally interact
with the system by being 'consumed' by agents), het-
erogeneous agent-based preferences for neutral-valued
spatial cues, and geometric constraints that occlude
agents' visibility of cues, rewards, and other agents.
Swarms governed by NeuroSwarms self-organize
into emergent, transitory configurations in position
and phase that directly recall spatial attractor dy-
namics (Zhang, 1996; Tsodyks, 1999; Samsonovich and
McNaughton, 1997; Hedrick and Zhang, 2016; Knierim
and Zhang, 2012) and sequential oscillatory phenom-
ena (O'Keefe and Recce, 1993; Foster and Wilson,
2007; Drieu et al., 2018; Monaco et al., 2019a) that
have been theorized to operate within hippocampal
circuits. We explicitly designed NeuroSwarms to com-
bine features of attractor maps and oscillatory comput-
ing using robust transformations (such as the spatial
kernels of distance converted to synaptic strengths in
12
Monaco et al.
Fig. 6 Dynamics of a multi-agent swarm in a large hairpin maze. Movie frame captures are shown for simulations with N = 300
agents in a rectangular environment (885×519 points including borders) partitioned into 5 hallways in a hairpin pattern. Three
hallways contain rewards which are substantially occluded from the other maze sections. Emergent formations are circled in red.
A. Frames from a pure swarming simulation, without reward or sensory cue influence. NeuroSwarms parameters: dmax = 1.5,
η = 1, ηr = 0, gc = 0, gr = 0, gs = 1; Table 1. B. Frames from a capturable-rewards simulation with 1:1 swarm/cue input
gains but no reward influence. NeuroSwarms parameters: dmax = 1.5, η = 1, ηr = 0, gc = 0.5, gr = 0, gs = 0.5; Table 1. C.
Frames from a capturable-rewards simulation with equalized swarm, reward, and cue input gains. NeuroSwarms parameters:
dmax = 1.5, ηs = 1, ηr = 1, gc = gr = gs = 1/3; Table 1. D. Multi-agent trajectories are shown from two 80 s simulations:
fixed rewards (top) and capturable rewards with drad = 10 points (bottom). Compare with multireward arena simulations in
Fig. 5A. NeuroSwarms parameters: dmax = 1.5, σ = 2, κ = 6.6, gc = 0.1, gr = 0.1, gs = 0.8; Table 1.
equations (3) and (4)). However, we did not antici-
pate how readily such a system would self-organize
into a variety of dynamic spatiotemporal structures
that recombined in complex patterns while supporting
navigation through our environments. A weakness of
the presented implementation was the use of a global,
shared oscillation without allowing for noise, drift,
or independent perturbations (cf. Zilli and Hasselmo,
2010; Monaco et al., 2011, 2019a). A more decentral-
ized approach might utilize resonant agent-oscillators
that self-organize local oscillations depending on avail-
able information, task requirements, or context. Such
bottom-up oscillations might aggregate into a global,
swarm-wide oscillation under certain conditions, which
should be studied in future models.
To leverage inertial, energetic, and cost benefits of
small-scale robots, critical future applications of au-
tonomous technologies may depend on coordinating
large numbers of agents with minimal onboard sens-
ing and communication resources. However, a critical
problem for autonomous multi-robot groups is that
state-of-the-art control schemes break down as robotic
agents are scaled down (decreasing agent resources)
and the numerical size of swarms is scaled up (increas-
ing communication and coordination requirements)
(Murray, 2007; Hamann et al., 2016; Yang et al., 2018;
Chung et al., 2018). NeuroSwarms addresses the hy-
pothesis that a similar distributed scaling problem may
have been solved by the evolved neural architecture
of mammalian brains. Compared to signal compre-
hension, signal production errors may be particularly
deleterious to large-scale, distributed and decentral-
ized computations (Salahshour et al., 2019). Thus,
onboard suites for future cognitive swarming platforms
based on NeuroSwarms principles should emphasize re-
liable transmission of low-bandwidth data packets (e.g.,
spikes or continuous phase signals). Low fidelity inputs
are more easily compensated by distributed processing;
Cognitive swarms with attractors and oscillators
13
Fig. 7 Multi-agent ring formations and trajectory sequences in the hairpin maze. Frames from a single simulation are shown
for elapsed simulation times (s) from left-to-right, top-to-bottom: 0.89, 1.85, 3.73, 5.81, 9.63, 12.06, 41.43, 48.99, 297.65.
NeuroSwarms parameters: duration = 300.0, dmax = 1.5, σ = 2, κ = 6.6, gc = 0.1, gr = 0.1, gs = 0.8.
i.e., sensor designs should emphasize energy and cost
to maximize deployment duration and swarm size.
In summary, we made the explicit analogy from
agents and swarms to neurons and neural circuits. This
analogy permitted the tools of theoretical neuroscience
to be leveraged in developing a model artificial spa-
tial system. The NeuroSwarms controller required two
features to support cognitive swarming: (1) an inter-
nal phase state, and (2) decoupling of physical location
from internal self-localization. The phase state natu-
rally encapsulated neural activation (cf. equation (12))
and could be used to drive spike generation, if desired,
in future models. Phase-based organization additionally
leveraged the expressive complexity of mobile oscilla-
tors revealed by the swarmalator formalism (O'Keeffe
et al., 2017; Monaco et al., 2019b). The separation of
position vs. self-localization allowed swarm motion dy-
namics to be interpreted as Hebbian learning in an oscil-
latory place-coding neural network (Section 2.2). Thus,
theorized hippocampal phenomena such as attractor
map formation and oscillatory sequence generation pro-
vide a framework for advances in decentralized swarm
control and, reciprocally, the theoretical neuroscience of
spatial navigation in complex, changing environments.
Acknowledgements The authors
thank Marc Burlina
and Marisel Villafane-Delgado for preliminary analyses
and Robert Chalmers for helpful discussions about the
manuscript.
Conflict of interest
The authors have no conflicts of interest to declare.
References
Babichev A, Dabaghian Y (2017) Transient cell assem-
bly networks encode stable spatial memories. Scien-
tific Reports 7(1):3959
Bellmund JLS, Gardenfors P, Moser EI, Doeller CF
(2018) Navigating cognition: Spatial codes for human
thinking. Science 362(6415)
Burgess N (2014) The 2014 nobel prize in physiology or
medicine: A spatial model for cognitive neuroscience.
Neuron 84(6):1120 -- 1125
Buzs´aki G (1989) Two-stage model of memory trace
formation: a role for "noisy" brain states. Neuro-
science 31(3):551 -- 70
14
Monaco et al.
Buzs´aki G (2005) Theta rhythm of navigation: link
between path integration and landmark naviga-
tion, episodic and semantic memory. Hippocampus
7(15):827 -- 40
Buzs´aki G, Moser EI (2013) Memory, navigation and
theta rhythm in the hippocampal-entorhinal system.
Nat Neurosci 16(2):130
Casali G, Bush D, Jeffery K (2019) Altered neural
odometry in the vertical dimension. Proceedings of
the National Academy of Sciences 116(10):4631 -- 4636
Chung SJ, Paranjape AA, Dames P, Shen S, Kumar
V (2018) A survey on aerial swarm robotics. IEEE
Transactions on Robotics 34(4):837 -- 855
Climer JR, Newman EL, Hasselmo ME (2013) Phase
coding by grid cells in unconstrained environments:
two-dimensional phase precession. Eur J Neurosci
38(4):2526 -- 41
Drieu C, Todorova R, Zugaro M (2018) Nested
sequences of hippocampal assemblies during be-
havior support subsequent sleep replay. Science
362(6415):675 -- 679
Eichenbaum H (2018) Barlow versus Hebb: When is
it time to abandon the notion of feature detectors
and adopt the cell assembly as the unit of cognition?
Neuroscience Letters 680:88 -- 93
Feng T, Silva D, Foster DJ (2015) Dissociation between
the experience-dependent development of hippocam-
pal theta sequences and single-trial phase precession.
J Neurosci 35(12):4890 -- 902
Fenton AA, Kao HY, Neymotin SA, Olypher A, Vayn-
trub Y, Lytton WW, Ludvig N (2008) Unmasking
the CA1 ensemble place code by exposures to small
and large environments: more place cells and multi-
ple, irregularly arranged, and expanded place fields
in the larger space. J Neurosci 28(44):11250 -- 62
Foster DJ (2017) Replay comes of age. Annu Rev Neu-
rosci 40:581 -- 602
Foster DJ, Wilson MA (2007) Hippocampal theta se-
quences. Hippocampus 17(11):1093 -- 1099
Gaussier P, Banquet JP, Cuperlier N, Quoy M, Aubin
L, Jacob PY, Sargolini F, Save E, Krichmar JL,
Poucet B (2019) Merging information in the entorhi-
nal cortex: what can we learn from robotics experi-
ments and modeling? Journal of Experimental Biol-
ogy 222(Suppl 1)
Gazi V, Passino KM (2011) Swarm stability and opti-
mization. Springer Science & Business Media
Gupta AS, van der Meer MAA, Touretzky DS, Redish
AD (2010) Hippocampal replay is not a simple func-
tion of experience. Neuron 65(5):695 -- 705
Hamann H, Khaluf Y, Botev J, Divband Soorati M,
Ferrante E, Kosak O, Montanier JM, Mostaghim S,
Redpath R, Timmis J, et al. (2016) Hybrid societies:
challenges and perspectives in the design of collec-
tive behavior in self-organizing systems. Frontiers in
Robotics and AI 3:14
Hartley T, Lever C, Burgess N, O'Keefe J (2014) Space
in the brain: how the hippocampal formation sup-
ports spatial cognition. Philos Trans R Soc Lond B
Biol Sci 369(1635):20120510
Hasselmo ME (2018) A model of cortical cognitive func-
tion using hierarchical interactions of gating matrices
in internal agents coding relational representations.
arXiv e-prints p arXiv:1809.08203
Hebb DO (1949) The Organization of Behavior: A Neu-
ropsychological Theory. Wiley and Sons, New York
Hedrick KR, Zhang K (2016) Megamap: flexible rep-
resentation of a large space embedded with nonspa-
tial information by a hippocampal attractor network.
Journal of Neurophysiology 116(2):868 -- 891
Jayakumar RP, Madhav MS, Savelli F, Blair HT,
Cowan NJ, Knierim JJ (2019) Recalibration of
path integration in hippocampal place cells. Nature
566(7745):533 -- 537
Jeewajee A, Barry C, Douchamps V, Manson D, Lever
C, Burgess N (2014) Theta phase precession of grid
and place cell firing in open environments. Philos
Trans R Soc Lond B Biol Sci 369(1635):20120532
Jensen O, Lisman JE (2000) Position reconstruction
from an ensemble of hippocampal place cells: con-
tribution of theta phase coding. J Neurophysiol
83(5):2602 -- 9
Johnson A, Redish AD (2007) Neural ensembles in CA3
transiently encode paths forward of the animal at a
decision point. J Neurosci 27(45):12176 -- 89
Knierim JJ (2006) Neural representations of location
outside the hippocampus. Learn Mem 13(4):405 -- 415
Knierim JJ, Hamilton DA (2011) Framing spatial cog-
nition: Neural representations of proximal and dis-
tal frames of reference and their roles in navigation.
Physiological Reviews 91(4):1245 -- 1279
Knierim JJ, Zhang K (2012) Attractor dynamics of spa-
tially correlated neural activity in the limbic system.
Annu Rev Neurosci 35:267 -- 85
Kunz L, Maidenbaum S, Chen D, Wang L, Jacobs J,
Axmacher N (2019) Mesoscopic neural representa-
tions in spatial navigation. Trends in Cognitive Sci-
ences
Levy WB, Steward O (1979) Synapses as associa-
tive memory elements in the hippocampal formation.
Brain Res 175(2):233 -- 45
Momennejad I, Otto AR, Daw ND, Norman KA (2018)
Offline replay supports planning in human reinforce-
ment learning. eLife 7:e32548
Monaco JD, Abbott LF (2011) Modular realignment of
entorhinal grid cell activity as a basis for hippocam-
Cognitive swarms with attractors and oscillators
15
pal remapping. J Neurosci 31(25):9414 -- 25
Monaco JD, Knierim JJ, Zhang K (2011) Sensory feed-
back, error correction, and remapping in a multiple
oscillator model of place-cell activity. Front Comput
Neurosci 5:39
Monaco JD, Rao G, Roth ED, Knierim JJ (2014) Atten-
tive scanning behavior drives one-trial potentiation
of hippocampal place fields. Nat Neurosci 17(5):725 --
731
Monaco JD, De Guzman RM, Blair HT, Zhang
K (2019a) Spatial
from
coupled rate-phase neurons. PLoS Comput Biol
15(1):e1006741
synchronization codes
Monaco JD, Hwang GM, Schultz KM, Zhang K (2019b)
Cognitive swarming: an approach from the theoreti-
cal neuroscience of hippocampal function. In: Micro-
and Nanotechnology Sensors, Systems, and Applica-
tions XI, International Society for Optics and Pho-
tonics, vol 10982, p 109822D
Moser EI, Paulsen O (2001) New excitement in cogni-
tive space: between place cells and spatial memory.
Curr Opin Neurobiol 11(6):745 -- 751
Moser EI, Kropff E, Moser MB (2008) Place cells, grid
cells, and the brain's spatial representation system.
Annu Rev Neurosci 31(1):69 -- 89
Muessig L, Lasek M, Varsavsky I, Cacucci F, Wills TJ
(2019) Coordinated emergence of hippocampal re-
play and theta sequences during post-natal develop-
ment. Current Biology 29(5):834 -- 840.e4
Murray RM (2007) Recent research in cooperative con-
trol of multivehicle systems. Journal of Dynamic Sys-
tems, Measurement, and Control 129(5):571 -- 583
Ocko SA, Hardcastle K, Giocomo LM, Ganguli S (2018)
Emergent elasticity in the neural code for space. Proc
Natl Acad Sci U S A 115(50):E11798 -- E11806
Oja E (1982) Simplified neuron model as a principal
component analyzer. Journal of Mathematical Biol-
ogy 15(3):267 -- 273
O'Keefe J, Dostrovsky J (1971) The hippocampus as a
spatial map: preliminary evidence from unit activity
in the freely-moving rat. Brain Res 34(1):171 -- 175
O'Keefe J, Nadel L (1978) The Hippocampus as a Cog-
nitive Map. Clarendon Press, Oxford, UK
O'Keefe J, Recce ML (1993) Phase relationship between
hippocampal place units and the EEG theta rhythm.
Hippocampus 3(3):317 -- 30
O'Keeffe KP, Hong H, Strogatz SH (2017) Oscilla-
tors that sync and swarm. Nature communications
8(1):1504
´Olafsd´ottir HF, Bush D, Barry C (2018) The role of
hippocampal replay in memory and planning. Curr
Biol 28(1):R37 -- R50
Papale AE, Zielinski MC, Frank LM, Jadhav SP, Redish
AD (2016) Interplay between hippocampal sharp-
wave-ripple events and vicarious trial and error be-
haviors in decision making. Neuron 92(5):975 -- 982
Pfeiffer BE, Foster DJ (2013) Hippocampal place-cell
sequences depict future paths to remembered goals.
Nature 497(7447):74 -- 9
Poll DB, Nguyen K, Kilpatrick ZP (2016) Sensory feed-
back in a bump attractor model of path integration.
J Comput Neurosci 40(2):137 -- 55
Poulter S, Hartley T, Lever C (2018) The neurobi-
ology of mammalian navigation. Current Biology
28(17):R1023 -- R1042
Renn´o-Costa C, Tort ABL (2017) Place and grid cells in
a loop: implications for memory function and spatial
coding. J Neurosci 37(34):8062 -- 8076
Rich PD, Liaw HP, Lee AK (2014) Large environments
reveal the statistical structure governing hippocam-
pal representations. Science 345(6198):814 -- 817
Salahshour M, Rouhani S, Roudi Y (2019) Phase tran-
sitions and asymmetry between signal comprehension
and production in biological communication. Scien-
tific Reports 9(1):3428
Samsonovich A, McNaughton BL (1997) Path integra-
tion and cognitive mapping in a continuous attractor
neural network model. J Neurosci 17(15):5900 -- 5920
Savelli F, Knierim JJ (2010) Hebbian analysis of
the transformation of medial entorhinal grid-cell in-
puts to hippocampal place fields. J Neurophysiol
103(6):3167 -- 83
Savelli F, Knierim JJ (2019) Origin and role of path
integration in the cognitive representations of the
hippocampus: computational insights into open ques-
tions. J Exp Biol 222(Pt Suppl 1)
Savelli F, Yoganarasimha D, Knierim JJ (2008) Influ-
ence of boundary removal on the spatial representa-
tions of the medial entorhinal cortex. Hippocampus
18(12):1270 -- 1282
Schiller D, Eichenbaum H, Buffalo EA, Davachi L, Fos-
ter DJ, Leutgeb S, Ranganath C (2015) Memory and
space: Towards an understanding of the cognitive
map. J Neurosci 35(41):13904 -- 13911
Tomov M, Yagati S, Kumar A, Yang W, Gershman S
(2018) Discovery of hierarchical representations for
efficient planning. bioRxiv 499418
Tsodyks M (1999) Attractor neural network mod-
els of spatial maps in hippocampus. Hippocampus
9(4):481 -- 9
Vanderwolf CH (1969) Hippocampal electrical activ-
ity and voluntary movement in the rat. Electroen-
cephalogr Clin Neurophysiol 26(4):407 -- 18
Wang C, Chen X, Lee H, Deshmukh SS, Yoganarasimha
D, Savelli F, Knierim JJ (2018) Egocentric coding of
16
Monaco et al.
external items in the lateral entorhinal cortex. Sci-
ence 362(6417):945 -- 9
Wang Y, Roth Z, Pastalkova E (2016) Synchronized ex-
citability in a network enables generation of internal
neuronal sequences. Elife 5
Wikenheiser AM, Redish AD (2015) Hippocampal
theta sequences reflect current goals. Nat Neurosci
18:289 -- 294
Yadav CK, Doreswamy Y (2017) Scale invariance in lat-
eral head scans during spatial exploration. Physical
Review Letters 118(15):158104
Yang GZ, Bellingham J, Dupont PE, Fischer P, Floridi
L, Full R, Jacobstein N, Kumar V, McNutt M, Merri-
field R, et al. (2018) The grand challenges of Science
Robotics. Science Robotics 3(14):eaar7650
Yartsev MM, Ulanovsky N (2013) Representation of
three-dimensional space in the hippocampus of fly-
ing bats. Science 340(6130):367 -- 72
Zhang K (1996) Representation of spatial orientation
by the intrinsic dynamics of the head-direction cell
ensemble: a theory. J Neurosci 16(6):2112 -- 26
Zilli EA, Hasselmo ME (2010) Coupled noisy spiking
neurons as velocity-controlled oscillators in a model
of grid cell spatial firing. J Neurosci 30(41):13850 -- 60
|
1403.5734 | 2 | 1403 | 2015-12-07T19:06:31 | Software Agents Interaction Algorithms in Virtual Learning Environment | [
"cs.MA",
"cs.CY"
] | This paper highlights the multi-agent learning virtual environment and agents communication algorithms. The researcher proposed three algorithms required software agents interaction in virtual learning information system environment. The first proposed algorithm is agents interaction localization algorithm, the second one is the dynamic agents distribution algorithm (load distribution algorithm), and the third model is Agent communication algorithm based on using agents intermediaries. The main objectives of these algorithms are to reduce the response time for any agents changes in virtual learning environment (VLE) by increasing the information exchange intensity between software agents and reduce the overall network load, and to improve the communication between mobile agents in distributed information system to support effectiveness. Finally the paper describe the algorithms of information exchange between mobile agents in VLE based on the expansion of the address structure and the use of an agent, intermediary agents, matchmaking agents, brokers and their entrepreneurial functions | cs.MA | cs | World of Computer Science and Information Technology Journal (WCSIT)
ISSN: 2221-0741
Vol. 4, No. 2, 18-25, 2014
Software Agents Interaction Algorithms in Virtual
Learning Environment
Dr. Zahi A.M. Abu Sarhan
Faculty of Information Technology: Software Engineering Department
Applied Science Private University
Amman, Jordan
Abstract— This paper highlights the multi-agent learning virtual environment and agent's communication algorithms. The
researcher proposed three algorithms required software agent’s interaction in virtual learning information system environment. The
first proposed algorithm is agents interaction localization algorithm, the second one is the dynamic agents distribution algorithm
(load distribution algorithm), and the third model is Agent communication algorithm based on using agents intermediaries. The
main objectives of these algorithms are to reduce the response time for any agents’ changes in virtual learning environment (VLE)
by increasing the information exchange intensity between software agents and reduce the overall network load, and to improve the
communication between mobile agents in distributed information system to support effectiveness. Finally the paper describe the
algorithms of information exchange between mobile agents in VLE based on the expansion of the address structure and the use of
an agent, intermediary agents, matchmaking agents ,brokers and their entrepreneurial functions.
Keywords- multi-agent system; agent interaction models; Intermediary Agents; Virtual Learning Environment; Brokering Agents;
Matchmaking Agents.
I.
INTRODUCTION
At present, most of information system in various fields
based on software agents technology, these systems can
implemented as distributed information systems as centralized
information systems, and one of these complex systems is
distance or electronic learning systems. For the e-learning
systems wildly used centralized technology development and
information systems operation. The situation is still likes the
90's of the last century, when there were a large number of
scattered, technically and semantically mixed databases, used
locally. Different kinds of attempts to create a central system
for utilization all information system resources to achieve
valuable functionality and flexibility have yielded little success.
The wildly used today's Internet technologies and software
agents techniques with web access, solves the problem
partially. In this way, there is a distributed access to
information and the information resources and by the way it
keeps it to be centralized, which makes them not always
sufficiently relevant, high redundancy, technological, and
semantic diversity [1].
In this context, it is urgent the task to think about creating a
fully decentralized peer information systems to support e-
learning process, enabling flexible integration into virtual and
logical platform intelligently accomplish requested task and
usefully invoking all available information resources, thus
creating a unified virtual learning platform for fully efficient
interaction of all learning subjects. Effective technology for
implementation distributed information systems of this class is
the mobile software agents’ technology [1, 2]. These systems
must not only distribute access to information, but also
decentralized data storage and processing, and solving
problems related with information resources semantic diversity.
The developed multi-agent system implements a virtual
learning environment (VLE), in which real learning processes
related with development and
implementation of all
comprehensive forms of academic technological innovation in
learning and education and any comprehensive ideas that can
be implemented as relevant information processes[4]. The
subjects of learning can be presented as software agents that
interact with each other in a single VLE and perform the
interest of its owners, forming an open multi-agent system with
a decentralized architecture.
Software agent is computer system, which is found in some
environment and is capable of autonomous action in this
environment in order to meet its design objectives [4]. Software
agents have characteristics that make them suitable for complex
functions. Such features
interaction,
reactivity, activity, intelligence and mobility [5, 6].
include: autonomy,
18
WCSIT 4 (2), 18 -25, 2014
minimizing the data losses possibility and the transmitting
messages between agents delay time. The proposed solution in
this paper based on the expansion of the agent address structure
and the information about its current location in the network
and local caching this information on the system hosts, which
will allow sending messages directly from the sender agent to
the recipient agent, as well as provide the ability to search
agents with the joint action through intermediary agents, using
their brokering services [11] and matchmaking services [12]
functions. Algorithm searching agents’ initiators executed in
the agents intermediaries address space.
III. MULTI-AGENT VIRTUAL LEARNING ENVIRONMENT
The modern virtual learning environment (VLE) is a
software system designed to support teaching and learning.
VLEs generally function on the World Wide Web, and,
therefore, they can be utilized both on and off-campus,
provided that the users are authorized and can access the
Internet [12]. This surmount over the restriction of traditional
face-to-face learning, and guarantees that learning is neither
limited to location nor time. VLEs can contains complex of
interactions between teachers, learners and learning contents,
by adding agents and environment to this structure, the VLE
can be presented as a complex of agents interaction and every
agent represents
learning structure
components, such as learner agent, teacher agent, content
agent, and every agent
to
accomplish requested task. However the set of interacted
agents can contain “learner –learner interaction agent, learner-
teacher interaction agent, learner-content interaction agent,
teacher-teacher interaction agent, teacher-content interaction
agent, content-content interaction agent, learner-environment
interaction agent,
interaction agent,
content-environment interaction agent”, VLEs become more
popular and included in many college organizations all over the
world. It is not only because of their versatility, but also
because they provide an extensive range of tools or features,
such as content distribution, evaluation, emails[14, 15].
interacts with other agents
teacher-environment
interest of
the
the
Based on
the
representation of VLE can be presented as a set of agents that
interact together, shown in figure 1.
learning process
components
the
in
This paper presents the software agents interaction models
and the algorithms for learning activities in the virtual learning
information systems environment, enabling faster response to
changes
the agents virtual environment simulating
innovative field by increasing the intensity of information
exchange between the agents and to reduce the overall load on
the network. Indicating ways to enhance the interaction among
mobile agents in an apportioned multi-agent system of
information to back up learners and teachers action.
II. PROBLEM STATEMENT
Multi-agent systems (MAS) technology - a new pattern of
information technology, focused on the sharing of scientific
and technological achievements and benefits which provide
ideas and methods of artificial intelligence (AI), the current
local and global computer networks, distributed databases and
distributed computing, hardware and software tools to support
the theory of distribution and transparency.
Relevance of distributed AI and MAS, in accordance with
the works [7, 9], determined by the complexity of diversity
modern organizational and technical systems, complexity, and
tasks distributions, large volumes of Information flow and
information processing
is
widely used in various fields that required solutions for
complex distributed tasks such as combined product design, re-
engineering information processes and the construction of
virtual systems, and e-commerce systems, distributed computer
programs development.
time. Agent-oriented approach
The greatest complexity in the theoretical studies and
practical modern MAS implementations are issues related to
the agents’ interaction processes in the collective solving
problems with high practical complexity and relevance, as each
agent solves specific subtask, has only a partial idea about the
general task and should continuously interact with other
agents. Therefore, at present the most relevant in the MAS
theory and technology is the issues related with creating
intelligent agents interaction models in the MAS.
In inter-host software agents interaction in a virtual learning
environment, arise a set of issues related with increasing the
network load, and reducing the information exchange intensity
between agents which depends on the speed and bandwidth of
the network connections. Increasing the request delivery time
and the needs time to find the correct agent for interaction. As a
suitable solution for these problems, proposed a method based
on the decomposition a common information space in which
agents operate in virtual platform (platform represents some
separate network host) and move intensively interacting agents
to these platforms in order to combining agents in coalition.
The proposed solution can be implemented in two mutually
complementary algorithms software agents’ interaction: the
algorithm
interaction
(Transformation the inter-host agent’s interaction to intra-host
interaction) and software agent’s dynamic allocation algorithm
(balancing load between system hosts) [10].
inter-host
software
of
agent’s
However, the architecture design of decentralized open
information systems, arise problem related with determining
the location of dynamically distributed mobile agents, that are
moving between network hosts over
time, as well as
Figure 1. Virtual learning environment based on multi-agents architecture.
19
Learner-Content Interaction AgentTecher-Content Interaction AgentLearner-Learner Interaction agentLearner-Teacher Interaction AgentTeacher-Teacher Interaction agentContent-Content Interaction agent WCSIT 4 (2), 18 -25, 2014
IV. AGENTS COMMUNICATION MODEL
the
direct
following
interaction
categories:
and moderating
Agents Communication Model in an open multi-agent
information system that supports e-learning process, according
to the work [14], can be classified by the interaction character
(point-to-point
in
communication),
(throw
intermediaries or server agents) interaction. In point to point
agents’ model communication message delivered directly from
agent sender to the agent receiver, even if the agent receiver is
a mobile agent (a message can be sent through a set of
computer network host’s). In the intermediary agent model
communication mainly uses agents brokering services or
matchmaking services agent’s intermediaries, which allow
agents to communicate with each other, this model provides
multipoint and anonymous agent’s interaction. These two types
of services have their advantages and disadvantages: brokering
services are more efficient, while matchmaking services are
more flexible [10].
Depending on the agent-receiver location in the network
agent’s communication model, according to [6, 9], may be also
classified as follows: inter-host communication and intra-host
communication. When the agent-sender and agent-receiver
located on the same network host, the messages transmission
between these both agents called intra-host communication.
When they were located on two different network hosts, the
messages transmission between agents delivered via the
network nods, and this type of communication called the inter-
host communication. Even if the Internet data rate has greatly
increased, inter-host agent’s communication still would take
greater time than interaction with intra-host communication.
Therefore,
inter-host
communication to intra-host communication, which would
reduce the time required for messages transmission and
increase the interaction intensity in the agent’s communication
process.
to reform agent’s
is advisable
it
The main disadvantages of presented agent’s interaction
models can be listed as follows:
– At inter-host interaction: the network over load increases,
and the information exchange intensity between agents reduced
due to increasing message delivery times, agents reaction for
dynamically changing events of environment or the absence of
any response to change the environment in message losing
case;
– At the agents’ migration between system hosts there was
a problem of the needed responder agent location determining
for interaction during any time moment, which in its turn
results to increasing in messages loss probability during their
transfer and reducing the effectiveness of synchronized agent’s
interaction.
V. AGENT-BASED ALGORITHMS VIRTUAL LEARNING
PLATFORMS FORMATION
The higher effect obtained from using the virtual learning
environment (VLE), based on the biggest of its internal volume
(the number of registered requests), Agents representing the
Learning subjects interests, and placement of information
databases hosts [17]. However, the naturally growth of the
20
system volume leads to increase the information elements
search task complexity, and the selection of learning structures
options due polynomial growth of alternatives quantity. In
order that the system no longer runs under its own unrestricted
growth, needed some self-organization algorithms, allowing
dynamically reorganize its internal structure to reduce the
amount of processed and transmitted over the communication
lines data during the learning offerings placement and retrieval
on the media hubs, and forming a potentially effective learning
structures.
Self-organization is the automatic generation within-VLE
virtual learning platforms (VLP), combining agents with
similar interests in groups. Formation the VLP based on a the
register distribution support method for peer-to-peer hosts with
the implicit treelike organization [9] in which as an organizing
structure tree uses the hierarchical learning domain model.
Learning platforms formation carried out by displaying the
agent’s purposes on the treelike conceptual domain model,
Subsequent localization the main part of search and other
agents requests inside the group and further analysis the
communications activity with each other. The
interests
similarity leads to the fact that the most active and informative
agents communication focused
learning platform,
whereas outside platform the information exchange is less
active, thus an exchange object generalized (smaller by
volume) agents learning offers or requests [18].
inside
the
in
to
transfer
reducing
inter-host
transform
This approach allows not only
total
communication amount between agents, but also due to the use
of mobile agents
intra-host
communication. This, in turn, reduces the network load in case
of the distributed system implementation. The main agent
based algorithms that enhance e-leaning systems efficiency are
agent interactions localization algorithm within hosts by
forming groups actively communicates agents (virtual learning
platforms) and dynamic
load redistribution by
implementing agents group migration algorithm. These
algorithms provide ultimately, the conversion agents inter host
interactions in intra-host. In order to convert the inter-host
interactions in intra host interactions, the system agents can be
distributed the network hosts depending on their interaction
nature and behavior. Since the agents interaction nature can be
continuously changed, the agents distribution on the network
hosts should be dynamic. In this case, the agents automatically
distributed on multiple network hosts, depending on their
interaction nature, some hosts may be overwhelmed by the
several migrating agents to them. Therefore, the proposed
agents dynamic distribution algorithm, that based on their
interaction behavior should be complemented by another
agents distribution algorithm which support uniform load
distribution between network hosts.
In the proposed multi-agent system [7], each host equipped
with an agent-based platform (agent representative), which is a
agents local execution environment, in which software agents
operate and interact with each other. Implemented agents
interaction algorithms involve a sequence of similar operating
phases and use common components that are present on each
agent platform.
WCSIT 4 (2), 18 -25, 2014
To arrange a dynamic agents distribution on each agent
system platform the following components should operate. The
Message Manager (MM), coordinating the messages exchange
between the system agents; and the System Monitor (SM),
periodically checking the load on the current network host; and
the Agent Distribution Manager (ADM), is responsible for the
agents dynamic distribution and analysis the models of agents
interaction behavior and character; and finally the Agent
Migration Manager (AMM), displacing agents on other
platforms (hosts) and managing agents migration between
network hosts.
A. Agents interaction Localization algorithm
The agent’s interaction localization algorithm will contain
four phases as shown in figure 2.
Analysis phase. The Agent’s Distribution Manager
(ADM) evaluates inter-host and intra-host system
agents’ communications intensity, involving in this
stage system monitor and messages manager. ADM
also uses all information as about agent-sender, as
about agent-receiver host. ADM periodically evaluates
communication
dependency
at
time
between the agent
and agents host
as follows:
(1)
Where
- the number of messages sent by the agent
to host
agents over the time period
- is the
coefficient, which characterizes the relative dependency of new
information in relation to the outdated (expired) and used to
ignore the temporary intensive interaction with the agents in
,
particular agent platform;
- is defined as the value
of the same communication dependency in the previous time
interval.
stage
Agents Distribution Phase. After a specified number of
analysis
the
communication dependency coefficient between the
current host agent n and all other hosts in the system.
between
Communication dependency coefficient
repetitions ADM calculates
agent and host
agents can be given by:
(2)
When
the maximum value of
the communication
dependency coefficient over a predefined edge
, ADM
current host includes under consideration the agent in a group
of agents, located on a remote system host:
(3)
Where
is an agent
,
indicates the agents group
,
and under
hereinafter understood operation that returns
the value of
, where the ratio of communication dependence
coefficient
takes the maximum value.
Interaction Phase. Before moving the selected agents
group from the host
to the receiver host
ADM
host
interacts with ADM host
. ADM
checks
the current memory host state, CPU utilization, and the
21
Figure 2. Agent’s interaction localization algorithm
ADM Host P1StartEvaluate inter-host and intra-host communications intensityCollect Information about Agent Sender And Agent Receiver HostEvaluate communication dependency Done?NoYesADM Host P1System MonitorMessage Manager Add Agent I to group G YesNoADM P1 Requests available recourses on Host P2 Calculate the communication dependency coefficient ADM Host P2Enough recourses on Host P2?ADM P2 permits migrationYesNoMigrate Agents group to host P2 EndADM Host P1Analysis PhaseDistribution phaseInteraction phaseMigration phaseijDt12,tTttij()()()ijijijikkR(t)Dt=+1-Dt-1R(t)ddaeö÷ç÷ç÷ç÷ç÷ç÷ç÷÷çèøåijDtij12,ttTijDt-1ijMij,ijijinDMjnDargmaxjijikikBMMGiikGkargjjijM1P2P1P2P2P WCSIT 4 (2), 18 -25, 2014
number of agents hosted in this host, by using system
monitor. If the host
has enough free system
between the agents and the locally formed agents groups on
different network hosts [10].
resources for new agents, ADM host
authorizes
Communication dependence of
between agent
the agents group migration from host
to host
.
Agents’ Migration Phase. Accomplishing the agents’
selection operation for migration and receiving positive
response from ADM receiver host, the sender host
ADM jointly with Agents Migration Manager (AMM)
initiates the selected agents’ group migration to the
receiver host.
Agents’ interaction localization algorithm assumes the
analysis of dynamic changes in interaction patterns (models)
between agents, but this algorithm can overloaded some system
hosts due to the large number of agents which can be moved to
the hosts.
When a host is overloaded, the system monitor detects this
state and activates the agents’ redistribution process, which is
not only based on individual agent movement between network
hosts, but the whole interacting agents group intensively
interacts with each other.
B. The agents dynamic distribution algorithm
The
algorithm
agents dynamic distribution
(load
distribution algorithm) among system hosts consists five
phases: Analysis phase agents grouping phase, agent groups
distribution phase, agents interaction phase and the agents
migration phase.
Analysis Phase. Each system host monitor periodically
checks the agent platform state in which agents
interacting on this host. And by calling special system
functions collects information about the host physical
resources such as (current CPU load and free memory
space). When the system monitor defines that the host
is overloaded, it will activate the agents’ distribution
procedure between system hosts. As a host overload
criterion can be offered the maximum number of
agents operated on the host.
When ADM gets notification about congestion from the
system monitor, it starts the local agents interaction monitor
procedure, preliminary in order to locally splitting interacted
and cooperated agents into groups.
The group formed from agents that have most similar
interests and goals, which are intensively, interact with each
other by exchanging messages. Information about agents and
agents groups registered in the special register. The information
analysis provided on the register, allows to evaluate the load on
the system hosts and determine the intensity between agents
and intergroup and inter-host communication levels, which, in
turn, allows to select loaded and unloaded hosts in the system,
and implement the dynamic agents redistribution and agents
groups among the system hosts, in other words the agents or
agents groups movement with certain characteristics from
overloaded hosts to less loaded hosts in the system, containing
similar potential agents interests. After a predetermining the
time interval ADM updates communication dependences
22
and agents group
formula:
at the moment
can be determined by the
(4)
Where
- is the number of messages sent by the
agent
to the agents group
- over the time period
,
- is the coefficient, which characterizes the relative
importance of new information in relation to the outdated
(expired) and used. In this case the expression
shows the number of messages sent by agent
jointly functioning on the common host.
to any agent
Agents Grouping Phase. After accomplishing a number
agent is
of monitoring phases repetitions each
overridden in local agents group with new index
and given by:
(5)
Where
indicates
- local group agents.
Monitoring and agents grouping phases are repeated several
times. After each agents grouping phase information about
local communication dependencies between agents zeroed.
Agent Groups Distribution Phase. After performing a
certain number repetitions of monitoring and agents
grouping phases, ADM depending on agent platform
status takes the decision to move the agents group to
other system host. The selection of moved agents based
on the communication dependence between agent
groups and the system hosts.
Communication Dependence
between
- agents group
and the
system host based on the summation of the
communication dependence between all members of agents
group and the system host:
(6)
Where
- all agents indexes set, belonging to
- agents
group, and
- communication dependence between
agent
and
system host in the current time
. The agents
group
host and can be selected by the following rule:
, which have a smallest dependence on the current
2P2P1P2PijDtijt11ijijijiKKRtDtDtRtijDtij12,ttTiKKRtii*j**(cid:160)jijijjargmaxMtA*jA*jijWijiijkjkAWtCtiAikjCtkjt*i WCSIT 4 (2), 18 -25, 2014
(7)
Where
is the current host (Agent platform) number [10].
The receiving agent platform
, selected by
agents
group determined using communication dependence between
agents group and
system host as follows:
(8)
Where
is the current host (Agent platform) number.
interaction with
Agents Interaction Phase. When in the network defined
the receiving host for agent groups, the sender host
ADM begins
the corresponding
receiving host ADM. If the receiving ADM allows
agents movement, sending host ADM starts the agents
migrating phase. Otherwise, the sending host ADM
polls other system hosts ADM as long as it finds a
suitable host for a scheduled agents migration
(movement). If no host is unable to take the agents
group, the interaction phase fails and after a certain
time period
(allocation)
algorithm on the sending host restarted, and the process
repeats again.
the agents distribution
Agents interaction phase of this algorithm is similar to the
agents interaction localization algorithm. Yet, the interaction
level between these two algorithms is very different: agents’
interaction localization algorithm occurs at the agent level,
whereas the algorithm load distribution between the system
hosts implemented at the agent groups level. If the agent
receiver-platform
system
resources for all agents belong to the selected local agents
group, ADM agent receiver-platform representation may
authorize the agents’ group movement. Otherwise, the system
receiver host rejects the request to move the selected agents
group; the host is unable to accept only part of agents group.
representation has
sufficient
Agents Migration Phase. When ADM sender host
receives a response with acceptance to move the agents
group from ADM receiver host, ADM sender host
initiates migration the selected local agents group to
the receiver host.
Each agent-based representation independently performs its
own agents distribution algorithm in accordance with the
available information about the load on the system host, where
it is located, and the agents interaction character functioning in
agents platform [10]. The agents dynamic distribution
algorithm shown in figure 3.
Figure 3. Agent’s dynamic distribution algorithm
C. Agent communication model based on using agents
intermediary
In open multi-agent systems with extremely dynamic
character (in any time may come in or/and go out new agents,
23
*,ijijjninWiargmaxWn*jii*j*(cid:160),(cid:160)jijjargmaxWjnnADM sender HostStartEvaluate agents platform stateCollect Information about Host physical resourcesIs the host overloaded?NoYesSystem MonitorADM sender Requests migration validation from ADM receiver Host Valid response?Request migration validation from another ADM Host YesNoMigrate Agents group to selected host EndADM receiver HostActivate Agents distribution procedureLocally split interacted and cooperated agents into groupsUpdate communication dependency Done?NoYesADMADMCreate local Agents group Get communication dependency equals to zeroCalculate the communication dependenceSelect Agent receiving hostSelect Agent Group For movmentAnalysis PhaseDistribution phaseInteraction phaseMigration phaseGrouping phase WCSIT 4 (2), 18 -25, 2014
and be connect new hosts), each individual agent cannot have
full information about all other agents in the same information
and communication environment. In this environment, a joint
operation of network services agents intermediaries (brokerage
and matchmaking services) are highly effective in finding
potential agents for interaction [19, 22]. Software agents can
find the other agents names using matchmaking services or
send messages to other agents using brokering services, using
their attributes, such as methods, operation modes, features or
nicknames instead of their real names. To register the agents
names in distributed virtual learning environment can be used
the dedicated server for agents’ names. Agent Intermediary
functions
its
architecture, or they can be implemented as an independent
specialized application called agents intermediaries [24].
implemented as a separate component
in
algorithms provided by agents’ initiators. This is the main
difference between the proposed method of using intermediary
agents and their features compared to the existing ones.
VI. CONCLUSION
This paper discusses separated agent-based algorithms for
Virtual learning environment information system effective
functioning and distributed hosts
interaction algorithms.
Described various
implementation approaches of agents
communication algorithms. Analyzed the existent software
agents communication models. Identified weaknesses of these
models, based on that proposed approaches to improve the
mobile agents cooperation in distributed Virtual learning
environments.
Interaction mediating between the agents includes brokers’
agents or matchmakers’ agents. Brokers forward the messages
given by the agent sender to the network hosts where the agent
receiver located, whereas matchmakers’ only provides the
agents intending to send messages, all information about the
agent receiver location, in other words the final message was
delivered to the agent receiver brokers, while the matchmakers
only help sender agents to deliver messages to the agent
receiver, providing
locations.
Considering the number of steps within transmitting massages,
it can be concluded that the brokerage services are more
effective than matchmakers, since brokerage services typically
require two steps to transmit massage, while matchmakers
requires three steps.
information about
their
Using shared memory area, that called data fragments
space, which controlled by intermediary agents, agents can
register their attributes together with the names in this area, and
they can communicate each other by using the attributes
information about other agents, this information can be
extracted from this area. The basis of most developed
intermediary agents is Linda-model [9, 10]. The proposed
model presented in figure 4.
Figure 4. Agent communication model based on using agents intermediary
The model allows software agents to use intermediary
agents, and the provided services by them to search agents with
the similar interests and provides the basis method of using
intermediary agents and their functions. The intermediaries
agents don’t only control the data fragments, but can be also
used as own agents search algorithms “joint activity” and
24
Based on the modification of existing agents interaction
models, developed algorithms for data exchange among mobile
software agents
the agents efficiency for
distributed data processing and reduce the total transmitted data
amount over the network:
increase
that
Agent communication algorithm based on the using
agents brokers and their functions;
in virtual
information
system activities
However, developed software agents interaction algorithms
learning
for
environments,
that allowing faster agents response for
environment changes in which an agent functioning, by
increasing information exchange intensity between agents and
reduce the overall load on the network, those algorithms are
Agents interaction localization algorithm transformation inter-
host communication to interact as intra-host communication;
and agents dynamic distribution algorithm (balancing load
between the system hosts).
Based on developed software agents interaction models and
algorithms, proposed the minimizing inter-host communication
in problem-oriented distributed systems method. The method is
based on software agents classification in semantic space,
presented as conceptual domain model, and agents inter-host
communication transformation in intra-host. Implementation
this method provides a
in communication
infrastructure load and improving availability coefficient for
software agents application services.
reduction
ACKNOWLEDGEMENT
The author is grateful to the Applied Science Private
University, Amman, Jordan, for the financial support granted to
cover the publication fee of this research article
REFERENCES
[1] А. Маслобоев, Мультиагентная
технология формирования
едином информационно-
виртуальных бизнес-площадок
коммуникационном пространстве развития инноваций, Publisher,
City, 2009.
в
[2] B.M. Balachandran, M. Enkhsaikhan, Developing multi-agent e-commerce
applications with JADE, in: Knowledge-Based Intelligent Information
and Engineering Systems, Springer, 2007, pp. 941-949.
[3] D. Tavangarian, M.E. Leypold, K. Nölting, M. Röser, D. Voigt, Is e-
learning the Solution for Individual Learning, Publisher, City, 2004.
Massages QueueMsg 1Msg 2Msg nData Fragments AreaTasks QueueTask 1Task 2Task n……MatchmakersData Fragment ManagerBrokersOR WCSIT 4 (2), 18 -25, 2014
[4] T. Erl, A. Karmarkar, P. Walmsley, H. Haas, L.U. Yalcinalp, K. Liu, D.
Orchard, A. Tost, J. Pasley, Web service contract design and versioning
for SOA, Prentice Hall, 2009.
[5 F.P. Brooks Jr, The design of design: Essays from a computer scientist,
Pearson Education, 2010.
[6] E. Platon, N. Sabouret, S. Honiden, Tag Interactions in MultiAgent
Systems: Environment Support, Publisher, City, 2005.
[7] S. Alouf, F. Huet, P. Nain, Forwarders vs. centralized server: an evaluation
of two approaches for locating mobile agents, Publisher, City, 2002.
[8] N. Carriero, D. Gelernter, Linda in context, Publisher, City, 1989.
[9] B. Chen, H.H. Cheng, J. Palen, Integrating mobile agent technology with
multi-agent systems for distributed traffic detection and management
systems, Publisher, City, 2009.
[10] А.В. Маслобоев, Модели и алгоритмы взаимодействия программных
агентов в виртуальной бизнес-среде развития инноваций, Publisher,
City, 2009.
[11] K. Malik, Use of knowledge brokering services in the innovation process,
in: Management of Innovation and Technology (ICMIT), 2012 IEEE
International Conference on, IEEE, 2012, pp. 273-278.
[12] T.A. Farrag, A.I. Saleh, H.A. Ali, Semantic web services matchmaking:
Semantic distance-based approach, Publisher, City, 2013.
[13] Z. Xu, Z. Yin, A. El Saddik, A Web Services Oriented Framework for
Dynamic E-Learning Systems, Publisher, City.
[14] A. Targamadze, R. Petrauskiene, Classification of distance learning
agents, Publisher, City, 2010.
[15] E.M. Van Raaij, J.J. Schepers, The acceptance and use of a virtual
learning environment in China, Publisher, City, 2008.
[16] A. Di Stefano, C. Santoro, Locating mobile agents in a wide distributed
environment, Publisher, City, 2002.
[17] W.F. McComas, Virtual Learning Environment, in: The Language of
Science Education, Springer, 2014, pp. 110-110.
[18] E. Sangineto, An Adaptive E-Learning Platform for Personalized Course
Generation, Publisher, City, 2008.
[19] S.L. Greenspan, E. Hadar, Using cloud brokering services for an
opportunistic cloud offering, in, Google Patents, 2013.
[20] S. Muthaiyah, L. Kerschberg, Brokering Web Services via a Hybrid
Ontology Mediation Approach Using Multi Agent Systems (MAS),
Publisher, City, 2010.
[21] P. Khanna, B. Babu, Cloud Computing Brokering Service: A Trust
Framework, in: CLOUD COMPUTING 2012, The Third International
Conference on Cloud Computing, GRIDs, and Virtualization, 2012, pp.
206-212.
[22] J.-b. LI, C.-h. LI, Research on the Development of the Cooperation
among Governments, Enterprises, Universities, Research Institutes,
Financial Units and Intermediary Agents [J], Publisher, City, 2010.
AUTHORS PROFILE
Dr. Zahi A. M. Abu Sarhan Received the
M.S. and PhD degrees in Computerized
Control Automated Systems
and
Progressive Information Technologies
from Kharkov National University of
Radio Electronics, Kharkov in 1998
and 2004, respectively. During 2004-
2008, I was an Assistant Professor at the
Economics and Administrative science/
MIS Department at Applied Science University. Since 2008, I am an Assistant
Professor at the Faculty of Information technology, Applied Science
University in Jordan. Research interests include: Information system
reengineering, Service oriented architecture, software agents, agents
theory, agents behavior.
25
|
0707.1558 | 1 | 0707 | 2007-07-11T05:05:57 | Autonomy with regard to an Attribute | [
"cs.MA"
] | This paper presents a model of autonomy called autonomy with regard to an attribute applicable to cognitive and not cognitive artificial agents. Three criteria (global / partial, social / nonsocial, absolute / relative) are defined and used to describe the main characteristics of this type of autonomy. A software agent autonomous with regard to the mobility illustrates a possible implementation of this model. | cs.MA | cs | Autonomy with regard to an Attribute
Eric Sanchis
Laboratoire de Gestion et Cognition
Université Paul Sabatier
115 route de Narbonne
31077 Toulouse, France
[email protected]
Abstract
This paper presents a model of autonomy called
autonomy with regard to an attribute applicable to
cognitive and not cognitive artificial agents. Three
criteria (global / partial, social / nonsocial, absolute /
relative) are defined and used to describe the main
characteristics of this type of autonomy. A software
agent autonomous with regard
to
the mobility
illustrates a possible implementation of this model.
1. Introduction
Since the second half of the Nineties, autonomous
agents are used more and more in the design and in the
implementation of complex computer systems. The
association of the concepts of agent and autonomy was
explored and commented on in many theoretical
works, primarily in the MultiAgent Systems (MAS)
area. A first point of view which was developed by
researchers such as Jennings and Wooldridge, or
Franklin and Graesser was to consider autonomy as a
global property, i.e. a property which applies to the
agent in its totality.
For Jennings and Wooldridge [7], “an agent is a
computer system situated in some environment, and
that
is capable of autonomous action
in
this
environment in order to meet its design objectives”.
The situated aspect means that the agent perceives its
environment and that it is able to modify it. Autonomy
means that the agent is able to act without the direct
intervention of human beings or other agents, and that
it is able to control its own actions and its internal
state. For the authors, autonomous agents are not an
innovation because systems like software daemons (ex:
xbiff) or a simple thermostat can be classified as
autonomous agents.
Franklin and Graesser [6] synthesize their point of
view on the agents with the following definition: “an
autonomous agent is a system situated within and a
part of an environment that senses that environment
and acts on it, over time, in pursuit of its own agenda
and so as to effect what it senses in the future”. For the
authors, the software agents are present in many fields
but any software program is not an agent: for example,
according to their definition a payroll program cannot
be considered as an agent. On the contrary, a
thermostat fully satisfies their definition of an agent,
i.e. an entity provided with autonomy.
The main purpose of the autonomy model that we
propose is to remove a major defect present in the two
models outlined previously: the global aspect of
autonomy. Indeed, a problem of global autonomy is
that it is carrying paradox and confusion [8]. Let us
illustrate this problem with an example. The definition
of an autonomous agent as an agent able to act without
the direct intervention of other agents and having a
control on its own actions and its internal state is not
entirely satisfactory. One can notice that even the
human agents do not have this control neither on their
internal state nor on all the actions they carry out.
However, a (partial) autonomy of a human agent
cannot be seriously questioned. This apparent paradox
seems to result from an insufficient decomposition of
the agent.
Moreover, as many researchers had already noticed,
autonomy is not an ordinary property but a complex
property. More precisely, our work
led us
to
distinguish two categories of properties of deeply
different nature: qualities and attributes [10]. Qualities
characterize properties with vague and elastic
contours, difficult (perhaps impossible) to have an
intimate knowledge. They are properties which content
changes according to the point of view which one can
have. They authorize multiple modelizations and
different interpretations, when these interpretations are
not contradictory. Some properties which can be
identified as qualities are autonomy and intelligence.
Attributes are properties intrinsically less difficult to
encircle than qualities. Examples of attributes are
mobility, replication or perception. They are generally
reduced to a mechanism with one or more well defined
procedures. Contrary to qualities, it is always possible
to say if an agent has or does not have a specific
attribute. Attributes play a significant role in the type
of autonomy which we propose.
Lastly,
in
the MAS field several models of
autonomy were designed for goal-directed cognitive
agents. For such an agent, a goal corresponds to a
cognitive representation which an agent has of the
world in which it is immersed. It results that these
models of autonomy are not appropriate to non
cognitive agents. The model of artificial autonomy
described in section 4 is sufficiently general to apply to
cognitive and non cognitive agents.
The paper is structured as follows. Section 2
presents the main characteristics of the autonomy
models generally associated to artificial agents. In the
next section, the significant features of these models
are used to define a set of criteria which will be used to
compare these models and to introduce the autonomy
with regard to an attribute. Lastly, section 4 will
describe the theoretical and practical aspects of this
model and some possible extensions.
2. Artificial Autonomy approaches
Autonomy is a typical example of what we have
called a quality. Indeed, autonomy characteristics seem
difficult to be synthesized with only one theory or
model. In fact, reducing autonomy to a single model
removes the essence of this property. This is why
many models of autonomy were proposed by
researchers in the Artificial Intelligence area. Various
aspects can be brought out from these models: the
organic autonomy of biological inspiration (Varela,
Bourgine), the social autonomy based on the power
relationships between agents (Castelfranchi, Scerri)
and the decisional autonomy founded on the choice of
the agent (Barber, Vendryès), only to quote the most
representative types of autonomy.
The autonomy models briefly outlined below
partially illustrate the significant disparity of the points
of view generated by this property.
The works of Luck and D’Inverno [9] on autonomy
is integrated into a more general theory on the
interactions between agents being based on
the
concepts of goal and motivation. The authors
distinguish two very different points of view about
autonomy: autonomy as the possibility for the agent to
generate its own goals and autonomy as relations of
dependence of an agent upon others. According to the
first point of view, autonomous agents generate their
own goals from their own motivations. According to
the second point of view, the autonomy of an agent is
modulated by the existence of relations of dependence
between agents. This distinction separates what we
will call thereafter social autonomy and nonsocial
autonomy.
Barber and al. [1], [2], consider autonomy as the
possibility for an agent to influence the decision-
making process in the resolution of a given problem.
Four levels of autonomy are defined. First level: the
agent carries out the orders which are provided to it by
another agent. Second level: the agent collaborates
with other agents in order to achieve a common goal.
Third level: the agent plans and carries out its own
actions. Lastly, on the fourth level, the agent plans the
actions for itself and for other agents. The agent is
completely autonomous when it completely manages
the decision-making process. It is partially autonomous
when it shares this decision-making with other agents.
Finally an agent which does not take part in this
process is heteronomous.
Castelfranchi and Falcone [3], [4], connect the
concept of autonomy to the concept of dependence.
They explain why on the one hand, autonomy is a
relational concept - relation between an active entity
and its environment or other active entities (social
autonomy) -, and why on the other hand this relation
also derives from the properties of the internal
architecture of the entity.
Adjustable autonomy describes the property of an
autonomous system to change its level of autonomy
among several levels while the system continues to
function. A level of autonomy indicates a particular
distribution between manual operations and automatic
operations
[5]. The distribution of autonomy
dynamically changes in order to optimize the global
performances of the system [12].
3. Criteria for a partial classification
In order to compare in a uniform manner the
autonomy with regard to an attribute and the preceding
models, three criteria of analysis were defined:
- Autonomy as a global or partial property of the
agent
- Autonomy as a social or nonsocial property
- Autonomy as an absolute or relative property.
Global or partial autonomy. Autonomy is said
global when the definition applies to the entire agent.
In the field of MAS, researchers generally use the
expression “autonomous agent” [6] [7]. Another
definition which typically illustrates this global aspect
of autonomy is: autonomy is the condition of a person
or a community who determines by itself the law to
which it is subjected. In this definition, this quality
also applies to the entity considered (agent, individual
and group) in its totality. On the other hand, the nature
of the law determined by the entity is not specified. We
interpret this general information in the following way:
as soon as an agent is able to create a law, it is globally
autonomous. By replacing the concept of law by the
concept of goal, we can qualify the model presented in
[9] as global autonomy.
Autonomy is partial when the property does not
apply any more to the entire agent but to a part of it.
The agent is in this case autonomous with regard to
something. The model presented in [2] can be
described as an example of partial autonomy because
the autonomy is defined for each agent goal and the
agent can be non autonomous (heteronomous) with
regard to a specific goal.
Social or nonsocial autonomy. Autonomy is
social when
described
as
its definition or
characterization explicitly refers to one (or several)
other(s) agent(s). It is nonsocial in the contrary case.
The (X, Y, B) triplet defined in [4] perfectly
characterizes social autonomy: the autonomy of agent
X from agent Y about goal B. The presence of agent Y
is essential to the various variations of this model
(collaborative autonomy, non collaborative autonomy,
etc). Models which use a decision-making process
shared by several agents, as defined in adjustable
autonomy [12] or in [2] are also and naturally models
of social autonomy. The agents considered in the
models of social autonomy are not of unspecified
nature but are cognitive agents, i.e. provided with a
certain type of intelligence.
The autonomy presented in [9] is nonsocial because
it is based on the possibility of an agent to generate a
goal without implying the presence of another agent.
However, it should be noted that a model of autonomy
described as nonsocial does not mean that this model is
inapplicable when several agents are in relation. This
qualifier simply means that the definition of the model
does not impose the presence of another agent.
Absolute or relative autonomy. Autonomy is
absolute when only one level of autonomy is defined:
the agent (or part of the agent) is autonomous or is not
autonomous [9].
levels of
is relative when several
Autonomy
autonomy are defined. The interest is to be able to
define a measurement of the autonomy of an agent. In
[2], the status of the agent can pass from simple
executant to decision maker.
It must be noticed that we distinguish carefully
types of autonomy and levels of autonomy. A type of
autonomy underlines the nature of this one: autonomy
of goals, autonomy of plan, autonomy from the
environment, etc. Our work on autonomy concerns a
specific type of autonomy: autonomy with regard to an
attribute.
Summary. The preceding criteria are useful
- To analyze existing models: the autonomy of an
agent presented in [2] is partial (because defined for
each goal), relative (several levels of autonomy were
identified) and social (some relations of dependence
exist between the agents). The model developed in [9]
is a global autonomy model (the agent is able or not
able to create a goal, its nature not being specified),
absolute (only one level is defined) and nonsocial (the
presence of another agent is not essential), and
- To design new models: we developed a model of
autonomy called autonomy with regard to an attribute.
This autonomy is partial, absolute and nonsocial. It is
(1) partial, because it applies only to a part of the
agent: a specific attribute, (2) absolute: the agent is
autonomous or is not autonomous with regard to this
attribute, (3) nonsocial: the model does not require the
presence of a second agent (Table 1).
Table 1. Partial taxonomy of some autonomy models.
global
X
partial
X
X
Barber
Luck
Sanchis
social
X
non social
absolute
relative
X
X
X
X
X
4. Autonomy with regard to an Attribute
Autonomy with regard to an attribute is a model of
decisional autonomy. It is stated in the following way:
an agent is autonomous with regard to an attribute A if
it can choose in a nondeterministic way a policy of use
p among several and if it can change this one during its
execution time.
An attribute corresponds to a simple and non
ambiguous property of an agent. A policy of use means
in our context a succession of actions which carries out
a particular and well identified functioning mode. Let
us illustrate the concepts of attribute and use policies
with two examples: replication and mobility.
Replication is an attribute which makes it possible
for an agent to create a clone of itself on the local or a
remote host. The diversity of replication policies can
(among other possibilities) come from the replication
rate (one and only one clone per site, not more than
one clone per site, at least one clone per site) and from
replication hosts (replication on all the sites accessible
to the agent, sites meeting a specific criterion).
Another example of attribute is mobility. It is this
attribute which will be used to illustrate the various
aspects of the model.
4.2. Model
Two elements characterize autonomy with regard to
an attribute:
- A set of operating modes of this attribute
- A two-component choice module.
Functioning modes. If mobility is taken as an
example of an attribute, one can distinguish various
migration policies:
- Navigation according to a route (the site of arrival
is different from the starting site)
- Circular navigation (the site of arrival is the
starting site)
- Navigation directed according
to a certain
criterion (ex: transfer to the less loaded host or to the
site having the greatest free disk space)
- Random navigation (the next site to reach is
chosen at random).
Circular navigation and navigation according to a
route consist of several elementary hops. Other
operating modes require only one transfer: directed
navigation and random navigation are mono-hop
strategies.
Broadly speaking, the set of policies can be static or
dynamic. It is static if the number or the nature of an
operating mode cannot be modified after the agent
creation. It is dynamic in the contrary case. To
simplify, we considered only static sets of policies.
Indeed, the implementation of a dynamic set of
policies can induce a substantial modification of the
choice module [11].
Choice Module. The decision-making process used
to choose a policy does not obligatorily imply a
possibility of reasoning, rational choice or cognitive
structure related to a particular model of intelligence.
Indeed, the module of choice splits up into two parts: a
deterministic component and a nondeterministic
component.
Provided with inputs and outputs, we will say that a
component is deterministic if to the same inputs always
correspond the same outputs. It is not deterministic if
at two different moments T and T ', the same inputs
can produce different outputs. The terms of inputs and
outputs were selected for simplicity reasons and must
be understood in our context in their most general
meaning (internal or external conditions, stimuli,
states, etc).
inputs
Deterministic Choice
outputs
inputs
Pi
Nondeterministic
Choice
outputs
Pj
Figure 1. Choice Module.
The deterministic part of the choice module is
provided with N functioning policies Pi (1≤i≤N, N>1).
With the mobility example, migration policies could be
the ones previously presented.
The nondeterministic part includes an additional
policy P0 corresponding to the empty policy. It means
that no policy is triggered by the agent. It models the
situation where, although the choice module of an
attribute A was activated, no policy of A is started:
there is an inhibition of the policy resulting from the
deterministic choice.
the
the activation of
That means
that after
nondeterministic part of the choice module, the policy
Pj (0≤j≤N) finally elected can be different from the
policy Pi previously selected by the deterministic
component.
Each part of the choice module has its own utility.
The presence of the deterministic module of choice
provides the agent with a coherent behavior (rational
choice) directed by the conditions of inputs. The
nondeterministic module of choice ensures the agent
the decoupling between
its
its autonomy and
monitoring. It protects the agent from all recurring
forms of explicit or implicit command.
If the agent included only the deterministic choice,
it would be a simple program; if the nondeterministic
choice were only present, the agent would not be
autonomous because
lacking
in
rationality and
condemned to a chaotic behavior.
It is to be noticed that within an agent, it is neither
obligatory nor necessary that the two components of
the choice module of an attribute must be spatially
adjacent: according to the general architecture of the
agent and
the properties
it
implements,
the
deterministic choice can be a part of another
component of the agent (for example, when several
deterministic parts of choice modules of different
attributes are located at the same place). Consequently,
Figure 1 must be understood as an abstract
representation of a choice module.
4.2. An implementation example
We have implemented a software agent immersed
into a real environment (network of Linux systems), an
agent autonomous with regard to the mobility and
having to carry out the same task on each visited site.
The purpose of this application was multiple:
- To characterize in a fine way the various
properties (qualities or attributes) present in the agent
- To define and implement several migration
policies
- To study the realization of a complete choice
module
and
deterministic
containing
(i.e.
nondeterministic components) as well as its relations
with the defined policies
- To verify that the absence of the agent migration
control by a user did not prevent the application from
functioning correctly
- To evaluate the potential of extension of this
autonomy model.
Agent task. The task carried out by the agent on
each visited site consists in collecting the users’ names
locally logged. Before finishing its execution, the
agent transmits to the user by electronic mail the result
of its various displacements, this result including the
temporarily inaccessible sites (stopped machines), the
sites prohibited to the agent (it does not have the
necessary authorizations or
the site presents an
inappropriate execution environment).
Navigation policies. The policies at the disposal of
the agent are the random transfer (P1) and the circular
navigation (P2). These moving policies were selected
for the following reasons:
- The random transfer offers the agent to randomly
draw the name of the next site to be visited. This
policy was introduced in order to differentiate it from
the nondeterministic choice: in the first case, the
random draw relates to the name of the next site to
reach (i.e. a parameter of the policy), in the second
case, the drawing of a policy among N (with N=2).
navigation
circular
- The
the
requires
implementation of a multi-hop mechanism which
implies the construction of a route and to keep up to
date the list of the sites to be visited. In this
configuration of navigation, the follow-up of the agent
is more complex because the "umbilical link" between
the agent and its launching site is broken. Lastly, it
makes it possible to implement the stop of a multi-hop
policy during its execution after the agent choice.
Choice module. The choice module breaks up into
two strictly ordered contiguous
levels and
is
the autonomous_choice
materialized by
function
(written in a script language). For simplicity reasons in
the design and in the functioning, the set of the
navigation policies and the choice module are static:
no policy is removed or added during the execution of
the agent and
the choice module (deterministic
component and nondeterministic component) remains
fixed during the lifespan of the agent.
The function autonomous_choice sequentially calls
deterministic_choice
two
functions,
and
nondeterministic_choice:
function autonomous_choice
{
deterministic_choice
nondeterministic_choice
}
The deterministic_choice
selects or
function
maintains the current navigation policy. There is a
selection of a new policy when the preceding policy
has
is
the current policy
finished, otherwise
maintained. After execution of this function, the policy
Pi (1≤i≤2) is positioned (random or circular). It will be
maintained or modified by the nondeterministic part.
The nondeterministic part is implemented by the
nondeterministic_choice function which provides a
random choice among the three following possibilities:
it preserves the policy fixed by the deterministic
component (with the probability Pr1) or it forces the
execution of the random policy (Pr2) or starts the
empty policy P0 (Pr3). This one means that no
navigation policy is activated. In our application, P0
causes the end of the execution of the agent. Indeed, as
the agent’s activity mainly consists in carrying out a
service (a task) on the local site then to move, it was
decided that when the empty policy would be started it
would cause the stopping of the agent on its local site.
In an application where the agent would be richer in
properties, the empty navigation policy would allow
the agent to continue its execution without moving, the
future of the agent depending on the properties it
integrates.
Qualities and attributes of the agent. As well as
autonomy (quality) and mobility (attribute), the agent
integrates two other attributes, these two being relative
to perception. The first is the perception by the agent
of a clone of itself, i.e. a residual "incarnation" of the
agent. The presence of this type of entity indicates a
dysfunction of the application (and not the presence of
a replication attribute into the agent). The second
attribute is the site perception.
The agent is autonomous with regard to the
mobility and non autonomous with regard to the other
two attributes relative to perception because for each
one of them, the agent integrates only one use policy
and no choice module.
4.3. Possible extensions
The work undertaken on autonomy with regard to
an attribute offers several tracks to be explored or
deepened. We will mention four of them:
- To apply the model of autonomy to an attribute
other than mobility. A candidate attribute already
mentioned is the replication of the agent. It is also
possible to build an agent autonomous with regard to
the task it has to carry out
- For a given attribute, to make the set of policies
eligible by the choice module dynamic. That induces a
more or
two
the
less major modification of
components of this module. If there is integration
within the agent of one or more policies which are by
design external to it, it can be interesting to provide the
agent with a software architecture which makes it
physically open [11]
- To design an agent autonomous with regard to two
different attributes. There are then two deterministic
components and two nondeterministic components. A
meticulous study of the interactions between the four
components will have then to be undertaken. The
problems to be resolved seem similar to those posed by
the fitting or the organization into a hierarchy of
autonomous subsystems within a broader entity,
problems already identified in the biological and social
fields [8]
- To integrate in the agent various models of
autonomy: autonomy with regard to an attribute and,
for example, a model of social autonomy.
5. Conclusion
As a quality, autonomy is a complex property,
difficult to encircle. Nevertheless, it is possible to
distinguish if an agent is or is not autonomous with
regard to a particular attribute. By being nonsocial, the
suggested autonomy model spares the introduction of
an additional source of complexity in the guise of a
second agent and with it, the complexity carried by the
interactions which it can produce with the first agent.
By
being
partial,
it
avoids
contradictory
interpretations: an agent
is autonomous or non
autonomous with regard to an attribute. Lastly, the
model being built on the concepts of choice and uses
policies, entities such as a thermostat or the xbiff
daemon cannot be considered within this model as
autonomous agents. Although provided with many
assets, this model (like the others) does not exhaust all
the richness of autonomy when it is considered as a
quality.
However, combining at the same time rational
choice and non determinism, this model offers a
conceptual framework making it possible to think how
to implement a sort of freedom of an artificial agent.
6. References
[1] Barber, K. S., Martin, C. E., Agent Autonomy:
Specification, Measurement, and Dynamic Adjustment.
Third International Conference on Autonomous Agents
(Agents 1999), Proceedings of the Autonomy Control
Software Workshop, May 1-5 1999, Seattle (Washington),
USA.
[2] Barber, K., Gamba, S., I. M., Martin, C. E., Representing
and Analyzing Adaptive Decision-Making Frameworks. In:
Hexmoor, H., Castelfranchi, C., Falcone, R. (eds.): Agent
Autonomy. Kluwer Academic Publishers, March 2003.
[3] Castelfranchi, C., Guarantees for autonomy in cognitive
agent architecture. In: Wooldridge, M., Jennings, N.R. (eds.):
Intelligent Agents: Theories, Architectures and Languages.
Lecture Notes in Artificial Intelligence, Vol. 890, Springer
Verlag, 1995.
[4] Castelfranchi, C., Falcone, R., From Automaticity to
Autonomy: The Frontier of Artificial Agents. In: Hexmoor,
H., Castelfranchi, C., Falcone, R. (eds.): Agent Autonomy.
Kluwer Academic Publishers, March 2003.
[5] Dorais, G., Bonasso, R. P., Kortenkamp, D., Pell, B.,
Schreckenghost, D., Adjustable Autonomy for Human-
Centered Autonomous Systems on Mars. Mars Society
Conference, August 1998.
[6] Franklin S., Graesser, A., Is it an Agent, or just a
Program? A Taxonomy for Autonomous Agents. In: Müller,
J. P., Wooldridge, M., Jennings, N. (eds.): Intelligent Agents
III – Proceedings of the Third International Workshop on
Agent Theories, Architectures and Languages (ATAL 96).
Lecture Notes in Artificial Intelligence, Vol. 1193, Springer
Verlag, 1996.
[7] Jennings, N., Wooldridge, M., Applications of Intelligent
Agents. In: Jennings, N. R., Wooldridge, M. (eds.): Agent
Technology: Foundations, Applications, and Markets.
Springer Computer Science, 1998.
[8] Koestler, A., The Ghost in the Machine, Hutchinson,
London, 1967.
[9] Luck, M., D'Inverno, M., Autonomy: A Nice Idea in
Theory. In: Castelfranchi and Lesperance (eds.): Intelligent
Agents VII: Proceedings of
the Seventh International
Workshop on Agent Theories, Architectures and Languages,
Lecture Notes in Artificial Intelligence, Vol. 1986, Springer
Verlag, 2001.
[10] Sanchis, E., Designing new Agent Based Applications
Architectures with the AGP Methodology. In: Proceedings of
the twelfth IEEE International Workshops on Enabling
Technologies: Infrastructure for Collaborative Enterprises
(WETICE 2003). IEEE Computer Society, 2003.
[11] Sanchis, E., A Systemic Framework for Open Software
Agents. In: Second International Workshop on Radical Agent
Concepts (WRAC 2005), Greenbelt (Maryland), USA,
Lecture Notes in Computer Science, Vol. 3825 (2006), p.
222-232.
[12] Scerri, P., Reed, N.E., Designing Agents for Systems
with Adjustable Autonomy. In: IJCAI-01, Workshop on
Autonomy, Delegation, and Control:
Interacting with
Autonomous Agents, August 6, 2001, Seattle (Washington),
USA
|
cs/0502094 | 1 | 0502 | 2005-02-27T18:16:30 | Coalition Formation: Concessions, Task Relationships and Complexity Reduction | [
"cs.MA"
] | Solutions to the coalition formation problem commonly assume agent rationality and, correspondingly, utility maximization. This in turn may prevent agents from making compromises. As shown in recent studies, compromise may facilitate coalition formation and increase agent utilities. In this study we leverage on those new results. We devise a novel coalition formation mechanism that enhances compromise. Our mechanism can utilize information on task dependencies to reduce formation complexity. Further, it works well with both cardinal and ordinal task values. Via experiments we show that the use of the suggested compromise-based coalition formation mechanism provides significant savings in the computation and communication complexity of coalition formation. Our results also show that when information on task dependencies is used, the complexity of coalition formation is further reduced. We demonstrate successful use of the mechanism for collaborative information filtering, where agents combine linguistic rules to analyze documents' contents. | cs.MA | cs | Coalition Formation: Concessions, Task Relationships and Complexity Reduction
Samir Aknine
LIP6, Paris 6
8 Rue du Capitaine Scott
75015, France
[email protected]
Abstract
formation problem
the coalition
to
Solutions
commonly assume agent rationality and, correspondingly,
utility maximization. This in turn may prevent agents from
making compromises. As shown
in recent studies,
compromise may
formation and
facilitate coalition
increase agent utilities. In this study we leverage on those
new results. We devise a novel coalition formation
mechanism that enhances compromise. Our mechanism
can utilize information on task dependencies to reduce
formation complexity. Further, it works well with both
cardinal and ordinal task values. Via experiments we
show that the use of the suggested compromise-based
coalition
formation mechanism provides significant
savings
communication
the computation and
in
complexity of coalition formation. Our results also show
that when information on task dependencies is used, the
complexity of coalition formation is further reduced. We
demonstrate successful use of
for
the mechanism
collaborative information filtering, where agents combine
linguistic rules to analyze documents' contents.
1 Introduction
Coalition formation mechanisms are often used as a
means for agent coordination. Coalition formation is
necessary when agents need to perform tasks which they
cannot carry out efficiently alone. Several coalition
formation models have been suggested to date (e.g.
[5,6,7]), based on var ious assumptions. Recently, a
solution that suggests that agents compromise their gains
to promote coalition formation was suggested [3]. In this
article, we address the coalition formation problem and
suggest compromise too. However, prior work assumes
that the value at which compromise is beneficial is
known, or can be arr ived at experimentally. In many real
applications where coalitions are necessary,
th is
assumption does not hold. In our coalition formation
mechanism, we do not assume
that
the optimal
compromise poin t is known in advance. Rather, we
provide agents with means to gradually arr ive at an
agreed compromise via a series of concessions. Such a
Onn Shehory
IBM Haifa Research Lab, Ha ifa University
Haifa 31905, Israel
[email protected]
solution is more flexible and more applicable to real
problems compared to previous solutions. Our coalition
formation method utilizes dependence relations among
tasks to guide agent search for coalitions. Searching the
coalition space based on task combinations serves as a
means for agents to arrive at compromise. Our approach,
in difference from other coalition formation solutions,
suggests that agents first attempt to reduce conflicts
amongst themselves, and only then form the coalition.
Conflict reduction in turn reduces negotiation time
between agents, as our experimental evaluation explicitly
shows. In addition, our solution considers the complex
relationships among tasks in the search for coalitions.
Such use of task complexity to improve the coalition
formation is not present in other solutions, as task
relationships are seldom considered.
Several reasons accredit the use of task structure
analysis for coalition search in the multi-agent system.
The presence of
task
relationships,
for
instance
dependencies between tasks, introduces a certain form of
dependence between the agents which will perform them.
Consequently, dependencies between
the coalitions
formed by the agents may arise. When these dependencies
exist, it would be preferable that agents identify them in
advance, as such knowledge can promote compromise
and simplify coalition formation. Thus, it should be
advantageous to manage dependencies between tasks
prior to their assignment to agents. This implies that
agents must reason about the tasks they will perform prior
to making a decision on the coalitions they join .
Dependencies between tasks can be revealed prior to
coalition formation via search. Leaving dependencies
between tasks unidentified does not avoid conflicts, it
merely postpones
their
resolution. When conflict
resolution is postponed to a late stage of the coalition
formation process, agents might form coalitions that they
will need to break when conflicts are found, thus require
further negotiation and re-formation.
Another significant reason for searching the coalition
space based on task dependencies is the reciprocal interest
of the agents. For instance, an agent A may take part in
the execution of a task T; task T may be of significance
for another agent B; Agent A might need B's support in
future tasks. If A and B know th is in advance, they are
motivated to be more cooperative, and in turn apply
compromise regarding the execution of T and other tasks.
In this article we illustrate the method we propose in
the context of a European project. The project provides a
multilingual system for the analysis and detection of racist
and revisionist content on the Internet. In this application
domain, agents are used
to dynamically combine
linguistic rules for document analysis. Agents, each
implementing a linguistic rule, form coalitions. The
coalition formation method consists of clustering tasks
into task partitions based on relationships identified
among the tasks. Task relationships we consider include
dependence, similarity, covering, etc. Task clustering
simplifies coalition formation and consequently reduces
the negotiation time between agents.
This article is organized as follows. Section 2
introduces the coalition formation problem and the
document filtering application context for which we solve
it. In section 3, we present the concepts of the coalition
formation model, first the principles of th is model, then
the behaviors of the agents in the model. Section 4
presents the results of the experiments carried out. We
then conclude and discuss future work in section 5.
2 The problem
In this section we formally define the coalition formation
problem we solve and describe the context of the solution.
2.1 Problem description
Consider a set of agents C = {C1, C2, ... ,Cn} and a set
of tasks T = {T1, T2, ... ,Tm}. The agents in C need to
execute the tasks in T and can negotiate over their
execution. Several relationships are defined on T. Tasks in
T can be combined into sets of tasks, and negotiation can
be performed over such sets. Each agent in C has its own
utility function that it tries to maximize. The goal of an
agent is to determine the coalitions to which it should
belong and the tasks or task sets to perform in these
coalitions, such that its utility is maximized.
2.2 Application context
The application addressed in this study concerns web
document filtering. The tracking of racist documents on
the Internet faces a number of obstacles, which make it
impossible to rely only on the classical keyword-based
approach, neural network techniques, etc. [1,4] Racist
discourse spans from hate speech
to more subtle
insinuations, with different themes: racist, revisionist,
anti-semitic, etc. Different genres are used in these
documents pseudo-scientific articles, pamphlets, etc.
From the analysis of large sets of racist, anti-racist and
non-racist documents, a number of criteria for identifying
racist content have been identified by the teams of
linguists working on the project: (1) Unique racial
expressions used only by racist people, for example
“Rahowa” standing for “Racial Holy War”; (2) Average
frequencies of certain words in racist documents differ
from their average in general documents. These words are
not necessarily racist ones but more common words (like
“their” or “white”); (3) Combined frequencies of certain
word pairs are relevant, e.g. the combination of “our”
with “civilization”, “race” or “religion”; (4) Suffixes like
“al”, “ence”, “ism” are good indicators for separating
racist and anti-racist documents.
One difficulty of the document filtering system we try
to build is that a single criterion is not sufficient for
indicating racism; convergence of several criteria is
required for such indication; however, this indication may
be valid only provided that there are no concomitant
indications of an ti-racism. Hence the number of criteria
(several
thousands),
their correlations and relative
relevance, increase the overall complexity. To address
this problem, we in troduce a solution based on the use of
a multi-agent system. In particular, we employ a coalition
formation model to support criteria combination for the
information retrieval and filtering problem at hand.
3 Coalitions of criteria agents
As stated above, the analysis of documents in the
problem domain addressed here employs multiple
evaluation criteria. In practice, each document should be
evaluated based on a combination of several linguistic
criteria. Since combinations vary across documents, and
the right combination for each document is not known in
advance,
there is a need for a dynamic, flexible
mechanism for criteria selection and combination. In this
work this need is addressed via multi-agent coordination.
We in troduce a set of criteria agents, each representing a
single linguistic criterion, and provide them with a
dynamic coordination mechanism – coalition formation –
to address the criteria combination problem. Following,
we present the coalition formation model used by the
criteria agents. Each agent, given a document, has to
produce a set of characteristics of that document, using
both
its own processing methods and
information
produced by other criteria agents. To this end, the
important contribution of our work is in the collaborative
use of information produced by multiple agents. This
collaboration improves the quality of document and site
classification compared to results achieved without such
collaboration.
3.1 Model preliminaries
Based on a single criterion, an agent cannot
individually provide a defin itive judgment on a document.
Hence, agents must dynamically join together to produce
defin itive judgments. In our solution, this is achieved by
agents forming coalitions to analyze documents. A
coalition is a temporary association between agents to
reach join t goals. For the application domain addressed
here, the purpose of a coalition of agen ts is to categorize a
document as racist, revisionist, anti-racist, etc.
The details of a coalition formation protocol depend
on the type of problem studied. For instance, varying trust
relations or agents' objectives, might require different
protocols. To enable the agents to form coalitions, most
current protocols (an exception is found in [2,3]) assume
that the utility functions of agents, which measure their
degree of satisfaction with each suggested solution, must
be comparable or the same. This means that agents must
be able to agree (for each task or task combination) on a
common utility function. This assumption is acceptable
for many multi-agent systems, in particular for economic
cases where utilities can often be calculated in terms of
profit. However, in many cases comparing the utilities of
agents, and even more so their aggregation, is nontrivial.
The numerical measurement of the utility of an agent is
already a strong assumption for itself.
In this work we propose a new coalition formation
model which does not require aggregation of agent
preferences and utilities. This model does not force the
agents to follow a particular order while participating in
the coalition formation process; it guaranties an equitable
processing of agents' choices. The protocol suggested is
particularly suitable for problems with complex tasks
(where there is a need for several agents and for
coalitions) and for dynamic scenaria where tasks may be
added and others cancelled or modified, and where the
agents have different utility functions. Agents are self-
interested, i.e. they do not necessarily trust each other.
However they respect all the commitments to which they
agreed. We assume that the utility functions of the agents
are not known by the other agents and do not need to be
cardinal, an ordinal utility is enough.
Our mechanism starts with each agent building partial
solutions which better account for
its preferences.
Individual partial solutions are then merged by agents, a
process in which agents make concessions. A complete
solution may eventually be arrived at by a group of
agents, thus a coalition is formed. In this model agents
seek gradual and
reciprocal concessions. Without
concessions, solutions will not be reached and the agents
will not reach their objectives. The concessions are made
gradually. Step by step, each agent is asked to make a
compromise that will set its preferences closer to the
others'. We re-emphasize here that making concessions is
a rational behavior (as suggested in [3]), as it increases
the expected utility of the conceding agent. In the
example problem domain we study, the criteria agents
should make concessions regarding the documents to
which they will be applied first. At the first stage of their
processing, these agents draw a list of documents they
prefer to analyze and give them an order of priority. At
the coalition formation stage, agents should modify their
preferences regarding either the documents to be analyzed
or their priorities. Prioritizing is necessary since applying
all the criteria to a document is infeasible, as there are
thousands of cri teria which could be applied.
3.2 Definitions
Before presenting
the details of
the coalition
formation protocol and agent behaviors, we introduce
some definitions. We initially define the concepts on
which the coalitions and the solutions agents search for
are based. We proceed with defining various types of
relationships between tasks.
Definition 1: Coalition. A set of agents that joined
together to perform a task or a combination of tasks.
Definition 2: Coalition structure. A set of coalitions
that, together, can address all of the tasks and task
combinations to be performed at a given moment. If a
coalition structure is approved by all agents, it is
considered a solution. A solution guarantees that all of the
tasks are executed by the agents. The set of tasks and task
combinations that the coalitions in the solution will
perform is called the support of the coalition structure.
Definition 3: Group of coalition structures. A set of
coalition structures that each provides a specific agent
with the same utility. For brevity, it will be referred to as
a group of structures or simply a group.
Definition 4: Context. A set of parameters which must be
stable during a negotiation step.
A context is particularly important in the application
domain we address, because the locations of documents
change rapidly. Example context parameters are date and
time.
Definition 5: Utility function. Measures the satisfaction
of the agent with, or its surplus from, its collective and
individual actions. It is used to represent the preferences
of the agents. It may be ordinal or cardinal. In our case,
measuring the utility of a structure of coalitions means
comparing it with a reference state. The reference state
will be the same one throughout the negotiation.
Definition 6: Reference state. In order for the agents to
know whether they should accept a coalition structure as a
solution, they need to be able to compare it with their
minimal guaranteed gain during the negotiation. This
minimum is the reference state. If there are already
formed coalitions, the reference is the current coalitional
state.
To guarantee a solution after a negotiation, the
reference state needs to be feasible and identical across all
of the criteria agents. Otherwise, solutions arrived at by
different
agents might be
inconsistent
amongst
themselves.
Definition 7: Acceptable coalition structure. A
coalition structure is acceptable for an agent if it is
preferred over, or equivalent, to the current reference
state.
Definition 8: Signature of a coalition structure Ei,
denoted Sig(Ei ), defines the set of criteria agents that have
approved this coalition structure.
Definition 9: Knowledge K(Ci ) of an agent Ci is the set
of coalition structures, and the corresponding signatures,
known to Ci .
Definition 10: Unacceptable coalition structures: a set of
coalition structures Out(Ci ) for which agent Ci, notifies
the other agents of it being unacceptable for it. An
unacceptable coalition structure cannot be proposed as a
solution by any agen t.
The search for the support of a solution considers
relationships between tasks. Below, we define some types
of these relationships and provide examples. First, we
introduce the relationships between tasks by considering
the agents which will perform them; then, we consider
relationships based on the structure of the tasks.
Definition 11: Total covering. Tasks Ti and Tj are in a
total covering relationship if the agents selected to carry
out Ti can also carry out Tj and inversely, and the tasks do
not conflict (resource conflicts, etc.).
In the case of total covering, tasks Ti and Tj could be
carried out either sequentially or in parallel, since possible
conflicts are avoided.
Definition 12: Inclusive covering. Tasks Ti and Tj are in
inclusive covering if a subset of the agents selected to
carry out one of the two tasks can carry out the other.
Definition 13: Partial covering. Tasks Ti and Tj are in
partial covering if a subset of the agents selected to carry
out Ti can also carry out a part of Tj and inversely.
Definition 14: Total complementary tasks. Tasks Ti and
Tj are totally complementary if each of them can reuse the
results of the other task.
As an example, consider two documents Di and Dj
indexed by
the same site. Based on preliminary
experiments we performed, we know that, (a) frequently,
documents found at the same site are similar to one
another (in the linguistic sense); (b) because of their
similarity, the results obtained when analyzing Di subject
to certain criteria, may be applicable to Dj too. For
instance, computing the frequencies of one or several
words in a document could be a shared result. This
introduces a complementary relationship.
Definition 15: Dependent tasks. Tasks Ti and Tj are
dependent if only one of the two tasks needs the results of
the other task to perform its execution.
The dependence relationship between tasks allows
grouping of agents based on their shared interests. For
example, in order to analyze a document Di at a given
site, several criteria are necessary, among which some
may be relevan t to the analysis of another document Dj at
the same site. Such criteria sharing may promote the
formation of a coalition for performing the two analysis
tasks of Di and Dj.
Definition 16: Competitive tasks. Tasks Ti and Tj are
competitive if Ti and Tj compute the same indicators using
different methods. Ti and Tj should not be par t of the
support of a same solution. Recognizing competitive tasks
makes it possible to prune the solutions search space.
3.3 The coalition formation method
The aim of th is method is to solve the agent coalition
formation problem without having to aggregate the
preferences of the agents, and to allow a dynamic and fast
reorganization of these coalitions according to changes in
the problem domain. The method we propose is based on
two concurrent behaviors of the agents: task analysis, and
negotiation. Task analysis consists of grouping of tasks
into combinations, based on relationships among the
tasks, where each combination is to be performed by a
coalition. To simplify and optimize the search among
those combinations of tasks, a binary tree is constructed,
and combinations are placed in its nodes. Additionally,
from each combination of tasks competitive combinations
(defined below) can be isolated. The advantage of the
suggested method is in directing the search of the support
of the solutions towards preferred tasks, thus reducing
search complexity.
Definition 17: Competitive combination of tasks. Two
combinations of tasks CT1 and CT2 are competitive for an
agent ai if its utility from CT1 ,U(ai, CT1), is equal to its
utility from CT2 , U(ai , CT2), and CT1 and CT2 do not
belong to the support of the same solution.
Definition 18: Tree of task combinations is constructed
based on combinations' preference. The preferred
combination of the agent that constructs the tree is placed
in the root. The rest of the nodes are populated in a
descending preference order. Two branches emanate from
each node N: a positive branch labeled with (+),
indicating that N is in the support of the solution; a
negative branch labeled with (-), which indicates that N is
out of the support of the solution.
The supports of the coalition structures in this tree are
read starting from the root and by retaining only the nodes
when branches emanating from them are labeled with a
positive sign. For instance, in Figure 3 the support S3 is
only formed of the combinations <{T1, T2, T3}, {T5, T6}>
since on the path starting from the root to the leaf node S3,
the node {T4, T5} is negatively labeled on the branch of
this path. Hence, this node does not belong to the support
S3.
The tree of combinations is incrementally constructed.
To develop this tree an agent selects the combinations that
it organized first in partitions. We define a partition of
task combinations as a set of combinations concerned
with one task and containing all the combinations where
this task appears. Each combination of tasks belongs to
only one partition (cf. Figure 2).
For instance, consider four agents C1,…,C4, eight tasks
T1,…,T8 and seven possible coalition structures E0,…,E6,
where E0 is the initial state. Of these structures, only three
are Pareto optima (E1, E3 and E6). Let Ui (E) be the utility
of agent Ci for the coalition structure E. We represent the
coalition structures in Figure 1 according to the utility that
they bring to each agent. To build its coalition structures,
agent C1 computes some task combinations. This is done
with respect to its preferences. For instance, C1 computes
the following preferred combinations: {T1, T2, T3}, {T1,
T2}, {T4, T5}, {T5, T6}, {T3, T4, T5}, {T6, T7} and {T6, T7,
T8}.
U2
E1
E5
E3
E4
E2
E0
Sub-optimal so lution
Optimal so lution
U1 : Agent C1 util ity U2 : Agent C2 util ity
E6
U1
Figure 1. Possible coalition structures in a utility
space
In this example, these combinations are used to build the
coalition structure E1 as explained hereafter. To simplify
the construction of, and the search in, the tree of the
supports for the solutions, these combinations are first
classified into several partitions (P1, P2 and P3 in Figure
2) and ordered in decreasing preference order. The
preference degree of each partition is computed based on
the preference degrees of the combinations it includes.
The first partition p1 has the highest preference degree
and is thus placed at the top. In building the tree, the
preferred combination of agen t C1 ,{T1, T2, T3}, is attached
to the root of its tree (cf. Figure 3). Then, the agent
creates
the next
for
two branches and searches
combination to be considered. In the positive branch, the
combination must be selected from a partition other than
P1, because combinations in p1 share at least the task T1,
and a support of a coalition structure should have only
one instance of each
task. The agent selects
the
combination {T4, T5} from P2. As for the negative branch,
the agent selects the combination {T1, T2} from P1. The
tree is further built in the same manner. For instance, the
following positive combination is {T6, T7} form P3.
Note that singleton combinations (containing one task)
are not immediately in tegrated into the tree. To avoid
building unnecessarily deeper
trees, singletons are
directly added to the support once formed. Searching the
tree and selecting the nodes from which positively labeled
branches emanate gives us already several possible
supports, for
instance, S3. S3
is formed of
two
combinations <{T1, T2, T3}, {T5, T6}>. If among the
singleton combinations, there are some that the agent
would like to perform, they are added to the support. E.g.,
if C1 is interested in tasks {T4} and {T7}, S3 becomes <{T1,
T2, T3}, {T5, T6}, {T4}, {T7}>.
Once the agent has identified the supports in the tree
of task combinations, and in particular its most preferred
coalition structures, it is ready to start the negotiation
phase. In particular, it will send coalition structures it has
found as part of coalition formation proposals to peer
agents. Each agent computes its own tree. Each tree may
provide several supports of different coalition structures,
and such supports may vary across trees. Th is in turn
the coalition
to
allows several different solutions
formation problem, as each task or combination of tasks
in each support can be allocated to different coalitions of
agents.
T1 , T2, T3
{P1}
{P2}
{P3}
{T1, T2, T3}
{T1, T2}
{T4, T5}
{T5, T6}
{T3,T4, T5}
{T6,T7}
{T6,T7, T8}
Figure 2. Partitions of
task combinations
-
+
T1, T2
-
+
T4, T5
-
+
…
…
T5, T6
T6, T7
+
-
T6, T7, T8
S3
+
S1
+
S2
Figure 3. Supports tree
3.3.1 Initiation of the negotiation. The negotiation
process we propose is based on three phases: initialization
of the negotiation and transfer of tasks, the core
negotiation, and transmission of the solution. There is no
pre-established order among the agents. Agents can
initiate negotiation and participate at any
time.
Negotiation may be in itiated when new documents to be
analyzed are sent by the search engines or by a peer
criterion agent. Each agent asks the other agents to send it
their tasks, prioritized. Upon such a request, each agent
computes and sends the vector of its conditional choices.
We assume that agents are cooperative in this respect. For
instance, in our application domain, tasks are documents
to be analyzed. An agent Ci may send a vector, e.g.,
Want(Ci )=(t1:D1D2D3, t2:D4 if Cj(cid:7507)D4 ,:D5). This vector
indicates that Ci wants to analyze documents D1, D2 or D3
first (at time t1). It wants to analyze document D4 second
(at time t2), if the necessary resources are provided by
agent Cj. It has no demands for document D5.
After task vectors with preferences are transmitted,
each agent holds the set of tasks and can compute
coalitions to be proposed or agreed to. The in itiator agent
computes
the preferred combinations of
tasks.
It
constructs a tree of supports as described above. Each
support in this tree will provide a set of coalition
structures. Indeed, for each task or combination of tasks
in this support, the initiator agent finds its preferred
agents which will perform them. The agent then gathers
these coalition structures in groups in order to initiate the
negotiation.
3.3.2 The negotiation model. Once the initiator agent
computes the coalition structures, it chooses an agent to
which these coalition structures will be sent (as proposals
for coalition formation). This choice is based on the
agent’s
strategies. The agent
also declares
its
unacceptable coalition structures. The initiator initially
sends its most preferred coalition structures, signed in
Sig(Ei ); it may iteratively send, in a decreasing order of
preference, its other coalition structures. This may be
continued until there are no more coalition structures at
least equivalent to the current state. However, before
sending a less preferred coalition structure, the agent may
wait until it receives a message from another agent either
about the former coalition structure it has proposed or
about a new coalition structure proposed by that agent.
Each of the other agents also computes its preferred
coalition structures. It then computes the coalition
structures that it will choose in second position and so on.
When an agent R receives a group of coalition
structures from a sender agent S, it sorts those coalition
structures in order of preference into homogeneous new
groups. In each of these groups, all coalition structures are
equivalent in terms of utility of S. R updates its
knowledge on the coalition structures in K(R). When the
utility of a new group is equivalent to the utility of a
group which is already known to R, these groups are not
merged by R, since for S their utility is not the same. R
only sorts coalition structures that are at least equivalent
to the reference state and the others are not considered. If
there is at least one coalition structure which is preferable
or equivalent to R's best choice, it forwards this coalition
structure CS to the next agent that it wishes to include in
the negotiation. R signs CS and adds it to Sig(CS). This
information indicates that CS has been approved by R and
by all the preceding senders as well. When R finds in the
set it receives unacceptable coalition structures, it has to
declare them unacceptable to the other agents in Out(R).
Out(R) also contains the coalition structures that R itself,
locally, identified as unacceptable.
Information on unacceptable structures is useful
because it prevents the need of other agents computing
coalition structures that will be systematically refused.
However, its transmission is expected only when agents
trust each other. With no trust, agents would not transmit
this information, to prevent the others from using the
information
instance, such
their strategies. For
in
information could enable agents to know which coalition
structures an agent may accept, thus gaining an advantage
in negotiation. When a combination of tasks in a support
of a coalition structure is unacceptable for an agent, it
should also indicate that. This combination of tasks is
then deleted from the partitions of tasks to prevent it from
being considered as part of a support of another coalition
structure. The tree is also revised in order to erase the
supports containing this combination of tasks.
In case that a coalition structure CS is acceptable for
agent A, but its utility is inferior to its top choice, A may
nevertheless decide to forward CS to the next agent B.
Consequently the number of agents having approved CS
grows and CS is thus reinforced. A may also decide to
temporarily block CS but indicates in K(A) that it has
received it, if it considers that there is still enough
negotiating time to reach a consensus. In this case the
agent sends to B the coalition structures it prefers.
A possible end point of the negotiation occurs when
an agent C receives a group of coalition structures
approved by all other agents. C sorts the coalition
structures into groups. If at least one of the coalition
structures is better than, or equivalent to, C's reference
state, and if negotiation time is about to expire, it can
consider the suggested structure as its best group. All the
coalition structures of th is group are Pareto optima, so C
can arbitrarily choose one of them as a solution for the
negotiation. In case that negotiation time has not expired
and provided that C has some undeclared groups of
coalition structures, it can continue the negotiation.
Once the last agent has identified a Pareto optimal
solution which is approved by all, it sends this coalition
structure to the other agents, which accept it as the
solution for
the negotiation, as
they have already
confirmed it in Sig(Ei ) and are thus committed to it.
3.3.3 Negotiation example. We illustrate the negotiation
model using the previous example. For simplicity, we
limit our discussion to two agents, C1 and C2. Agent C1
initiates
the negotiation and
it builds
its
tree of
combinations of tasks. Then C1 generates a set of
coalition structures it considers acceptable. It then sorts
them into equivalent groups of coalition structures:
G1(E6) ; G2(E4;E3) ; G3(E2) ; G4(E0;E1). In each group,
coalition structures have the same preference (cf. Figure
4). E5 is not sorted as the reference state (E0) is better.
G4
E1
E5
E0
U1=0
G3
E2
G2
E3
E4
G1
E6
U1
Figure 4. Group of coalition structures of agent C1
Groups G1, G2, G3 and G4 are acceptable to agent C1
as they correspond to a state which is as satisfactory as
the initial reference state, or better. In the same way,
agent C2 also searches for its preferred acceptable
coalition structures and sorts them into equivalent groups
of coalition structures (Figure 5): G’1(E1); G’2(E2;E3);
G’3(E4); G’4(E5); G’5(E0). E6 is not sorted as the reference
state E0 is better.
G’4
E5
G’3
G’2
E3
E2
E4
G’5
E0
U2=0
E6
G’1
E1
U2
Figure 5. Group of coalition structures of agent C2
Groups G’1, G’2, G’3, G’4 and G’5 are acceptable to
agent C2 as they correspond to a state which is as
satisfactory as the initial reference state, or better. Thus
C1 starts by sending its first preferred group G1. Agent C2
starts by receiving G1 and evaluates it (cf. Figure 6). The
unique coalition structure E6 in G1 is unacceptable for C2
because it leads to a state less satisfactory than the initial
state. Agent C2 does not send this coalition structure. It
notifies C1 that it rejects G1(E6). If one or more coalitions
in the coalition structure E6 is appropriate for C2, it may
indicate it to C1.
U2
U2 =0
E6
Figure 6. First group received by agent C2
Agent C1 must then wait for a new proposal from
agent C2. Thus C2 sends its preferred set, i.e. G’1(E1). This
group contains only coalition structure E1 which is
acceptable to agent C1 but which corresponds at the same
time to its least preferred choice. For E1 the utility of C2 is
the same as that of its reference state E0 (cf. Figure 5).
Agent C1 does not reject th is coalition structure but it still
has a possibility to propose a new group to agent C2. C1
thus sends its second preferred group G2(E3; E4).
U1
U1 =0
E1
E0
Figure 7. First group received by agent C1
Agent C2 receives G2 and separates its two coalition
structures into two groups. E3 of G2 now belongs to its
group G’2 which corresponds to its second preferred
group. As for coalition structure E4, it belongs to group
G’3 its third choice. E3 is acceptable. As all the other
agents have already par ticipated in the negotiation, agent
C2 cannot send it to others. Coalition structure E3 of G’2
can thus be a solution. Agent C2 has no other Pareto
optimal choices left. Either it sends E3 to agent C1 in order
to indicate to it the final result of the negotiation, or it
waits until C1 gives way on coalition structure E1 which is
also Pareto optimal considering the fact that agent C2 has
already refused one of the coalition structures that agent
C1 proposed to it. The negotiation between the two agen ts
inevitably fin ishes on one of the two Pareto optima and
before expiration of the pre-set negotiation time. Note that
classical game theoretic analysis might bring the two
agents to an equilibrium where no coalition is formed
(and both lose). Yet, as shown in [3], and reinforced in
our work, it is in the best in terest of agents, even if they
are self-interested,
to compromise. Implementing a
compromise strategy, one of these agents will normally
yield to allow for a solution to be arrived at.
4 Experimental evaluation
To evaluate it, the model suggested in this study was
implemented and experimented with. The code was
written in Java and run on a Windows 98 host, 1.4 GHZ
processor and 256 MB RAM. We in itially performed
experiments on
the basis of
some dependence
relationships. In order to evaluate the protocol, we have
analyzed
its performance by observing
several
parameters: the number of exchanged messages, the size
of these messages (the number of coalition structures they
contain), the number of coalition structures that have been
evaluated and the negotiation runtime. The results
obtained are summarised in the following figures. It
should be noted that each point in the graphs is the
average of 10 tests carried out under the same conditions.
Runt i me (ms)
4000
F ig ure 8 : Neg o t iat io n runt ime
3500
3000
2500
2000
1500
1000
500
0
# agents
10
9
8
7
6
5
4
3
2
Figure 8 shows negotiation runtimes obtained in
milliseconds with the number of agents varying between 2
and 10. The negotiation runtime increases in this figure.
This is due to the h igher number of proposals that agents
would compute and exchange. However the search time
remains acceptable even when the number of agents
grows. Figure 9 shows the number of coalition structures
sent by the agents. For instance in our experiments, with 4
agents and 4 documents, 18 coalition structures were
transmitted, and for 10 agents with 10 documents, only
5058 coalition structures have been sent, compared to
45927 coalition structures expected in a case where
dependencies are not considered (the latter was computed
offline). The number of messages sent varies considerably
according to the incompatibility of the preferences of the
agents.
Fi gur e 9: Number of coal i t i on st r uc t u r es sent dur i ng a negot iat ion
# st r uc t u r es
6 00 0
500 0
4 00 0
3 00 0
2 00 0
100 0
0
# ag ent s
2
3
4
5
6
7
8
9
10
Figure 10 shows the number of evaluated coalition
structures and allows measuring the effectiveness of the
search when handling of dependence relationships is
employed. Had agents not considered dependencies, for
instance in the case of 4 agents and 4 tasks, the number of
coalition structures they should have evaluated would
have been 6561. As the graph shows, the actual number is
by far smaller. This holds for larger numbers of agents
too: 10 agents have only examined 8983 coalition
structures, compared to 45927 possible ones when
dependencies are not handled.
# st ruct ur es
10000
8000
6000
4000
2000
0
# agents
Figur e 10: Number o f eva luat ed coal i t i on st r uc t u r es du r ing a negot iat ion
2
3
4
5
6
7
8
9
10
task
the use of
the gains from
To evaluate
dependencies as part of the coalition formation process,
we performed two experiments where in each the same
tasks were provided to the agents, however in one
experiment they used task dependencies and in the other
they did not. Comparative results between these two
experiments are presented in Figure 11. As seen there, the
gains of coalition formation with the handling of task
dependencies are, on average, about 13% higher than the
gains without it. This gain in computation results from
task
combinations
that
are
facilitated by
task
dependencies. By using task combinations, the agents
reduce the number of coalit ions they need to consider.
For instance, without the use of the combinations, an
agent that has 4 tasks must examine the 4 corresponding
coalitions. But with task combinations, for the same 4
tasks, the number of coalitions may decrease. E.g., if the
tasks are gathered in two task combinations, the agent will
have to examine only 2 corresponding coalitions.
# structures
20000
Figure 11: Total number of evaluated coalitions with and without
handling dependenc ies
15000
10000
5000
0
# agents
2
3
4
5
6
7
8
9
10
Note that although the results presented here refer to
10 agents, we have performed additional experiments in
larger systems, exceeding 50 agents. The subset presented
is merely aimed at providing the essence of the results.
5 Conclusion
Classical solutions to the coalition formation problem
assume cardinal ordering of task values. Additionally, it is
commonly assumed that agents behavior rationally, and
should thus not compromise their utility. Further, existing
solutions hardly analyze complex task dependencies.
Recently, it was shown [3] that compromise strategies in
coalition formation may dominate other strategies. In this
study we leverage on this new result. We advocate that,
for some problem domains, compromise is necessary not
only to facilitate the formation of coalitions; it is also
necessary to reduce the complexity of coalition formation.
We further claim that task dependencies can by utilized to
prune the coalition formation search space. As we show in
our experiments, the use of the compromise-based
coalition formation model we present provides significant
savings
the computation and communication
in
complexity of coalition formation. Our results also show
that when information on task dependencies is used, the
complexity of coalition formation is further reduced.
This study should be extended to address several issues.
Firstly, the sizes of the systems examined were relatively
small. Since in our experiments we observed a steep
growth in computational costs, scale-up of the proposed
mechanism may face difficulties, in spite of it performing
much better than a naïve approach. Secondly, the
mechanism should be check in other domains. Although it
was specifically designed, and proved successful, for
tracking racist documents, it is desirable to prove it
applicable to other domains. In future work we in tend to
pursue these directions. The current results are already
very promising, as were able to utilize a MAS technique –
coalition formation – for solving an important real-world
problem.
References
[1] Cardie, C. Empirical Methods
Information
in
Extraction, AI Magazine, 18(4):65-80, 1997.
[2] Kraus, S., Shehory, O., Tasse, G. Coalition Formation
with Uncertain Heterogeneous Information, AAMAS,
1-8, 2003.
[3] Kraus, S., Shehory, O., Tasse, G., The Advantages of
Compromising
in Coalition Formation with
Incomplete Information, AAMAS, 588-595 2004.
[4] Letsche, T. Berry, M.W. Large-scale Information
Retrieval with Latent Semantic Indexing. Information
Sciences, 100:105–137, 1997.
[5] Sandholm T.W., Lesser V.R. Coalitions among
Computationally Bounded Agents, AI, 99-137, 1997.
[6] Sandholm T.W., Larson K., Andersson M., Shehory
O., Tohmé F. Coalition Structure Generation with
Worst Case Guarantees, Artificial Intelligence, 111,
209-238, 1999.
[7] Shehory O, Kraus S. Methods for Task Allocation via
Agent Coalition Formation, AI, 165-200, 1998.
|
1802.01730 | 1 | 1802 | 2018-02-05T23:30:55 | Local Wealth Redistribution Promotes Cooperation in Multiagent Systems | [
"cs.MA",
"cs.AI",
"cs.GT"
] | Designing mechanisms that leverage cooperation between agents has been a long-lasting goal in Multiagent Systems. The task is especially challenging when agents are selfish, lack common goals and face social dilemmas, i.e., situations in which individual interest conflicts with social welfare. Past works explored mechanisms that explain cooperation in biological and social systems, providing important clues for the aim of designing cooperative artificial societies. In particular, several works show that cooperation is able to emerge when specific network structures underlie agents' interactions. Notwithstanding, social dilemmas in which defection is highly tempting still pose challenges concerning the effective sustainability of cooperation. Here we propose a new redistribution mechanism that can be applied in structured populations of agents. Importantly, we show that, when implemented locally (i.e., agents share a fraction of their wealth surplus with their nearest neighbors), redistribution excels in promoting cooperation under regimes where, before, only defection prevailed. | cs.MA | cs | Local Wealth Redistribution Promotes Cooperation in
Multiagent Systems
Flávio L. Pinheiro
Collective Learning Group, The MIT Media Lab
Massachusetts Institute of Technology, 22 Ames Street,
Cambridge, Massachusetts
[email protected]
Fernando P. Santos
INESC-ID and Instituto Superior Técnico,
Universidade de Lisboa
2744-016 Porto Salvo, Portugal
[email protected]
8
1
0
2
b
e
F
5
]
A
M
.
s
c
[
1
v
0
3
7
1
0
.
2
0
8
1
:
v
i
X
r
a
ABSTRACT
Designing mechanisms that leverage cooperation between agents
has been a long-lasting goal in Multiagent Systems. The task is
especially challenging when agents are selfish, lack common goals
and face social dilemmas, i.e., situations in which individual inter-
est conflicts with social welfare. Past works explored mechanisms
that explain cooperation in biological and social systems, provid-
ing important clues for the aim of designing cooperative artificial
societies. In particular, several works show that cooperation is able
to emerge when specific network structures underlie agents' in-
teractions. Notwithstanding, social dilemmas in which defection
is highly tempting still pose challenges concerning the effective
sustainability of cooperation. Here we propose a new redistribu-
tion mechanism that can be applied in structured populations of
agents. Importantly, we show that, when implemented locally (i.e.,
agents share a fraction of their wealth surplus with their nearest
neighbors), redistribution excels in promoting cooperation under
regimes where, before, only defection prevailed.
CCS CONCEPTS
• Computing methodologies → Multi-agent systems; Coop-
eration and coordination;
KEYWORDS
Emergent behaviour; Social networks; Social simulation; Simulation
of complex systems; Cooperation
1 INTRODUCTION
Explaining cooperation among selfish and unrelated individuals has
been a central topic in evolutionary biology and social sciences [23].
Simultaneously, the challenge of designing cooperative Multiagent
Systems (MAS) has been a long standing goal of researchers in
artificial intelligence (AI) [10, 19]. More than thirty years ago it was
already clear that "Intelligent agents will inevitably need to interact
flexibly with other entities. The existence of conflicting goals will
Proc. of the 17th International Conference on Autonomous Agents and Multiagent Systems
(AAMAS 2018), M. Dastani, G. Sukthankar, E. Andre, S. Koenig (eds.), July 2018, Stockholm,
Sweden
© 2018 International Foundation for Autonomous Agents and Multiagent Systems
(www.ifaamas.org). All rights reserved.
This is the author's version of the work. It is posted here for your personal use. Not
for redistribution. The definitive Version of Record was published in Proceedings
of Proc. of the 17th International Conference on Autonomous Agents and Multiagent
Systems (AAMAS 2018), M. Dastani, G. Sukthankar, E. Andre, S. Koenig (eds.), July 2018,
https://doi.org/doi.
need to be handled by these automated agents, just as it is routinely
handled by humans." [10].
In Cooperative multiagent interactions, agents need to collabo-
rate towards common goals, which introduces challenges associated
with coordination, communication and teamwork modeling [19, 28].
Self-interested interactions, in contrast, require the design of indi-
rect incentive schemes that motivate selfish agents to cooperate
in a sustainable way [7, 19]. Cooperation is often framed as an
altruistic act that requires an agent to pay a cost (c) in order to
generate a benefit (b) to another. Refusing to incur in such a cost is
associated with an act of defection and results in no benefits gener-
ated. Whenever the benefit exceeds the cost (b > c) and plays occur
simultaneously, agents face the Prisoner's Dilemma, a decision-
making challenge that embodies a fundamental social dilemma
within MAS [21]: rational agents pursuing their self-interests are
expected to defect, while the optimal collective outcome requires
cooperation. If defection is the likely decision of rational agents,
however, how can we justify the ubiquity of cooperation in the real
world? Evolutionary biology has pursued this fundamental question
by searching for additional evolutionary mechanisms that might
help to explain the emergence of cooperative behavior [22, 23].
Some of these mechanisms allowed to develop solutions that found
applications in computer science, such as informing about ways of
incentivizing cooperation in p2p networks [9, 11], wireless sensor
networks [2], robotics [47] or resource allocation and distributed
work systems [43] – to name a few.
Network reciprocity is one of the most popular mechanisms to
explain the evolution of cooperation in social and biological systems
[24, 26, 30–32, 38]. In this context, populations are structured and
interactions among agents are constrained. These constraints are
often modelled by means of a complex network of interactions.
Applications of this mechanism have been explored in the design
of MAS that reach high levels of cooperation [1, 17, 29, 34]. Despite
these advances, cooperation on structured populations is still hard
to achieve when considering social dilemmas with high levels of
temptation to defect. Additional complementary mechanisms are
required.
Here we consider that agents contribute a percentage of their
surplus (defined below), which is later divided among a Benefi-
ciary Set of other agents. In this context, we aim at answering the
following questions:
operation?
• Does redistribution of wealth promote the evolution of co-
• How should Beneficiary Sets be selected?
• What are the potential disadvantages of such a mechanism?
AAMAS'18, July 2018, Stockholm, Sweden
Flávio L. Pinheiro and Fernando P. Santos
Using methods from Evolutionary Game Theory (EGT) [44] and
resorting to computer simulations, we explore how wealth redis-
tribution impacts the evolution of cooperation on a population of
agents without memory (i.e. unable to recall past interactions) and
rationally bounded (i.e. lacking full information on payoff struc-
ture of the game they are engaging). We assume that agents resort
to social learning through peer imitation, which proves to be a
predominant adaptation scheme employed by humans [36]. Also,
we consider that strategies are binary – Cooperate and Defect –
opting to focus our attention on the complexity provided by 1)
heterogeneous populations, 2) the redistribution mechanism and 3)
the self-organizing process of agents when adapting over time. The
role of larger strategy spaces (such as in [29, 34, 41]) lies outside
the scope of the present work.
With redistribution, we show that cooperation emerges in a
parameter region where previously it was absent. Moreover, we
show that the optimal choice of redistributing groups consists in
picking the nearest neighbors (local redistribution). This result fits
with a local and polycentric view of incentive mechanisms [27, 46]
in MAS, which may not only be easier to implement but, as we
show, establish an optimal scale of interaction in terms of eliciting
cooperation.
2 RELATED WORK
The problem of Cooperation is a broad and intrinsically multidisci-
plinary topic, which has been part of the MAS research agenda for
a long time [10, 19]. In the realm of evolutionary biology, several
mechanisms were proposed to explain the evolution of cooperation
[22]. Kin selection [13], direct reciprocity [45], indirect reciprocity
[25, 42] and network reciprocity [26, 38] constitute some of the
most important mechanisms proposed. Remarkably, these mech-
anisms have been applied in AI in order to design MAS in which
cooperation emerges. For example, Waibel et al. associated kin se-
lection with evolutionary robotics [47]; Griffiths employed indirect
reciprocity to promote cooperation in p2p networks while Ho et al.
investigated the social norms that, through a system of reputations
and indirect reciprocity, promote cooperation in crowdsourcing
markets [12, 16]. Similarly, Peleteiro et al. combined indirect reci-
procity with complex networks to design a MAS where, again,
cooperation is able to emerge [29]. On top of that, Han applied
EGT – as performed in our study – in order to investigate the role
of punishment and commitments in multiagent cooperation, both
in pairwise [14] and group interactions [15]. Regarding alterna-
tive agent-oriented approaches to sustain cooperation in MAS, we
shall underline the role of electronic institutions [4, 8] whereby
agents' actions are explicitly constrained so that desirable collective
behaviors can be engineered.
The role of population structure and network reciprocity is, in
this context, a prolific area of research. In [31] it was shown that
complex networks are able to fundamentally change the dilemma
at stake, depending on the particular topology considered [18, 31];
Ranjbar-Sahraei et al. applied tools from control theory in order to
study the role of complex networks on the evolution of cooperation
[34]. Importantly, the role of dynamic networks – i.e., agents are
able to rewire their links – was also shown to significantly improve
the levels of cooperation, especially in networks with a high average
degree of connectivity [32, 39]. A survey on the topic of complex
networks and the emergence of cooperation in MAS can be accessed
in [17].
Previous works found that cooperation in structured population
substantially decreases when the temptation to defect increases
(see Model for a proper definition of Temptation). Thereby, here
we contribute with an additional mechanism of cooperation on
structured populations. We consider a mechanism of redistribu-
tion, inspired in the wealth redistribution mechanisms that prevail
in modern economic/political systems, mainly through taxation.
We are particularly interested in understanding how to sample re-
distribution groups in an effective way. In this context, we shall
underline the works of Salazar et al. and Burguillo-Rial, in which
a system of taxes and coalitions was shown to promote coopera-
tion on complex networks [37] and regular grids [5]. While [37]
and [5] do an excellent job showing how coalitions – leaded by a
single agent – emerge, here we consider a simpler/decentralized
model (e.g. no leaders are considered and taxes are redistributed
rather than centralized in a single entity) and focus our analysis on
showing that local redistribution sets are optimal. Our approach
does not require additional means of reciprocity, memory, leader-
ship, punishment or knowledge about features of the network. We
cover a wide range of dilemma strengths and explicitly show when
the local redistribution promotes cooperation by itself. Notwith-
standing, the analysis performed in [37] and [5] surely provides
important insights to address in future works, on how to explicitly
model the adherence to beneficiary sets and guarantee their stabil-
ity. Also, while here we assume an egalitarian redistribution over
each individual in the Beneficiary Set, we shall note that different
redistribution heuristics may imply different levels of allocation
fairness [33]. In this context, a recent work introduces the concept
of Distributed Distributed Justice [20] and shows that local inter-
actions may provide a reliable basis to build trust and reputation
between agents, which can be used to regulate, in a decentralized
way, the levels of justice in agents' actions. This way, it is reward-
ing to note that local interactions not only constitute an optimal
scale to form cooperative Beneficiary Sets (as we show, see below),
but also provide the convenient interaction environment to allow
justice in contributions to be sustained.
3 MODEL
3.1 Three Stage Redistribution Game
Here we propose a sequential game dynamics made of three stages.
Focusing on an arbitrary agent i, these stages can be described as
follows:
(1) Agent i participates in a one-shot game (here a Prisoner's
Dilemma) with all his/her neighbors j. From each interaction
j, he/she obtains a payoff πi, j. After all interactions, agent i
accumulates a total payoff Πi =
j πi, j;
(2) Next, agent i contributes a fraction α of his/her payoff sur-
plus (Πi − θ) to be redistributed. The group that benefits
from agent i contribution is called Beneficiary Set i (Bi).
(3) Finally, agent i receives his/her share from each Beneficiary
Set that he/she is part of.
We refer to α as the level of taxation, as it defines the fraction of
the surplus that agents contribute, while θ is the threshold level of
Local Wealth Redistribution Promotes Cooperation in Multiagent Systems
AAMAS'18, July 2018, Stockholm, Sweden
Figure 1: Solutions for the two-person game with wealth re-
distribution. Each curve indicates the critical taxation levels
(α∗) above which the nature of the social dilemma changes,
for different payoff thresholds (θ) and as a function of the
Temptation parameter (T ).
payoff that defines the surplus. By definition, agents with negative
payoff cannot contribute (i.e., θ > 0); they might, however, receive
benefits from the Beneficiary Sets. Each agent i contributes only to
one Beneficiary Set Bi from which they cannot be part of, that is,
agents do not receive from the Beneficiary Set they contribute to.
A central question of this work is how to select Bi for each i. As
we show, this decision has a profound and non-trivial impact on
the overall cooperation levels in the system.
3.2 The Prisoner's Dilemma Game
In general, all the possible outcomes of a two-strategy two-player
game, in which two agents engage in a one-shot interaction that
requires them to decide – independently and simultaneously –
whether they wish to Cooperate (C) or to Defect (D), can be sum-
marized in a payoff matrix, such as
C D
C R
D T
S
P
which reads as the payoff obtained by playing the row strategy
when facing an opponent with the column strategy. Here, R rep-
resents the Reward payoff for mutual cooperation and P the Pun-
ishment for mutual defection. When one of the individuals Defects
and the other Cooperates, the first receives the Temptation payoff
(T ) while the second obtains the Sucker's payoff (S). In this manu-
script we consider that agents interact according to the Prisoner's
Dilemma (PD). Agents are said to face a PD whenever the relation-
ship between the payoffs is such that T > R > P > S [44]. In such
a scenario, rational agents seeking to optimize their self-returns
are expected to always Defect. However, since the best aggregated
outcome would have both players cooperating (2R > 2P), agents are
said to face a social dilemma: optimizing self-returns clashes with
optimizing the social outcome. In this sense, mutual cooperation
is Pareto Optimal and contributes to increase both average payoff
(over mutual defection) and egalitarian social welfare (over unilat-
eral cooperation) [6]. It is noteworthy to mention that other situa-
tions – with different optimal rational responses – arise when the
Figure 2: Graphical depiction of the specific structures used
in this work. a) Homogeneous Networks correspond to a
structure in which all nodes have the same degree. b) Het-
erogeneous Networks are characterized by a high variance
among the degree of nodes. The color of each node indicates
its degree: blue tones represent lower degree and red tones
higher degree. Panel c) and d) show, respectively, the degree
distributions of the Homogeneous and Heterogeneous net-
works under analysis. In particular, we use scale-free net-
works as representatives of heterogeneous structures; these
have a degree distribution that decays as a power law.
parameters take a different relationship [21]: the Stag Hunt game
when R > T > P > S; the Snowdrift Game when T > R > S > P;
the Harmony Game when R > T > S > P; or the Deadlock Game
when T > P > R > S, to name a few. Notwithstanding, the PD is
by far the most popular metaphor of social dilemmas [44] and the
one that presents the biggest challenge for cooperation to emerge.
For these reasons, PD shall be the main focus of study in this man-
uscript. We further simplify the parameter space by considering
that R = 1, P = 0, S = 1 − T and 1 < T ≤ 2 with the game being
fully determined by the Temptation value (T). In that sense, higher
temptation creates more stringent conditions for the emergence of
cooperation.
3.3 Prisoner's Dilemma with Wealth
Redistribution
As an introductory example, let us start by analyzing the particular
case of two interacting agents (i and j) in a one-shot event. In this
case, the Beneficiary Sets of each agent (Bi and Bj) are composed
only by the opponent. Wealth/payoff redistribution can thus be
analyzed by considering a slightly modified payoff matrix, that
takes into account the second and third stages. The resulting payoff
matrix becomes
C
1
C
D T − α(T − θ)
D
1 − T + α(T − θ)
0
1.9Temptation parameter, T1.0critical level of taxation, ⍺* = 0.90 = 0.99 = 0.75 = 0.50 = 0.25 = 0.000.80.60.40.20.02.01.81.71.61.51.41.31.21.11.0a)homogeneous networkHeterogeneous Networkb)c)d)degree (z)fraction of nodesfraction of nodesdegree (z)D(z)⇡z single peakover hzi AAMAS'18, July 2018, Stockholm, Sweden
Flávio L. Pinheiro and Fernando P. Santos
where θ is the payoff threshold and α is the level of taxation. The
rationale to arrive at this payoff structure is the following: whenever
both players choose to act the same way the payoff remains the
same as their contributions (from taxes) and benefits (from receiving
the contributions of their opponent) cancel out. A Defector playing
against a Cooperator sees his payoff of T subtracted by an amount
α(T − θ) while not receiving any benefit, since the Cooperator has
negative payoff and does not contribute. Likewise, the Cooperator
is exempt from contributing but receives an additional contribution
of α(T − θ), which represents the amount taxed to the Defector.
To inspect whether wealth redistribution changes the nature of
the social dilemma (i.e. from a Prisoner's Dilemma to another type
of game) we have to inspect whether there is a difference in the
relationship between the payoffs R and T or P and S. This sums up
to solving a single inequality,
(1)
T − α(T − θ) < 1
which results in the critical values of α,
T − 1
(T − θ)
∗
α
>
(2)
Hence, depending on the choice of θ and for a given T , α∗ is the
minimum level of taxation required to observe a change in the
nature of the game faced by agents. It is straightforward to notice
that the nature of the game changes from a Prisoner's Dilemma to
an Harmony Game as the relationship moves from T > R > P > S
to R > T > S > P. Figure 1 shows α∗ for different values of T and θ.
Clearly, in well-mixed populations and under the simple scenario of
a MAS composed by two agents, the redistribution mechanism has
the simple effect of reshaping the payoff matrix, trivially changing
the nature of the dilemma. Such a trivial conclusion cannot be
drawn with large populations playing on networks, where we will
show that different ways of assigning the Beneficiary Sets have a
profound impact on the ensuing levels of cooperation.
3.4 Structured Populations
Let us consider a population of Z agents in which agents correspond
to the nodes/vertices of a complex network, while links dictate who
interacts with whom. The structure reflects the existence of con-
straints that limit interactions between agents. These constraints
can arise from spatial or communication limitations.
The number of interactions that each agent i participates in de-
fines his/her degree zi. The distribution of degrees, D(z), describes
the fraction of agents that has degree z. In this work we consider
two structures: Homogeneous Random Graphs [40, 41] and Scale-
Free Barabási Networks [3].
Homogeneous Random Graphs are generated by successively
randomizing the ends of pairs of links from an initially regular
graph (e.g. Lattice or Ring). The resulting structure has a random
interaction structure but all nodes in the network have the same
degree. Figure 2a) depicts graphically an example of such structures
and Figure 2c) the corresponding Degree distribution.
Scale-free networks are generated by an algorithm of growth
and preferential attachment [3]. This algorithm is as follows: 1)
start from three fully connected nodes; 2) add, sequentially, each
of the Z − m remaining nodes; 3) each time a new node is added,
it connects to m pre-existing nodes, selecting preferentially nodes
Figure 3: Level of Cooperation on Homogeneous Random
Networks (a) and Heterogeneous (Scale-free) Networks (b).
Each plot shows the level of cooperation under a different
combination of taxation level, α, and Temptation, T . In all
cases the fitness threshold is fixed at θ = R = 1.0. Blue indi-
cates regions where Cooperation dominates, Red delimits re-
gions dominated by Defectors. Top bars above each panel in-
dicate the level of cooperation in the absence of wealth redis-
tribution, as a function of the Temptation payoff parameter.
The level of cooperation is computed by estimating the ex-
pected fraction of cooperators when the population reaches
a stationary state. To that end we run 104 independent sim-
ulations that start with 50% cooperators and 50% defectors.
Population size of Z = 103 and intensity of selection β = 1.0.
with higher degree. Here we have used m = 3 The resulting network
is characterized by a heterogeneous degree distribution (one which
decays as a power law), in which the majority of the nodes have few
connections while a few have many. Figure 2b) shows a graphical
example of such structure and Figure 2d) the degree distribution.
In the following we explore the case of networks with Z = 103
z zD(z) = 4. During the
nodes and average degree of ⟨z⟩ =
a)1.00.90.80.60.70.50.30.40.20.00.11.01.11.21.31.41.51.61.71.81.92.0Temptation Parameter, Tlevel of taxation, ⍺ = 1.00b)1.00.90.80.60.70.50.30.40.20.00.11.01.11.21.31.41.51.61.71.81.92.0Temptation Parameter, Tlevel of taxation, ⍺ = 1.00HomogeneousHeterogeneouslevel of cooperation0.501.000.000.250.75level of cooperation0.501.000.000.250.75 Local Wealth Redistribution Promotes Cooperation in Multiagent Systems
AAMAS'18, July 2018, Stockholm, Sweden
Figure 4: Level of cooperation on Heterogeneous (a) and Ho-
mogeneous (b) populations for different values of the payoff
threshold (θ) as a function of the Temptation payoff param-
eter (T ). Gray Dashed line shows the results obtained in the
absence of a wealth redistribution scheme. Population size
of Z = 103 and intensity of selection β = 1.0.
simulations we make use of 20 independently generated networks
of each type.
3.5 Games on Networks
We study the expected level of cooperation attained by the popu-
lation. We estimate this quantity through computer simulations.
The level of cooperation corresponds to the expected fraction of
cooperators in a population that evolved after 2.5 × 106 iterations.
We estimate this quantity by averaging the observed fraction of
cooperators at the final of each simulation, over 104 independent
simulations.
Each simulation starts from a population with an equal compo-
sition of Cooperators and Defectors, which are randomly placed
along the nodes of the network. In between each update round,
each agent i plays once with all his/her zi nearest neighbors (i.e.,
agents they are directly connected with). The accumulated payoff
over all interactions an agent i participates can be computed as
i T − σi(1 − T)(nD
i )
i + nC
Πi = nC
(3)
where nD
(nC
i ) is the number of neighbors of i that Defect (Co-
i
operate) and σi is equal to 1 if i is a Cooperator and 0 otherwise.
Figure 5: Comparison between the effects of assigning the
nearest neighbors of an agent i to the corresponding Bene-
ficiary Set Bi (dark blue line) and when agents are assigned
at random to Bi (light blue), on the level of cooperation in
the domain of the Temptation payoff parameter, T. Panel a)
shows the results on Heterogeneous populations and panel
b) the impact on Homogeneous populations. Population size
of Z = 103 and intensity of selection β = 1.0.
From the accumulated payoff, agents contribute to a pool a fraction
α of the surplus Π − θ. The fitness fi of an agent i results from
subtracting from his/her accumulated payoff his/her contributions
plus the share he/she obtains from each of the Beneficiary Sets j
he/she participates in. We shall underline that, while T is the same
for all agents (that is, the dilemma is the same for everyone in the
population), heterogeneous populations introduce an additional
complexity layer by implying that different agents may vary in the
maximum values of accumulated payoff that they are able to earn.
This can be formalized as
fi = (1 − α)(Πi − θ) +
δi, j
Z
j
α(Πj − θ)
Bj
(4)
where δi, j is equal to one if i is part of the Beneficiary Set towards
which j contributes and zero otherwise, while Bj denotes the size
of set Bj.
Evolution in the frequency of strategies adopted in the popula-
tion happens through a process of imitation or social learning. At
each iteration a random agent, say i, compares his fitness with the
a)b)1.11.21.31.41.51.61.71.81.92.0Temptation Parameter, T1.01.11.21.31.41.51.61.71.81.92.0Temptation Parameter, T1.01.00.80.60.40.20.0Level of Cooperation1.00.80.60.40.20.0Level of CooperationHomogeneous[⍺=0.8]Heterogeneous[⍺=0.5]Legend = 0.0 = 0.4 = 0.8 = 1.2 = 1.6 = 2.0No WealthRedistributiona)1.11.21.31.41.51.61.71.81.92.0Temptation parameter, T1.01.00.80.60.40.20.0level of cooperationNearest Neighbors(d = 1)Random Groupb)1.11.21.31.41.51.61.71.81.92.0Temptation parameter, T1.01.00.80.60.40.20.0level of cooperationHeterogeneous[ = 0.80, ⍺ = 0.50]Homogeneous[ = 1.00, ⍺ = 0.90]Nearest Neighbors(d = 1)Random Group AAMAS'18, July 2018, Stockholm, Sweden
Flávio L. Pinheiro and Fernando P. Santos
fitness of a neighbor, say j. Depending on the fitness difference, i
adopts the strategy of j with probability
p =
1
1 + Exp(−β(fj − fi))
(5)
The meaning of this sigmoid function can be understood as follows:
if j is performing much better than i, then i updates his/her strategy,
adopting the strategy of j. Conversely, if j is performing much worse,
i does not update the strategy. The parameter β, often called the
intensity of selection and akin to a learning rate, dictates how sharp
is the transition between these two regimes, as fj − fi approaches
zero. Large β means that individuals act in a more deterministic
way, updating strategies at the minimum difference; small β means
that individuals are prone to make imitation mistakes.
4 RESULTS
4.1 Wealth Redistribution and the Level of
Cooperation in Structured Populations
In this section we start by analyzing the scenario in which the Bene-
ficiary Set of each agent i corresponds to his/her nearest neighbors.
Hence, the size of the Beneficiary set of i is Bi = zi. These are
also the agents from whom he/she interacts with and obtains a
payoff from. Figure 3 shows the achieved levels of cooperation
when the payoff threshold is set to θ = R = 1.0, as a function of the
Temptation payoff (T ) and the level of taxation (α). Figure 3a shows
the results on Homogeneous networks, and Figure 3b on Hetero-
geneous. We find that, for a fixed payoff threshold (θ), increasing
the level of taxation results in an increase in the levels of cooper-
ation. This effect diminishes with an increase in the Temptation
(T ). That is, when increasing T the minimum value of α necessary
to promote cooperation increases as well. The same behavior is
observed in both structures. However, there is a larger degree of
cooperation on Heterogeneous networks, where there is always a
level of taxation for a given Temptation that guarantees a 100% level
of cooperation. Hence, in order for cooperation to be evolutionary
viable on homogeneous networks, more stringent conditions are
necessary, e.g. higher tax levels.
Figure 4 shows how the level of cooperation depends on varia-
tions of the fitness threshold (0 ≤ θ ≤ 2.0, in intervals of 0.4) while
keeping a fixed level of taxation (α = 0.5) under different levels
of the Temptation payoff (T ). Figure 4a shows the results obtained
for Heterogeneous networks and panel b) the results on Homoge-
neous structures. For a constant level of taxation, α, decreasing the
payoff threshold, θ, increases the range of Temptation, T , under
which cooperation can possibly evolve. This is the case in both
types of structures. However, once again, the effect is more limited
in homogeneous populations.
Both Figure 3 and 4 highlight the positive impact of a local wealth
redistribution mechanism in the enhancement of cooperation. It
also puts in evidence that the success of such mechanism depends
on the volume of payoff that is redistributed. Ultimately, this can
be done by either increasing the level of taxation, α or decreasing
the payoff threshold, θ, that defines the taxable payoff.
Figure 6: Panel a) compares how extending beneficiary sets,
from the nearest neighbors (d = 1) to nodes at a distance up
to d = 4 links away, impacts the level of cooperation on Het-
erogeneous networks. Panel b) shows how extended benefi-
ciary sets impact the level of cooperation on Homogeneous
networks. In both cases extending the set of beneficiaries
has a negative a negative impact in the levels of cooperation.
Population size of Z = 103 and intensity of selection β = 1.0.
4.2 Randomized Beneficiary Set
Next we explore to which extent the results obtained depend on the
way agents are being assigned to each Beneficiary Set. To that end,
we compare two cases: i) nearest set assignment – the Beneficiary
Set of each agent corresponds to her/his nearest neighbors, as above;
and ii) random set assignment – agents are assigned at random
to each Beneficiary Set. The number of agents assigned to each
set is equal to the degree of the contributing agent, in both cases,
which guarantees that the collected payoffs from each agent are
distributed among the same number of individuals in both i) and
ii).
Figure 5a and b show the results obtained, respectively, on Het-
erogeneous and Homogeneous networks. We consider θ = 0.5,
α = 0.9 and explore the domain 1.0 ≤ T ≤ 2. Dark blue curves
show the results obtained under the nearest set assignment and
light blue curves the results obtained under a random set assign-
ment. The results show that the ability of a wealth redistribution
mechanism lies in the redistribution of the taxed payoff among the
agents that are spatially related. A random assignment of agents
a)1.11.21.31.41.51.61.71.81.92.0Temptation parameter, T1.01.00.80.60.40.20.0level of cooperationNearest Neighbors(d = 1)d=4b)1.11.21.31.41.51.61.71.81.92.0Temptation parameter, T1.01.00.80.60.40.20.0level of cooperationLegendd = 1d = 2d = 3d = 4Homogeneous[ = 1.0, ⍺ = 0.90]Heterogeneous[ = 0.8, ⍺ = 0.50]Legendd = 1d = 2d = 3d = 4Nearest Neighbors(d = 1)d = 2d = 3d = 2 Local Wealth Redistribution Promotes Cooperation in Multiagent Systems
AAMAS'18, July 2018, Stockholm, Sweden
conclusion that wealth redistribution is only efficient when agents
return, in form of taxes, a share of the accumulated payoffs to the
agents they have engaged with. We shall underline that here both
distance and size of Bi play a role on the obtained results, while
in the previous section the size of Bi was kept constant for each i
across the different treatments, thus disambiguating the effect of
Bi size and distance on the resulting cooperation levels.
4.4 What is the cost of wealth redistribution?
Figure 7a and b shows the fixation times of populations when θ =
1.0 along the domain bounded by 0.0 ≤ α ≤ 1.0 and 1.0 ≤ T ≤ 2.0.
The fixation times correspond to the expected number of gener-
ations (i.e., sets of Z potential imitation steps) for the population
to reach a state in which only one strategy is present in the popu-
lation. These plots map directly into Figure 3a and b, allowing to
compare the relative fixation times of regions with high/low levels
of cooperation.
We observe that the evolution of cooperation is associated with
an increase in the fixation times. This increase can be in some
situations an order of magnitude higher. The regions that exhibit
larger fixation times lie in the critical boundary that divides areas of
defectors and cooperators dominance (Figure 3). Hence, promoting
cooperation by redistributing wealth also requires a longer waiting
time for the population to reach a state of full cooperation. However,
setting higher taxation values than the bare minimum necessary for
the emergence of cooperation allows populations to reach fixation
quicker.
Figure 7: Panel a) shows the fixation times (in generations)
on homogeneous networks. Panel b) shows the fixation
times (in generations) in heterogeneous networks. A gener-
ation corresponds to Z iteration steps and the fixation times
indicate the expected time that the population takes to ar-
rive to a state dominated by Cooperators or Defectors when
starting from a state with equal abundance of both strate-
gies. Population size of Z = 103 and intensity of selection
β = 1.0.
drastically decreases the levels of cooperation obtained in both
networks. But to which extent do the Beneficiary Sets need to be
constrained spatially?
4.3 Extended Beneficiary Set
To answer the previous question, we explore the case in which all
nodes (up to a distance of d links) are assigned to the Beneficiary
Set of a focal agent i; when d = 1 the previous results are thereby
obtained.
Figure 6a and b show the results up to d = 4 on Heterogeneous
and Homogeneous networks respectively. In both cases, we see
that an expansion in the size of the Beneficiary Set leads to a de-
crease in the levels of cooperation. This result further reinforces the
4.5 Multiple Contribution Brackets
In the real world, taxes are unlikely to be defined by a single thresh-
old (θ) that separates agents who contribute from those that do not.
In reality taxes are progressive, in the sense that taxation levels
(α) increase with increasing level of income (in this case accumu-
lated payoff). In this section we implement a similar approach and
inspect the impact of increasing the number of taxation brackets.
Let us consider that, instead of a single threshold we now have B
taxation brackets divided by B−1 threshold levels. For each bracket
we define αb as the effective tax and θb as the bottom threshold of
bracket b, where b ∈ {0, 1, 2, ..., B − 1, B}.
By definition B = 0 corresponds to the case in which no taxes
are collected, and the redistribution of wealth is absent. Moreover,
B = 1, implies the existence of a single bracket were all individuals
would contribute, a case that we do not explore in this manuscript.
B = 2, corresponds to the case in which there are two brackets,
which is the scenario that we have explored until now.
We consider the case in which taxation increases linearly with
increasing brackets. Let us define θb = bθ/B. Individuals in bracket
b have their payoff surplus taxed by αb = (b − 1)α/B when their
accumulated payoff falls into θb < Π ≤ θb+1 for b < B − 1. For
b = B the tax level is αb = α and affects all individuals with Π > θ.
As an example, for B = 3 each bracket would be characterized
by the following tax levels
b = 0) αb = 0 for all individuals with Π ≤ θ/3;
b = 1) αb = α/3 for all individuals with θ/3 < Π ≤ 2θ/3;
b = 2) αb = 2α/3 for all individuals with 2θ/3 < Π ≤ θ;
b = 3) αb = α for all individuals with Π > θ.
a)1.00.90.80.60.70.50.30.40.20.00.11.01.11.21.31.41.51.61.71.81.92.0Temptation Parameter, Tlevel of taxation, α =1.0b)1.00.90.80.60.70.50.30.40.20.00.1Temptation Parameter, Tlevel of taxation, α = 1.0HomogeneousHeterogeneousnumber of generations1.01.11.21.31.41.51.61.71.81.92.03161003210number of generations10003161001032 AAMAS'18, July 2018, Stockholm, Sweden
Flávio L. Pinheiro and Fernando P. Santos
5 CONCLUSION
To sum up, we show that wealth redistribution embodies an effective
mechanism that significantly helps cooperation to evolve. It works
by fundamentally changing the nature of the dilemma at stake: by
appropriately choosing the level of taxation (α) and payoff threshold
(θ) it is possible to shift from a Defector dominance to a Cooperator
dominance dynamics. Moreover, we find that in Heterogeneous
populations allow us to ease the redistribution mechanism – that is,
imposing lower taxation rates and/or lower taxable surplus values
when compared with Homogeneously structured populations.
Additionally, we show, for the first time, that different assign-
ments of Beneficiary Sets significantly impact the ensuing levels of
cooperation. Local Beneficiary Sets, where agents receive the con-
tributions from their direct neighbors, constitute a judicious choice
when compared with Beneficiary Sets that are formed by 1) agents
randomly picked from the population or 2) by including agents at
higher distances. Naturally, a Local wealth redistribution scheme
may not only prove optimal in terms of achieved cooperation levels,
but also reveal much simpler to implement, by exempting the need
of central redistribution entities and by minimizing the number of
peers that agents need to interact with. We shall highlight, how-
ever, that promoting cooperation through a wealth redistribution
mechanism bears longer fixation times, in terms of the number of
iterations required to achieve overall cooperation.
Here we assume that the redistribution mechanism is externally
imposed. Agents are not able to opt out from the taxation scheme.
Given that this mechanism increases the overall cooperation and
average payoff in the system, an argument for its acceptance - by
rational agents - can be formulated based on the infamous veil of
ignorance proposed by John Rawls [35]: Agents should decide the
kind of society they would like to live in without knowing their
social position. Agents would, this way, prefer a cooperative soci-
ety where redistribution exists, provided that here average payoff
is maximized. Notwithstanding, future research shall analyze the
role of more complex strategies that give opportunity of agents to
voluntarily engage (or not) in the proposed redistribution scheme.
Alongside, effective mechanisms that discourage the second or-
der free riding problem (i.e., free riding by not contributing to the
redistribution pot, while expecting others to do so) shall be exam-
ined. Future works shall also evaluate whether alternative taxation
schemes are prone to be more efficient than the one proposed here.
In all these cases, an evolutionary game theoretic framework –
such as the one here developed – constitutes a promising toolkit to
employ.
6 ACKNOWLEDGMENTS
The authors acknowledge the useful discussions with Francisco C.
Santos, Jorge M. Pacheco and Aamena Alshamsi. F.L.P. is thankful
to the Media Lab Consortium for financial support. F.P.S. acknowl-
edges the financial support of Fundação para a Ciência e Tecnologia
(FCT) through PhD scholarship SFRH/BD/94736/2013, multi-annual
funding of INESC-ID (UID/CEC/50021/2013) and grants PTDC/EEI-
SII/5081/2014, PTDC/MAT/STA/3358/2014.
Figure 8: Relative wealth inequality after the redistribution
step, in a heterogeneous population dominated by coopera-
tors and for different combinations of taxation level (α) and
threshold (θ). We quantify the relative wealth inequality af-
ter the redistribution step as the ratio between the variance
of the fitness distribution (V arf , i.e. variance in gains across
the population after redistribution) and the variance of the
accumulated payoff distribution (V arp, i.e. variance in gains
before redistribution). Population size of Z = 103 and inten-
sity of selection β = 1.0.
In this way we use θ and α as the upper level bound and only
parameters in this condition.
We find that variations in the number of taxation brackets (B=3,4,5)
have only a marginal impact in the overall levels of cooperation
observed when compared with the scenarios studied so far (B=2).
4.6 Wealth Inequality
Finally, we discuss the effect of wealth redistribution on fitness
inequality. First, it is important to highlight that the observed levels
of inequality depend, by default, on the distribution of strategies and
network degree. In homogeneous structures, if every agent adopts
the same strategy – either Defectors or Cooperators – everyone
obtains the same fitness. In heterogeneous structures, a Cooperation
dominance scenario bounds the feasible equality levels, given the
degree distribution of the population. In fact, some agents engage
in more interactions than others and Beneficiary Sets have different
sizes, depending on the particular connectivity of agents. We shall
focus on this scenario.
We compare the variance of fitness (i.e. gains after the redistribu-
tion step) and the variance of accumulated payoff (i.e. gains before
the redistribution step) in order to quantify the relative inequality
after we apply the proposed redistribution mechanism. In particular,
we use the ratio between the variance of fitness and the variance
of accumulated payoff as a metric of resulting wealth inequality.
Figure 8 shows how higher levels of θ and α reduce the resulting
inequality. In fact, while increasing payoff threshold limits taxation
to the richer agents, increasing level of taxation increases the flow
of fitness from rich agents to their Beneficiary Sets. In the most
strict case – high θ and α – the variance of the fitness distribution
is reduced to as low as 7% of the accumulated payoff distribution.
���������������level of taxation, ⍺payoff threshold, 2.00.01.02.03.04.00.10.51.52.53.50.20.30.40.50.60.70.80.91.0Varf / Varpdecreasing inequality Local Wealth Redistribution Promotes Cooperation in Multiagent Systems
AAMAS'18, July 2018, Stockholm, Sweden
International Foundation for Autonomous Agents and Multiagent Systems, 669–
676.
[30] Flávio L. Pinheiro and Dominik Hartmann. 2017. Intermediate Levels of Network
Heterogeneity Provide the Best Evolutionary Outcomes. Scientific Reports 7, 1
(2017), 15242.
[31] Flavio L Pinheiro, Jorge M Pacheco, and Francisco C Santos. 2012. From local to
global dilemmas in social networks. PloS ONE 7, 2 (2012), e32114.
[32] Flávio L Pinheiro, Francisco C Santos, and Jorge M Pacheco. 2016. Linking
individual and collective behavior in adaptive social networks. Physical Review
Letters 116, 12 (2016), 128702.
[33] Jeremy Pitt, Julia Schaumeier, Didac Busquets, and Sam Macbeth. 2012. Self-
organising common-pool resource allocation and canons of distributive justice.
In Self-Adaptive and Self-Organizing Systems (SASO), 2012 IEEE Sixth International
Conference on. IEEE, 119–128.
[34] Bijan Ranjbar-Sahraei, Haitham Bou Ammar, Daan Bloembergen, Karl Tuyls,
and Gerhard Weiss. 2014. Theory of cooperation in complex social networks. In
Proceedings of AAAI'14. AAAI Press.
[35] John Rawls. 2009. A theory of justice. Harvard University Press.
[36] Luke Rendell, Robert Boyd, Daniel Cownden, Marquist Enquist, Kimmo Eriksson,
Marc W Feldman, Laurel Fogarty, Stefano Ghirlanda, Timothy Lillicrap, and
Kevin N Laland. 2010. Why copy others? Insights from the social learning
strategies tournament. Science 328, 5975 (2010), 208–213.
[37] Norman Salazar, Juan A Rodriguez-Aguilar, Josep Ll Arcos, Ana Peleteiro, and
Juan C Burguillo-Rial. 2011. Emerging cooperation on complex networks. In
Proceedings of the 2011 International Conference on Autonomous Agents and Multi-
agent Systems. International Foundation for Autonomous Agents and Multiagent
Systems, 669–676.
[38] Francisco C Santos and Jorge M Pacheco. 2005. Scale-free networks provide a
unifying framework for the emergence of cooperation. Physical Review Letters
95, 9 (2005), 098104.
[39] Francisco C Santos, Jorge M Pacheco, and Tom Lenaerts. 2006. Cooperation
prevails when individuals adjust their social ties. PLoS Computational Biology 2,
10 (2006), e140.
[40] Francisco C Santos, JF Rodrigues, and Jorge M Pacheco. 2005. Epidemic spreading
and cooperation dynamics on homogeneous small-world networks. Physical
Review E 72, 5 (2005), 056128.
[41] Fernando P Santos, Jorge M Pacheco, Ana Paiva, and Francisco C Santos. 2017.
Structural power and the evolution of collective fairness in social networks. PloS
ONE 12, 4 (2017), e0175687.
[42] Fernando P. Santos, Jorge M. Pacheco, and Francisco C. Santos. 2018. Social
norms of cooperation with costly reputation building. In AAAI'18. AAAI Press.
[43] Sven Seuken, Jie Tang, and David C Parkes. 2010. Accounting Mechanisms for
Distributed Work Systems. In AAAI'10. AAAI Press.
[44] Karl Sigmund. 2010. The calculus of selfishness. Princeton University Press.
[45] Robert L Trivers. 1971. The evolution of reciprocal altruism. The Quarterly Review
of Biology 46, 1 (1971), 35–57.
[46] Vítor V Vasconcelos, Francisco C Santos, and Jorge M Pacheco. 2015. Cooperation
dynamics of polycentric climate governance. Mathematical Models and Methods
in Applied Sciences 25, 13 (2015), 2503–2517.
[47] Markus Waibel, Dario Floreano, and Laurent Keller. 2011. A quantitative test of
Hamilton's rule for the evolution of altruism. PLoS Biology 9, 5 (2011), e1000615.
REFERENCES
[1] Stéphane Airiau, Sandip Sen, and Daniel Villatoro. 2014. Emergence of conven-
tions through social learning. Autonomous Agents and Multi-Agent Systems 28, 5
(2014), 779–804.
[2] Ian F Akyildiz, Weilian Su, Yogesh Sankarasubramaniam, and Erdal Cayirci. 2002.
Wireless sensor networks: a survey. Computer Networks 38, 4 (2002), 393–422.
[3] Réka Albert and Albert-László Barabási. 2002. Statistical mechanics of complex
networks. Reviews of Modern Physics 74, 1 (2002), 47.
[4] Josep Ll Arcos, Marc Esteva, Pablo Noriega, Juan A Rodríguez-Aguilar, and
Carles Sierra. 2005. Engineering open environments with electronic institutions.
Engineering Applications of Artificial Intelligence 18, 2 (2005), 191–204.
[5] Juan C Burguillo-Rial. 2009. A memetic framework for describing and simulating
spatial prisoner's dilemma with coalition formation. In Proceedings of AAAI'09.
AAAI Press, 441–448.
[6] Ulle Endriss and Nicolas Maudet. 2003. Welfare engineering in multiagent
systems. In International Workshop on Engineering Societies in the Agents World.
Springer, 93–106.
[7] Eithan Ephrati and Jeffrey S Rosenschein. 1996. Deriving consensus in multiagent
systems. Artificial Intelligence 87, 1-2 (1996), 21–74.
[8] Marc Esteva, Bruno Rosell, Juan A Rodriguez-Aguilar, and Josep Ll Arcos. 2004.
AMELI: An agent-based middleware for electronic institutions. In Proceedings of
the Third International Joint Conference on Autonomous Agents and Multiagent
Systems. IEEE Computer Society, 236–243.
[9] Michal Feldman and John Chuang. 2005. Overcoming free-riding behavior in
peer-to-peer systems. ACM SIGecom Exchanges 5, 4 (2005), 41–50.
[10] Michael R Genesereth, Matthew L Ginsberg, and Jeffrey S Rosenschein. 1986.
Cooperation without communication. In Proceedings of AAAI'86. AAAI Press.
[11] Philippe Golle, Kevin Leyton-Brown, Ilya Mironov, and Mark Lillibridge. 2001.
Incentives for sharing in peer-to-peer networks. In Electronic Commerce. Springer,
75–87.
[12] Nathan Griffiths. 2008. Tags and image scoring for robust cooperation. In Proceed-
ings of the 2008 International Conference on Autonomous Agents and Multi-agent
Systems. International Foundation for Autonomous Agents and Multiagent Sys-
tems, 575–582.
[13] William D Hamilton. 1964. The genetical evolution of social behaviour. Journal
of Theoretical Biology 7, 1 (1964), 17–52.
[14] TA Han. 2016. Emergence of Social Punishment and Cooperation through Prior
Commitments. In Proceedings of AAAI'16. AAAI Press, 2494–2500.
[15] TA Han, Luís Moniz Pereira, Luis A Martinez-Vaquero, and Tom Lenaerts. 2017.
Centralized vs. Personalized Commitments and their influence on Cooperation
in Group Interactions. In Proceedings of AAAI'17. AAAI Press.
[16] Chien-Ju Ho, Yu Zhang, Jennifer Vaughan, and Mihaela Van Der Schaar. 2012.
Towards social norm design for crowdsourcing markets. In AAAI'12 Technical
Report WS-12-08. AAAI Press.
[17] Lisa-Maria Hofmann, Nilanjan Chakraborty, and Katia Sycara. 2011. The evolu-
tion of cooperation in self-interested agent societies: a critical study. In Proceed-
ings of the 2011 International Conference on Autonomous Agents and Multi-agent
Systems. International Foundation for Autonomous Agents and Multiagent Sys-
tems, 685–692.
[18] Genki Ichinose, Yoshiki Satotani, and Hiroki Sayama. 2017. How mutation
alters fitness of cooperation in networked evolutionary games. arXiv preprint
arXiv:1706.03013 (2017).
[19] Nicholas R Jennings, Katia Sycara, and Michael Wooldridge. 1998. A roadmap of
agent research and development. Autonomous Agents and Multi-agent Systems 1,
1 (1998), 7–38.
[20] David Burth Kurka and Jeremy Pitt. 2016. Distributed distributive justice. In
Self-Adaptive and Self-Organizing Systems (SASO), 2016 IEEE 10th International
Conference on. IEEE, 80–89.
[21] Michael W Macy and Andreas Flache. 2002. Learning dynamics in social dilemmas.
Proceedings of the National Academy of Sciences 99 (2002), 7229–7236.
[22] Martin A Nowak. 2006. Five rules for the evolution of cooperation. Science 314,
[23] Martin A Nowak. 2012. Evolving cooperation. Journal of Theoretical Biology 299
5805 (2006), 1560–1563.
(2012), 1–8.
Nature (2005).
[24] Martin A Nowak and Robert M May. 1992. Evolutionary games and spatial chaos.
Nature 359, 6398 (1992), 826–829.
[25] Martin A Nowak and Karl Sigmund. 2005. Evolution of indirect reciprocity.
[26] Hisashi Ohtsuki, Christoph Hauert, Erez Lieberman, and Martin A Nowak. 2006.
A simple rule for the evolution of cooperation on graphs. Nature 441, 7092 (2006),
502.
[27] Elinor Ostrom. 2015. Governing the commons. Cambridge University Press.
[28] Liviu Panait and Sean Luke. 2005. Cooperative multi-agent learning: The state of
the art. Autonomous Agents and Multi-agent Systems 11, 3 (2005), 387–434.
[29] Ana Peleteiro, Juan C Burguillo, and Siang Yew Chong. 2014. Exploring indirect
reciprocity in complex networks using coalitions and rewiring. In Proceedings of
the 2014 International Conference on Autonomous Agents and Multi-agent Systems.
|
1002.0169 | 1 | 1002 | 2010-02-01T01:38:32 | Moment-Based Analysis of Synchronization in Small-World Networks of Oscillators | [
"cs.MA",
"cs.CE",
"cs.DM",
"nlin.AO"
] | In this paper, we investigate synchronization in a small-world network of coupled nonlinear oscillators. This network is constructed by introducing random shortcuts in a nearest-neighbors ring. The local stability of the synchronous state is closely related with the support of the eigenvalue distribution of the Laplacian matrix of the network. We introduce, for the first time, analytical expressions for the first three moments of the eigenvalue distribution of the Laplacian matrix as a function of the probability of shortcuts and the connectivity of the underlying nearest-neighbor coupled ring. We apply these expressions to estimate the spectral support of the Laplacian matrix in order to predict synchronization in small-world networks. We verify the efficiency of our predictions with numerical simulations. | cs.MA | cs | Moment-Based Analysis of Synchronization in Small-World Networks
of Oscillators
Victor M. Preciado and Ali Jadbabaie
0
1
0
2
b
e
F
1
]
A
M
.
s
c
[
1
v
9
6
1
0
.
2
0
0
1
:
v
i
X
r
a
Abstract- In this paper, we investigate synchronization in
a small-world network of coupled nonlinear oscillators. This
network is constructed by introducing random shortcuts in a
nearest-neighbors ring. The local stability of the synchronous
state is closely related with the support of the eigenvalue distri-
bution of the Laplacian matrix of the network. We introduce,
for the first time, analytical expressions for the first three
moments of the eigenvalue distribution of the Laplacian matrix
as a function of the probability of shortcuts and the connectivity
of the underlying nearest-neighbor coupled ring. We apply these
expressions to estimate the spectral support of the Laplacian
matrix in order to predict synchronization in small-world
networks. We verify the efficiency of our predictions with
numerical simulations.
I. INTRODUCTION
In recent years, systems of dynamical nodes intercon-
nected through a complex network have attracted a good deal
of attention [20]. Biological and chemical networks, neural
networks, social and economic networks [9], the power grid,
the Internet and the World Wide Web [8] are examples of
the wide range of applications that motivate this interest
(see also [15], [4] and references therein). Several modeling
approaches can be found in the literature [8], [22], [1]. In
this paper, we focus our attention on the so-called small-
world phenomenon and a model proposed by Newman and
Strogatz to replicate this phenomenon.
Once the network is modeled, one is usually interested in
two types of problems. The first involves structural prop-
erties of the model. The second involves the performance
of dynamical processes run on those networks. In the latter
direction, the performance of random walks [12], Markov
processes [6], gossip algorithms [5], consensus in a network
of agents [16], [10], or synchronization of oscillators [21],
[17], are very well reported in the literature. These dynamical
processes are mostly studied in the traditional context of
deterministic networks of relatively small size and/or regular
structure. Even though many noteworthy results have been
achieved for large-scale probabilistic networks [13]– [2],
there is substantial reliance on numerical simulations.
The eigenvalue spectrum of an undirected graph contains
a great deal of information about structural and dynamical
properties [7]. In particular, we focus our attention on the
spectrum of the (combinatorial) Laplacian matrix uniquely
associated with an undirected graph [3]. This spectrum
This work was supported by ONR MURI N000140810747, and AFOR's
The
authors
complex networks program.
and Sys-
tems Engineering, University of Pennsylvania, 3451 Walnut Street,
{preciado,jadbabai}@seas.upenn.edu
are with the Department of Electrical
contains useful information about, for example, the number
of spanning trees, or the stability of synchronization of a
network of oscillators. We analyze the low-order moments
of the Kirchhoff matrix spectrum corresponding to small-
world networks.
The paper is organized as follows. In Section II, we
review the master stability function approach. In Section
III, we derive closed-form expressions for the low-order
moments of the Laplacian eigenvalue distribution associated
with a probabilistic small-world network. Our expressions
are valid for networks of asymptotically large size. Section
IV applies our results to the problem of synchronization
of a probabilistic small-world network of oscillators. The
numerical results in this section corroborate our predictions.
II. SYNCHRONIZATION OF NONLINEAR OSCILLATORS
In this section we review the master-stability-function
(MSF) approach, proposed by Pecora and Carrol in [17],
to study local stability of synchronization in networks of
nonlinear oscillators. Using this approach, we reduce the
problem of studying local stability of synchronization to the
algebraic problem of studying the spectral support of the
Laplacian matrix of the network. First, we introduce some
needed graph-theoretical background.
A. Spectral Graph Theory Background
In the case of a network with symmetrical connections,
undirected graphs provide a proper description of the net-
work topology. An undirected graph G consists of a set of N
nodes or vertices, denoted by V = {v1, ...,vn}, and a set of
edges E, where E ∈ V ×V . In our case, (vi,v j) ∈ E implies
(v j,vi) ∈ E, and this pair corresponds to a single edge with no
direction; the vertices vi and v j are called adjacent vertices
(denoted by vi ∼ v j) and are incident to the edge (vi,v j). We
only consider simple graphs (i.e., undirected graphs that have
no self-loops, so vi 6= v j for an edge (vi,v j), and no more
than one edge between any two different vertices). A walk
on G of length k from v0 to vk is an ordered set of vertices
(v0,v1, ...,vk) such that (vi,vi+1) ∈ E, for i = 0,1, ...,k − 1; if
n k = n 0 the walk is said to be closed.
The degree di of a vertex vi is the number of edges incident
to it. The degree sequence of G is the list of degrees, usually
given in non-increasing order. The clustering coefficient,
introduced in [22], is a measure of the number of triangles in
a given graph, where a triangle is defined by the set of edges
{(i, j) , ( j,k) , (k,i)} such that i ∼ j ∼ k ∼ i. Specifically, we
define clustering as the total number of triangles in a graph,
T (G) , divided by the number of triangles in a complete (all-
to-all) graph with N vertices, i.e., the coefficient is equal to
T (G).(cid:0)N
3(cid:1) .
It is often convenient to represent graphs via matrices.
There are several choices for such a representation. For
example, the adjacency matrix of an undirected graph G,
denoted by A(G) = [ai j], is defined entry-wise by ai j = 1
if nodes i and j are adjacent, and ai j = 0 otherwise. (Note
that aii = 0 for simple graphs.) Notice also that the degree
di can be written as di = (cid:229) N
j=1 ai j. We can arrange the
degrees on the diagonal of a diagonal matrix to yield the
degree matrix, D = diag (di). The Laplacian matrix (also
called Kirchhoff matrix, or combinatorial Laplacian matrix)
is defined in terms of the degree and adjacency matrices
as L(G) = D(G) − A(G). For undirected graphs, L(G) is a
symmetric positive semidefinite matrix [3]. Consequently, it
has a full set of N real and orthogonal eigenvectors with real
non-negative eigenvalues. Since all rows of L sum to zero, it
always admits a trivial eigenvalue l 1 = 0, with corresponding
eigenvector v1 = (1,1, ...,1)T .
The moments of the Laplacian eigenvalue spectrum are
central to our paper. Denote the eigenvalues of our N × N
symmetric Laplacian matrix L(G) by 0 = l 1 (G) ≤ ... ≤
l N (G). The empirical spectral density (ESD) of L(G) is
defined as
m G (l ) =
d (l − l i) ,
1
N
N(cid:229)
i=1
where d (·) is the Dirac delta function. The k-th order moment
of the ESD of L(G) is defined as:
N(cid:229)
l i(G)k
qk(G) =
1
N
i=1
(which is also called the k-th order spectral moment1).
In the following subsection, we illustrate how a network
of identical nonlinear oscillators synchronizes whenever the
Laplacian spectrum is contained in a certain region on
the real line. This region of synchronization is exclusively
defined by the dynamics of each isolated oscillator and the
type of coupling [17], [11]. This simplifies the problem of
synchronization to the problem of locating the Laplacian
eigenvalue spectrum.
B. Synchronization as a Spectral Graph Problem
Several techniques have been proposed to analyze the
synchronization of coupled identical oscillators. In [23],
well-known results in control theory, such as the passivity
criterion, the circle criterion, and a result on observer design
are used to derive synchronization criteria for an array
of identical nonlinear systems. In [19],
the authors use
contraction theory to derive sufficient conditions for global
synchronization in a network of nonlinear oscillators. We
pay special attention to the master-stability-function (MSF)
1Given that our interest is in networks of growing size (i.e., number of
nodes N), a more explicit notation for m and qk would perhaps have been
m (N) and q(k)
k . However, for notational simplicity, we shall omit reference
to N in there and other quantities in this paper.
approach, [17]. This approach provides us with a criterion
for local stability of synchronization based on the numerical
computation of Lyapunov exponents. Even though quite
different in nature, the mentioned techniques emphasize the
key role played by the graph eigenvalue spectrum.
In this paper we consider a time-invariant network of
N identical oscillators, one located at each node, linked
with 'diffusive' coupling. The state equations modeling the
dynamics of the network are
xi = f (xi) + g
N(cid:229)
j=1
ai jG (x j − xi) , i = 1, ...,N
(1)
where xi represents an n-dimensional state vector corre-
sponding to the i-th oscillator. The nonlinear function f (·)
describes the (identical) dynamics of the isolated nodes. The
positive scalar g can be interpreted as a global coupling
strength parameter. The n × n matrix G
represents how
states in neighboring oscillators couple linearly, and ai j are
the entries of the adjacency matrix. By simple algebraic
manipulations, one can write down Eq. (1) in terms of the
Laplacian entries, L(G) = [li j], as
xi = f (xi) − g
N(cid:229)
j=1
li jG x j, for i = 1, ...,N.
(2)
We say that the network of oscillators is at a synchronous
equilibrium if x1(t) = x2(t) = ... = xN(t) = f (t), where
f (t) represents a solution for x = f (x). In [17], the authors
studied the local stability of the synchronous equilibrium.
Specifically, they considered a sufficiently small perturbation,
denoted by e i(t), from the synchronous equilibrium, i.e.,
xi(t) = f (t) + e i(t).
After appropriate linearization, one can derive the following
equations to approximately describe the evolution of the
perturbations:
e i = Df (t) e i(t) − g
n(cid:229)
j=1
li, jG
e j(t), for i = 1, ...,N.
(3)
where Df (t) is the Jacobian of f (·) evaluated along the
trajectory f (t). This Jacobian is an n × n matrix with time-
variant entries. Following the methodology introduced in
[17], Eq. (3) can be similarity transformed into a set of linear
time-variant (LTV) ODEs of the form:
i (G)) G
x i= [Df (t) + (gl
]x i, for i = 1, ...,N,
(4)
where {l i (G)}1≤i≤N is the set of eigenvalues of L (G). Based
on the stability analysis presented in [17], the network of
oscillators in (1) presents a locally stable synchronous equi-
librium if the corresponding maximal nontrivial Lyapunov
exponents of (4) is negative for i = 2, ...,N.
Inspired in Eq. (4), Pecora and Carroll studied in [17]
the stability of the following parametric LTV-ODE in the
parameter s :
x = [Df (t) + s
]x ,
(5)
G
where Df (t) is the linear time-variant Jacobian in Eq. (3).
The master stability function (MSF), denoted by F (s ), is
defined as the value of the maximal nontrivial Lyapunov
exponent of (5) as a function of s . Note that F (s ) de-
pends exclusively on f (·) and G
, and is independent of the
coupling topology, i.e., independent of L (G). The region of
synchronization is, therefore, defined by the range of s > 0
for which F (s ) < 0. For a broad class of systems,
the
MSF is negative in the interval s ∈ [0,s max] ≡ S (although
more generic stability sets are also possible, we assume,
for simplicity, this is the case in subsequent derivations). In
order to achieve synchronization, the set of scaled nontrivial
Laplacian eigenvalues, {gl
i}2≤i≤N, must be located inside
the region of synchronization, S. This condition is equivalent
to: gl
2 > 0 and gl N < s max.
We illustrate how to use of the above methodology in the
following example:
Example 1: Study the stability of synchronization of a
ring of 6 coupled Rossler oscillators [14]. The dynamics of
each oscillator is described by the following system of three
nonlinear differential equations:
xi = − (yi + zi) ,
yi = xi + a yi,
zi = b + zi (xi − c) .
The adjacency entries, ai j, of a ring graph of six nodes
are ai, j = 1 if j ∈ {(i + 1) mod 6, (i − 1) mod 6}, for i =
1,2, ...,6, and ai j = 0 otherwise. The dynamics of this ring
of oscillators are defined by:
xi
yi
zi
− (yi + zi)
xi + a yi
=
b + zi (xi − c)
+ g
where we have chosen to connect the oscillators through
their xi states exclusively. Our choice is reflected in the
structure of the 3 × 3 matrix, G
, inside the summation in
Eqn. (6).
j∈R(i)
x j − xi
0
0
(6)
Numerical simulations of an isolated Rossler oscillator
unveil the existence of a periodic trajectory with period
T = 5.749 when the parameters in Eqn. (6) take the values
a = 0.2, b = 0.2, and c = 2.5 (see Fig. 1). We denote this
periodic trajectory by f (t) = [f x (t) ,f y (t) ,f z (t)]. In our
specific case, the LTP differential equation (5) takes the
following form:
−1 −1
0
a
0
+ s
c
1
0
0
0
0
0
0
0
0
x ,
(7)
0
1
f z (t)
x =
where the leftmost matrix in the above equation represents
the Jacobian of the isolated Rossler evaluated along the
periodic trajectory f (t) , and the rightmost matrix represents
.
In Fig. 2, we plot the numerical values of the maximum
Floquet exponent of Eqn. (7) for s ∈ [0,15], discretizing at
intervals of length 0.2. This plot shows the range in which
Fig. 1.
Periodic trajectory with period T = 5.749 in a Rossler oscillator
when the parameters in Eqn. (6) take the values a = 0.2, b = 0.2, and c = 2.5.
Fig. 2. Numerical values of the maximum Floquet exponent of Eqn. (7)
for s ∈ [0,15], discretizing at intervals of length 0.2.
the maximal Floquet exponent is negative. This range of
stability is S = (0,s ∗), for s ∗ ≈ 4.7. The MSF criterion
introduced in [17] states that the synchronous equilibrium
is locally stable if the set of values {g l
i (G)}i=2,...,n lies
inside the stability range, S. For the case of a 6-ring
configuration, the eigenvalues of L (G) are {0,1,1,3,3,4},
so the set {gl
i}i=2,...,n is {g ,g ,3g ,3g ,4g } . Therefore, we
achieve stability for g ∈ (0,s ∗/l n (G)), where in our case
s ∗/l n (G) ≈ 1.175.
We now illustrate this result with several numerical sim-
ulations. First, we plot in Fig. 3.a the temporal evolution of
the xi states of the 6-ring when g = 1.0. Observe how, since
g ∈ (0,1.175), we achieve asymptotic synchronization. On
the other hand, if we choose g = 1.3 /∈ (0,1.175), the time
evolution of the set of oscillators does not converge to a
common trajectory (see Fig. 3.b); instead, the even and odd
nodes settle into two different trajectories.
In the next subsection, we propose an approach to estimat-
ing the support of the eigenvalue distribution of large-scale
probabilistic networks from low-order spectral moments.
This allows us to predict synchronization in a large-scale
Chung-Lu network.
(cid:229)
G
In Fig. a, we plot the temporal evolution of the xi states of the 6-ring when g = 1.0. In Fig.b, we plot the time evolution of the set of oscillators
Fig. 3.
for g = 1.3 /∈ (0,1.175).
III. SPECTRAL ANALYSIS OF SMALL-WORLD
NETWORKS
In this section we study the Laplacian eigenvalue spectrum
of a variant of Watts-Strogatz small-world network [22].
After describing the model, we use algebraic graph theory to
compute explicit expressions for the Laplacian moments of
a small-world network as a function of its parameters. Our
derivations are based on a probabilistic analysis of the ex-
pected spectral moments of the Laplacian for asymptotically
large small-world networks.
A. Small-World Probabilistic Model
We consider a one-dimensional
i.e., vi
(cid:8)v j : j ∈ [(i − k)modN, (i + k)modN](cid:9). Then,
{v1, ...,vN}, with periodic boundary conditions,
a ring, and connect each vertex vi
est neighbors,
lattice of N vertices,
i.e., on
to its 2k clos-
is connected to the set of nodes
instead of
rewiring a fraction of the edges in the regular lattice as
proposed by Watts and Strogatz [22], we add some random
'shortcuts' to the one-dimensional lattice. These shortcuts are
added by independently assigning edges between each pair of
nodes (i, j) , 1 ≤ i < j ≤ N with probability p. The resulting
small-world graph is intermediate between a regular lattice
(achieved for p = 0) and a classical random graph (achieved
for p = 1). In general, small-world networks share properties
with both the regular grid and the classical random graph for
0 < p < 1. In particular, they show the following apparently
contradictory features:
(i) most nodes are not neighbors of one another (such as
in a regular grid), and
(ii) most nodes can be reached from every other node by
a small number of steps (such as in a random graph).
An interesting property observed in this model was the
following: for small probability of rewiring, p ≪ 1,
the
number of triangles in the network is nearly the same as that
of the regular lattice, but the average shortest-path length is
close to that of classical random graphs. In the rest of the
paper we assume we are in the range of p in which this
property holds, in particular, we will prescribe p to be r/N,
for a given parameter r.
In the coming sections we shall study spectral properties
of the Laplacian matrix associated to the above small-world
model. In our derivations we will need the probabilistic
distribution for the degrees. It is well known that, for asymp-
totically large graphs, the degree distribution of a classical
random graph with average degree r is a Poisson distribution
with rate r. Hence, the degree distribution of the above small-
world network is
Pr(di = d) =( 0,
rd−2ke−r
(d−2k)! ,
for d < 2k,
for d ≥ 2k,
(8)
which corresponds to a Poisson with parameter r 'shifted' 2k
units. The Poisson distribution is shifted to take into account
the degree of the regular 2k-neighbors ring superposed to the
random shortcuts.
Furthermore, it is well known that the clustering coeffi-
cient (or, equivalently, the number of triangles) of the regular
2k-neighbors rings is very lightly perturbed by the addition
of random shortcuts for p = r/N. In particular, one can prove
the following result:
E[T ] = (1 + o(1))
1
3
N(cid:18)2k
2(cid:19),
(9)
3 N(cid:0)2k
where the dominant term, 1
number of triangles in a 2k-neighbors ring with N nodes.
2(cid:1), corresponds to the exact
In the following section, we shall derive explicit ex-
pressions for the first low-order spectral moments of the
Laplacian matrix associated with the small-world model
herein described. Even though our analysis is far from
complete,
in that only low-order moments are provided,
valuable information regarding spectral properties can be
retrieved from our results.
B. Algebraic Analysis of Spectral Moments
In this section we deduce closed-form expressions for the
first three moments of the Laplacian spectrum of any simple
graph G. First, we express the spectral moments as a trace
using the following identity:
qk (G) =
1
N
N(cid:229)
i=1
l i (G)k =
1
N
tr L (G)k .
(10)
This identity is derived from the fact that trace is conserved
under diagonalization (in general, under any similarity trans-
formation). In the case of the first spectral moment, we obtain
q1 =
1
N
tr (D − A) =
1
N
N(cid:229)
i=1
di.
where d is the average degree of the graph. For analytical
and numerical reasons, we define the normalized Kirchhoff
moment as
where T = 1
3
network.
(cid:229) N
i=1 ti is the total number of triangles2 in the
It is worth noting how our spectral results are written in
terms of two widely reported measurements, [15]: the degree
sequence and the clustering coefficient (which provides us
with the total number of triangles.) This allows us to compute
low-order spectral moments of many real-world networks
without performing an explicit eigenvalue decomposition.
(11)
C. Probabilistic Analysis of Spectral Moments
In this section, we use Eq. (15) to compute the first
three expected Laplacian moments of the small-world model
under consideration. The expected moments can be computed
if we had explicit expressions for the moments of the
degrees, E[di], E[d2
i ], and the expected number
of triangles, E[T ]. Since we know the degree distribution (8)
for this model, the moments of the degrees can be computed
to be:
i ], and E[d3
(16)
E[di] = r + 2k,
E[d2
E[d3
i ] = r2 + (1 + 4k) r + 4k2,
i ] = r3 + 15 (3 + 6k) r2 +(cid:0)1 + 6k + 12k2(cid:1) r − 8k3.
We can therefore substitute the expressions (9) and (16) in
Eq. (15) in order to derive the following expressions for the
(non-normalized) expected Laplacian moments for N → ¥
:
E[q1] → r + 2k,
E[q2] → r2 + (4k + 2) r + 4k2 + 2k,
E[q3] → r3 + (6k + 6) r2 +(cid:0)12k2 + 18k + 4(cid:1) r
+8k2 + 8k3 + 2k.
(17)
In the following table we compare the numerical values
of the Laplacian moments corresponding to one random
realization of the model under consideration with the an-
alytical predictions in (17). In particular, we compute the
moments for a network of N = 512 nodes with parameters
p = r/N = 4/N and k = 3. It is important to point out that the
indicated numerical values are obtained for one realization
only, with no benefit from averaging.
Moment order
Numerical realization
Analytical expectations
Relative error
3rd
1st
1,467.6
10.14
10
1,431
1.38% 2.53% 2.49%
2nd
116.96
114
In the next subsection, we use an approach introduced in
[18] to estimate the support of the eigenvalue distribution
using the first three spectral moments. In coming sections,
we shall use this technique to predict whether the Laplacian
spectrum lies in the region of synchronization.
2A triangle is defined by a set of (undirected) edges {(i, j) ,( j,k) ,(k,i)}
such that i ∼ j ∼ k ∼ i.
qk =
1
N
N(cid:229)
i=1(cid:0)l i/d(cid:1)k
=
1
N dk tr (D − A)k .
The fact that D and A do not commute forecloses the pos-
sibility of using Newton's binomial expansion on (D − A)k.
On the other hand, the trace operator allows us to cyclically
permute multiplicative chains of matrices. For example,
tr (AAD) =tr (ADA) =tr (DAA). Thus, for words of length
k ≤ 3, one can cyclically arrange all binary words in the
expansion of (11) into the standard binomial expression:
a
a (cid:19) (−1)
dkN
a
tr(cid:16)A
Dk−a (cid:17) ,
for k ≤ 3.
(12)
Also, we can make use of the identity tr(cid:0)Aa Dk−a (cid:1) =
i=1 (Aa
(cid:229) N
to write
qk =
qk =
k(cid:229)
a =0(cid:18) k
)ii dk−a
k(cid:229)
N(cid:229)
i=1(cid:18) k
a =0
i
a
a (cid:19) (−1)
dkN
dk−a
i
a
(A
)ii ,
for k ≤ 3.
(13)
Note that this expression is not valid for k ≥ 4. For example,
for k = 4, we have that tr(AADD) 6=tr(DADA) .
We now analyze each summand in expression (13) from
a graph-theoretical point of view. Specifically, we find a
closed-form solution for each term tr(cid:0)AiD j(cid:1), for all pairs
1 ≤ i + j ≤ 3, as a function of the degree sequence and the
number of triangles in the network. In our analysis, we make
use of the following results from [3]:
Lemma 2: The number of closed walks of length a
in a
graph G, joining node i to itself, is given by the i-th diagonal
entry of the matrix Aa .
Corollary 3: Let G be a simple graph. Denote by ti the
number of triangles touching node i. Then,
(A)ii = 0, (cid:0)A2(cid:1)ii = di, and (cid:0)A3(cid:1)ii = 2ti.
(14)
After substituting (14) into (13), and straightforward al-
gebraic simplifications, we obtain the following exact ex-
pression for the low-order normalized spectral moments of
a given Kirchhoff matrix K:
qk =
1,
1
1
N d2(cid:0)(cid:229) N
N d3(cid:2)(cid:0)(cid:229) N
i=1 d2
i=1 d3
i + (cid:229) N
i=1 di(cid:1) ,
i + 3(cid:229) N
i=1 d2
i(cid:1) − 6 T(cid:3) ,
for k = 1,
for k = 2,
for k = 3,
(15)
D. Piecewise-Linear Reconstruction of the Laplacian Spec-
trum
Our approach, described more fully in [18], approximates
the spectral distribution with a triangular function that ex-
actly preserves the first three moments. We define a triangular
distribution t (l ) based on a set of abscissae x1 ≤ x2 ≤ x3 as
h
x2−x1
h
(l − x1) ,
(x2−x3) (l − x3) ,
0,
for l ∈ [x1,x2) ,
for l ∈ [x2,x3] ,
otherwise.
where h = 2/ (x3 − x1). The first
distribution, as a function of the abscissae, are given by
three moments of this
(x1 + x2 + x3) ,
(18)
Fig. 4.
Comparison between the histogram of the eigenvalues of one
random realization of the Laplacian matrix of a small-world model with
parameters N = 512, p = 4/N and k = 3, and the triangular function that
fits the expected spectral moments.
t (l ) :=
M1 =
M2 =
M3 =
1
3
1
1
6(cid:0)x2
10(cid:0)x3
+ x2
1 + x2
2 + x2
1 + x2
2x3 + x3
1x2 + x2
3 + x2
2 + x2
1x3 + x3
3 + x1x2 + x1x3 + x2x3(cid:1) ,
3x2 + x1x2x3(cid:1) .
3x1 + x2
2x1
Our task is to find the set of values {x1,x2,x3} in order to fit
a given set of moments {M1,M2,M3}. The resulting system
of algebraic equations is amenable to analysis, based on the
observation that the moments are symmetric polynomials3.
Following the methodology in [18], we can find the abscissae
{x1,x2,x3} as roots of the polynomial:
x3 − P
1x2 + P
2x − P
3 = 0,
where
1 = 3 M1,
2 = 9 M2
3 = 27 M3
1 − 6 M2,
1 − 36 M1M2 + 10 M3.
The following example illustrates how this technique pro-
vides a reasonable estimation of the Laplacian spectrum for
small-world Networks.
Example 4: Estimate the spectral support of the small-
world model described in Subsection III-A for parameters
N = 512, p = 4/N and k = 3. In subsection III-C we
computed the expected spectral moments of this particular
network to be {M1 = 10,M2 = 114,M3 = 1,431}. Thus, we
apply the above technique with these particular values of the
moments to compute the following set of abscissae for the tri-
angular reconstruction {x1 = 1.577,x2 = 8.662,x3 = 19.76}.
In Fig. 4 we compare the triangular function that fits the
expected spectral moments with the histogram of the eigen-
values of one random realization of the Laplacian matrix. We
also observe that any random realization of the eigenvalue
histograms of the Laplacian is remarkably close to each
other. Although a complete proof of this phenomenon is
beyond the scope of this paper, one can easily proof using
the law of large numbers that the distribution of spectral
moments in (15) concentrate around their mean values.
3A symmetric polynomial on variables (x1,x2,x3) is a polynomial that is
unchanged under any permutation of its variables.
(19)
(20)
Fig. 5. Comparison between the values of the triangular abscissae x1 and
x3 with the extreme points of the Laplacian spectral support, l 2 and l n, for
a small-world network with N = 512 nodes, k = 3, and p in the range of
values [1/N : 0.01/N : 10/N] .
We observe that the above estimation is valid for a large
range in the values of the parameters. For example, in Fig.
5, we compare the values of the triangular abscissae x1 and
x3 with the extreme points of the Laplacian spectral support,
l 2 and l n, for a small-world network with N = 512 nodes,
k = 3, and p in the range of values [1/N : 0.01/N : 10/N]. It
is important to point out that, in this case too, the numerical
values for the eigenvalues are obtained for one realization
only, with no benefit from averaging. In the next section,
we propose a methodology which uses results presented
in previous sections to predict the local stability of the
synchronous state in a small-world network of oscillators.
IV. ANALYTICAL ESTIMATION OF SYNCHRONIZATION
In this section we use the expressions in (17) and the
triangular reconstruction in the above subsection to predict
synchronization in a large small-world network of coupled
nonlinear oscillators. Specifically, we study a network of
coupled Rossler oscillators, as those in Example 1. We build
our prediction based on the following steps:
1) Determine the region of synchronization following
the technique presented in Subsection II-B. As illus-
trated in Example 1, the set of scaled eigenvalues
{gl
}i=2,...,N must lie in a certain region of stability,
(K)
i
P
P
P
S, to achieve synchronization (in our example S =
(0,s ∗ ≈ 4.7)).
2) Compute the expected spectral moments of the Lapla-
cian eigenvalue spectrum for a given set of parameters
using the set of Eqns. in (17).
i
3) Estimate the support of the Laplacian eigenvalue spec-
trum, {l (K)
}i=2,...,N, using the methodology presented
in Subsection III-D. From Example 4, we have that
sl = 1.57 and su = 19.76 are good estimates of the
lower and upper extremes of the spectral support,
respectively.
4) Compare the region of stability in Step 1 with the
i.e.,
estimation of the spectral support
(1.57 g ,19.76 g ).
in Step 3,
Following the above steps, one can easily verify that our
estimated spectral support, (1.57 g ,19.76 g ), lies inside the
region of stability, (0,s ∗ ≈ 4.7), for 0 < g < 4.7/19.76 ≈
0.238. Therefore, the small-world network of 512 coupled
Rossler oscillators is predicted to synchronize whenever the
global coupling strength satisfies g ∈ (0,0.238).
A. Numerical Results
In this section we present numerical simulations support-
ing our conclusions. We consider a set of identical 512
Rossler oscillators (as the one described in Example 1)
interconnected through the Small–World network defined
in Example (p = 4/N and k = 3). Using the methodology
proposed above, we have predicted that the synchronous state
of this system is locally stable if the coupling parameter g
lies in the interval (0,0.238). We run several simulations with
the dynamics of the oscillators presenting different values of
the global coupling strength g . For each coupling strength,
we present two plots: (i) the evolution of the 512 x-states of
the Rossler oscillators in the time interval 0 ≤ t ≤ 40, and (ii)
the evolution of xi (t)− ¯x(t) for all i, where ¯x (t) = 1
i xi (t).
N
Since our stability results are local, we have to carefully
choose the initial states for the network of oscillators. For
our particular choice of parameters, the (isolated) Rossler
oscillator presents a stable limit cycle. For our simulations,
we have chosen as initial condition for each oscillator in
the network a randomly perturbed version of a particular
point of this stable limit cycle. This particular point is s0 =
(3.5119, −3.5332,0.2006). We have chosen the perturbed
initial state for the i-th oscillator to be s0 + ei, where ei is a
uniformly distributed random variable in the 3-dimensional
cube [−2,2]3, and ei is independent of e j for i 6= j.I
In our first simulation, we use a coupling strength g =
0.1 ∈ (0,0.238); thus, we predict the synchronous state to
be locally stable. Fig. 6 (a) and (b) represents the dynamics
x-states for the 512 oscillators in the small-world network.
In this case, we observe a clear exponential convergence of
the errors to zero. In the second simulation, we choose g =
0.3 /∈ (0,0.238); thus, we predict the synchronous state to
be unstable. In fact, we observe in Figs. 7.a and 7.b how
synchronization is clearly not achieved.
V. CONCLUSIONS AND FUTURE RESEARCH
In this paper, we have studied the eigenvalue distribu-
tion of the Laplacian matrix of a large-scale small-world
networks. We have focused our attention on the low-order
moments of the spectral distribution. We have derived ex-
plicit expressions of these moments as functions of the
parameters in the small-world model. We have then applied
our results to the problem of synchronization of a network
of nonlinear oscillators. Using our expressions, we have
studied the local stability of the synchronous state in a large-
scale small-world network of oscillators. Our approach is
based on performing a triangular reconstruction matching the
first three moments of the unknown spectral measure. Our
numerical results match our predictions with high accuracy.
Several questions remain open. The most obvious extension
would be to derive expressions for higher-order moments of
the Kirchhoff spectrum. A more detailed reconstruction of
the spectral measure can be done based on more moments.
VI. ACKNOWLEDGMENTS
The first author gratefully acknowledges George C. Vergh-
ese and Vincent Blondel for their comments and suggestions
on this work.
REFERENCES
[1] A. L. Barab´asi, and R. Albert, "Emergence of Scaling in Random
Networks," Science, vol. 285, pp. 509-512, 1999.
[2] M. di Bernardo, F. Garofalo, and F. Sorrentino, "Effects of Degree
Correlation on the Synchronization of Networks of Oscillators," In-
ternational Journal of Bifurcation and Chaos, vol. 17, pp. 3499-3506,
2007.
[3] N. Biggs, Algebraic Graph Theory, Cambridge University Press,
second edition, 1993.
[4] S. Boccaletti S., V. Latora, Y. Moreno, M. Chavez, and D.-H. Hwang,
"Complex Networks: Structure and Dynamics," Physics Reports, vol.
424, no. 4-5, pp. 175-308, 2006.
[5] S. Boyd, A. Ghosh, B. Prabhakar and D. Shah, "Randomized Gossip
Algorithms," IEEE Trans. Inf. Theory, vol. 52, pp. 2508-2530, 2006.
[6] P. Bremaud, Markov Chains: Gibbs Fields, Monte Carlo Simulation,
and Queues, Springer, 2001.
[7] F.R.K. Chung, Spectral Graph Theory, AMS: CBMS series, vol. 92,
1997.
[8] S.N. Dorogovtsev, and J.F.F. Mendes, Evolution of Networks: From
Biological Nets to the Internet and WWW, Oxford University Press,
2003.
[9] M.O. Jackson, Social and Economic Networks, Princeton University
Press, 2008.
[10] A. Jadbabaie, J. Lin, and A. Morse, "Coordination of Groups of Mobile
Autonomous Agents Using Nearest Neighbor," IEEE Trans. Autom.
Control, vol. 50, no.1, 2003.
[11] X. Li, and G. Chen, "A Time-Varying Complex Dynamical Network
Model and Its Controlled Synchronization Criteria," IEEE Trans.
Autom. Control, vol. 50, no.1, 2005.
[12] L. Lov´asz, "Random Walks on Graphs: A Survey," Combinatorics,
Paul Erdos is Eighty (vol. 2), pp. 1-46, 2003.
[13] J. Lu, X. Yu, G. Chen and D. Cheng, "Characterizing the Synchro-
nizability of Small-World Dynamical Networks," IEEE Trans. Circuits
Syst. I, vol. 51, pp. 787-796, 2004.
[14] S.C. Manrubia, A.S. Mikhailov, and D. Zanette, Emergence of Dy-
namical Order, World Scientific, 2004.
[15] M.E.J. Newman, "The Structure and Function of Complex Networks,"
SIAM Review vol. 45, pp. 167-256, 2003.
[16] R. Olfati-Saber, J.A. Fax, and R.M. Murray, "Consensus and Cooper-
ation in Networked Multi-Agent Systems," Proc. IEEE, vol. 95, pp.
215-233, 2007.
(cid:229)
Fig. 6. We plot the dynamics of the x-states for 512 Rossler oscillators (as the one described in Example 1) interconnected through the Small–World
network with p = 4/N and k = 3, in Fig.a. In Fig.b, we observe a clear exponential convergence of the errors towards zero.
Fig. 7. We plot the dynamics of the x-states for the 512 Rossler oscillators for g = 0.3 /∈ (0,0.238) in Fig.a. We observe in Fig.b. how the errors do not
converge to zero.
[17] L.M. Pecora, and T.L. Carroll, "Master Stability Functions for Syn-
chronized Coupled Systems," Phys. Rev. Lett., vol. 80, no. 10, pp.
2109-2112, 1998.
[18] V.M. Preciado, Spectral Analysis for Stochastic Models of Large-Scale
Complex Dynamical Networks, Ph.D. dissertation, Dept. Elect. Eng.
Comput. Sci., MIT, Cambridge, MA, 2008.
[19] J.-J.E. Slotine and W. Wang, "A Study of Synchronization and Group
Cooperation using Partial Contraction Theory," in Cooperative Control
(S. Morse, N. Leonard, and V. Kumar, eds.), Lecture Notes in Control
and Information Science, vol. 309, Springer-Verlag, 2004.
[20] S.H. Strogatz, "Exploring Complex Networks," Nature, vol. 410, pp.
268-276, 2001.
[21] S.H. Strogatz, Sync: The Emerging Science of Spontaneous Order,
New York: Hyperion, 2003.
[22] D.J. Watts, and S. Strogatz, "Collective Dynamics of Small World
Networks," Nature, vol 393, pp. 440-42, 1998.
[23] C.W. Wu, "Synchronization in Arrays of Coupled Nonlinear Systems:
Passivity, Circle Criterion, and Observer Design," IEEE Trans. Circuits
Syst. I, vol. 48, pp. 1257-1261, 2001.
|
1911.01165 | 1 | 1911 | 2019-11-04T12:42:14 | A Comprehensive Study on Pedestrians' Evacuation | [
"cs.MA",
"physics.soc-ph"
] | Human beings face threats because of unexpected happenings, which can be avoided through an adequate crisis evacuation plan, which is vital to stop wound and demise as its negative results. Consequently, different typical evacuation pedestrians have been created. Moreover, through applied research, these models for various applications, reproductions, and conditions have been examined to present an operational model. Furthermore, new models have been developed to cooperate with system evacuation in residential places in case of unexpected events. This research has taken into account an inclusive and a 'systematic survey of pedestrian evacuation' to demonstrate models methods by focusing on the applications' features, techniques, implications, and after that gather them under various types, for example, classical models, hybridized models, and generic model. The current analysis assists scholars in this field of study to write their forthcoming papers about it, which can suggest a novel structure to recent typical intelligent reproduction with novel features. | cs.MA | cs | Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
A Comprehensive Study on Pedestrians' Evacuation
https://doi.org/10.3991/ijes.v7i4.11767
Danial A. Muhammed ()
University of Sulaimani, Sulaymaniyah, Kurdistan, Iraq
[email protected]
Soran Saeed
Sulaimania Polytechnic University, Sulaymaniyah, Kurdistan, Iraq
Tarik A. Rashid
University of Kurdistan Hewler, Hewler, Kurdistan, Iraq
Cite as: Danial Muhammed, Soran Saeed, Tarik A. Rashid (2019). A Comprehensive Study on
Pedestrians' Evacuation, International Journal of Recent Contributions from Engineering,
Science & IT (iJES), Vol. 7, No. 4 : DOi: 10.3991/ijes.v7i4.11767
Abstract -- Human beings face threats because of unexpected happenings,
which can be avoided through an adequate crisis evacuation plan, which is vital
to stop wound and demise as its negative results. Consequently, different typical
evacuation pedestrians have been created. Moreover, through applied research,
these models for various applications, reproductions, and conditions have been
examined to present an operational model. Furthermore, new models have been
developed to cooperate with system evacuation in residential places in case of
unexpected events. This research has taken into account an inclusive and a 'sys-
tematic survey of pedestrian evacuation' to demonstrate models methods by fo-
cusing on the applications' features, techniques, implications, and after that
gather them under various types, for example, classical models, hybridized
models, and generic model. The current analysis assists scholars in this field of
study to write their forthcoming papers about it, which can suggest a novel
structure to recent typical intelligent reproduction with novel features.
Keywords -- Emergency evacuation; Participants' emergency behavior; Evac-
uation time; Environment; Evacuation models.
1
Introduction
Some emergency cases will not be controlled easily and will be obstacles in front
of the evacuation process because of perplexity, dread and even uncertainty and un-
easiness to mass dwellers [1]. There are many factors that affect evacuation processes
such as the surrounding, how they react with each other, and various environmental
conditions, that is why many evacuation approaches occur like protective, preventive
rescue and constructive evacuation [2]. The issue of evacuation proceeds to the
1
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
crowd's movement and it is affected by the physical and social environment, such as
the high degree of danger, pressure, and lack of data, which is a mixture of environ-
mental hazards; population demographics and the attendee's conduct [3]. A crowd is
gathering of a group of people [4] that has many features, during simulation a number
of potential behaviors are anticipated [5], and simulation is a way of guessing behav-
iors through answering for the "what-if" conditions [6]. Hence, Crowd evacuation is a
way to aping the behavior of participants in the same situation [4], in the last 20 years
many types of research have been done, practicing evacuation have been considered
to lower the damages; deaths and injuries in emergency situations involving pedestri-
ans [7-13]. The sole purpose of the investigations is to improve a managing emergen-
cy situation that is why many models are enhanced to see how people react in differ-
ent scenarios in emergencies [14-20]. One way of finding solutions is modeling thus
after the careful examination because of the originality of the procedure it has been
renamed a model [4]. The models could be divided into three groups; classical, hy-
bridized and generic models, each of which is subdivided. These models have been
used to explore crowd evacuation in normal and emergency situations.
This research has been conducted with the aim of First is to gather a huge number
of papers about different 'applications' conducted on the evacuation of the pedestri-
ans. The second is to examine characteristics, 'techniques' and the indications of
various applications. Third, is to learn from the first two aforementioned points to
decide to scheme a novel smart and dependable model to pretend attendees' 'appear-
ing emergency' conducts and evacuation efficacy when an area is in need of emer-
gency evacuation. Thus, the current research sheds light on the pedestrian evacuation
specifically and generally. Yet it is a chance for the scholars to gain relevant infor-
mation with ease and decide about how their forthcoming papers be directed.
The current paper will contribute to the literature by benefiting from the existing
studies to assist scholars to utilize it in different cases. It can guide them to master
their forthcoming expected studies, firstly. It can assist the researchers and those who
specialize in design relying on the obtained results which were obtained through in-
vestigation of the past studies with the intention of making a decision in a better way
for designing and implementing a novel smart 'simulation model' comprising recent
capacities, secondly.
The following is the structure of this research: In the second section, 'evacuation
models for the crowd' is presented. In the third section, the previous models and the
relevant methods in various applications are demonstrated. The final section contains
the conclusion and suggestions for further research.
2
Evacuation Models for Crowd
The crowd is performing a bunch of people together [4]. It is the only condition in
which reverse could be the alternative of panic rather than being touched [21]. In
1895 presenting La Psychologie des foules by LeBon was the starting of researchers'
concern for crowd dynamics [22]. On the other hand, Helbing in 1991 presented a
widespread work, which was one of the attempts for such reason to display the motion
2
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
of pedestrians [23]. Scientists, until 2001, had previously created relating models
going for alleviating clog and obstacle wonders dependent on experimental infor-
mation [24]. Meanwhile, different fields studied crowd dynamics when pedestrians'
dynamic was presented [25]. Irregular movement, the effect of congestion and the
occurrence of own-arrangement were specified [26]. Numbers of applications were
simulated via this crowd evacuation model [25-30]. For displaying crowd evacuation
from building various methods developed, such as cellular automata method, social
force method, lattice gas method and agent-based method [31-33]. Hence, these
methods based on the ability to know the detail of the individuals in the crowd cov-
ered via three different models macroscopic model, mesoscopic model and micro-
scopic model [34, 35]. On the other hand, different hybridized methods were devel-
oped, such as zone based, layer based, and sequentially based [36]. Another method
generic framework was presented from the previous mentioned hybridized models
[34].
Figure 1 shows an overview of the developed models for crowd evacuation. The
crowd models are categorized into three main models; these are a classical model, a
hybridized model, and a generic model. Each model has its own approaches to inves-
tigate the flow of people and their behaviors during the evacuation process.
Fig. 1. An overview of the developed models for crowd evacuation
3
People BehaviourSocio-Psycological studiesMathematical Simulation StudiesExperimental StudiesPedestrian CrowdSwarm SimulationCrowd FormationTraffic simulation Crowd EvacuationHybridize ModelClassical ModelGeneric ModelMicroscopicModelMesoscopicModelMacroscopicModelLayer BasedModelZone BasedModelSequential Based ModelTransiTUMgreat density and flow of people(homogenous)Great Size and individual movementPrecise information of individual movement (heterogeneous) FluidCombination of CA and Gas Kinetic approacheSocial Force (SF)Agent Based (AB)Game TheoreticalExperimentation With animalLattice Gas (LG)Cellular Automata (CA) Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
3
Models and Their Approaches in Different Applications
In this section, the models and approaches are described, and emphasis on high-
lighting the features, techniques, and implications of current simulation models. Table
1 shows information for future use.
3.1
Classical Models
The classical model can be divided into three different models; macroscopic,
mesoscopic and microscopic. Each model was to design different approaches to know
how humans move and behave during movements from one place to another of the
specified area. These models are described in the following subsections:
Macroscopic Model: Macroscopic is one of the classical models and with such a
model flow of people is noticed and individual features are neglected due to dealing
with the homogenous people. Figure 2 illustrates the macroscopic model. In the mac-
roscopic model, the fluid dynamic was designed.
Fig. 2. Macroscopic model
In previous decades, fluid-like characteristics had been represented as a pedestrian
crowd. There were numbers of connections between fluid and pedestrians, for in-
stance, movement on all sides of the obstructions shows follow "streamlines", so, it
was not an unexpected situation, especially, such as the premature models of pedestri-
ans, which is vehicular dynamics took motivation hydrodynamics or gas-kinetic theo-
ry [37-40]. Henderson believed that a person on foot flocks acts comparably to gases
or fluid [41]. Bradley estimated that the Navier -- Stokes conditions administering
smooth movement could be utilized to depict movement in groups at high densities
[42]. Helbing et al. abridged that at medium and high densities, the movement of a
person on foot flocks demonstrated some hitting analogies with the movement of
fluid. For example, the impressions of people on foot in snow seem to be like stream-
lines of liquids or, once more, the surges of walkers through standing groups are prac-
tically equivalent to riverbeds [43]. Liquid powerful models portray how thickness
and speed change after some time with the utilization of halfway differential condi-
tions [44].
In 2002, Hughes designed a continuum theory for the flow of pedestrians. The
present hypothesis is intended for the advancement of general strategies to compre-
hend the movement of vast groups. Nevertheless, it is additionally helpful as a prog-
nostic instrument. The manner anticipated by these conditions of movement is com-
4
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
pared with the existing reaction for the Jamarat Bridge close Mecca, Saudi Arabia
[45]. In 2003, Hughes built up a continuum model distinct from a classical fluid in
light of the property that a group has the ability to think, fascinating new physical
thoughts are associated with its investigation. This property made many intriguing
applications scientifically controllable. To do this, models were given in which the
hypothesis had been utilized to give conceivable help with the yearly Muslim Hajj, to
comprehend the Battle of Agincourt, and shockingly, to find obstructions that really
increment the stream of people on foot over that when there are no hindrances present
[46]. Moreover, in 2004, Colombo and Rosini displayed a continuum show for the
person on foot stream to represent normal highlights of this sort of stream, namely, a
few impacts of fear. Specifically, this model depicts the conceivable over compres-
sions in a group and the fall in the surge through an entryway of a freezing swarm
stick. They considered the circumstance where a gathering of individuals needs to
leave a passageway through an entryway. On the off chance that the maximal surge
permitted by the entryway is low, the progress to freezing in the group moving toward
the entryway may almost certainly cause an emotional decrease in the real outflow,
reduction the outflow much more [47].
Microscopic Model: There are several old models of which microscopic is one.
Within which everything is realized accurately such as full information about individ-
ual and individual manners. Nevertheless, it is not perfect in examining the huge
number of attendees. Figure 3 illustrates the microscopic model. Several objects are
designed in microscopic such as cellular automata, lattice gas, social force, agent-
based, game theory, and experimental approaches. Details about cellular automata and
its applications are demonstrated below:
Fig. 3. Microscopic model
Cellular Automata
The accurate invention of physical methods in which 'time and space' are a remote
and liable set of dissimilar values being approved inside the corporeal dimensions is
called cellular automata. Cell automation includes a normal identical network, which
is to some extent perpetual in grade with a different variable at every position (cell).
The status of each cellular automation is mostly based on the approximations of the
total reasons for all sites. The development of cellular automation in distinct stages,
with the speculation of the 'variable' at a site undergoing influence of the reasons at
endpoints 'in its neighborhood on the' start of the previous procedure. For the area of
5
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
site (cell) it is essential to take into account two things: the 'site' and the 'neighboring
locales'. The causes at all the sites are to be up to date together at the same time in
order, in the light of the speculations of all the causes in their neighborhood at the
start of the previous process, and for the distinctive preparation of local instructions in
the corporeal abilities. They have been connected and reintroduced for a wide assort-
ment of purposes and alluded to by an assortment of names, including tessellation
automata homogeneous structures, cell structures, cellular structures, and iterative
arrays. Von Neumann and Ulam were the ones who introduced cell automata first,
which they called it cell spaces, like imaginable idealizing of 'organic' outlines (Von
Neumann, 1963, 1966), which has a unique enthusiasm behind showing 'natural self-
multiplication'. For a variety of reasons they have been linked and reinstated and
referred to by a variety of names such as tessellation automata homogeneous struc-
tures, cell structures, cellular structures, and iterative arrays [48].
In the last two decades, cellular automata models have been created to consider an
evacuating group of individuals under different circumstances. These models can be
categorized into two groups. The first depends on the associations among situations
and walkers. For example, in 2002, Perez et al. illustrated a cellular automata model
to study pedestrian exit dynamics that distributed within a single room and content to
leave through an experienced way out at the earliest possible time. The possible direc-
tion of pedestrians' movement was cardinal movement (or forward, backward, left,
and right) which relied on the empty neighbor grid and specify coercion to their cor-
poreal ability relation with neighbors and movement in conformity with ordinary
instructions. This investigation presented the arching behavior due to the jamming
effect at the way outs. Moreover, in the simulation of the way out output it observed
various features, such as flowing and disorderly intervention. Furthermore, widths of
the rooms' way out that create the possibility of pedestrians' exit at the same time
caused to pedestrians leave the room in various sizes of bursts [49].
In 2002, Kirchner and Schadschneider demonstrated leaving process imitation uti-
lized a new presented cellular automaton model for pedestrian dynamics. The idea of
chemotaxis utilized in this model, which used a bionic method to define communica-
tion between pedestrians. In this research, some relatively simple conditions were
examined, for instance, leaving a big room with a single or a couple of way outs. In
addition, it was found that changing in dimensions of the model can define manner
from regular to panic in various forms. Furthermore, it is discovered that for accom-
plishing best leaving times an appropriate amalgamation of herding manner and utili-
zation of way out familiarity was essential [50]. In 2003, Kirchner et al. enhanced a
new proposed cellular automaton model for pedestrian dynamics with adding friction
parameters. This research investigated the effect of pedestrians' collisions. Friction
parameter applied to prevent the possibility of moving conflicted participants into the
same space at the same time step. Hence, this type of conflict is possible, but elimi-
nating such situation is crucial in the precise definition of the dynamics. Besides,
creating local compression among pedestrians in the model due to the friction pa-
rameter made the model have a role in an area with great density. Large room with a
single exit door used for the evacuation simulations' experimentations. From the
6
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
experimentation's result, it was discovered that the friction parameter in arching be-
havior participated in both of the quantitative influences and qualitative fluctuate [51].
In 2005, Yang et al. utilized a two-dimensional Cellular Automata model in mim-
icking leaving the process with kin manner. Within the real leaving process several
attraction occurrences, such as confusion, congestion, assembly, step back and wait-
ing, pretended due to the difference in constructing of the building, the organizing
participants, choosing a path and the interesting for the kin manner. From the simula-
tion results noticed some times walking in mass could be safe, there was no difference
in leaving with one exit door and two exit doors from the aspect of kin manner, leav-
ing efficiency greatly affected by sub-groups number and size of the sub-groups, as
well as waiting and steps back decreased leaving efficiency [52]. In 2005, Li et al.
presented a distinctive procedure based on human behavior to make the rules more
practical for pedestrian movement, and then via assuming the bi-direction walker
motion in a corridor bottom-up and top-down walkers' motion was illustrated, as well
as probability of the influence of walkers' swapping locations was identified [53].
In 2006, Zhao et al. offered a two-dimensional cellular automata model to imitate
participant leaving through exit dynamics. This study emphasized on two features,
way out width and door partition. Hence some convenient aspects appeared, such as
the width of the way out ought to be higher than a critical value, and door partition
ought to be medium not too large and not too small. Moreover, One way out's width
increment resulted in reducing the flow out for each unit width, nonetheless entire
flow out greater than before. Whole way out's flow out was a cumulative nonlinear
function of the way out width. Furthermore, way out width assessment did not effect
on door partition's best value, and way out design had better be balanced. These as-
pects improve the efficiency of building design [54]. In 2006, Georgoudas et al. uti-
lized a two-dimensional cellular automata model and applied a computational intelli-
gent technique to examine pedestrians dynamic within a wide space. In addition, this
recommended model differs from the previous ones when the model treated with a
heterogeneous crowd. In fact, replying heterogeneous parts to the instruction, artifi-
cially arranged manner in the crowd, and made pedestrian attain one of the way outs.
Finally, characteristics of pedestrians' actions utilized to examine different assump-
tions such as pedestrians' collision during the leaving process, collective effects, sus-
pending issues, and fixed and movable obstructions [55].
In 2007, Varas et al. utilized a two-dimensional cellular automaton model to imi-
tate the process of leaving from a single and double door classroom with complete
ability. In this study, to each grid, the structure of the room, obstructions spreading,
floor field were measured. Moreover, the effect of panic as a counted dimension was
calculated which % 5 possibilities of not moving. The model applied the random
selection to cope with collision issue. Therefore, the proposed model changed into
non-deterministic via these characteristics. From the simulation result, it was clearly
observed that the best locations of the door and evacuation efficiency were not en-
hanced by substituting double door with two distinct doors. Finally, for the evacuation
time, a number of persons and way out width were considered due to suggesting sim-
ple scaling law [56].
7
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
In 2007, Yamamoto et al. demonstrated a real-coded cellular automata (RCA)
model dependent on real-coded lattice gas to simulate the left from an area with one
way out of different widths. The method of changing the pedestrian's position was
exposed. In the former developed cellular automata model, the movement was partly
simple, while they mimicked the over straight line movement and avoiding the per-
verse direction. Hence, observing the precise duration of evacuation was difficult. In
this developed model, pedestrians allowed to move in the desired direction and meas-
ured the rational duration of the evacuation. From the simulation results, congestion
observed at the exit of the big room. In fact, a critical number of pedestrian investi-
gated who made congestion.
The correlation between the number of people in the room and the total evacuation
time was achieved. Two regions region 1 and region 2 were tested. Inside region 1,
however the number of pedestrians increased, the evacuation time remained steady.
In contrast, inside region 2, people in the room needed more time to evacuate when an
initial number of people increased [57].
In 2011, Alizadeh put forward a CA model to examine the procedure of evacuation
in a place which was provided with 'obstructions' which had various configuration of
the place, like places of exit and obstruction, 'the width of the exit, light' of the place,
psychological status of the evacuee and the dispersal of the people gathered. Its influ-
ence was clearly seen in the process of evacuation. A restaurant and a classroom were
taken as a case of this model. The way the evacuees distributed, 'location and with of
the door on of the evacuation' discussed and production of the model was made ready
for comparison with some motionless models [58]. In 2014, Guo, Ren-Yong made a
model relied on 'CA with a better separation of the area and advanced speed of walk-
ing' to show going away of people who are walking from a place with one door for
exit. Two factors affected the shape of the people gathered during the experiments;
'the advanced speed of walking and the separation of area' interval of people at dif-
ferent places and the efficacy of the people who left their houses shown through
clocks. Moreover, the connection of 'width and flow of the exit' was demonstrated
through this archetype [59].
In 2015, Li and Han proposed a model for simulating pedestrian evacuation relied
on widened cellular automata to support various behavioral tendencies in people.
Understanding and violence were two of the selected social tendencies to be looked at
through this archetype. When examining this simulation, social constraints and pedes-
trians flow orders were confirmed. The results of the study show that evacuation time
does not increase with an individual's knowledge and does not decrease when the
individual's condition is noticed as aggressiveness. It is quite obvious that when the
individual avoids aggressiveness in his conduct, the best type of evacuation will be
recorded [60]. In 2018, Kontou et al. made a model of crowd evacuation on cellular
automata (CA) parallel computing tool to simulate and evaluate manners and different
features of pedestrians in the evacuation area; including disables. The simulation
process was made in a school where disables existed. A center of education in Xanthi,
which contained disable people, was selected for the simulation process. With observ-
ing and prevailing earthquake, the school organized security application; the total
8
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
time of the emptying was noted. Lastly, suggested archetype through the experimental
data validated and there was a suitability implication to the particular location [61].
Lattice Gas Models
In 1982 by Fredkin and Toffoli and in 1983 by Wolfram Lattice gases were pro-
moted, which is a unique instance of cellular automata [62-64]. The individual on the
grid of lattice gas models is measured as an active element. Possibility and measure-
ment were considered to help these models to investigate individuals' crowd charac-
teristics [44]. Individuals are fixed with L × W in this model, one individual is for one
location. Based on executing a biased random walk with no back steps, the individu-
als move to a special direction, and available locations are allowed solely [65].
In 2001, Tajima et al., used lattice gas models of biased-random walkers to pre-
tend walker channel stream at a bottleneck under the open boundaries. Then they
noticed changing free flow into chocking flow under serious appearance density,
filling flow proportion and changing the measure of density, and connection between
flow rate and scaling law [66]. In 2002, Itoh and Nagatani presented a lattice gas
model of pedestrians to simulate moving of the gathering of people between two halls
through a door, and they noticed an ideal admission time for moving the viewers. The
time has effect on the gate when visitors want to enter the hall because with decreas-
ing the admission time to under optimal time jammed state occurs and the viewers
have no ability to enter the inner hall while arrival pedestrians stopped by the depar-
ture pedestrians from entering [67]., in 2003 by Helbing et al. They made experiments
and simulations for leaving process from a classroom. For the experimentation, they
utilized video cameras for recording leaving students from a classroom and leaving
time of each student recorded. Alternatively, for the simulation, they applied lattice
gas model of pedestrian flows to compare with the experimentation results. They
noticed disorganization specification empirically is well repeated in the evacuation
process, and initial location has a major role in leaving time. Jamming or queuing
state at the exit has a great effect on increasing leaving time [68].
In addition, this lattice gas model was utilized to think of group evacuation under
various circumstances. For instance, in 2004, Nagai et al., made experimentations and
simulations to present leaving the process in a room without visibility with a number
of exits. In the experimentations, sightless students wearing eye covers imitate indi-
viduals in a room without visibility. Additionally, the video camera was applied to
record the evacuation of confused students, and then the student path and leaving time
were evaluated. On the other hand, students' detected manners were mimicked via the
extended lattice gas model wherein sightless students are simulated through biased
random walkers. Further, the mean value of the leaving time and students' leaving
dynamic patterns were measured and made a comparison with the output of the exper-
imentations. In this study, the exact emphasis was on the leaving time distribution
[69].
In 2005, Nagai et al., made experimentations and simulations to show two types of
counter-flow of students going on all fours in an open boundaries channel. In the
experimentation, video camera utilized for recording and capacities of each student
9
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
arrival times were calculated. Experimentally features of counter-flow were elucidat-
ed.
This research made a comparison between pedestrian counter-flow and the coun-
ter-flow of students on all fours. In this study, lattice gas simulation was applied to
imitate the experiment, and a biased random walker utilized to pretend students crawl-
ing [70]. In 2006, Song et al., they made up a new lattice gas model ''multi-grid mod-
el'' due to presenting force concept of social force model into a lattice gas model.
Therefore, better lattice participated which made the walkers reside in more than one
grid, and constructed instructions for walkers and walkers and structures. This new
model was used to simulate leaving walkers from a big room with an exit door, thus,
the effect of collaboration force and drift factor on leaving time were evaluated. Final-
ly, a common limitation of the two factors on the leaving process was discovered
[71].
In 2007, Fukamachi and Nagatani Studied sidle influence on counter-flow pedes-
trian and investigated manners of sidle walkers within the crowd in the counter-flow
pedestrian. Individuals within the crowd change his moving into sideways in order to
be far from congestion and barrier. In this study, the influence of sidle investigated
with the enhanced biased random walk model. Three models were demonstrated; 1)
face to face usual pace, 2) only sidelong pace, 3) during crowd and barriers change
usual walk into sideways, and get back to usual walk when mass left. They noticed
the usual pace was slower compare to sidelong pace due to rising congestion, and
emergent jamming state in the transference points. In the model number 3, jam clus-
ter nearby middle of the channel extremely fluctuating near the jamming transference
point [72], also, various ways joined with lattice gas models to deal with leaving pro-
cess research.
In 2012, Guo et al. created a varied lattice gas model via utilizing both models of
cellular automata (CA) and mobile lattice gas model (MLG model) to simulate evacu-
ation processes during an emergency. Inside this model concept of local population
density presented, and in this new model, this concept with a factor of exit crowded
degree applied to update rule. Besides, drift D that is a significant parameter has an
impact on the evacuation process and can be changed with considering the presented
concept. The nonlinear function of the corresponding distance used to define commu-
nications, such as friction, attraction, and repulsion between every two walkers and
walkers with building dividers. When spaces between them get smaller, repulsion
forces increase severely. Simple characteristics of pedestrian evacuation, such as
clogging and arching phenomena could be taken from numerical examples [73].
In 2013, Guo et al. offered an agent-based and fire and pedestrian interaction (FPI)
model to investigate the leaving process during an existing emergency. It was thought
that the environmental temperature field creates an effect on probability direction of
the movement. Besides, the multi-grid method was applied to define decreasing speed
by low transparency in the fire and pedestrian interaction (FPI). Hence, the authors
created an extended heterogeneous lattice gas (E-HLG), model. Inside this new model
factor of altitude was added to define the height location of lattice locations. Due to
the model and experimentations, characteristics of the left in a terrace classroom were
studied. Outputs from the extended HLG model were close to the experiments. In
10
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
addition, leaving controlled due to the altitude factor, and the different decision of
choosing evacuation paths and annoying high-temperature field causes to local jam-
ming and clogging [74].
In 2016, song et al., created an evacuation scene based on cellular automata and a
lattice gas model to simulate behaviors of selfless and selfish for the pedestrians dur-
ing the evacuation and competitiveness behaviors, meanwhile to present the influence
of them on pedestrians' strategies. Furthermore, some experimentation performed on
the width of the building exit door and analyzed. Outputs of the simulation tests
demonstrated that individuals with self-behavior caused more deficiency and rise
evacuation duration. Conversely, sympathy caused to decrease evacuation duration
and more collaborators. Finally, an important factor for the duration of the evacuation
was the exit door width. When the size was less than six cells of the size of 50 x 50,
evacuation time increased, conversely, the time was seriously decreased with increas-
ing the width. However, this would be no noticeable when the door exit width much
more increased [75].
Social Force Model
In 1995, Helbing and Molnar presented that pedestrian movements can comply
with 'social forces'. The movement of the pedestrian is controlled by the accompany-
ing principle impacts, which are first, pedestrian needs to achieve a specific goal.
Secondly, pedestrian keeps a specific separation from different people on foot. The
third one is that pedestrian additionally keeps a specific separation from the edge of
obstructions, for example, dividers. Fourthly, a pedestrian is some of the time is
pulled in by different people or objects [76].
In 2000, to simulate fear conditions Helbing et al. built an alternative social force
model. In this model combination of physical forces and socio-psychological with
people mass behaviors referred [77]. In 2002, Zheng et al made a combination of the
social force model (Helbing et al., 1995) and neural network as a model to simulate
different conditions of walkers for collective behaviors [78]. In 2005, Parisi and
Dorso applied the social force model, which presented by Helbing and assistants in
2000 to permits investigation of various levels of fear in evacuation within a single
exit door room. In this research, presenting concept 'faster is slower' participated in
changing manners and the rising chance of congestion suspensions, and also this con-
cept with the blocking clusters made a robust connection, whereas, size effect for exit
door was concisely debated [79].
In 2006, Seyfried et al. applied the adapted social force model which was present-
ed by Helbing and assistants in 1995 to specifically examine the effect of various
methodologies for communication between walkers, which are self-driven objects
moving in a continuous space on the velocity-density relation's output. Consequently,
they noticed it is possible to simulate the usual arrangement of the fundamental dia-
gram when individual space and current speed are increased. Additionally, they pre-
sent distant force has an effect on velocity-density relation [80]. Besides, the social
force models are joined with different models to examine swarm departure. In 2006,
Lin et al., suggested a system for evacuation crowds during emergencies via dynamic
model to propose a standard framework for future studies. Carrying out standard func-
11
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
tions was the main task, and emphasized on framework constancy and capable of
being extended to propose new required tasks in the future. This study tried to en-
hance framework independency and better system execution. Hence, experimenta-
tions were executed in a certain construction to evaluate the effectiveness of the
crowd evacuation, thus, in the experimentations results specifically interpreted crowd
manner, construction, and density of the people were mean of the effect [81]. Later in
2007, utilized the social force model to make pedestrians be dynamic, and then exam-
ined 200 pedestrians leaving process from a room during a panic situation. Parameter
such as υd, which was denoted aspiring speed for pedestrians to move, was applied to
control panic levels. In this study, the effect of "faster is slower" concept according to
the attempts with applying various forces was known. When υd efficiency of the
evacuation starting to reduction swiftly and flaw rate reaches to the peak, exponential
mass distribution changes into ''U-shaped'' [82].
In 2008, Guo and Huang proposed a mobile lattice gas model based on utilizing
both social force model and lattice gas model benefits. The model specifies each pair
walkers' communication and structure partition with pedestrian communication due to
space and pace size in motion. The output of this emergency simulation model
demonstrated 1) walkers' evacuation simple features, such as arching and clogging
behavior and practically produced a mean of the evacuation time 2) load computation
not as much as social force model and gain duration of the evacuation more precisely
[83].
In 2011, Okaya and Takahashi utilized a BDI model to simulate the behavior of
communications in the crowd, which usually occurred during evacuation. Inside such
a model, evacuation behaviors were influenced due to people interactions. Hence,
Helbing's social force model adapted in order to consider the intentions of the pedes-
trians. The output of the experiment's simulation demonstrated that as usual interac-
tions among pedestrians due to congestion made the evacuation take a longer time. In
addition, evacuation of family members together increased evacuation duration. Addi-
tionally, evacuation behaviors were influenced by directing the evacuation process
[84]. In 2014, Hou et al. applied a modified social force model to simulate the influ-
ence of the number and location of the evacuation guiders on evacuation dynamics in
partial visibility rooms. Inside this model, guiders who are qualified can identify the
exit location precisely, and others compliance with the guiders' locations and instruc-
tions. Experimentations' output reveals for one exit, one or two guiders put a signifi-
cant impact. Alternatively, for more than one exit position without adequate benefit
from the whole exit, the evacuation gets slower. Consequently, it was obvious to
increase the effect of guider on making evacuation faster, a number of exits with the
number of evacuation guiders should be equal and guiders properly inside the room
centralized of the multi-exits [85].
In 2017, Han and Liu applied a modified social force model involving an infor-
mation transmission mechanism to simulate behaviors of walkers, when the majority
walkers were unfamiliar with the evacuation location during a disaster. This improved
model considers the approach of preventing collision and disappearing information.
The difference between this adapted model and the previous model was this altered
model defines the way of finding and selecting the correct direction, and the previous
12
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
model was applied to eliminate the pedestrians collide. The output of the simulation
demonstrated that due to information transmission mechanism walkers could deter-
mine the right motion direction, although walkers' real behavior could be simulated
when emergency exists. Furthermore, there were different outcomes from the simula-
tion was obtained to enhance the evacuation. Firstly, utilizing all exit door via the
occupied extensively reduce time and rise efficiency of the evacuation. Secondly,
using exits with more width completely causes the decreasing time of the evacuation
and enhancing evacuation efficiency. Thirdly, in the start of evacuation walkers were
restricted to select exits with greater width with fewer densities for their evacuation
route. Lastly, at the start of the evacuation process essential directing was vital [86].
Agent-Based Model
ABMs are computational models that assemble social structures from the '' bottom-
up '', by reproducing people with virtual agents and making promise associations out
of the task of principles that run connections among operators [87]. Bonabeau main-
tained the perspective of the following manner. In describing agent, the manner of
mutual fear is an occurrence, which is growing due to the generally complex individ-
ual-level manner and cooperation among agents. Therefore, the agent-based model
(ABM) appeared to be perfectly suited to give significant prudence into the method
and prerequisites for fear and overcrowd by incoordination [88]. Nearly a a couple of
decades, the ABM method has been utilized to contemplate crowd evacuation in dif-
ferent circumstances. ABMs compare to other methods, such as cellular automata,
social force, lattice gas or fluid-dynamic models are commonly more computationally
costly. Besides, dealing with heterogeneous people is considerably easier due to
ABMs' capacity to enable every agent to have distinctive manners [44].
In 2004, Zarboutis and Marmaras demonstrated a method of modeling and simulat-
ing a metro system with an existing fire state in a tunnel. The ability of the simulation
method to search for an effective strategy in the rescue was debated. This system
included various subsystems and made multifaceted adjustable system. Hence, they
established an agent-based simulation to support fitting dynamic illustration of the
difficulties in the designed area, and 1) made the serious reliance and robustness of
the system be captured by the designer, 2) described the intended rescue plan via the
determined characteristics by the designer, 3) evaluated their proficiency. From the
experimentations, output demonstrated different arrangements with different situa-
tions for metro personnel's activities [89]. In 2005, Braun et al., dependent on social
force model exhibited an agent-based model (Helbing et al., 2000) to simulate effects
of various floors, walls, and obstructions on agents, and cooperation among agents in
emergency conditions. In the model, XML script was utilized to define the probability
of reproducing various scenarios. Danger incident exhibited and visualized, multifac-
eted environment measured alarms organization were spread and considered through
the environment, which led to reducing dead agents' numbers [90].
In 2006, Toyama et al., dependent on cellular automata suggested an agent-based
model (ABM) denote various walkers' features, for example, room geometry
knowledge, speed; gender, obstacle avoidance behavior, and herding behavior. They
investigated the effect of various designs, various number of pedestrians in groups,
13
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
and features on pedestrians dynamics and system's macroscopic manners [91]. In
2007, Pelechano et al. introduced High-Density Autonomous Crowds (HiDAC) mod-
el, which was a multi-agent model. This model relies upon psychological and geomet-
rical instructions while it is a parameterized social force model. It very well may be
adjusted to mimic different kinds of the crowd, extending from the high-density
crowd under quiet conditions (exit from a cinema after a movie) to extreme fear cir-
cumstances (fire leaving) [92].
In 2012, Simo et al. demonstrated a model, which employed the social force model
extensively, and agents' movement defined via Newtonian manner to simulate coun-
ter-flow conditions for mediators' behavior, which attempt to eliminate the colliding
with approaching mediators. Inside this model, mediators noticed the moving path of
head mediators, and their activities were selected based on that observation [93]. In
2012, Ha and Lykotrafitis applied self-moving particles system and movement was
directed via social force model to simulate how various conditions, such as, multifac-
eted structure of the building, size of the room exit, main exit size, friction factor, and
preferred speed influence on the evacuation time in which conditions enhancement
occurred in evacuation efficiency. The output of the simulation demonstrated evacua-
tion in a single room with small size for the exit, defined high speed takes longer time
due to occurring overcrowding. With increasing, door size overcrowding disappeared.
Friction had a significant effect on overcrowding, with minimizing the factor of fric-
tion overcrowding quickly disappeared.
On the other hand, for a floor with two rooms, one main exit door, and a hallway,
a smaller the door room size may enhance evacuation efficiency. Because the number
of the evacuee agents would be smaller from the room into the hallways, thus, the
number of evacuees for the main exit would be smaller and agents can go out without
occurring serious congestion. Conversely, with increasing room door size and agents'
speed of the evacuation time decreased, while serious congestion occurred near the
main exit door [94]. In 2018, Poulos et al. employed an agent-based evacuation model
to simulate the school's staff and nearly 1500 children of an inclusive evacuation
process executed for the whole city. This study emphasizes on kindergarten to 12th-
grade school and examines the movements of various mediators. This simulation
certified via comparing the real event, which shot video of the event, and expected a
result from the developed model simulation, errors between the real and expected was
the only % 7.6. Hence, output said that utilizing a mathematical model in evacuation
for adapting logistical issues in an emergency arrangement is fair [95].
Game Theoretic Models
On the off chance that the intelligent choice procedure of the evacuees is reasona-
ble, a game theoretic methodology can be embraced to display the choice circum-
stance [96]. In a game, the evacuees survey the majority of the accessible choices and
select the elective that augments their utility. Every evacuee's last utility adjustments
will rely upon the activities picked by all evacuees. Game is the determination of a
cooperative state via a group of individuals, conceivable approaches of every individ-
ual, and the group of all conceivable utility adjustments. For one leave, the competi-
tive behavior of the walkers in crisis departure could be deciphered in a game hypo-
14
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
thetical manner [97]. For a few ways out, Lo et al., built up a non-agreeable game
hypothesis display for the dynamic leave choice procedure of evacuees. The model
inspects how the reasonable communicating conduct of the evacuees will influence
the clearing designs. For the leave determination process, a blended procedure is con-
sidered as the likelihood of leave decision. The blended methodology Nash Equilibri-
um for the amusement depicts the balance for the evacuees and the blockage condi-
tions of ways out [96].
In 2006, Lo et al., presented an original method to show the dynamic process of
participants' way out choice. In a space with the extensive crowd, density participants
manage his/her plan of evacuation on crowd's action, movement space to the way out,
and influence of environmental motivation and way out familiarity. This plan is due
to observing activities of participants and environmental conditions via other partici-
pants and reply to their familiarity to choose their egress path. This research presented
a non-cooperative game theory framework. The model looked at how evacuation
patterns and leaving time of an area with numbers of the way out influenced via the
participants' collaborating manner. A merged procedure is considered for a way out
determination as to the way out the choice possibility. The merged procedure Nash
Equilibrium for the game defines the stability for the participants and the overcrowd-
ing conditions of way outs [96].
Approaches Based on Experiments with Animals
The utilization of creatures is another methodology for examining swarm depar-
ture. Tests in real departure freeze are troublesome, particularly with people in view
of conceivable moral and even legitimate concerns. The elements of evacuation re-
striction are not totally comprehended in light of the fact that reviews have been
largely kept to numerical recreations [98].
In 2003, Saloma et al., examined the dynamics of outflow fear in mice getting
away from a water pool to a dry stage through an exit entryway. The investigation
demonstrated how the engineering of the space in which they are kept affected in the
manner of fearing groups. The investigation output exposed that for a basic inspecting
interim their outflow manners concurred with the numerically anticipated exponential
and power-law frequency distributions of the leave burst measure notwithstanding for
brief time spans [98]. In 2005, Altshuler et al. used ants as a model of pedestrians to
demonstrate herding behavior. In the experimentations, the ants were applied in one
cell with equally placed two exits. Ants nearly used both exits to abandon in the same
way when the situation was normal, however, with setting panic due to expeller liquid
select one of the exits more severely. Hence, they modified the former hypothetical
model, which comprised herding related to the panic element as the main element to
simulate noticed dynamics outflow in detail. Moreover, from the experimentation
outputs, emerged hypothetical models, preferred that during the leaving the process
with the existing panic situation there are some common characteristics of the social
manners between humans and ants [99].
15
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
Mesoscopic Model: Mesoscopic is one of the classical models and in these model
movements of large size of people are investigated and somehow individual features
are specified, Figure 4 illustrates the mesoscopic model. In mesoscopic, cellular au-
tomata and gas Kinetic approaches are combined. The following describes cellular
automata and gas Kinetic approach:
Fig. 4. Mesoscopic Model
Cellular Automata and Gas Kinetic
Cellular Automata (CA) with Gas Kinetic methodology made a mesoscopic
model, which utilized into the motion of individuals observation. Besides, this model
presents and imitates the great size of the group [100]. CA is a model, which is divid-
ed into numbers of grids; every grid has adjacent and different state [4]. In addition,
CA to interact with simulating the departure of agents depends on separation, distribu-
tion, and utilizes an irregular way. CA thinks about the collecting manner of the
agents. A key part of the CA display is more suitable to speak to pedestrian stream in
perspective of its straightness, flexibility, and effectiveness [101].
3.2
Hybridized Models
Via using both macro and micro models of the classical models, a model, such as
hybridized models designed and it can be divided into three different models; zone
based, layer based, and sequential based. These models deal with the area of the evac-
uation and motion of the participants during the evacuation. These models are de-
scribed in the following subsections:
Zone-Based Model: In this methodology, the area of simulation is partitioned into
numerous zones. In light of use needs, each zone is reproduced either for the micro-
scopic or macroscopic model. Zone imitation under macroscopic procedure gives in
the general stream of the group, though zone mimicked with microscopic model of-
fers singular dimension practices perception. Largely, the proposed procedures run
the two models all the while on pre-defined zones [102-104].
In 2011, Wei et al. utilized Hybrid Grid Simulation Infrastructure to simulate a
leaving process of a high-density mass in an urban area. Three heterogeneous models
had been built that were a computational microscopic crowd model, a pedestrian
agent model, and a vehicle agent model to depict different parts and features of the
16
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
big and compound simulated situation. From the output exhibited that suggested in-
frastructure is a feasible and capable method for big and compound simulation system
[103]. In 2011, Sewall et al., for shared simulation of largescale vehicle traffic for
virtual universes and enlarged airborne maps exhibited a different strategy, which
dynamically combines continuum and discrete approaches. Accepting these two
unique approaches at the same time in various areas takes into account an adaptable
simulation system where the client can without much of a stretch and naturally ex-
change quality and effectiveness at runtime. They had employed this method to the
mimic of extensive systems of vehicle traffic dependent on artificial engineered urban
situations, real-world information, and accomplished more prominent than continuous
execution [105].
In 2012, Anh et al. showed a hybrid modeling method for evacuation simulation to
increase the speed of pedestrians' movement and worked on the arrangement problem
of both micro and macro models. Initial outputs demonstrated that to simulate leaving
strategy in road network via the hybrid model more effective than via micro model
alone [102]. In 2012, Xiong et al. wanted to use the benefits of macroscopic and mi-
croscopic models together. These two models in a simulation worked concurrently
and performed within various shared special partitions. Through each step of simula-
tion execution, each model had self-governing. Nevertheless, for crowd entering to
opposite partition was allowed for each of the models. Models could join at edge
border via using the various interactions of aggregation and disaggregation to swap
information. From the output, it was exhibited that this hybrid model was more effi-
cient than the microscopic model, and made quality enhancement compare to the
macroscopic model [104].
Layer Based Model: Accepting way of applying both micro and macro methods
partly into various layers is another method to deal with mass imitation. These applied
methods are used in the whole area of the imitation in order to determine plane mass
movement and additionally motion forms of the agents in the mass. This new method
for both distinct layers does the arrangement of the global path, evasion of local ob-
stacle and other wanted manners of the mass [106-108]. Inside this proposed method,
the macro method applied to mimic mass motion in accordance set of rules in the first
layer and the mass motion manner from this layer goes to the second layer as input.
Hereafter, in the second layer micro method is using to mimic motion individuals
independently and with rising density protect the cost-effectiveness.
In 2008, Banerjee et al. displayed an augmentation of the layered knowledge strat-
egy that is prevalent in the amusement business for adaptable group recreation. They
noticed a few navigation behaviors could be applied effectively in this system. The
central preferred standpoint of this system is broadened capacity, where new behav-
iors can be included by including separate layers, without influencing the current
layers. The edge rates have been observationally demonstrated to be adequate to han-
dle huge groups continuously. A few angles have been distinguished where this simu-
lation framework can be moved forward [107]. A natural way exhibited to deal with
direct recreation of virtual groups using objective coordinated route capacities. The
methodology effectively showed for a wide assortment of reproduction to produce
17
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
diverse plainly visible practices and characteristic looking movement designs, resolve
clog and perform an objective coordinated route. Mostly, this methodology offers a
straightforward, yet incredible technique to direct or control swarm reenactments.
In 2011, Patil et al. exhibited a natural way and employed navigation functions to
deal with the direct simulation of virtual groups. This approach effectively showed to
produce diverse macroscopic behaviors and natural looking movement designs, re-
solve to clog and perform goal-directed navigation. This procedure was an incredible
technique for controlling and directing the crowd simulation process [108]. In 2012,
Tissera et al. exhibited a hybrid simulation model to check behavior patterns in an
emergency leaving. Both environmental (EsM) and pedestrian (PsM) sub-models
shared inside the hybrid model. Constructing a synthetic location occupied with inde-
pendent cooperative agents due to the combination of the model with the computa-
tional procedure. Authors made sequence investigations; for instance, check the envi-
ronment to the individuals leaving that were behaviors was available for the "adjacent
door". After that, check the effect of familiarity of the individuals into the environ-
ment, outside motivation to instruct the individuals was utilized to the other conceiva-
ble outflow exit. The behavior of people reacting to this improvement is expected to
"get out the entryway quicker" [106].
Sequential Based Model: Like layer based hybrid models, another methodology is
a sequential hybrid procedure, which additionally runs both large scale (macro) and
small-scale (micro) models for the entire group. It first runs a large-scale model to
direct the motion forms of group and after that applies a small-scale model to the
same group for watching the individual manners. It executes the two models in a
successive way where a synchronization technique is required to exchange the group
state between the two modes [109,110].
In 2011, Park et al. demonstrated a hybrid framework for crowd simulation due to
applying both continuum-based and agent-based methodologies. The model catches
the dynamics of the crowd in a vast group besides the individual behaviors of every
agent. From outputs of the performance demonstrated that their methodology made a
good balance in big and great determination simulated environment. Another promis-
ing benefit of this methodology is map field development due to the capability of the
model to link to any possibilities for the map field, which gives the ability to enlarge
the continuum-based simulation with non-lattice based paths [36]. In 2013, Xiong et
al. suggested a hybrid model due to utilizing both macroscopic and microscopic mod-
els to simulate crowd in dynamic environments. Movement tendency for the crowd
was simulated via the macroscopic model. On the other hand, determining the veloci-
ty and moving direction was due to the microscopic model. According to the outputs
of the simulation appeared there is a good performance to show the features of crowd
movement and human [109].
3.3 Generic Model
Due to crowd density, the Simulation of an application nearby requires using most
suitable software, the needed dimension of individual manners (corporeal, mental and
18
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
collective), and execution time. Simulation software projects are reliant on fundamen-
tal models that cannot be changed according to end client necessity. Hence, the gener-
ic model would be an important need to give the ability to choose models on user
selection for detecting different crowd dynamics [111]. The following describes the
transit approach:
TransiTUM Model: The latest attempt exhibited to build up a conventional struc-
ture for multiscale coupling of walker imitation models for transition zones [112].
Grouping different models, such as mesoscopic and microscopic models need the
autonomous of these models. Besides that, essential parameters, such as speed, cur-
rent location, subsequent goal, max speed and so on could be moved between them
via assisting a data file. The displayed model concentrated on autonomous of related
models and in this manner can be connected to any mix of mesoscopic and micro-
scopic models. With the assistance of an outer information record, models can openly
exchange essential parameters (speed, current location, subsequent goal, max speed
and so on) between themselves. It has employed the idea of transit area and relaxation
zones to flawlessly move the people from one model to another. Therefore, walkers
can enter from any points. Nonetheless, this starter progress in the direction of con-
ventional coupling and multi-point entry to transition zone needs further examination.
Table 1. Highlighting previous models and approaches' features, techniques and
implications of current simulation models (where MT: Models Type, MTH: Methods, MD:
Models, PNT: Participants, SST: Simulation State, IOB: Investigated Occurrences and
Behaviours)
T
M
F
E
R
H
T
M
D
M
T
N
P
T
S
S
B
O
I
[66]
LGM
Normal
caused to saturated flow rate and transition density
free flow to the choking flow, bottleneck width
c
i
p
o
c
s
o
r
c
i
M
[70]
LGM
[71]
LGM
[67]
[68]
LGM
LGM
[69]
LGM
[72]
[73]
LGM and
SFM
MLGM and
CAM
[74]
E-HLGM
scale
Emer-
gency
Nor-
mal/Em
ergency
Normal
Counterflow of people crawling on all fours, speed,
jamming transition, and pattern formation
Sidle effect, counterflow jamming transition and
pattern formation
Shifting of the audience between two halls, and
Jamming transition
Jamming (queuing), and effect on increasing escape
y
c
n
e
g
r
e
m
e
time
Communication of escape times and the exit
configuration, explain blind people feature proper-
ties
Determine simple features of the social force
model, such as, clogging and arching,
Impact of concept of local population density on
drift D within the evacuation, friction, attraction,
and repulsion
Altitude factor added caused to control evacuation,
choosing evacuation paths, annoying high tempera-
s
u
o
n
e
g
o
m
o
h
s
u
o
e
n
e
g
o
r
e
t
e
h
19
s
n
o
i
t
a
c
i
l
p
p
A
l
e
d
o
M
s
a
G
e
c
i
t
t
a
L
s
l
e
d
o
M
l
a
c
i
s
s
a
l
C
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
[75]
[52]
LGM and
CAM
CAM
and
SFM
[54]
CAM
[49]
[56]
[51]
[53]
[55]
[50]
[57]
CAM
CAM
CAM
CAM
CAM
CAM
CAM
and
LGM
c
i
p
o
c
s
o
r
c
i
M
s
u
o
n
e
g
o
m
o
H
[58]
CAM
emergency
[59]
CAM
normal
s
u
o
e
n
e
g
o
r
e
t
e
h
s
u
o
e
n
e
g
o
r
e
t
e
h
c
i
p
o
c
s
o
r
c
i
M
s
u
o
n
e
g
o
m
o
H
[60]
CAM
[61]
CAM
[89]
[90]
[92]
[91]
[93]
[94]
ABM
ABM
and
SFM
ABM
and
SFM
ABM
and
CAM
ABM
and
SFM
ABM
and
ture field caused to local jamming and clogging
Selfless and selfish for the pedestrians during
evacuation
Impact of kin behavior on enhancing evacuation
efficiency
Effect of exit dynamics on flux, arching
Arching, streaming, disruptive
Effect of obstacles
Clogging, friction effects, arching effects
Bi-direction movement, moving up and down,
self-organization
Clogging and mass behavior, arching, fixed and
moveable obstacles
Herding behavior
Clogging, movement pedestrians (oblique direction
to the grid)
Impact of distribution of the evacuees, location, and
width of the door on time of the evacuation argued
the shape of the crowd, duration of the individuals
at various positions, efficiency of the evacuees
expressed via two time indicators, the association
between width and flow of the exit
Familiarity and aggressive, evacuation time
disable children, evacuation time
y
c
n
e
g
r
e
m
E
-
r
e
m
E
y
c
n
e
g
Normal
e
m
E
n
e
g
r
y
c
Normal
normal
emergency
Emer-
gency
Emer-
gency
Normal/
Emer-
gency
Nor-
mal/Em
ergency
normal
emer-
gency
Rule-based behavior
Behaviours (risk, decision, escape)
Respectful behaviors when desired (Queuing, Decision),
agent interaction (pushing behavior, panic propagation,
impatience, real time reactions to changes in the environ-
ment)
Obstacle avoidance and herding behaviors, pedestrian group
different features led to different escape probabilities
Counter flow conditions, observing the moving path of head
mediators and selecting activities based on that observation
Effect of different features on occurring overcrowding and
evacuation time
20
n
o
i
t
a
c
i
l
p
p
A
l
e
d
o
M
a
t
a
m
o
t
u
A
r
a
l
u
l
l
e
C
s
n
o
i
t
a
c
i
l
p
p
A
l
e
d
o
M
d
e
s
a
B
t
n
e
g
A
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
SFM
[95]
ABM
emer-
gency
Movements of various mediators. Supporting that
mathematical model in evacuation for adapting logistical
issues in the emergency arrangement is fair
Emer-
gency
Normal/
Emer-
gency
Normal
Emer-
gency
Emer-
gency
Emer-
gency
Emer-
gency
Provide insight for panic and jamming, Clogging, faster-is-
slower, mass behavior, pressure, beginning of panics by
counter flows and impatience
Crowd impatience, the existing optimal ratio of impatient
persons to patient persons of pedestrians made the move-
ment be quick and easy
Impact of required space and remote action on the funda-
mental diagram
Faster-is-slower, clogging, the impact of the exit door size
Faster-is-slower, cluster mass distribution
Separating data into a block, mass behavior, crowd's evac-
uation efficiency
Simple features of pedestrian evacuation (arching, clogging),
average evacuation time
communications in the crowd during the evacuation
Effect of Evacuation leaders on the evacuation process
y
c
n
e
g
r
e
m
e
Unfamiliar pedestrian behaviors with the evacuation location
s
n
o
i
t
a
c
i
l
p
p
A
l
e
d
o
M
e
c
r
o
F
l
a
i
c
o
S
[77]
SFM
[78]
SFM
[80]
SFM
[79]
SFM
[82]
[81]
[83]
SFM
SFM
and
ABM
SFM
and
LGM
[84]
SFM
[85]
SFM
[86]
SFM
with
the
infor-
mation
trans-
mis-
sion
mech-
anism
c
i
p
o
c
s
o
r
c
i
M
s
u
o
n
e
g
o
m
o
H
s
u
o
e
n
e
g
o
r
e
t
e
h
s
u
o
n
e
g
o
m
o
H
s
u
o
e
n
e
g
o
r
e
t
e
h
Fluid Dynamic Model Applications
[45]
[46]
[47]
FDM
FDM
FDM
-
o
r
c
i
M
c
i
p
o
c
s
-
g
o
m
o
H
s
u
o
n
e
Normal
Normal
Emer-
gency
Theory Model for game Applications
The motion of large crowds
crowd in motion, ''Thinking fluids'' behavior
Impact of over compression
[96]
Theory
Model for game
c
i
p
o
c
s
o
r
c
i
M
s
u
o
n
e
g
o
m
o
H
Eme
rgen
cy
exit choice, the impact of evacuees' interaction on evacuation
pattern and clearance time of a multi-exit zone
Animal Experimentations Model Applications
21
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
[98]
Mice investiga-
tion
[99]
Ant investigation
c
i
p
o
c
s
o
r
c
i
M
Emer
s
gency
u
o
n
e
g
o
m
o
H
Emer
gency
Self-organized queuing, diffusive flow, scale-free behavior in
escape panic
Panic condition, herding behavior
[105]
[103]
[102]
[104]
[107]
[108]
[106]
[110]
[109]
d
e
s
a
b
-
e
n
o
Z
d
e
s
a
b
-
r
e
y
a
L
d
e
s
a
b
-
l
a
i
t
n
e
u
q
e
S
d
n
a
c
i
p
o
c
s
o
r
c
i
m
c
i
p
o
c
s
o
r
c
a
m
s
u
o
e
n
e
g
o
r
e
t
e
h
s
u
o
e
n
e
g
o
r
e
t
e
h
-
c
a
m
d
n
a
c
i
p
o
c
s
o
r
c
i
m
c
i
p
o
c
s
o
r
c
i
p
o
c
s
o
r
c
a
m
d
n
a
c
i
p
o
c
s
o
r
c
i
m
Zone-based Model Applications
Normal
Leaving strategy in the road network
s
u
o
e
n
e
g
o
r
e
t
e
h
y
c
n
e
g
r
e
m
e
y
c
n
e
g
r
e
m
e
l
a
m
r
o
N
Normal
Leaving process of a high density mass urban area
emergen-
cy
Normal
increase speed of pedestrians' movement
Improve efficiency and enhancing quality
Layer-based Model Applications
navigation behaviors, new behaviours can be included by including separate
layers, handle huge groups continuously
clogging and perform goal-directed navigation
behaviour patterns in an emergency evacuation, environment, familiarity,
and external motivation impacts
Sequential-based Model Applications
group behaviors and complex behaviors
crowd tendency, determine velocity, moving direction
TransiTUM Model Applications
[111]
Transi-
TUM
microscopic
and
mesoscopic
with having
external data
file
s
u
o
e
n
e
g
o
m
o
h
l
a
m
r
o
N
Solving pedestrian transition between models, they can
transit in different point.
s
l
e
d
o
M
e
z
i
d
i
r
b
y
H
l
e
d
o
m
c
i
r
e
n
e
G
4
Discussion and Conclusion
Although the developed models relied on pedestrian evacuation methods have
been used widely and completed, it is yet necessary to create models that precisely
'simulate people evacuation time' and existing manners in the time of emergency
evacuation. It becomes necessary to review the past studies and make a distinction
between dissimilar causes that influenced the manners of participants when emergen-
cy cases happen and similarly on participants' evacuation time.
22
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
This paper is a comprehensive and systematic survey of pedestrian evacuation
which sheds light on 'applications' characteristics, methods, and implications of mod-
els methods. There are different categories for these models' methods and generic
models. Forthcoming studies may benefit from this sorting of models and shedding
light on their dissimilar features in several conditions and as a sort of guide to direct
future studies. Finally, researchers can get benefit from results so as to be able to
make better decisions in the system of evacuation and it will be a great support to
enhance novel simulation models with novel abilities.
Relying on the literature review, the forthcoming studies will use a developed re-
cent smart model. Moreover, the way fire is designed and implemented will be looked
at. Other characteristics like the way fire spread out through the residential places as
related to its location will be part of this, too. Additionally, the influence that the fire
leaves on the agent's conducts will be explicated.
To finalize, the consequence of the fire and smoke will be written in the suggested
model numbers of the demise, wounded or suffocated people.
5
Acknowledgment
The study is fully funded by the University of Sulaimani (UOS). The authors
would like to thank the UOS for providing facilities and equipment for this research
work. The authors would like to thank the editorial office of the journal of Interna-
tional Journal of Recent Contributions from Engineering, Science & IT (iJES) for
processing and reviewing of the manuscript.
References
6
[1] J.E. Almeida, Z. Kokkinogenis, RJ. Rossetti, "NetLogo Implementation of an Evacuation Scenario,"
arXiv preprint arXiv:1303.4695, 2013.
[2] K. Christensen, M. Sharifi, A. Chen, "Considering individuals with disabilities in a building evacua-
tion: An agent-based simulation study," In92nd Annual Meeting of the Transportation Research
Board, Washington, DC, pp.11-13, Jan. 2013.
[3] D. Duives and H. Mahmassani, "Exit choice decisions during pedestrian evacuations of buildings,"
Transportation Research Record, vol. 2316, no. 1, pp.84-94, 2012.
[4] N. Bakar, M. Majid, K. Ismail, "An overview of crowd evacuation simulation," Advanced Science
Letters, vol. 23, no. 11, pp.11428-11431, 2017.
[5] H. Liu, B. Liu, H. Zhang, L. Li, X. Qin, G. Zhang, "Crowd evacuation simulation approach based on
navigation knowledge and two-layer control mechanism," Information Sciences, vol. 436, no. 436-
437, pp.247-267, 2018.
[6] M. Majid, P. Siebers, U. Aickelin, "Modelling reactive and proactive behaviour in simulation: a case
study in a university organization," arXiv preprint arXiv:1307.1073, 2013.
[7] C. Burstedde, K. Klauck, A. Schadschneider, J. Zittartz, "Simulation of pedestrian dynamics using a
two-dimensional cellular automaton," Physica A: Statistical Mechanics and its Applications, vol. 295,
no. 3-4, pp.507-525, 2001.
[8] Y. Tajima, T. Nagatani, "Scaling behavior of crowd flow outside a hall," Physica A: Statistical Me-
chanics and its Applications, vol. 292, no. 1-4, pp. 545-54, 2001.
23
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
[9] A. Kirchner, A. Schadschneider, "Simulation of evacuation processes using a bionics-inspired cellular
automaton model for pedestrian dynamics," Physica A: statistical mechanics and its applications, vol.
312, no. 1-2, pp.260-276, 2002.
[10] H. Klüpfel, M. Meyer-König, J. Wahle, M. Schreckenberg, "Proceedings of the Fourth International
Conference on CA for Research and Industry," 2000.
[11] D. Helbing, I. Farkas, T. Vicsek, "Simulating dynamical features of escape panic," Nature, vol. 407,
no. 6803, pp. 487, 2000.
[12] D. Helbing, I. Farkas, T. Vicsek, "Freezing by heating in a driven mesoscopic system," Physical re-
view letters, vol. 84, no. 6, pp. 1240, 2000.
[13] D. Helbing, M. Isobe, T. Nagatani, K. Takimoto, "Lattice gas simulation of experimentally studied
evacuation dynamics," Physical review E, vol. 67, no. 6, pp. 067101, 2003.
[14] S. Gwynne, ER. Galea, M. Owen, PJ. Lawrence, L. Filippidis, "A review of the methodologies used in
the computer simulation of evacuation from the built environment," Building and environment, vol.
34, no. 6, pp. 741-749, 1999.
[15] G. Santos, BE. Aguirre, "A critical review of emergency evacuation simulation models". In Proceed-
ing of Conference "Building Occupant Movement During Fire Emergencies". Gaithersburg, Mary-
land: National Institute of Standards and Technology, (forthcoming). June 2004.
[16] B. Zhan, DN. Monekosso, P. Remagnino, SA. Velastin, LQ. Xu, "Crowd analysis: a survey," Machine
Vision and Applications, vol. 19, no. 5-6, pp. 345-357, 2008.
[17] X. Zheng, T. Zhong, M. Liu, "Modeling crowd evacuation of a building based on seven methodologi-
cal approaches," Building and Environment, vol. 44, no. 3, pp. 437-45, 2009.
[18] M. Kobes, I. Helsloot, B. De Vries, JG. Post, "Building safety and human behaviour in fire: A litera-
ture review," Fire Safety Journal, vol. 45, no. 1, pp. 1-11, 2010.
[19] S. Zhou, D. Chen, W. Cai, L. Luo, MY. Low, F. Tian, VS. Tay, DW. Ong, BD. Hamilton, "Crowd
modeling and simulation technologies," ACM Transactions on Modeling and Computer Simulation
(TOMACS), vol. 20, no. 4, pp. 20, 2010.
[20] J. Radianti, OC. Granmo, N. Bouhmala, P. Sarshar, A. Yazidi, J. Gonzalez, "Crowd models for emer-
gency evacuation: A review targeting human-centered sensing," 46th Hawaii International Conference
on System Sciences, IEEE, pp. 156-165, 2013.
[21] E. Canetti, Crowds and power, Macmillan, 1984.
[22] GL. Bon, The Crowd: A Study of the Popular Mind, Kitchener, Ont: Batoche, 2001.
[23] D. Helbing, "A mathematical model for the behavior of pedestrians," Behavioral science, vol. 36, no.
4, pp. 298-310, 1991.
[24] Y. Tajima, K. Takimoto, T. Nagatani, "Scaling of pedestrian channel flow with a bottleneck," Physi-
ca A: Statistical Mechanics and its Applications, vol. 294, no. 1-2, pp. 257-268, 2001.
[25] D. Helbing, P. Molnar, "Social force model for pedestrian dynamics," Physical review E, vol. 51, no.
5, pp. 4282, 1995.
[26] D. Helbing, I. Farkas, T. Vicsek, "Simulating dynamical features of escape panic," Nature, vol. 407,
no. 6803, pp. 487, 2000.
[27] D. Helbing, "A fluid dynamic model for the movement of pedestrian," arXiv preprint cond-
mat/9805213, 1998.
[28] D. Helbing, L. Buzna, A. Johansson, T. Werner, "Self-organized pedestrian crowd dynamics: Experi-
ments, simulations, and design solutions," Transportation science, vol. 39, no. 1, pp. 1-24, 2005.
[29] D. Helbing, A. Johansson, HZ. Abideen, "Dynamics of crowd disasters: An empirical study," Physical
review E, vol. 75, no. 4, pp. 046109, 2007.
[30] A. Johansson, D. Helbing, HZ. Al-Abideen, S. Al-Bosta, "From crowd dynamics to crowd safety: a
video-based analysis," Advances in Complex Systems, vol. 11, no. 04, pp. 497-527, 2008.
[31] ] X. Zheng, T. Zhong, M. Liu, "Modeling crowd evacuation of a building based on seven methodolog-
ical approaches," Building and Environment, vol. 44, no. 3, pp. 437-445, 2009.
24
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
[32] D. Helbing, IJ. Farkas, P. Molnar, T. Vicsek, "Simulation of pedestrian crowds in normal and evacua-
tion situations," Pedestrian and evacuation dynamics, vol. 21, no. 2, pp. 21-58, 2002.
[33] A. Schadschneider, W. Klingsch, H. Klüpfel, T. Kretz, C. Rogsch, A. Seyfried, "Evacuation dynam-
ics: Empirical results, modeling and applications," arXiv preprint arXiv:0802.1620, 2008.
[34] DH. Biedermann, PM. Kielar, O. Handel, A. Borrmann, "Towards TransiTUM: A generic framework
for multiscale coupling of pedestrian simulation models based on transition zones," Transportation
Research Procedia, vol. 2, pp. 495-500, 2014.
[35] RM. Tavares, "Evacuation processes versus evacuation models: "Quo Vadimus"?," Fire Technology,
vol. 45, no. 4, pp. 419-430, 2009.
[36] K. Ijaz, S. Sohail, S. Hashish, "A survey of latest approaches for crowd simulation and modeling us-
ing hybrid techniques," In17th UKSIMAMSS International Conference on Modelling and Simulation,
pp. 111-116, 2015.
[37] N. Bellomo, B. Piccoli, A. Tosin, "Modeling crowd dynamics from a complex system viewpoint,"
Mathematical models and methods in applied sciences, vol. 22, no. supp02, pp. 1230004, 2012.
[38] D. Helbing, "A fluid dynamic model for the movement of pedestrians," arXiv preprint cond-
mat/9805213, 1998.
[39] LF. Henderson, "On the fluid mechanics of human crowd motion," Transportation research, vol. 8,
no. 6, pp. 509-515, 1974.
[40] M. Chraibi, A. Tordeux, A. Schadschneider, A. Seyfried, "Modelling of Pedestrian and Evacuation
Dynamics," Encyclopedia of Complexity and Systems Science, pp. 1-22, 2018.
[41] LF. Henderson, "The statistics of crowd fluids," nature, vol. 229, no. 5284, pp. 381, 1971.
[42] GE. Bradley, "A proposed mathematical model for computer prediction of crowd movements and
their associated risks," Engineering for crowd safety, Proceedings of the International Conference on
Engineering for Crowd Safety. Amsterdam: Elsevier, pp. 303-311, 1993.
[43] D. Helbing, IJ. Farkas, P. Molnar, T. Vicsek, "Simulation of pedestrian crowds in normal and evacua-
tion situations," Pedestrian and evacuation dynamics, vol. 21, no. 2, pp. 21-58, 2002.
[44] X. Zheng, T. Zhong, M. Liu, "Modeling crowd evacuation of a building based on seven methodologi-
cal approaches," Building and Environment, vol. 44. no. 3, pp. 437-445.
[45] RL. Hughes, "A continuum theory for the flow of pedestrians," Transportation Research Part B:
Methodological, vol. 36, no 6, pp. 507-535, 2002.
[46] RL. Hughes, "A continuum theory for the flow of pedestrians," Transportation Research Part B:
Methodological, vol. 36, no 6, pp. 507-535, 2003.
[47] RM. Colombo, MD. Rosini, "Pedestrian flows and non‐classical shocks," Mathematical Methods in
the Applied Sciences, vol. 28, no. 13, pp. 1553-1567, 2005.
[48] S. Wolfram, "Cellular automata and complexity: collected papers," CRC Press, 2018.
[49] GJ. Perez, G. Tapang, M. Lim, C. Saloma, "Streaming, disruptive interference and power-law behav-
ior in the exit dynamics of confined pedestrians," Physica A: Statistical Mechanics and its Applica-
tions, vol. 312, no. 3-4, pp. 609-618, 2002.
[50] A. Kirchner, A. Schadschneider, "Simulation of evacuation processes using a bionics-inspired cellular
automaton model for pedestrian dynamics," Physica A: statistical mechanics and its applications,
2002.
[51] A. Kirchner, K. Nishinari, A. Schadschneider, "Friction effects and clogging in a cellular automaton
model for pedestrian dynamics," Physical review E, vol. 67, no. 5, pp. 056122, 2003.
[52] LZ. Yang, DL. Zhao, J. Li, TY. Fang, "Simulation of the kin behavior in building occupant evacuation
based on cellular automaton," Building and Environment, vol. 40, no. 3, pp. 411-415, 2005.
[53] L. Jian, Y. Lizhong, Z. Daoliang, "Simulation of bi-direction pedestrian movement in corridor," Phys-
ica A: Statistical Mechanics and its Applications, vol. 354, pp. 619-28, 2005.
[54] Z. Daoliang, Y. Lizhong, L. Jian, "Exit dynamics of occupant evacuation in an emergency," Physica
A: Statistical Mechanics and its Applications, vol. 363, no. 2, pp. 501-511, 2006.
25
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
[55] LG. Georgoudas, GC. Sirakoulis, IT. Andreadis, "A simulation tool for modelling pedestrian dynam-
ics during evacuation of large areas," In IFIP International Conference on Artificial Intelligence Ap-
plications and Innovations, Springer, Boston, MA, pp. 618-626, 2006.
[56] A. Varas, MD. Cornejo, D. Mainemer, B. Toledo, J. Rogan, V. Munoz, JA. Valdivia, "Cellular au-
tomaton model for evacuation process with obstacles," Physica A: Statistical Mechanics and its Ap-
plications, vol. 382, no. 2, pp. 631-642, 2007.
[57] K. Yamamoto, S. Kokubo, K. Nishinari, "Simulation for pedestrian dynamics by real-coded cellular
automata (RCA)", Physica A: Statistical Mechanics and its Applications, vol. 379, no. 2, pp. 654-660,
2007.
[58] R. Alizadeh, "A dynamic cellular automaton model for evacuation process with obstacles," Safety Sci-
ence, vol. 49, no. 2, pp. 315-323, 2011.
[59] RY. Guo, "New insights into discretization effects in cellular automata models for pedestrian evacua-
tion," Physica A: Statistical Mechanics and its Applications, vol. 400, pp. 1-11, 2014.
[60] D. Li, B. Han, "Behavioral effect on pedestrian evacuation simulation using cellular automata," Safety
science, vol. 80, pp. 41-55.
[61] P. Kontou, IG. Georgoudas, GA. Trunfio, GC. Sirakoulis, "Cellular Automata Modelling of the
Movement of People with Disabilities during Building Evacuation," In 2018 26th Euromicro Interna-
tional Conference on Parallel, Distributed and Network-based Processing (PDP), IEEE, pp. 550-557,
2018.
[62] E. Fredkin, T. Toffoli, "Conservative logic," International Journal of theoretical physics, vol. 21, no.
3-4, pp. 219-253, 1982.
[63] S. Wolfram, "Statistical mechanics of cellular automata," Reviews of modern physics, vol. 55, no. 3,
pp. 601, 1983.
[64] S. Wolfram, "Universality and complexity in cellular automata," Physica D: Nonlinear Phenomena,
vol.10, no. 1-2, pp. 1-35, 1984.
[65] DM. Shi, BH. Wang, "Evacuation of pedestrians from a single room by using snowdrift game theo-
ries," Physical Review E, vol. 87, no. 2, pp. 022802, 2013.
[66] Y. Tajima, K. Takimoto, T. Nagatani, "Scaling of pedestrian channel flow with a bottleneck," Physica
A: Statistical Mechanics and its Applications, vol. 294, no. 1-2, pp. 257-268, 2001.
[67] T. Itoh, T. Nagatani, "Optimal admission time for shifting the audience," Physica A: Statistical Me-
chanics and its Applications, vol. 313, no. 3-4, pp. 695-708, 2002.
[68] D. Helbing, M. Isobe, T. Nagatani, K. Takimoto, "Lattice gas simulation of experimentally studied
evacuation dynamics," Physical review E, vol. 67, no. 6, pp. 067101, 2003.
[69] R. Nagai, T. Nagatani, M. Isobe, T. Adachi, ". Effect of exit configuration on evacuation of a room
without visibility," Physica A: Statistical Mechanics and its Applications, vol. 343, pp. 712-724,
2004.
[70] R. Nagai, M. Fukamachi, T. Nagatani, "Experiment and simulation for counterflow of people going
on all fours," Physica A: Statistical Mechanics and its Applications, vol. 358, no. 2-4, pp. 516-528.
[71] W. Song, X. Xu, BH. Wang, S. Ni, "Simulation of evacuation processes using a multi-grid model for
pedestrian dynamics," Physica A: Statistical Mechanics and its Applications, vol. 363, no. 2, pp. 492-
500, 2006.
[72] M. Fukamachi, T. Nagatani, "Sidle effect on pedestrian counter flow," Physica A: Statistical Mechan-
ics and its Applications, vol. 377, no. 1, pp. 269-278, 2007.
[73] X. Guo, J. Chen, Y. Zheng, J. Wei, "A heterogeneous lattice gas model for simulating pedestrian
evacuation," Physica A: Statistical Mechanics and its Applications, vol.391, no. 3, pp. 582-592, 2012.
[74] X. Guo, J. Chen, S. You, J. Wei, "Modeling of pedestrian evacuation under fire emergency based on
an extended heterogeneous lattice gas model," Physica A: Statistical Mechanics and its Applications,
vol. 392, no. 9, pp. 1994-2006, 2013.
26
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
[75] X. Song, L. Ma, Y. Ma, C. Yang, H. Ji, "Selfishness-and selflessness-based models of pedestrian
room evacuation," Physica A: Statistical Mechanics and its Applications, vol.447, pp. 455-466, 2016.
[76] D. Helbing, P. Molnar, "Social force model for pedestrian dynamics," Physical review E, vol. 51, no.
5, pp. 4282, 1995.
[77] D. Helbing, I. Farkas, T. Vicsek, "Simulating dynamical features of escape panic," Nature, vol. 407,
no. 6803, pp. 487, 2000.
[78] M. Zheng, Y. Kashimori, T. Kambara, "A model describing collective behaviors of pedestrians with
various personalities in danger situations," InProceedings of the 9th International Conference on Neu-
ral Information Processing, IEEE, vol. 4, no. ICONIP'02, pp. 2083-2087, 2002.
[79] DR. Parisi, CO. Dorso, "Microscopic dynamics of pedestrian evacuation," Physica A: Statistical Me-
chanics and its Applications. Vol. 354, pp. 606-618.
[80] A. Seyfried, B. Steffen, T. Lippert, "Basics of modelling the pedestrian flow," Physica A: Statistical
Mechanics and its Applications, vol. 368, no. 1, pp. 232-238, 2006.
[81] Q. Lin, Q. Ji, S. Gong, "A crowd evacuation system in emergency situation based on dynamics mod-
el," InInternational Conference on Virtual Systems and Multimedia, Springer, Berlin, Heidelberg, pp.
269-280, 2006.
[82] DR. Parisi, CO. Dorso, "Morphological and dynamical aspects of the room evacuation process," Phys-
ica A: Statistical Mechanics and its Applications, vol. 385, no.1, pp. 343-355, 2007.
[83] RY. Guo, HJ. Huang, "A mobile lattice gas model for simulating pedestrian evacuation," Physica A:
Statistical Mechanics and its Applications, vol. 387, no 2-3, pp. 580-586.
[84] M. Okaya, T. Takahashi, "BDI agent model based evacuation simulation," InThe 10th International
Conference on Autonomous Agents and Multiagent Systems, International Foundation for Autono-
mous Agents and Multiagent Systems, vol. 3, pp. 1297-1298, 2011.
[85] L. Hou, JG. Liu, X. Pan, BH. Wang, "A social force evacuation model with the leadership effect,"
Physica A: Statistical Mechanics and its Applications, vol. 400, pp. 93-99, 2014.
[86] Y. Han, H. Liu, "Modified social force model based on information transmission toward crowd evac-
uation simulation," Physica A: Statistical Mechanics and its Applications, vol. 469, pp. 499-509,
2017.
[87] RL. Goldstone, MA. Janssen, "Computational models of collective behavior," Trends in cognitive sci-
ences, vol. 9, no. 9, pp.424-430, 2005.
[88] E. Bonabeau, "Agent-based modeling: Methods and techniques for simulating human systems," Pro-
ceedings of the national academy of sciences, vol. 99, no. suppl 3, pp. 7280-7287, 2002.
[89] N. Zarboutis, N. Marmaras, "Searching efficient plans for emergency rescue through simulation: the
case of a metro fire," Cognition, Technology & Work, vol. 6, no. 2, pp. 117-126, 2004.
[90] A. Braun, BE. Bodmann, SR. Musse, "Simulating virtual crowds in emergency situations," InProceed-
ings of the ACM symposium on Virtual reality software and technology, ACM, pp. 244-252, 2005.
[91] MC. Toyama, AL. Bazzan, "Da Silva R. An agent-based simulation of pedestrian dynamics: from lane
formation to auditorium evacuation," InProceedings of the fifth international joint conference on Au-
tonomous agents and multiagent systems, ACM, pp. 108-110, 2006.
[92] N. Pelechano, JM. Allbeck, NI. Badler, "Controlling individual agents in high-density crowd simula-
tion," InProceedings of the 2007 ACM SIGGRAPH/Eurographics symposium on Computer anima-
tion, Eurographics Association, pp. 99-108, 2007.
[93] S. Heliövaara, T. Korhonen, S. Hostikka, H. Ehtamo, "Counterflow model for agent-based simulation
of crowd dynamics," Building and Environment, vol. 48, pp. 89-100, 2012.
[94] V. Ha, G. Lykotrafitis, "Agent-based modeling of a multi-room multi-floor building emergency evac-
uation," Physica A: Statistical Mechanics and its Applications, vol. 391, no. 8, pp. 2740-2751, 2012.
[95] A. Poulos, F. Tocornal, JC. de la Llera, J. Mitrani-Reiser, "Validation of an agent-based building
evacuation model with a school drill," Transportation Research Part C: Emerging Technologies, vol.
97, pp. 82-95, 2018.
27
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
[96] SM. Lo, HC. Huang, P. Wang, KK. Yuen, "A game theory based exit selection model for evacuation,"
Fire Safety Journal, vol. 41, no. 5, pp. 364-369, 2006.
[97] A. Kirchner, H. Klüpfel, K. Nishinari, A. Schadschneider, M. Schreckenberg, "Simulation of competi-
tive egress behavior: comparison with aircraft evacuation data," Physica A: Statistical Mechanics and
its Applications, vol. 324, no. 3-4, pp. 689-697, 2003.
[98] C. Saloma, GJ. Perez, G. Tapang, M. Lim, C. Palmes-Saloma, "Self-organized queuing and scale-free
behavior in real escape panic," Proceedings of the National Academy of Sciences, vol. 100, no. 21, pp.
11947-11952, 2003.
[99] E. Altshuler, O. Ramos, Y. Núñez, J. Fernández, AJ. Batista-Leyva, C. Noda, "Symmetry breaking in
escaping ants," The American Naturalist, vol. 1666, no. 6, pp. 643-649, 2005.
[100] YQ. Jiang, W. Zhang, SG. Zhou, "Comparison study of the reactive and predictive dynamic models
for pedestrian flow," Physica A: Statistical Mechanics and its Applications, vol. 441, pp. 51-61, 2016.
[101] T. Huan-Huan, D. Li-Yun, X. Yu, "Influence of the exits' configuration on evacuation process in a
room without obstacle," Physica A: Statistical Mechanics and its Applications, vol. 420, pp. 164-178,
2015.
[102] NT. Anh, ZJ. Daniel, NH. Du, A. Drogoul, VD. An, "A hybrid macro-micro pedestrians evacuation
model to speed up simulation in road networks," InInternational Conference on Autonomous Agents
and Multiagent Systems, Springer, Berlin, Heidelberg, pp. 371-383, 2011.
[103] X. Wei, M. Xiong, X. Zhang, D. Chen, "A hybrid simulation of large crowd evacuation," In2011
IEEE 17th International Conference on Parallel and Distributed Systems, IEEE, pp. 971-975, 2011.
[104] M. Xiong, M. Lees, W. Cai, S. Zhou, MY. Low, "Hybrid modelling of crowd simulation. Procedia
Computer Science," vol. 1, no. 1, pp. 57-65, 2010.
[105] J. Sewall, D. Wilkie, MC. Lin, "Interactive hybrid simulation of large-scale traffic," ACM Transac-
tions on Graphics (TOG), vol. 30, no. 6, pp. 135, 2011.
[106] PC. Tissera, AM. Printista, E. Luque, "A hybrid simulation model to test behaviour designs in an
emergency evacuation," Procedia Computer Science, vol. 9, pp. 266-275, 2012.
[107] B. Banerjee, A. Abukmail, L. Kraemer, "Advancing the layered approach to agent-based crowd simu-
lation," In2008 22nd Workshop on Principles of Advanced and Distributed Simulation, IEEE, pp.
185-192, 2008.
[108] S. Patil, J. Van Den Berg, S. Curtis, MC. Lin, D. Manocha, "Directing crowd simulations using navi-
gation fields," IEEE transactions on visualization and computer graphics, vol. 17, no. 2, pp. 244-254,
2011.
[109] M. Xiong, S. Tang, D. Zhao, "A hybrid model for simulating crowd evacuation," New Generation
Computing, vol. 31, no. 3, pp. 211-235, 2013.
[110] SI. Park, Y. Cao, F. Quek, "Large Scale Crowd Simulation Using A Hybrid Agent Model," Motion in
Games, 2011.
[111] K. Ijaz, S. Sohail, S. Hashish, "A survey of latest approaches for crowd simulation and modeling us-
ing hybrid techniques," In17th UKSIMAMSS International Conference on Modelling and Simulation,
pp. 111-116, 2015.
[112] DH. Biedermann, PM. Kielar, O. Handel, A. Borrmann, "Towards TransiTUM: A generic framework
for multiscale coupling of pedestrian simulation models based on transition zones," Transportation
Research Procedia, vol. 2, pp. 495-500, 2014.
7
Authors
Danial Abdulkareem Muhammed received the B.Sc. degree in the field of com-
puter science from the University of Sulaimani, Iraq, in 2009, and the master's degree
(Hons.) in the field of software systems and internet technology from the University
28
Paper -- A Comprehensive Study on Pedestrians' Evacuation: A Review
of Sheffield, U.K., in 2012. He is currently pursuing the Ph.D. degree with the Uni-
versity of Sulaimani, working in the field of Artificial Intelligence in the subject of
Building Evacuation under Different Situations. He worked as an Assistant Program-
mer for nearly two years, and then continued his study in the University of Sheffield.
After that, he joined the Department of computer science, University of Sulaimani,
where he is currently affiliated with.
Soran AB. M. Saeed received the PhD degree in Computer Science (AI-CBR)
2006 University, London, UK. In 2001, He was representative of Sulaimani Universi-
ty for Scientific and Cultural affairs, Current position: Vice President for Scientific
Affairs and Higher Education at Sulaimani Polytechnic University (SPU). He is Head
the Board of eCourt System at Sulaimani Court in Iraq Developing by AKTORS
company from Estonia. His research interests are Artificial Intelligence, E-
Commerce, Information Security, Business Technology, Research Methodology, and
Software Engineering.
Tarik Ahmed Rashid received the Ph.D. degree in computer science and infor-
matics from the College of Engineering, Mathematical and Physical Sciences, Uni-
versity College Dublin (UCD), in 2006, where he was a Postdoctoral Fellow of the
Computer Science and Informatics School, from 2006 to 2007. He joined the Univer-
sity of Kurdistan Hewlêr, in 2017. His research interests include three fields. The first
field is the expansion of machine learning and data mining to deal with time series
applications. The second field is the development of DNA computing, optimization,
swarm intelligence, and nature inspired algorithms and their applications. The third
field is networking, telecommunication, and telemedicine applications.
29
|
1805.08629 | 1 | 1805 | 2018-05-20T22:50:30 | Correlation Clustering Based Coalition Formation For Multi-Robot Task Allocation | [
"cs.MA",
"cs.RO"
] | In this paper, we study the multi-robot task allocation problem where a group of robots needs to be allocated to a set of tasks so that the tasks can be finished optimally. One task may need more than one robot to finish it. Therefore the robots need to form coalitions to complete these tasks. Multi-robot coalition formation for task allocation is a well-known NP-hard problem. To solve this problem, we use a linear-programming based graph partitioning approach along with a region growing strategy which allocates (near) optimal robot coalitions to tasks in a negligible amount of time. Our proposed algorithm is fast (only taking 230 secs. for 100 robots and 10 tasks) and it also finds a near-optimal solution (up to 97.66% of the optimal). We have empirically demonstrated that the proposed approach in this paper always finds a solution which is closer (up to 9.1 times) to the optimal solution than a theoretical worst-case bound proved in an earlier work. | cs.MA | cs | Correlation Clustering Based Coalition Formation
For Multi-Robot Task Allocation
Ayan Dutta1, Vladimir Ufimtsev2, Asai Asaithambi1
8
1
0
2
y
a
M
0
2
]
A
M
.
s
c
[
1
v
9
2
6
8
0
.
5
0
8
1
:
v
i
X
r
a
Abstract- In this paper, we study the multi-robot
task
allocation problem where a group of robots needs to be allocated
to a set of tasks so that the tasks can be finished optimally. One
task may need more than one robot to finish it. Therefore the
robots need to form coalitions to complete these tasks. Multi-
robot coalition formation for task allocation is a well-known
NP-hard problem. To solve this problem, we use a linear-
programming based graph partitioning approach along with
a region growing strategy which allocates (near) optimal robot
coalitions to tasks in a negligible amount of time. Our proposed
algorithm is fast (only taking 230 secs. for 100 robots and 10
tasks) and it also finds a near-optimal solution (up to 97.66%
of the optimal). We have empirically demonstrated that the
proposed approach in this paper always finds a solution which is
closer (up to 9.1 times) to the optimal solution than a theoretical
worst-case bound proved in an earlier work.
I. INTRODUCTION
Due to the inherent complexity of real world tasks (e.g.,
search and rescue, fire extinguishing, information collection)
and limited capabilities of the available robots in the market,
it is almost impossible for one robot to finish a complex
task. As a result, cooperation among the robots is one of
the basic requirements for any task completion. One form
of cooperation among robots is coalition formation, where
a subset of available robots forms a team that is assigned
to a specific task. In this paper, we do not study how the
task is subdivided among the members of a robot coalition.
Instead, we study how a set of N robots can be optimally
divided into M coalitions to complete M tasks. Most of the
previous research in coalition formation deals with software
agents [6], [8], [12]. But real-world complexities, such as the
on-board computational power of the robots and constrained
communication, tend to limit the number of robots these al-
gorithms can handle. The approach of this paper successfully
handles as many as 100 robots.
The problem of coalition formation for multi-robot task
allocation can be described as follows: given a set of M
tasks and N robots (N > M), how to partition the robot
group optimally into coalitions where each coalition will be
assigned to a unique task. It has been proven that multi-
robot coalition formation problem for task allocation is NP-
hard to both solve exactly and to approximate within a
factor of O(M 1−),∀ > 0 [13]. Therefore, solving the
coalition formation problem for multi-robot task-allocation
in a reasonable amount of time while retaining the quality
1School of Computing at
{a.dutta, asai.asaithambi}@unf.edu
the University of North Florida, USA.
2Dept. of Mathematics & Computer Science at the East Central Univer-
sity, USA. [email protected]
of the formed coalitions is an extremely challenging problem.
In this paper, we propose an efficient algorithm for coalition
formation by a group of mobile robots for task allocation so
that the algorithm can run on the robots' on-board computers.
We present a novel value function for the tasks which
is used to assign only the required number of robots to
each task. We also present a distance-based cost function
to minimize the travel distances by the robots to the tasks.
Our approach employs a clustering-based coalition formation
methodology [1]. Clustering is a technique of gathering
'common' elements under the same label. We exploit this
idea to allocate nearby robots considered as 'common' to the
same cluster, centered around a specific task, while distant
robots are assigned to different tasks. Results show that our
approach finds a near-optimal solution (up to 97.66% of the
optimal) in a negligible amount of time.
II. RELATED WORK
Autonomous robots need to cooperate with each other
to complete complex tasks at hand. Forming teams (or,
coalitions) for efficient task completion is a computation-
ally hard problem. One of the earliest studies on task
allocation via coalition formation is due to Shehory and
Kraus [14], in which the authors have proposed a greedy
algorithm that
is guaranteed to find a solution within a
factor of (k + 1) of the optimal where k is maximum size
of any coalition formed. Coalition formation by a multi-
agent system has been studied extensively in the following
decade. Optimal [9], [10] and near-optimal [11] solutions
for coalition formation have been proposed. Most of these
proposed algorithms employ a search-based technique to find
the best solution. Although coalition formation algorithms
have been developed frequently in the last decade, very few
of them are targeted for multi-robot/agent task allocation
[13]. Taxonomies of coalition formation algorithms for task
allocation are proposed in [5], [4]. Following the taxonomy
in [4], our work in this paper can be classified as addressing
a single-task robot and multi-robot task problem. In other
words, each robot is capable of completing only one task
at a time and each task requires more than a single robot
to finish. A distributed solution which formulates the coali-
tion formation problem for multi-agent task allocation as
a distributed set partitioning problem is proposed in [15].
In [13], the authors have proposed a modified version of
the algorithm proposed in [14] and the complexity of their
algorithm, (O(N 3
2 M )), is polynomial compared to that of
Shehory and Kraus [14], which is O(N kM ), exponential on
the size of the largest coalition. However, both [13], [14]
report similar sub-optimality guarantees.
Our proposed solution in this paper generates coalitions
using a correlation clustering technique [1], [3]. It
is a
very commonly used technique in machine learning and
pattern recognition. In correlation clustering, highly corre-
lated points, robots in our case, are assigned to the same
clusters whereas the points with low correlation are allocated
to different clusters. One cluster is formed centering on
one specific task and the result of this clustering process
is equivalent to the generation of non-overlapping coalitions
of robots.
III. PROBLEM DEFINITION AND NOTATIONS
available with the robots. We assume that(cid:80)
Let R = {r1, r2, .., rN} denote a set of N robots. Each
robot is characterized by a tuple (cid:104)Pi, θi(cid:105) where Pi and θi
respectively denote the position and the orientation of the
robot ri. We assume that each robot is able to localize
itself in the environment using an on-board GPS. Let T =
{t1, t2, .., tM} denote a set of M tasks (N > M). Any task
tj is characterized by a tuple (cid:104)Pj, Oj(cid:105) where Pj and Oj
respectively denote the task location and optimal number of
robots needed to finish that task. The value of Oj for each
task tj is pre-defined and this information is assumed to be
1≤j≤M Oj = N.
The robots are homogeneous in nature, i.e., any robot can be
exchanged with any other robot. The environment is assumed
to be a rectangle of size length × width is discretized into
a set of square cells (denoted by cell) and one cell can only
be occupied by at most one robot or one task at a time.
A coalition c ⊆ R is a team of robots assigned to one task.
Without loss of generality, we sometimes call a coalition a
cluster. A coalition structure CS, defined as a partition, can
be thought of as a set of non-overlapping clusters which
covers all the robots. Let CS = {c1, c2,··· , cM} denote
a coalition structure with ci assigned to task ti for i =
1, 2,··· , M. Let ζ denote the set of all possible partitions,
and hence CS ∈ ζ.
Value Function As the robots are homogeneous, the effec-
tiveness of any robot coalition depends solely on the size of
the coalition. We define V al : CS → R, a value function
that assigns a virtual reward to a coalition and is defined as
V al(ci) = O2
i − (Oi − ci)2.
(1)
V al ensures that if the coalition ci assigned to a certain task,
ti, has Oi members in it, i.e., ci = Oi, then that coalition
earns the maximum possible value. If the size of the coalition
is greater or less than the associated Oi, then the value of the
coalition is not the highest. On the other hand, if ci > 2Oi,
then the value of the coalition becomes negative. This makes
sure that none of the formed coalitions is too large in size if
it is not required by the pre-defined O value. We define the
value of a coalition structure as the summation of values of
all the coalitions in it, i.e., V al(CS) = (cid:80)∀ci∈CS V al(ci).
(cid:80)
Note that the maximum value of any coalition structure can
be mathematically computed as follows: M AX V AL =
j .
1≤j≤M O2
Cost Function The robots are initially randomly placed in
the environment. When a robot is assigned to a task as part of
a coalition, it needs to move to the task location to complete
the task. Each robot spends a certain amount of energy (in
terms of battery power) to move from one point to another.
This is represented using the proposed cost-distance func-
where
tion, defined as costdist(ri, tj) =
d denotes the Euclidean distance between two locations. We
next define the quantity fval(ri, tj) = 1 − costdist(ri, tj) to
represent the probability of a pair of robots or robot-task pair
being in the same coalition. From here, we can calculate the
'similarity' between a task and a robot [3] as
length2+width2+1
d(Pi,Pj )
√
(cid:18) fval(ri, tj)
1 − fval(ri, tj)
(cid:19)
w(ri, tj) = ln
.
(2)
We use the same function to represent the similarity between
two robots ri, rj. A higher value of w indicates that the
members of the pair of robots or the robot-task pair are
'similar' and they should be in the same coalition, while
a lower value of w would mean that they are 'dissimilar'
and should be in different coalitions. To ensure that no two
tasks are part of the same coalition, we define their similarity
to be highly negative.
Problem Objective The problem objective is to find a
set of coalitions for all the tasks such that the generated
coalition structure has the minimum cost, while its value is
the maximum. For each coalition ci ∈ CS assigned to task
ti, a cohesion function is defined as follows:
Co(ci) =
w(rj, ti) +
w(rj, rl)
(cid:88)
rj∈ci
and the cohesion quality of CS is
CQ(CS) =
Co(ci).
(cid:88)
rj ,rl∈ci,j(cid:54)=l
(cid:88)
∀ci∈CS
Now we can formally define the multi-robot task allocation
problem as follows:
Definition 1: Given a set of N robots and M tasks and
each task ti requiring Oi number of robots to finish it, find
the coalition structure CS∗ containing M coalitions (to be
assigned to M tasks) where:
CS∗ = arg max
CS∈ζ
CQ(CS),
also satisfying
V al(CS∗) = M AX V AL.
IV. COALITION FORMATION ALGORITHM
FOR TASK ALLOCATION
The total number of possible coalition structures (parti-
tions) is exponential in the number of robots. For N robots,
and a fixed size M, 1 ≤ M ≤ N, the total number of coali-
tion structures containing exactly M non-empty coalitions is
given by the Stirling number of the second kind: S(N, M ) =
1
M !
coalition structures grows exponentially in the number of
robots. With the goal of reducing the complexity of finding
the optimal coalition structure, we use the framework of [3]
(cid:1)(M − i)N . Thus the number of possible
i=0(−1)i(cid:0)M
(cid:80)M
i
which models the set of robots and tasks as a weighted
complete graph. The robots and tasks are represented by
vertices of the graph and edge weights correspond to the
'similarity' of a pair of robots or robot-task being in the same
coalition. The cohesion quality of a given coalition structure
(partition of robots into coalitions) is calculated by summing
the edge weights of all edges that are between robots in
the same coalition. If two robots are in different coalitions,
the weight of the edge between them is not included in the
sum. To actually generate a coalition structure, we use a
graph partitioning algorithm to split the vertices (robots) into
groups (coalitions) under the constraint that the generated
coalition structure has close to optimal cohesion quality.
A. Linear Programming Formulation for Graph Partitioning
For our purposes, G = (A, E, w) will be an undirected,
weighted, complete (fully-connected) graph. A is the set of
vertices which corresponds to the set of robots R and set
of tasks T i.e. A = T ∪ R, and E is the edge set which
consists of all possible pairs of robots and tasks from A (thus
(cid:1)). The edge weight function w : E → R is
E =(cid:0)R+T
as defined in Eq. (2).
2
For any given coalition structure CS, the penalty is defined
by
P en(CS) = P enp(CS) + P enm(CS),
(3)
where P enp(CS) corresponds the sum of positive edge
weights across different coalitions and P enm(CS) corre-
sponds to the sum of negative edge weights within the same
coalitions. More specifically:
P enp(CS) =
w(ri, rl) +
w(ri, tj)
(cid:88)
w(ri,tj )>0
ri∈ck1
k1(cid:54)=j
(cid:88)
(4)
w(ri, tj)
P enm(CS) =
w(ri, rl) +
w(ri,rl)<0
ri,rl∈cj
w(ri,tj )<0
ri∈cj
(5)
Note that the subscript of a coalition matches the subscript
of the task it is assigned to i.e. coalition cj corresponds
to task tj. The penalty incorporates both positive weighted
edges between different coalitions and negative weighted
edges that are part of the same coalition. Through the max-
imization of the sum of edge weights within coalitions, the
optimal coalition structure is obtained, considering only the
function CQ(CS), without the V al(CS) function. Through
minimization of the penalty, the cohesion quality function
(CQ(CS)) of the coalition structure is maximized [2].
As specified in [3], for each edge e = (ai, aj) ∈ E, where
ai, aj ∈ A, binary variables xai,aj ∈ {0, 1} for a clustering
(coalition structure) CS are defined as: xai,aj = 0 ↔ ∃cl ∈
CS : ai, aj ∈ cl (i.e., ai, aj are in the same coalition) and
xai,aj = 1 ↔ ∃ck1 , ck2 ∈ CS, k1 (cid:54)= k2 : ai ∈ ck1 , aj ∈ ck2
(i.e., ai, aj are in different coalitions). We will use xai,aj
and xe interchangeably from here on. The P en(CS) is then
(cid:88)
k1(cid:54)=k2
(cid:88)
w(ri,rl)>0
ri∈ck1 ,rl∈ck2 ,
reformulated using the following non-negative constants:
me =
pe =
0
(cid:26) w(e)
(cid:26) w(e)
(cid:88)
0
e∈E
if w(e) < 0
if w(e) ≥ 0
if w(e) > 0
if w(e) ≤ 0
(cid:88)
e∈E
P en(CS) is then given as:
P en(CS) =
pexe +
me(1 − xe)
(6)
As stated previously, finding a coalition structure with
minimal penalty is equivalent to finding the structure with
maximal cohesion quality CQ(CS). This problem is given
as the following 0-1 integer linear program:
min: (cid:80)
pexe + (cid:80)
e∈E
e∈E
me(1 − xe)
constraints:
xai,aj ∈ {0, 1},∀ai, aj ∈ A, i (cid:54)= j
xai,aj + xaj ,ak ≥ xai,ak ,∀ai, aj, ak ∈ A, i (cid:54)= j (cid:54)= k
xai,aj = xaj ,ai∀ai, aj ∈ A, i (cid:54)= j
The second constraint
is the triangle inequality, while
the third is the symmetry constraint. These ensure that a
valid coalition structure is generated from the solution.
Since this problem is NP-complete, it is relaxed to a linear
program with the same objective function and the following
constraints:
[3], [2]:
xai,aj ∈ [0, 1],∀ai, aj ∈ A, i (cid:54)= j
xai,aj + xaj ,ak ≥ xai,ak ,∀ai, aj, ak ∈ A, i (cid:54)= j (cid:54)= k
xai,aj = xaj ,ai∀ai, aj ∈ A, i (cid:54)= j
Algorithm 1 shows the pseudo-code for the coalition
structure formation based only on the cohesion quality. This
process runs in polynomial time (in N) and gives a O(log N )
approximation (see [3]). Although this problem can be solved
in polynomial time, the solution may be non-integer i.e.
fractional. Note that, if 0 < xai,aj < 1, then there is no
definite answer and we can think of xai,aj as the probability
that ai, aj are in different coalitions. In this case, there is
extra work to be done in order to determine whether or
not ai, aj are in the same coalition. This 'rounding off'
procedure is explained in the next section (IV.B Region
Growing). It may also happen that some robots are not
assigned to any coalition where there is a task. They may
be in their own singleton coalition, or may be in a cluster
with other robots but no task. In this case, extra work also
has to be done to assign the robots to the best possible
coalition. In fact, in such a case, there will be coalitions
with tasks that will not have a sufficient amount of robots
to complete the task, since in our formulation it is assumed
that the total number of robots needed to complete all tasks
is exactly N. The region growing technique explained in the
next section can also be used in these scenarios. Another
situation to consider is that even if an integer solution is
obtained, the value of the coalition structure found may not
be the maximum (M AX V AL), meaning some coalitions
will have too many or too few robots. In this case also, we
use the region growing algorithm to optimize the value.
Algorithm 2: Region Growing algorithm for value opti-
mization and assigning unassigned robots to tasks
Input: CS: Current coalition structure (result of the linear
Algorithm 1: Coalition structure formation based on the
CQ function
Input: R: A set of robots;
T : A set of tasks.
Output: CS: A coalition structure;
Rua: A set of unassigned robots.
1 Rua ← ∅
2 for each (ai, aj) ∈ A, (A = T ∪ R) do
3
4 Set the linear program constraints after calculating the penalty
Calculate w(ai, aj).
function (Eq. 6)
5 Obtain a solution that satisfies the above-mentioned
constraints by solving the linear program.
6 if 0-1 integer solution is obtained then
7
Whenever xai,aj = 0, group ai, aj into the same
coalition. This will create a valid coalition structure CS
(due to the symmetric and triangle inequality
constraints).
if V al(CS) (cid:54)= M AX V AL then
Use the Region Growing algorithm (Algorithm 2)
else
return CS.
8
9
10
11
12 else
13
14
Add the robots, for which all the edges including them
yields a fractional solution, to Rua.
Use the Region Growing algorithm (Algorithm 2).
B. Region Growing
The coalition structure found by the Algorithm 1 maxi-
mizes the cohesion quality of CS, but does not take the value
of CS into consideration. For this reason, it might so happen
that one of the coalitions formed in this stage is unnecessarily
large and as a consequence, while other coalitions may be
smaller in size than required. For instance, let us suppose that
in a warehouse, M stacks of boxes need to be moved from
one place to the other. In this case, each stack of boxes needs
four robots to carry it, because otherwise, either the stack will
fall or it is probably too heavy for a fewer number of robots.
On the other hand, if there are too many robots assigned to
one task, then resources will be wasted. In this example, if
the robot-coalition size is less than four, then the coalition
is useless. In order to address such scenarios effectively, our
objective function (Definition 1) requires the value of the
coalition structure to be maximized, after minimizing the
cost of forming it. Therefore, the region growing algorithm
aims to optimize the value of the coalition structure found
by the linear programming approach.
The region growing algorithm is executed under one or
both of the following conditions: first, the solution found in
the previous stage is fractional; or second, for the coalition
structure (CS) found by the linear programming solution,
V al(CS) (cid:54)= M AX V AL. In the region growing process, a
virtual ball (centered around a task) is iteratively grown. This
programming solution);
Rua: unassigned robots.
Output: CS(cid:48): The final coalition structure.
of robots assigned to them.
ci ← current coalition from CS formed for task ti
while ci < Oi do
1 Tsort ← Sort the tasks T in descending order of the number
2 for ti ∈ Tsort do
3
4
5
6
7
8
9
rad ← length(cell)
Grow a virtual ball of radius rad around ti
if rj ∈ Rua AND dist(rj, ti) ≤ rad then
ci ← ci ∪ rj /*CS is updated to CS(cid:48)*/
Rua ← Rua \ rj
10
11
12
13
rad ← rad + length(cell)
if ci > Oi then
Remove the farthest robots rk ∈ ci s.t. ci = Oi
/*CS is updated to CS(cid:48)*/
Rua ← Rua ∪ rk
ball decides which robots are ultimately clustered together
for a particular task and which robots are removed from
a cluster previously formed during the linear programming
phase. This can happen if one coalition size was initially too
large resulting from the previous solution. A ball is grown
for each task. The region growing algorithm terminates when
each robot is allocated to some task (i.e., assigned to a
cluster).
Let Rua ⊆ R denote the set of unassigned robots. One
robot can be unassigned from any task because of one of
the following two reasons: first, a fractional solution has
been found in the previous round for this particular robot; or
second, the cluster previously formed is unnecessarily large
for the task. Before the start of the region growing algorithm,
the set Rua is initialized with all robots unassigned to any
cluster in the previous stage.
In the region growing algorithm (shown as Algorithm 2),
a virtual ball of a certain radius (rad) is grown for each
task (with the task as its center) iteratively. Note that rad is
initialized to one cell length. In other words, the ball will
encompass all the robots which are one cell-away from the
task ti. In the next iteration, the radius is increased to two-
cell length. If ti's virtual ball has already engulfed Oi robots
in it, then we stop growing the ball any further for this task.
If there were more than Oi robots assigned to this task,
then those robots are declared unassigned now and added
to the set Rua. Note that, the virtual ball of any task can
engulf not only the already allocated robots to it in the linear
programming phase, but also the robots which are part of the
set Rua.
Lemma 1: The worst-case time complexity of the region
growing algorithm is O(M N ).
Proof: The worst-case time complexity of the region growing
algorithm is easily seen to be O(M N ) as follows. In step
1, the M tasks are sorted in the descending order of the
number of robots assigned. This will be of complexity
O(M log M ). The time complexity for the rest of the al-
gorithm can be determined by observing that Rua, the
number of unassigned robots must be reduced to zero. Of
course, since Rua ≤ N, if s iterations are required (by
growing the region with increasing radii each time) for each
coalition ci, i = 1, 2,··· , M,
the modification of Rua
and ci together will take at most sM N = O(M N ) time.
Thus, the time complexity of the region growing algorithm
is O(M log M + M N ). Since N > M, we conclude the
complexity is O(M N ).
Lemma 2: Each task ti will get Oi number of robots
M(cid:80)
j=1
assigned to it.
Proof: First, note that,
Oj = N. This also means that,
the total number of extra robots (based on O-value) assigned
to some tasks is equal to the total number of less robots
assigned to the rest of the tasks. Therefore, if any task tj is
assigned more robots than Oj, then there is definitely one
task tk for which ck < Ok. When the tasks are sorted in
descending order, the first task t1 in the list Tsort will have
either exactly O1 or more robots assigned to it. If c1 = O1,
then we move on to t2. If c1 > O1, then we detach the extra
(c1 − O1) robots from it and put them into the set Rua. If
cj < Oj for any task tj ∈ Tsort, j > 1, then robots from
Rua will be assigned to tj. Thus, every task, tj, will have
exactly Oj robots assigned to it, i.e., cj = Oj,∀tj ∈ T .
V. EXPERIMENTS
A. Settings
We have implemented our algorithms using the Java
programming language, and ran tests on a desktop computer
(Intel i7-7700 processor, 16GB RAM). We varied the number
of robots (N) between [10, 100] in steps of 10 and selected
the number of tasks (M) from the set [2, 4, 6, 8, 10]. Remem-
bering that the maximum value of S(N, M ) could grow ex-
ponentially with increasing M, We have kept the number of
tasks at a maximum level of 10 so that the number of possible
partitions to consider does not become prohibitively large.
Additionally, note that if the task count was greater than
50% of the number of robots, that robot-task pair was not
considered. The distinct 2D locations of the robots and the
tasks were randomly generated from U({[1, 100], [1, 100]}).
Also, we considered all possible ways of assigning optimal
numbers of robots for the tasks at hand. Thus, for each pair
(N, M ) we considered for experimentation, we generated all
possible Oi's by partitioning N into exactly M parts using
integer partitioning. For example, with N = 10 and M = 2,
the set of Oi's generated and tested are {{9,1}, {8,2},
{7,3}, {6,4}, {5,5}}. The results presented here represent
the averages of the results obtained over 10 runs with each
of these settings.
B. Results
In this section, we discuss our main findings from the
experiments.
Comparison with the optimal: We have implemented a
(a)
(b)
Fig. 1.
(a) Runtime comparison on log-scale with a brute-force algorithm
[7] (4 tasks); (b) Distance-based cost comparison with the optimal solution
(higher is better – 1 being the best-case). Dotted lines indicate the theoretical
bounds [13] and the solid lines indicate the performances of our proposed
approach.
brute-force algorithm [7] that finds the optimal solution
which can be compared with the solution produced by our
proposed approach. As our proposed strategy always finds
a solution with the maximum value, we found the coalition
structure using the brute-force method which has the max-
imum value (using Eq.1) and the minimum cost among all
the maximum-valued coalition structures. We could test this
algorithm for up to 12 robots and 4 tasks after which it
became prohibitive on our test machine. Two metrics have
been compared: 1) runtime and 2) total distance-cost among
the robots and the tasks they are assigned to. The result
is shown in Figure 1. As expected, this result (Fig. 1.(a))
shows that the brute-force algorithm takes considerably more
time than our proposed approach (up to 1630 times for 12
robots and 4 tasks). On the other hand, using our proposed
approach, the robots need to travel almost similar amount
of distances compared to the optimal solution. For example,
with 6 robots and 2 tasks, the total distance traveled by the
robots using the optimal solution was 296.49m.; using our
approach it was 303.58m.; and this indicates a 97.66% near-
optimal result. Moreover, our proposed approach performs
near-optimally in terms of finding the coalition structure
with the lowest distance-cost measurement while keeping
the value of the coalition structure optimal (Fig. 1.(b)). As
the more distance traveled by the robots would result in
higher battery expenditure, without loss of generality, we
may claim that our proposed approach would eventually be
able to bound the battery expenditure at a near-optimal.
We also compare this distance-cost ratio to the optimal
with a theoretical worst-case bound proved in [14], [13]. The
plot of this theoretical bound (maxi∈[1,m] Oi + 1) is shown
in Fig. 1.(b). This figure shows that our method always finds
a significantly better solution in terms of closeness to the
optimal in each of the test cases. The maximum and the
minimum difference of these two ratios are found to be 9.1
times (for 12 robots and 2 tasks) and 3.61 times (for 8 robots
and 4 tasks).
Performance of our approach: Next, we show how the
performance of our proposed approach scales with a large
set of robots and tasks. First, we test how the runtime of
our proposed algorithm scales for up to 100 robots and 10
78910111213Number of robots100101102103104Run time ratio to brute-force4681012Number of robots00.20.40.60.81Cost ratio to optimalM=2, our appraochM=2, theoretical boundM=4, our appraochM=4, theoretical bound tasks. As can be observed in Fig. 2.(a), the maximum time
taken by our approach is about 230 secs. for 100 robots
being assigned to 10 tasks. Note that, for this setting, the
astronomical number of possible coalition structures is 2.75×
1093.
(a)
(b)
(a) Runtime of our approach; (b) Normalized average distance-cost
Fig. 2.
the robots.
tance ((cid:80)
ri,tj
In Fig. 2.(b), we show how the normalized average dis-
costdist/N) traveled by the robots for moving
from their initial locations to their allocated task locations
changes with different number of robots and tasks. In this
figure, an almost-static trend can be noticed while the max-
imum difference between any two cases being about 0.06.
This shows that the average distance traveled by the robots
does not change significantly with varying N and M. We
are also interested to see how much gain we make in terms
of the value of the a coalition structure by using the region
growing algorithm. Remember that the linear programming
component does not take the O-values into account while
forming the best coalition structure. Therefore, we cannot
guarantee that this coalition structure will have the value
M AX V AL. It is evident from the result (Fig. 3.(a)) that we
always gain a significant amount of coalition structure value
by using the region growing algorithm (up to 3.2 × 105%).
In Fig. 3.(b), we see that the value of the coalition structure
produced as a final output is always the optimal thus showing
the importance of the region growing algorithm.
(a)
(b)
Fig. 3.
(a) Increment in coalition structure value from the linear program-
ming solution by using the region growing algorithm; (b) Comparison with
the optimal value (1 being the best-case).
We have empirically demonstrated that although we min-
imize the cost and maximize the value of the coalition
structures in two successive procedures, our approach could
still form a coalition structure (i.e., a set of coalitions
assigned to the tasks) which not only has the optimal value,
the cost of it is also very close to the optimal. Also, this
near-optimal result is produced in a negligible amount of
time given the notoriously intractable nature of the problem.
VI. CONCLUSION AND FUTURE WORK
We have proposed a multi-robot coalition formation algo-
rithm for task allocation inspired by the idea of correlated
clustering. Our proposed approach first finds a coalition
structure with the minimum cost using a linear programming-
based graph partitioning formulation and next, using a region
growing approach,
it optimizes the value of this found
coalition structure. We have empirically shown that our
proposed approach can yield a near-optimal solution within
an insignificant amount of time. This approach also performs
significantly better compared to a previously proposed theo-
retical bound. In the future, we plan to make this approach
distributed so that we can avoid the single point of failure.
Also, we plan to implement this approach on a group of
robots in a real-world setting.
REFERENCES
[1] N. Bansal, A. Blum, and S. Chawla. Correlation clustering. Machine
Learning, 56(1-3):89–113, 2004.
[2] P. Dasgupta, V. Ufimtsev, C. Nelson, and S. Hossain. Dynamic recon-
figuration in modular robots using graph partitioning-based coalitions.
In Proceedings of the 11th International Conference on Autonomous
Agents and Multiagent Systems-Volume 1, pages 121–128, 2012.
[3] E. D. Demaine and N. Immorlica. Correlation clustering with partial
In Approximation, Randomization, and Combinatorial
information.
Optimization.. Algorithms and Techniques, pages 1–13. Springer, 2003.
[4] B. P. Gerkey and M. J. Matari´c. A formal analysis and taxonomy of
task allocation in multi-robot systems. The International Journal of
Robotics Research, 23(9):939–954, 2004.
[5] H. C. Lau and L. Zhang. Task allocation via multi-agent coalition
In Tools with
formation: Taxonomy, algorithms and complexity.
Artificial Intelligence, 2003. Proceedings. 15th IEEE International
Conference on, pages 346–350. IEEE, 2003.
[6] T. Michalak, J. Sroka, T. Rahwan, M. Wooldridge, P. McBurney, and
N. R. Jennings. A distributed algorithm for anytime coalition structure
In Proceedings of the 9th International Conference on
generation.
Autonomous Agents and Multiagent Systems, pages 1007–1014, 2010.
[7] M. Orlov. Efficient generation of set partitions. Engineering and
Computer Sciences, University of Ulm, Tech. Rep, 2002.
[8] T. Rahwan and N. Jennings. An improved dynamic programming
algorithm for coalition structure generation. In Proceedings of the 7th
international joint conference on Autonomous agents and multiagent
systems, pages 1417–1420, 2008.
[9] T. Rahwan, S. Ramchurn, N. Jennings, and A. Giovannucci. An
anytime algorithm for optimal coalition structure generation. J. Artif.
Intell. Res. (JAIR), 34:521–567, 2009.
[10] T. Rahwan, S. D. Ramchurn, V. D. Dang, A. Giovannucci, and N. R.
In AAAI,
Jennings. Anytime optimal coalition structure generation.
volume 7, pages 1184–1190, 2007.
[11] T. Rahwan, S. D. Ramchurn, V. D. Dang, and N. R. Jennings. Near-
optimal anytime coalition structure generation. In IJCAI, volume 7,
pages 2365–2371, 2007.
[12] S. Ramchurn, M. Polukarov, A. Farinelli, C. Truong, and N. Jennings.
Coalition formation with spatial and temporal constraints. In Proceed-
ings of the 9th International Conference on Autonomous Agents and
Multiagent Systems, pages 1181–1188, 2010.
[13] T. C. Service and J. A. Adams. Coalition formation for task allocation:
Theory and algorithms. Autonomous Agents and Multi-Agent Systems,
22(2):225–248, 2011.
[14] O. Shehory and S. Kraus. Methods for task allocation via agent
coalition formation. Artificial Intelligence, 101(1):165–200, 1998.
[15] P. T. Tosic and G. A. Agha. Maximal clique based distributed coali-
tion formation for task allocation in large-scale multi-agent systems.
MMAS, 4:104–120, 2004.
020406080100Number of robots050100150200250Run time (seconds)M=2M=4M=6M=8M=10020406080100Number of robots0.250.30.350.4Average cost (normalized)M=2M=4M=6M=8M=10020406080100Number of robots00.511.522.533.5Increment in value(%)105M=2M=4M=6M=8M=1046810Number of Tasks102030405060708090100Number of Robots00.511.52Ratio to Optimal Value |
1605.01853 | 1 | 1605 | 2016-05-06T08:07:55 | A computational intuition pump to examine group creativity: building on the ideas of others | [
"cs.MA"
] | This paper presents a computational approach to modelling group creativity. It presents an analysis of two studies of group creativity selected from different research cultures and identifies a common theme ("idea build-up") that is then used in the formalisation of an agent-based model used to support reasoning about the complex dynamics of building on the ideas of others. | cs.MA | cs | Citation: Sosa, R. & Connor, A.M. (2015) A computational intuition pump to examine group
creativity: building on the ideas of others. Proceedings of the 2015 IASDR Conference:
Interplay 2015. Brisbane, 2-5th November 2015.
A computational intuition pump to examine
group creativity: building on the ideas of
others
Ricardo Sosa, Design and Creative Technologies, Auckland University of
Technology, [email protected]
Andy M. Connor, Design and Creative Technologies, Auckland University of
Technology, [email protected]
Abstract
This paper presents a computational approach to modelling group creativity. It presents an
analysis of two studies of group creativity selected from different research cultures and
identifies a common theme ("idea build-up") that is then used in the formalisation of an
agent-based model used to support reasoning about the complex dynamics of building on the
ideas of others.
Agent-based simulation; creative teams; research methods
Understanding and managing the group dynamics that lead to creative collaborations and
creative teamwork is an important part of developing innovation capability (Francis &
Bessant, 2005). Of particular interest are multidisciplinary and multicultural settings
(Fruchter, 2001). Several problems are associated with methodically studying a process such
as creativity, elusive by definition (Cardoso de Sousa, 2007). This paper applies agent-based
simulation (ABS) as a way to model multi-level principles related to creativity and
innovation (Watts & Gilbert, 2014). It illustrates how computational models can be relevant
for mixed-method studies (Creswell, 2009) by modelling qualitative and quantitative aspects
of group creativity, in particular the critical process of building upon the ideas of others.
Introduction
Aiming to develop a deeper and more structured reasoning about principles of group
creativity, this paper adopts agent-based simulation (ABS) as a way to support rigorous
thinking and argumentation. ABS is applied here as an 'intuition pump', defined as a
resource to aid reasoning about complex subjects by harnessing intuition (Dennett, 2014).
Using ASB as an 'intuition pump' allows researchers to setup, run and analyse computational
systems as a way to support intuitive reasoning about difficult problems, to assess and clarify
arguments and to reveal meaningful ideas about group creativity. Rather than seeking
validity, or concrete recommendations for the management of creativity, ABS is used here as
a tool to "pump an intuition" and to help researchers and practitioners say "Aha! Oh, I get
it!" (Dennett, 2014).
The work presented in this paper builds on the tradition of research on creative collaboration
(Diehl & Stroebe, 1987) and creative teams (Osborn, 1963). Its main motivation is to
contribute to the theoretical understanding of what characterises creative teamwork, and to
the practical formulation of evidence-based strategies to form and manage creative groups
(Paulus & Nijstad, 2003).
Qualitative and quantitative research cultures
In creativity research two cultures or traditions exist based on two types of methods broadly
defined as qualitative and quantitative (Mahoney & Goertz, 2006). Table 1 presents the main
underlying differences between these two research cultures, related to how studies are
framed, how participants are selected, and how data is collected and analysed.
Table 1: Differences between Qualitative and Quantitative Research Cultures -adapted from (Mahoney & Goertz, 2006).
Quantitative methods used in creativity research
Qualitative methods used in creativity research
Cross-case analysis
Within-case analysis
Exemplary creative subjects selected
Randomised subjects selected; sampling
Statistical analysis
Descriptive analysis
Effects of independent variables
Causes of observations
Semantic treatment of concepts (definitions)
Numerical treatment of concepts (measurement)
Tends to symmetric explanations
Tends to asymmetric explanations
Deep data, often captures long processes
Wide scope data, often constrained to short processes
Personal experience is critical; tolerance for
implicitness
Seeks thorough explicitness
This paper aims to connect these two research cultures by building computational
simulations to support reasoning about creativity and innovation (Watts & Gilbert, 2014).
Both quantitative and qualitative lenses are used here, focusing specifically on the process of
("idea build-up", inspired by studies that represent the two research cultures and which
coincide in giving high importance to the building of new ideas upon the ideas of others.
Background
In recent decades a rich body of knowledge has been advanced related to the formation,
management and performance of creative groups (Diehl & Stroebe, 1987; Kohn, Paulus, &
Choi, 2011; Paulus & Nijstad, 2003). Valuable evidence exists to guide the practice of group
creativity including the effects of incentives and motivation (Shepperd, 1993), the role of
cognitive and cultural diversity in creative teams (Shin, Kim, Lee, & Bian, 2012; Stahl,
Maznevski, Voigt, & Jonsen, 2010), and the links between creative teams and social capital
(Han, Han, & Brass, 2014), to name a few. Nonetheless, many issues remain open about
group creativity, often with findings that are inconclusive or contradictory, correlations with
no meaningful theoretical explanations, a general lack of multi-level connections for
example between cognitive, group and social processes, virtually no connections across
disciplinary traditions, and a lack of validation of studies in real settings (Sutton &
Hargadon, 1996).
Cellular automata and agent-based simulations have been used in the study of creativity and
innovation from a variety of research agendas, ranging from illustration and pedagogical, to
aids for designing public policy and forecasting future scenarios (Watts & Gilbert, 2014).
Axelrod presented a model of cultural transmission (1997) that has been used subsequently
to model group creativity and innovation (Sosa & Gero, 2005; Watts & Gilbert, 2014). Other
agent models have been proposed for the study of systemic creativity (Kahl & Hansen,
2015), creative cities (Malik, Crooks, Root, & Swartz, 2015), the effects of peer-review in
innovation (Sobkowicz, 2015), creative leadership (Leijnen & Gabora, 2010), the impact of
intellectual property on innovation (Haydari & Smead, 2015), and idea generation in multi-
level neural models (Iyer et al., 2009).
A number of authors working with computational creativity have articulated the need to
include social evaluation in an otherwise solipsistic tradition of modelling individual
creativity (Grace, Maher, Fisher, & Brady, 2014; Jordanous, 2012; Sosa & Gero, 2015) Key
principles across five theoretical domains have been linked to formulate a framework for the
study of systemic creativity that includes social evaluation (Sosa, Gero, & Jennings, 2009),
and a functional framework was proposed to support multi-level modelling (Sosa & Gero,
2015). This paper adopts a multi-dimensional approach to examine idea generation of
groups, focusing on the process by which team members build on the idea of others (Kohn et
al., 2011).
The following section presents an analysis of two prominent studies of creativity from the
quantitative and qualitative research cultures where similarities and differences are
highlighted, and a common issue of interest is extracted to guide the development of a multi-
agent simulation of group creativity.
Building on the ideas of others
Two related studies of group creativity from different research traditions are analysed here.
An experimental study of idea generation in teams (Girotra, Terwiesch, & Ulrich, 2010)
represents the quantitative culture, whilst a four-year case study of creative collaboration
between product designers represents the qualitative culture (Elsbach & Flynn, 2013). Table
2 summarises the main similarities, differences, and identifies a common theme in these
studies. The similarities include research questions aimed at collaborative work by creative
designers in the context of organisational structures, and a particular focus on examining the
sharing and building on ideas. The differences range from the research methods applied, the
selection of participants and the design task conditions, the sources and type of data, and
how the analysis is carried. Whilst these studies are incommensurable and follow very
different criteria of scholarship (Mahoney & Goertz, 2006), both identify a common theme
related to the collaborative behaviour of building on the ideas of others, which we study in
detail in this paper aided by computational simulations described below. This issue is tackled
in different ways between research cultures, in the quantitative camp a rating system is
applied by trained judges to measure build-up scores, whilst in the qualitative tradition, a
theory-based framework is built from the data in an iterative process validated by the own
designers participating in the study.
Table 2. Analysis of two studies of group creativity from different research cultures
Key
similarities
Key
differences
(Girotra et al., 2010)
Research question: "How might individual work
in combination with working together (hybrid
groups) offer advantages over a pure team
structure?"
Motivation: "We build theory that relates
organizational phenomena to four different
variables that govern the underlying statistical
process of idea generation and selection."
Hypothesis: "The quality of the best ideas
generated and selected by a hybrid group is
higher than that of a team."
Results: "We show that idea generation in
teams is more likely to lead to ideas that build
on each other."
Limitations: "Our subjects' limited time,
resources, and prior exposure to the problem-
solving context limit our ability to perfectly
mimic a real situation. Furthermore, our
subjects were not placed in teams that had
developed a deep working relationship."
Research method: "Laboratory experiment
which compares group structures with respect
to each of the four variables individually, and
which measures their collective impact on the
quality of the best idea."
Data collection: "A within-subjects design, each
subject generates ideas under both the
treatments -team and hybrid."
Participants: "44 subjects from an upper-level
(undergraduate) product design elective
course (who) received training in idea
generation techniques. Subjects were
randomly divided in groups of four."
Design problem: "Each team is given 30
minutes to complete an idea generation
challenge. In the hybrid treatment, 10 minutes
to work individually and an additional 20
minutes to share and discuss their ideas and
to develop new ideas. Challenges: "1) A
manufacturer of sports and fitness products is
interested in new product concepts that might
be sold to students in a sporting goods retailer.
2) A manufacturer of dorm and apartment
products is interested in new product concepts
that might be sold to students in a home-
products retailer."
(Elsbach & Flynn, 2013)
Research question: "How do specific
collaboration activities relate to the self-
concepts of creative workers in corporate
contexts?"
Motivation: "We hope to shed light on the
reasons why some creative collaborations in
organizations succeed, while others fail."
Hypothesis: "We propose that affirming self-
concepts may conflict with collaborative
behaviours expected of creative workers."
Results: "Our analysis indicated that idea-
giving behaviours affirmed the personal
identities of most designers. The second
pattern indicated that idea-taking behaviours
threatened personal identities of most
designers."
Limitations: "Further research is needed to
identify the extent to which our findings can
generalize across different organizations and
different organizational cultures."
Research method: "Qualitative methods,
including interviews and non-participant
observation."
Data collection: "Over a four-year period: 40
open-ended interviews and approximately 100
hours of observation."
Participants: "40 designers (35 men, 5 women;
average age = 39.5 years; average time
working at the corporation = 12.7 years).
Participants held the titles of staff designer (7),
project designer (16), or designer (17)."
Design problem: "Designers in this division
were required to come up with approximately
1000 new toy designs a year. These original
designs sometimes involved relatively minor
modifications to existing designs (a new toy
car design), but often involved the creation of
completely new toy concepts."
Evaluation: "These designers viewed
themselves, and were viewed by management,
as creative workers. The director also told us
that collaborative teamwork was the norm on
all design projects, and that designers were
expected to collaborate with everyone on a
project team. All of the designers reported that
they were required to work extensively in
collaborative teams and that they considered
Evaluation: "An accurate measurement of idea
quality is central to our work (…) we use two
approaches: a quality evaluation tool, which
collects about 20 ratings per idea, and a
purchase-intent survey, which captures about
40 consumer opinions about their intent to
purchase a product. A (judging) panel received
formal training in the valuation of new
products. To verify the reliability of these
ratings, we constructed Kappa and AC1
statistics for each of the two idea domains."
Units of analysis: "Average quality of ideas,
number of ideas, variance in the quality of
ideas, and the ability of groups to discern the
quality of ideas. Our metric for effectiveness is
the quality of ideas selected as the best."
Data analysis: "All hypotheses related to idea
quality are tested using both business value
and purchase intent as measures of quality.
We use an analysis of variance (ANOVA) of
the judges' ratings given each idea. We include
controls for the four-person group of
individuals generating the ideas and the rater
who provided the rating. This is because there
are substantial differences in ability across the
groups, and because there are systemic
differences in how the scales were used by
different raters."
Definition: "One person can build on the ideas
of another in a way that increases the quality
of the ideas. As far as we know there has not
been any theoretical or empirical support for
the claim that these ideas are better than ideas
that are generated independently. In our study,
we directly explore the role of buildup. To do
this, we hired three independent judges to
code the substance of ideas on three
dimensions: the type of product, the principal
sporting activity associated with the product,
and the key benefit proposition of the proposed
product To construct our buildup metric, we
compare the classification of two consecutively
generated ideas. For example, if the idea
shares all three dimensions with the preceding
idea, it earns a buildup score of 3. More
generally, the buildup score is the number of
dimensions on which an idea shares a value
with the idea generated immediately
previously. We average this buildup score
across the three independent judges."
Findings: "We show that idea generation in
teams is more likely to lead to ideas that build
on each other (however) we find that ideas that
build on a previous idea are worse, not better,
on average. We found that differences in
performance across individuals are large and
highly significant."
Managerial implications: "If the interactive
buildup is not leading to better ideas, an
themselves to be creative workers due to their
job requirements to produce original toy
designs."
Units of analysis: "Two primary types of
personal identities of designers: (1) 'artistic'
personal identities, that included the self-
categorizations of 'creator', 'controller', and
'idealist'; and (2) 'problem solving' personal
identities, that included the self-categorizations
of 'pragmatist', 'refiner', and 'enabler'."
Data analysis: "Data analysis followed an
iterative approach, moving back and forth
between theory- development, data review,
and literature review. We analysed our data in
four stages. We asked (the designers) if our
placement made sense. All 30 of the designers
confirmed our choice of self-categorization
dimensions and personal identity
categorizations. Further, many of them
confirmed our placements of their colleagues.
Based on this feedback, we felt very confident
in our descriptions of these 30 designers'
personal identity categorizations and that our
coding scheme would produce accurate
results."
Definition: "We used past research as a guide
to label instances randomly selected from the
interview data. We came to agree upon a
framework of six common collaborative
behaviours described by designers: two
related to giving ideas (i.e. offering ideas and
promoting ideas), three related to taking ideas
(i.e. soliciting ideas, considering ideas, and
incorporating ideas), and one involved taking
and giving (co-creating) concurrently."
Findings: "Our findings suggest that the
positive impact of the collaborative behaviour
of incorporating ideas may depend, in part, on
the types of ideas being incorporated."
Managerial implications: "Promoting the
behaviour of idea-taking (not idea-giving) may
be what is critical to improving creative
collaborations. These findings make clear that
a one-size-fits-all approach to designing and
promoting creative collaborations is unlikely to
work. "Our findings suggest that, rather than
rewarding employees for offering 'the most
creative idea', organizations might offer
rewards to employees who are able to
effectively incorporate others' ideas into their
own work. Further, training programmes may
need to be updated to include education on
how to effectively take ideas from others, in
addition to how to give ideas. Effectively
leading creative collaborations may mean
preventing any single group member from
'owning' an idea in the early stages."
Common
theme:
"Building
on the
ideas of
others"
organization might be better off relying on
asynchronous idea generation by individuals."
The importance of "idea build-up" is recognised in both studies, albeit they approach it from
different angles, and recognise the need for further research on this theme. How to reason
about "idea build-up" and how to tackle it in systematic studies? How do 'idea-taking' and
'idea-giving' shape creative collaboration? How may the definition of "idea build-up" help
interpret the outcomes of ideation teams? We propose here a mixed-method approach using
computational modelling where a variety of lenses can be worn to look at idea build-up.
An agent-based simulation of creative collaboration
An influential simulation model of idea transmission is extended here that aims at examining
"just how much of cultural emergence and stability can be explained without resorting to
centralized authority" (Axelrod, 1997). In that model of cultural convergence (i.e.,
imitation), agents share information by exchanging values between neighbours. These
stochastic systems gradually converge either to a single value (system ergodicity), or they
form 'regions' or clusters when agent interaction is conditioned by compatibility
(neighbouring agents that share at least one value at initial time). This simple model has
been widely extended, including in studies of creativity and innovation (Araujo & Mendes,
2009; Kiesling, Günther, Stummer, & Wakolbinger, 2012; Leydesdorff, 2002; Sosa & Gero,
2005; Watts & Gilbert, 2014).
3.1 The Model
Axelrod's description of "the entire dynamics of the system" consists of two steps1: pick a
random agent to be active and one of its neighbours, and with probability equal to their
similarity (shared values), these agents interact by selecting at random a feature on which the
active agent and its neighbour differ (if any), as a result the active agent changes its trait on
this feature to the neighbour's trait on this feature (Axelrod, 1997). The model is usually
analysed consisting of hundreds of agents arranged in a two-dimensional grid initialised with
randomised cultural values, where 'culture' is defined by an array of integer variables (of
size = feature length) and the domain of these variables (traits) typically {0, 1, 2… 9}.
Notice that in the original formulation of this model, all active agents are defined as idea-
taking agents (Elsbach & Flynn, 2013).
The extension to this model presented here consists of incorporating a type of change agency
based on dissent or divergent behaviour. Applying the MDC framework of creativity (Sosa
& Gero, 2015), dissent is modelled at the societal level (MDC-S) as a standard behaviour
where every agent that identifies total group convergence attempts to introduce a new idea
1 "At random, pick a site (cell) to be active, and pick one of its neighbors. With probability equal to their
cultural similarity, these two sites interact. An interaction consists of selecting at random a feature on which the
active site and its neighbour differ (if there is one) and changing the active site's trait on this feature to the
neighbor's trait on this feature" (Axelrod, 1997)
by changing its variables to random values (with a low probability of success of 0.1%).
Inspired by the two ideation studies analysed above, we distinguish between idea-giving and
idea-taking behaviours at the group level (MDC-G) by defining a ratio from each type in the
formation of groups at initial time in the simulation. The specific aim here is to understand
the effects of group composition across different ratios of 'idea takers' and 'idea givers'
interacting.
Previous studies have shown that incorporating dissent or divergent behaviour in the original
Axelrod model triggers cycles of collective convergence -or consensus on a dominant value
followed by sudden episodes of collective change where groups adopt new values (Sosa &
Gero, 2005). This paper investigates the effects of different ratios of idea-taking and idea-
giving agents in a group. To this end, we repeatedly run the model to obtain average values
for groups where all agents are 'idea-takers' (as the original formulation), and gradually
introduce idea-giving agents until groups where all agents are 'idea-givers'. One specific
dependent variable is measured: the number of collective changes of the dominant value,
which we call a revolution (Kuhn, 1996). In this context, a revolution is registered whenever
the entire group of agents adopts a value that was introduced by a dissenting agent. The
model analysed here consists of 9 agents in a 3x3 torus grid. Following Axelrod's lexicon,
ideas are encoded in 6 features, with 9 traits, with 'Von Neumann' neighbourhoods, and
initialised in fully converged mode, i.e., all agents adopting the same ideas at initial state.
Results are obtained by averaging 103 cases running for 104 simulation steps.
Results
Group composition is inspected from "all takers" (9t) to "all givers" (9g) in the group, in
increments of one member each time. The effects of group composition are not linear -as
shown in Figure 1. First, revolutions (change episodes) are highest in groups where all active
agents are idea-takers, or "rev_9t" in Figure 1 (as in Axelrod's original model). Introducing
idea-giving agents in the group affects the indicator of group creativity in this model by first
causing a dip in revolutions when one idea-giver agent is introduced ("rev_8t1g" in Figure
1), then increasing as a majority of agents engage in idea-giving behaviour.
Figure 1. Quantitative analysis of the idea-taking and idea-giving model (averages)
A way to interpret these results is to consider that a "9t" scenario is only a theoretical
construct, as research shows that all members of creative design teams engage in idea-giving
(Elsbach & Flynn, 2013). It is also worth noting that the increase in group creativity also
creates a different type of revolutions: when new ideas gain dominance, it is possible to
analyse whether they are adopted as introduced by the change agent, or whether they are the
product of combinatorial contributions from multiple team members. When only a few idea-
giving agents are present in the team and the level of revolutions is low, change agents tend
to retain 'ownership' of new ideas, whereas in groups with more idea-giving agents where
revolutions take place more often, winning ideas are almost by definition result of 'idea
build-up'. This seems like a fertile ground for co-creation, but it could also lead to negative
effects such as groupthink (Rose, 2011) in the absence of appropriate facilitation strategies.
These quantitative results give a general overview of the effects of idea-taking and idea-
giving when many cases are averaged. However, it is far from clear from looking at these
results what may explain the significant dip in group creativity with the introduction of one
idea-giving team member. Why are (hypothetical) teams of all idea-takers able to trigger
collective change more easily than teams where only one idea-giver agent is inserted? And
why is this result reversed as more idea-givers are introduced in a team?
To examine these questions, we adopt a qualitative lens to look at a specific case across three
conditions: when all team members engage in idea-taking ("rev_9t"), when only one idea-
giver joins the team ("rev_8t1g"), and when three idea-giver agents interact ("rev_6t3g").
Figure 2 shows a specific revolution episode from each of these three conditions that
illustrates the type of revolutions that occur. All episodes start with a new idea being
introduced breaking group convergence: each frame shows a simulation step, nine agents in
a torus 3x3 grid, each agent has 3 features and 9 possible traits per feature. In rev_9t, all
agents have adopted idea [000] when agent #4 introduces [626]. In rev_8t1g, all agents hold
[000] when agent #7 introduces [238], and in rev_6t3g, agent #2 introduces [682]. In rev_9t,
the second feature of the new idea spreads through the group, reaching dominance as [020]
in a few steps. In rev_8t1g, the new idea starts spreading as three composite ideas [200],
[030], and [230], and all agents switch to one of these variants, except for agent #0, which
sticks to [000], i.e., this is the agent fixed on idea-giving behaviour. In rev_6t3g, the new
idea [682] spreads as [600], [002], and [602], ultimately gaining dominance as [002] in the
group.
Figure 2. Qualitative analysis of the idea-taking and idea-giving model (one selected case)
These results are suggestive of the complexities of group creativity: when all agents in a
team engage in idea-taking, the probability that a new value be spread is high because at any
given iteration step, when agents interact with their neighbours, active agents take the new
'dissenting' value from neighbours. However, when an idea-giving agent is introduced, it
controls the process and becomes a 'gatekeeper' by virtue of being the only agent in the
group to give ideas to others. Such monopoly is broken as soon as another team member has
access to idea-giving capabilities, as both of them take ideas from others when they are in a
neighbouring (inactive) position of the other 'idea-giving' (active) agent. In teams of 'all
givers', the 'taking' role is performed by all agents when in the role of neighbours of active
agents.
Thus we can see that, other than the original Axelrod conditions of fully imitative team
members, this model shows that teams with only a single idea-giving member support very
low levels of group creativity. The model also shows that a critical mass of idea-giving
behaviour may be sufficient to reach high levels of group creativity, after which more idea-
giving fails to increase group creativity beyond what appears to be a 'glass ceiling' or natural
group capacity. Moreover, the model shows why this occurs: teams with a single dominant
idea-giving role suffer from a bottleneck enforced by an agent that indirectly adopts a
'change resistance' role. Rather than removing that type of agent from a team, our model
suggests that adding another idea-giving agent helps to balance the interactions and rapidly
increases the creative potential of the group. After a critical mass of about half the team,
adding more idea-giving agents may not have an impact on group creativity, and it may only
be beneficial if an increase in co-creation strategies is preferred.
Although these results cannot be generalised beyond this model, it is a good illustration of
the sensitivity of creative collaboration, and the large effects that even apparently small
changes can have. There are caveats to these results, with fundamental managerial
implications: whilst the dip between "all idea-takers" and "one idea-giver" is significant,
only 58% of cases show a decrease in number of revolutions, whilst 29% remain constant
and 13% in fact show an increase in this indicator of group creativity. This should be
considered a strength rather than a weakness of this simple model, as its stochastic nature is
reminiscent of the variability and unpredictable nature of creativity across cases.
Discussion
We have applied agent-based simulations as 'intuition pumps' to think quantitatively and
qualitatively about group creativity, and in particular 'idea build-up', or the building of new
ideas on the ideas of others. Big claims about the outcomes of this model in relation to real
world teams are explicitly avoided, as their role here is only to guide our intuitions. The
main lessons to think about creative groups are: a) group formation can be critical when
initial conditions define ranges of possible outcomes, and effects can be expected to be non-
linear; b) the balance between idea-giving and idea-taking is a delicate one in creative
collaborations, and a range of consequences should be considered both in research and
practice, such as idea ownership and idea decomposition; c) finally, this work shows that
agent-based models are useful as intuition pumps to reason about complex situations such as
creativity and innovation.
References
Araujo, T., & Mendes, R. V. (2009). Innovation and Self-Organization in a Multi-Agent Model. Advances in Complex
Systems, 12(02), 233–253. http://doi.org/10.1142/S0219525909002180
Axelrod, R. (1997). The Dissemination of Culture: A Model with Local Convergence and Global Polarization. Journal
of Conflict Resolution. http://doi.org/10.1177/0022002797041002001
Cardoso de Sousa, F. (2007). Still the elusive definition of creativity. International Journal of Psychology: A Bio
Psychosocial, 2, 55–82.
Creswell, J. W. (2009). Research design: Qualitative, quantitative, and mixed methods approaches. Research Design
qualitative quantitative and mixed methods approaches (Vol. 3rd). http://doi.org/10.1016/j.math.2010.09.003
Dennett, D. C. (2014). Intuition Pumps And Other Tools for Thinking. W. W. Norton & Company. Retrieved from
https://books.google.com/books?hl=en&lr=&id=9SduAwAAQBAJ&pgis=1
Diehl, M., & Stroebe, W. (1987). Productivity loss in brainstorming groups: Toward the solution of a riddle. Journal of
Personality and Social Psychology. http://doi.org/10.1037/0022-3514.53.3.497
Elsbach, K. D., & Flynn, F. J. (2013). Creative Collaboration and the Self-Concept: A Study of Toy Designers.
Journal of Management Studies, 50(4), 515–544. http://doi.org/10.1111/joms.12024
Francis, D., & Bessant, J. (2005). Targeting innovation and implications for capability development. Technovation,
25(3), 171–183. http://doi.org/10.1016/j.technovation.2004.03.004
Fruchter, R. (2001). Dimensions of teamwork education. International Journal of Engineering Education, 17(4 and 5),
426–430.
Girotra, K., Terwiesch, C., & Ulrich, K. T. (2010). Idea Generation and the Quality of the Best Idea. Management
Science, 56(4), 591–605. http://doi.org/10.1287/mnsc.1090.1144
Grace, K., Maher, M. L., Fisher, D., & Brady, K. (2014). Modeling expectation for evaluating surprise in design
creativity. In Design Computing and Cognition'14 (pp. 189–206). Springer International Publishing.
Han, J., Han, J., & Brass, D. J. (2014). Human capital diversity in the creation of social capital for team creativity.
Journal of Organizational Behavior, 35(1), 54–71. http://doi.org/10.1002/job.1853
Haydari, S., & Smead, R. (2015). Does Longer Copyright Protection Help or Hurt Scientific Knowledge Creation?
Journal of Artificial Societies and Social Simulation2, 18(2), 23.
Iyer, L. R., Doboli, S., Minai, A. A., Brown, V. R., Levine, D. S., & Paulus, P. B. (2009). Neural dynamics of idea
generation and the effects of priming. Neural Networks : The Official Journal of the International Neural
Network Society, 22(5-6), 674–86. http://doi.org/10.1016/j.neunet.2009.06.019
Jordanous, A. (2012). A Standardised Procedure for Evaluating Creative Systems: Computational Creativity
Evaluation Based on What it is to be Creative. Cognitive Computation, 4(3), 246–279.
http://doi.org/10.1007/s12559-012-9156-1
Kahl, C. H., & Hansen, H. (2015). Simulating Creativity from a Systems Perspective: CRESY. Journal of Artificial
Societies and Social Simulation, 18(1), 4. Retrieved from http://jasss.soc.surrey.ac.uk/18/1/4.html
Kiesling, E., Günther, M., Stummer, C., & Wakolbinger, L. M. (2012). Agent-based simulation of innovation
diffusion: A review. Central European Journal of Operations Research. http://doi.org/10.1007/s10100-011-
0210-y
Kohn, N. W., Paulus, P. B., & Choi, Y. (2011). Building on the ideas of others: An examination of the idea
combination process. Journal of Experimental Social Psychology, 47(3), 554–561.
http://doi.org/10.1016/j.jesp.2011.01.004
Kuhn, T. S. (1996). The structure of scientific revolutions. The University of Chicago Press, 3rd Ed. Retrieved from
papers2://publication/uuid/C5D275F2-36F4-44E2-BA6E-BCEF8D99CE3B
Leijnen, S., & Gabora, L. (2010). An agent-based simulation of the effectiveness of creative leadership. In
Proceedings of the Annual Meeting of the Cognitive Science Society. Portland, Oregon.
Leydesdorff, L. (2002). The complex dynamics of technological innovation: a comparison of models using cellular
automata. Systems Research and Behavioral Science, 19(6), 563–575.
Mahoney, J., & Goertz, G. (2006). A tale of two cultures: Contrasting quantitative and qualitative research. Political
Analysis, 14(3), 227–249. http://doi.org/10.1093/pan/mpj017
Malik, A., Crooks, A., Root, H., & Swartz, M. (2015). Exploring Creativity and Urban Development with Agent-
Based Modeling. Journal of Artificial Societies and Social Simulation, 18(2), 12.
Osborn, A. F. (1963). Applied Imagination: Principles and procedures of creative problem solving. Oxford.
Paulus, P. B., & Nijstad, B. A. (2003). Group Creativity: Innovation through Collaboration. Work (Vol. 12).
http://doi.org/citeulike-article-id:975273
http://doi.org/10.1016/S1572-0977(06)12002-6
Rose, J. D. (2011). Diverse Perspectives on the Groupthink Theory – A Literary Review. Emerging Leadership
Journeys, 4, 37–57. Retrieved from
http://www.regent.edu/acad/global/publications/elj/vol4iss1/Rose_V4I1_pp37-57.pdf
Shepperd, J. A. (1993). Productivity loss in performance groups: A motivation analysis. Psychological Bulletin.
http://doi.org/10.1037/0033-2909.113.1.67
Shin, S. J., Kim, T. Y., Lee, J. Y., & Bian, L. (2012). Cognitive team diversity and individual team member creativity:
A cross-level interaction. Academy of Management Journal, 55(1), 197–212.
http://doi.org/10.5465/amj.2010.0270
Sobkowicz, P. (2015). Innovation Suppression and Clique Evolution in Peer-Review-Based, Competitive Research
Funding Systems: An Agent-Based Model. Journal of Artificial Societies and Social Simulation2, 18(2), 13.
Sosa, R., & Gero, J. S. (2005). A Computational Study of Creativity in Design: The Role of Society. Artificial
Intelligence for Engineering Design, Analysis and ManufacturingAI EDAM.
http://doi.org/10.1017/S089006040505016X
Sosa, R., & Gero, J. S. (2015). Multi-dimensional creativity: a computational perspective. International Journal of
Design Creativity and Innovation, 1–25. http://doi.org/10.1080/21650349.2015.1026941
Sosa, R., Gero, J. S., & Jennings, K. (2009). Growing and destroying the worth of ideas. In Proceeding of the seventh
ACM conference on Creativity and cognition - C&C '09 (p. 295). New York, New York, USA: ACM Press.
http://doi.org/10.1145/1640233.1640278
Stahl, G. K., Maznevski, M. L., Voigt, A., & Jonsen, K. (2010). Unraveling the effects of cultural diversity in teams: A
meta-analysis of research on multicultural work groups. Journal of International Business Studies.
http://doi.org/10.1057/jibs.2009.85
Sutton, R., & Hargadon, A. (1996). Brainstorming Groups in Context : Effectiveness in a Product Design Firm.
Administrative Science Quarterly, 41(4), 685–718. http://doi.org/10.2307/2393872
Watts, C., & Gilbert, N. (2014). Simulating Innovation: Computer-based Tools for Rethinking Innovation. Edward
Elgar Publishing.
Author Biographies
Ricardo Sosa
Ricardo combines a creative background as a designer with a passion for the study of
computational systems. He studies creativity and innovation principles through multi-agent
social systems and is involved in the development of facilitation practices for team ideation
and for participatory decision making. Ricardo partners with colleagues across disciplines
including: robotics, social science, cognitive science, architecture, arts, engineering,
business, public health, and computer science: https://colab.aut.ac.nz/staff/ricardo-sosa
Andy Connor
Andy is a mechanical engineer by training but has a breadth of experience in mechatronics,
software engineering, computer science and more recently in creative technologies. Andy
has a broad range of research interests that include automated design, computational
creativity, education, evolutionary computation, machine learning and software engineering:
https://www.aut.ac.nz/profiles/creative-technologies/senior-lecturers/andy-connor
|
1603.07766 | 1 | 1603 | 2016-03-24T21:59:32 | Simulation Platform for Multi Agent Based Manufacturing Control System Based on The Hybrid Agent | [
"cs.MA"
] | Agent based distributed manufacturing control and scheduling systems are subsets of new manufacturing systems. Multi agent systems (MAS) not only drive design and engineering control solutions but also influence flexibility, agility, and re-configurability, which makes MASs a better centralized systems than its traditional counterparts. However, implementation of all MASs in the real factories are timely, also extremely costly. A simulation environment that would allow independent development and testing of the services and business processes of the related manufacturing hardware is needed. This paper presents the design and implementation of a userfriendly simulation platform for multi agent based manufacturing control systems by considering the shop floor level. The proposed simulation platform can simulate the software level of the factory by considering the hardware level of the mentioned factory. An example of the simulation platform is presented for a flexible manufacturing system, which is located in EMU CIM lab. | cs.MA | cs | CIE45 Proceedings, 28-30 October 2015, Metz / France
SIMULATION PLATFORM FOR MULTI AGENT BASED MANUFACTURING CONTROL SYSTEM
BASED ON THE HYBRID AGENT
Ali Vatankhah barenji1, Amir Shaygan2, Reza Vatankhah Barenji2
1Department of Mechanical Engineering Eastern Mediterranean University, Famagusta, North
Cyprus, Turkey
2Department of Industrial Engineering Hacettepe University, Beytepe Campus Ankara, Turkey
[email protected]
ABSTRACT
Agent based distributed manufacturing control and scheduling systems are subsets of new
manufacturing systems. Multi agent systems (MAS) not only drive design and engineering control
solutions but also influence flexibility, agility, and re-configurability, which makes MASs a better
centralized systems than its traditional counterparts. However, implementation of all MASs in the
real factories are timely, also extremely costly. A simulation environment that would allow
independent development and testing of the services and business processes of the related
manufacturing hardware is needed. This paper presents the design and implementation of a user-
friendly simulation platform for multi agent based manufacturing control systems by considering
the shop floor level. The proposed simulation platform can simulate the software level of the
factory by considering the hardware level of the mentioned factory. An example of the simulation
platform is presented for a flexible manufacturing system, which is located in EMU CIM lab.
Keywords: Multi-agent distributed manufacturing control system, Agility, Simulation
1. INTRODUCTION
The era of "agile manufacturing", "lean manufacturing" and "Intelligent manufacturing" in
recent years, has brought worldwide competition among the manufacturing enterprises. This
competition between the stockholders in global market intensively returns to scheduling and
control system performance of the production system. Recent developments in MASs in
manufacturing control field can solve the existing problems. In order for the aforementioned
problems to be solved, more flexibility and agility are required as mentioned by Marik et al. [1].
MASs offers an alternate approach to design and managing of the control systems by providing
modularity, robustness and autonomy. Jennings et al. [2] has indicated that at least 25% of the
manufacturing problems can be solved.
Jennings [2] also mentions that the software required for centralized methods is more complex
than of those needed by the agents based approaches in terms of facilitating development,
debugging and maintenance. However, the main problems of MASs are analyzing, testing, and
validating of the behavior of the agent-based systems. Analyzing the MASs in manufacturing
control not only depends on the software level but also on the hardware level in the shop floor.
The design phase which takes place before the deployment of the resulted design in the real
operation is usually an arduous and time consuming task. Therefore, simulation tools are required
CIE45 Proceedings, 28-30 October 2015, Metz / France
in order to support the correction of misunderstandings and errors before implementation as
mentioned by Luke et al.[3].
Several environments are reported for the simulation of multi agent systems in the literature such
as Vrba [4-6]. Barbosa et al. [7] used an agent based modeling platform for the simulation of multi
agent manufacturing systems .However; these platforms are developed for specific cases and
according to the application particularities. This means that simulating the behavior of agent
based manufacturing systems for newly developed systems needs significant effort. In addition,
there is a lack of standard platform with the ability to analyze multi agent based manufacturing
control system in both hardware and software levels. This paper proposes an architecture for
simulating multi agent based manufacturing control system in the shop floor based on Hybrid
Agent (HA). The HA is in simultaneous communication with MASs and shop floor control system.
The remainder of this paper is structured as follows: Section 2 describes the architecture of
simulation platform. Section 3 explains existing multi agent system. Section 4 explains a case
study and discusses the results. Finally, Section 5 concludes the paper and presents future work.
2. ARCHITECTURE OF THE SIMULATION PLATFORM
Virtual reality is one of the ways to simulate manufacturing control systems by considering shop
floor. However due to the extended development time associated with the virtual resources of
each machine, analyzing this type of system in an actual factory setting is infeasible according to
Lin et al.[8]. Additionally, based on Barenji [9],when comparing alternate scheduling systems, it
is extremely intricate to replicate a similar condition for tentative tests .A prototype simulation
that acts as an actual system can be used to overcome this problem. Moreover, it is indistinctly
linked to the simulation platform or the actual system to control the actual system. In this
research, an integrated simulation system i.e., "Simulation Platform" has been developed,
following the approach of AV Barenji et al. [10] for testing MASs. The architecture of simulation
platform is shown in Fig1.
Figure 1 Architecture of Simulation Platform
CIE45 Proceedings, 28-30 October 2015, Metz / France
The Simulation Platform consist of three main parts:
1. The Hardware Simulation Agent (HSA): HSA is a software which can design the hardware
level of the factory by considering characterized physical actions that take place in the
manufacturing environment. Examples of HSA are Arena, CPN tool, etc.
2. Multi agent based manufacturing control system: This part is based on the "Manufacturing
Execution System" (MES) which is responsible for operation scheduling, production
dispatching, work in process status, data acquisition, etc.
3. Hybrid agent: HAs are computational processes capable of moving throughout a network,
interacting with foreign hosts, gathering information on behalf of the user, and returning
to the user after performing their assignments. Characteristic of the HA are : a) Hybrid
agent is capable of real time communication with other agent, b) HA is able to understand
the decision taken by MASs, c) Hybrid agent is capable of communicating with HSA via XML
(Extensible Markup Language).
The Multi agent based manufacturing control system are used to implement different kinds of
MAS architecture with hybrid agents. Furthermore, the HSA characterizes the physical actions
that take place in the manufacturing environment. This part is supported by many kinds of
software such as Arena, PN tools, Flexisim and etc. The most characteristics of HSA software just
ability to communication with XML. The primary feature of a simulation platform is to support
the examination of any control system when applied to a discrete manufacturing system. Colored
Petri Nets (CPN) which is a mathematical and graphical language for the design, specification,
simulation, and verification of discrete systems can be named as an example of hardware
simulation platform (HSP) software based on Tsinarakis et al. [11]. CPN and their extensions are
widely used in modeling discrete-event dynamic systems, including production systems and
networks. One of these tools that is used broadly is the CPN Tools which is a simulation software
for the altering, simulating and analyzing Colored Petri Nets. This application can communicate
via XML.
3. MULTI-AGENT SYSTEMS
An RFID (Radio Frequency Identification System) based multi-agent scheduling and control
architecture has been developed in authors previous works named RFIDMASs in Barenji et al. [12-
15].It should be mentioned that RFID are systems that transfer data with the aid of
electromagnetic fields in order to automatically trace and identify the tagged entities. The
proposed multi-agent system is designed as a network of software agents that interact with each
other and with the system actors. These agents are categorized as shop management agents,
agent managers, shop monitoring and command agents, station control agents, station monitoring
agents, agent machine interfaces, and manufacturing resource agents. In addition to the agents
forming the architecture, two databases exist in the architecture; a shop database and a station
database. Furthermore, based on Barenji et al.[16], ontology (a capability-based knowledge
model) is required for proper communication between the agents.
CIE45 Proceedings, 28-30 October 2015, Metz / France
To illustrate the functionality of the proposed simulation platform for performance measurement
of RFIDMASs, the Flexible Manufacturing Systems Laboratory (FMS Lab.) located in the Eastern
Mediterranean University (EMU) is used as the case study.
4. CASE STUDY
The FMS laboratory of EMU consists of three stations: Station 1 is a machine tending station, which
consists of a CNC milling machine and a five-axis vertically articulated robot. Station 2 is an
assembly and quality control station, gluing machine, and laser-scan micrometer device. Station
3 is an automatic storage and retrieval system (AS/RS). A conveyer integrates the stations for
performing material handling within the cell. RFIDMAS is selected as manufacturing execution
system and CPN Tools version 4 is chosen as HSA. Fig2 shows the CPN model of the FMS lab.
Figure 2 CPN Model of the FMS Lab
4.1 Hybrid agents
As mentioned in the part 2, hybrid agents are computational processes which are capable of
maneuvering in a network, communicating with alien hosts, collecting data vicariously for the
user, and returning to the user after performing its duty. Furthermore, hybrid agents have the
ability of administrating the information existing in networks. Some of hybrid agents' features
and capabilities are the real time inter-agent communication as well as their capability to
comprehend the decision making process done by RFIDMASs. Additionally they possess the ability
to exchange information with other software using XML. Fig 3 shows the sequence diagram for
starting a new task in the system and illustrates the interaction between agents.
Figure 3 Sequence Diagram for Starting a New Task in the System
New task is sent to the HA via an XML file in order to communicate with shop-level agents. Shop-
level agents consist of three types of agents, namely SMA, SMCA, and AM. A shop-level agent sends
CIE45 Proceedings, 28-30 October 2015, Metz / France
a request to a database agent for obtaining new data, and the database agent responds to this
request. If the answer is positive, shop-level agents send a further request to the station agents
checking for availability and possibility of doing a new job. If the answer to this request is positive
as well, the shop-level agent accepts this job, sends a comment to the database agents, and the
database agents send requirement data to the station agents. The station agent performs its duty
via communication between sub-agents.
The hardware simulation platform (HSP), as mentioned before, defines the tangible actions taking
place in the manufacturing environment. The communication model is set up in CPN Tools
software, which creates the real time communication between the simulation part and multi
agent control part. The XML codes for description of agents, MASs and resources are shown in the
Fig4.
Figure 2 XML Codes for Description of Agents, MASs and Resources
The experiments is carried out by considering two different scenarios: (A) no disturbances and a
well-functioning system and (B) incidents of failures in station one e.g. CNC, with a probability
of 20%. In the experiments, setup time is not considered and it is assumed that negotiation
failure would never occur. A robot and conveyor perform the transportation operations and
orders are queued and executed in the arrival of the order. Each action of transportation takes
8 seconds. The process time for the CNC for each part is 10 seconds. Additionally, the process
time of the assemble machine for the final product is 15 seconds. Each individual book of orders
involves the production of 3 parts: 1 bodies, 1 handle, and 1 cover. The simulation-based
reported test reflects 1000 book orders.
Performance indicators:
Manufacturing lead time: The total time required to manufacture an item, including order
preparation time, queue time, setup time, process time, move time, inspection time, and
put-away time.
Throughput: An indicator of the productivity of a manufacturing system, defined here as
the number of items produced per time unit.
Repeatability: The mean value of the standard deviation of the percentage of utilization
of all resources of the system over several runs.
CIE45 Proceedings, 28-30 October 2015, Metz / France
5. RESULTS AND DISCUSSION
The results from the simulation platform allowed us to draw some conclusions concerning the
operation of the RFIDMASs in the FMS by taking into account the hardware and softer level. The
system was found to function robustly and as specified in both normal operations and in the
presence of disturbances. Furthermore, the re-configurability of the system is demonstrated by
its accurate reactions to the introduction, removal, and modification of manufacturing
components. In particular, it is shown that when a resource control agent broke down or was
removed from the system, other agents continued to find alternative solutions for executing the
production plan. Fig 5 shows the results for the stable scenario, without the occurrence of
unexpected disturbances. In the stable scenario, the RFIDMASs system for the FMS yielded smaller
values for the manufacturing lead-time (198) and higher values for the throughput (68), in
comparison to those obtained using a conventional control system. The better performance of the
proposed system is resulted by the cooperation of autonomous entities, i.e., an agent manger
that elaborates optimized production plans.
Figure 3 Stable Scenario
The results for the second scenario is summarized in Fig6. The first conclusion drawn from these
simulation results is that the values of all performance indicators decreased in the presence of
disturbances. An analysis of the lead times and throughputs confirmed that the RFIDMASs
nonetheless yielded better performance than the conventional control systems.
Disturbances increase the entropy and unpredictability of a manufacturing control system. The
implementation of a multi-agent scheduling approach improves system performance by improving
the system's ability to respond to disturbances, as indicated by the smaller values of the
manufacturing lead-time and higher values of the throughput than those obtained using the
conventional scheduling control approach. The results indicate that the proposed system can
achieve good productivity even when resource interruptions increase and that it can respond to
resource breakdowns effectively. Analysis of the experimental results confirms that the use of
the proposed multi-agent scheduling approach results in better resource utilization than the
conventional approach in both stable and unstable scenarios.
CIE45 Proceedings, 28-30 October 2015, Metz / France
Figure 4 Disturbances Scenario
6. CONCLUSION
This paper proposes simulation platform for MASs based on the hybrid agent communication via
XML coded with other software. A simulation-based performance assessment is performed to
investigate the effectiveness of the multi-agent scheduling approach in comparison to a
conventional scheduling approach. A case study of a flexible assembly system in a medium-sized
factory was conducted. The multi-agent scheduling and production control system was tested
over an integrated experimental test bed developed based on a simulation model of the real
factory, integrated with an external multi-agent-based control system platform. The results of
the simulation platform for the case study show that MASs for FMS create more re-configurability
and agility. Future work will include the development of more user-friendly engineering tools for
modeling and simulation of multi-agent manufacturing control system software tools and Petri
net engineering platforms and achievement of a smooth migration from virtual scenarios to real
systems.
7. REFERENCES
[1] V. Marik and D. McFarlane, "Industrial adoption of agent-based technologies," IEEE Intelligent
Systems, vol. 20, pp. 27-35, 2005.
[2] N. R. Jennings and S. Bussmann, "Agent-based control systems," IEEE control systems, vol.
23, pp. 61-74, 2003.
[3] S. Luke, C. Cioffi-Revilla, L. Panait, K. Sullivan, and G. Balan, "Mason: A multiagent
simulation environment," Simulation, vol. 81, pp. 517-527, 2005.
[4] P. Vrba, "MAST: manufacturing agent simulation tool," in Emerging Technologies and Factory
Automation, 2003. Proceedings. ETFA'03. IEEE Conference, 2003, pp. 282-287.
[5] P. Vrba and V. Marík, "Capabilities of dynamic reconfiguration of multiagent-based industrial
control systems," Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions
on, vol. 40, pp. 213-223, 2010.
[6] Barenji, Ali Vatankhah, and Canberk Değirmenci. "Robot Control System based on Web
Application and RFID Technology." In MATEC Web of Conferences, vol. 28, p. 04001. EDP Sciences,
2015.
CIE45 Proceedings, 28-30 October 2015, Metz / France
[7] J. Barbosa and P. Leitão, "Simulation of multi-agent manufacturing systems using Agent-
Based Modelling platforms," in Industrial Informatics (INDIN), 2011 9th IEEE International
Conference on, 2011, pp. 477-482.
[8] Barenji, Ali Vatankhah, Reza Vatankhah Barenji, and Majid Hashemipour, "Flexible testing
platform for employment of RFID-enabled multi-agent system on flexible assembly line." Advances
in Engineering Software 91 (2016): 1-11.
[9] R. V. Barenji, "Towards a capability-based decision support system for a manufacturing
shop," in Collaborative Systems for Reindustrialization, ed: Springer, 2013, pp. 220-227.
[10] P. Leitão, J. M. Mendes, A. Bepperling, D. Cachapa, A. W. Colombo, and F. Restivo,
"Integration of virtual and real environments for engineering service-oriented manufacturing
systems," Journal of Intelligent Manufacturing, vol. 23, pp. 2551-2563, 2012.
[11] G. J. Tsinarakis, N. Tsourveloudis, and K. P. Valavanis, "Modular Petri net based modeling,
analysis, synthesis and performance evaluation of random topology dedicated production
systems," Journal of Intelligent manufacturing, vol. 16, pp. 67-92, 2005.
[12] A. V. Barenji, "An RFID-based distributed control system for flexible manufacturing system,"
Eastern Mediterranean University (EMU)-Doğu Akdeniz Üniversitesi (DAÜ), 2013.
[13] A. V. Barenji, R. V. Barenji, and M. Hashemipour, "A frameworks for structural modelling
of an RFID-enabled intelligent distributed manufacturing control system," South African Journal
of Industrial Engineering, vol. 25, pp. 48-66, 2014.
[14] R. V. Barenji, A. V. Barenji, and M. Hashemipour, "A multi-agent RFID-enabled distributed
control system for a flexible manufacturing shop," The International Journal of Advanced
Manufacturing Technology, vol. 71, pp. 1773-1791, 2014.
[15] A. Vatankhah Barenji, R. Vatankhah Barenji, and M. Hashemipour, "Structural modeling of
a RFID-enabled reconfigurable architecture for a flexible manufacturing system," in Smart
Objects, Systems and Technologies (SmartSysTech), Proceedings of 2013 European Conference
on, 2013, pp. 1-10.
[16] R. Vatankhah Barenji, M. Hashemipour, and D. A. Guerra-Zubiaga, "A framework for
modelling enterprise competencies: from theory to practice in enterprise architecture,"
International Journal of Computer Integrated Manufacturing, pp. 1-20.
[17] Barenji, Ali Vatankhah, Reza Vatankhah Barenji, and Majid Hashemipour. "Flexible testing
platform for employment of RFID-enabled multi-agent system on flexible assembly line." Advances
in Engineering Software 91 (2016): 1-11.
|
1805.03103 | 1 | 1805 | 2018-05-08T15:24:50 | Ordinal Approximation for Social Choice, Matching, and Facility Location Problems given Candidate Positions | [
"cs.MA",
"cs.GT"
] | In this work we consider general facility location and social choice problems, in which sets of agents $\mathcal{A}$ and facilities $\mathcal{F}$ are located in a metric space, and our goal is to assign agents to facilities (as well as choose which facilities to open) in order to optimize the social cost. We form new algorithms to do this in the presence of only {\em ordinal information}, i.e., when the true costs or distances from the agents to the facilities are {\em unknown}, and only the ordinal preferences of the agents for the facilities are available. The main difference between our work and previous work in this area is that while we assume that only ordinal information about agent preferences in known, we know the exact locations of the possible facilities $\mathcal{F}$. Due to this extra information about the facilities, we are able to form powerful algorithms which have small {\em distortion}, i.e., perform almost as well as omniscient algorithms but use only ordinal information about agent preferences. For example, we present natural social choice mechanisms for choosing a single facility to open with distortion of at most 3 for minimizing both the total and the median social cost; this factor is provably the best possible. We analyze many general problems including matching, $k$-center, and $k$-median, and present black-box reductions from omniscient approximation algorithms with approximation factor $\beta$ to ordinal algorithms with approximation factor $1+2\beta$; doing this gives new ordinal algorithms for many important problems, and establishes a toolkit for analyzing such problems in the future. | cs.MA | cs |
Ordinal Approximation for Social Choice, Matching,
and Facility Location Problems given Candidate Positions
Elliot Anshelevich and Wennan Zhu
May 9, 2018
Abstract
In this work we consider general facility location and social choice problems, in which sets
of agents A and facilities F are located in a metric space, and our goal is to assign agents to
facilities (as well as choose which facilities to open) in order to optimize the social cost. We form
new algorithms to do this in the presence of only ordinal information, i.e., when the true costs or
distances from the agents to the facilities are unknown, and only the ordinal preferences of the
agents for the facilities are available. The main difference between our work and previous work in
this area is that while we assume that only ordinal information about agent preferences in known,
we know the exact locations of the possible facilities F. Due to this extra information about
the facilities, we are able to form powerful algorithms which have small distortion, i.e., perform
almost as well as omniscient algorithms but use only ordinal information about agent preferences.
For example, we present natural social choice mechanisms for choosing a single facility to open
with distortion of at most 3 for minimizing both the total and the median social cost; this factor
is provably the best possible. We analyze many general problems including matching, k-center,
and k-median, and present black-box reductions from omniscient approximation algorithms with
approximation factor β to ordinal algorithms with approximation factor 1 + 2β; doing this gives
new ordinal algorithms for many important problems, and establishes a toolkit for analyzing
such problems in the future.
1
Introduction
Many important problems involve assigning agents to facilities. For example, assigning patients
to hospitals, students to universities, people to houses, etc. The target of assignment problems is
usually to minimize social cost or maximize social welfare. When we consider the social cost of
assignment problems, it is natural to assume the agents prefer facilities that are "closer" to them in
some sense, thus the social cost of an agent is often represented by the distance between the agent
and the facility it is assigned to. Besides the cost of distances, there are many other cost functions
and constraints for different problems; for example, in the capacitated facility assignment problem,
each facility has a maximum number of agents it can accommodate.
In this work we consider general facility location problems, in which sets of agents A and facilities
F are located in a metric space, and our goal is to assign agents to facilities (as well as choose which
facilities to open) so that agents are assigned to facilities which are close to them. For example, F
may be possible locations for opening new stores, and the goal may be that all agents have a store
near them, or that the sum of agent distances to the stores they are assigned to is small, etc. This
setting also captures many social choice problems, in which the facilities correspond to candidates,
and the goal would be to choose a single candidate (and assign all agents to this candidate) so that
the distances from the agents to the chosen candidate are small. Here the distances correspond to
1
spatial preferences, i.e., the metric space represents the ideological space in which a more preferred
candidate would be closer to me; see [3, 18] for discussion of such spatial preferences in social
choice. Our setting also captures matching and many related problems, in which we would open all
facilities, but are only able to assign one agent to each facility, thus forming a matching between
agents and facilities; facilities here could correspond to houses or items, for example.
If the distances between agents and facilities are known, then we can calculate the optimal
solution for these assignment problems. Note that many of the facility location problems are NP-
Complete, but at least it is possible to compute optimum assignments of agents to facilities (or the
optimum candidates to select for social choice settings) given unlimited computational resources.
For many of the settings we mentioned above, however, it is unlikely that we know the exact
distances from the agents to the facilities. For social choice these distances would correspond to
the cardinal preferences of voters for candidates, for example, "My cost for candidate X winning is
exactly 2.35." It is far more common that only ordinal preferences of the agents for the candidates
are known, i.e., "I prefer X to Y". Similarly, when trying to form a matching, or even in general
facility location problems where we survey the agents to find out their preferences, it is much
easier to elicit ordinal preferences ("I prefer to be matched with X over Y") over precise numerical
preferences. These observations have recently led to a large body of work using the utilitarian
approach, in which we assume that some latent numerical costs or utilities exist, but we only
know the ordinal preferences of the agents, not their underlying numerical costs. See for example
[3, 4, 11, 16, 21, 23, 29] for the social choice setting, [1, 5–7] for matching and other graph problems,
and [13] for facility location. These works focus on measuring the distortion of various algorithms:
a measure of how well an algorithm behaves when using only ordinal information, as compared
to the optimum algorithm which has access to the true underlying numerical information. More
formally, the distortion [3, 28] of an assignment is defined as the worst-case ratio of its social cost
to the social cost of the optimal solution.
As in the work mentioned above, we assume that only ordinal information about the distances
between agents and facilities is known. However, although the locations and numerical preferences
of the agents are usually difficult to obtain, the locations of facilities are mostly public information.
The locations of political candidates in ideological space can be reasonably well estimated based
on their voting records and public statements. When forming a survey about new stores to open,
we may not know exactly how much the customers would prefer one store over the other since the
customer locations may be private, but the locations of the possible stores themselves are public
knowledge. The main difference between our work and previous work in this area is that we assume:
While only ordinal information about agent preferences in known, we know the exact locations of
the possible facilities F.
As we discuss below, this extra information about the locations of the facilities relative to each
other allows us to produce much stronger algorithms, and show much nicer bounds on distortion.
In fact, in many cases, we do not even need the full information about the locations of the facilities.
The main message of this paper is that having a small amount of information about the candidates
in social choice settings, or the facilities in facility location, allows us to obtain solutions which are
provably close to optimal for a large class of problems even though the only information we have
about the agent preferences is ordinal, and thus it is impossible (even given unlimited computational
resources) to compute the true optimum solution.
2
1.1 Our Contributions
We begin by looking at the social choice setting, in which we have agents A and candidates F in
a metric space, and we are given an ordinal ranking of each agent for the candidates. This setting
was considered in e.g., [3,4,16,21,23,24,29]. In particular, for the objective of minimizing the total
distance from the agents to the chosen candidate, [3] showed that Copeland and similar voting
mechanisms always have distortion of at most 5, while no deterministic voting mechanism can
achieve a worst-case distortion of less than 3. Finding a deterministic mechanism with distortion
less than 5 has been an open problem for several years [23]. In this paper, we show that if we
know the exact locations of the candidates in addition to the ordinal ranking of the agents, then
there is a simple algorithm which achieves a distortion of 3, and no better bound is possible. In
other words, while we do not know the true distances from agents to candidates, we can compute
an outcome which is a 3-approximation no matter what the true distances are, as long as they are
consistent with the ordinal preferences given to us. Moreover, this approximation is possible even
if for each agent we are only given their favorite (i.e., top-choice) candidate: there is no need for
the agents to submit a full preference ranking over all the alternatives.
We also study other objective functions in addition to minimizing the total distance from agents
to the chosen alternative. We give a natural deterministic voting mechanism which has distortion
at most 3 for objectives such as minimizing the median voter cost, the egalitarian objective of
minimizing maximum voter cost, and many other objectives. This mechanism achieves all these
approximation guarantees simultaneously, and moreover it does not need the exact locations of the
candidates: it suffices to be given an ordinal ranking of the distances from each candidate to each
other candidate. In other words, this mechanism is especially suitable for the case when candidates
are a subset of voters, as our mechanism will obtain the ordinal ranking of each voter for all the
candidates, and this is the only information which would be required. Note that [3] proved that
no deterministic mechanism can achieve a distortion of better than 5 for the median objective; the
reason why we are able to achieve a distortion of 3 here is precisely because we also know how each
candidate ranks all the other candidates, in addition to how each voter ranks all the candidates.
We then proceed to our general facility assignment model. We are given a set of agents and a set
of facilities in a metric space. The distances between facilities are given, but the distances between
agents and facilities are unknown; instead we only know ordinal preferences of agents over facilities
which are consistent with the true underlying distances. There could be arbitrary constraints on
the assignment, such as facility capacities, or constraints enforcing that some agents cannot be (or
must be) assigned to the same facility, etc. A valid assignment is to assign each agent to a facility
without violating the constraints. We consider many different social cost functions to optimize.
For a general class of cost functions (essentially ones which are monotone and subadditive), we
give a black-box reduction which converts an algorithm for the omniscient version of this problem
(i.e., the version where the true distances are known) to an ordinal algorithm with small distortion.
Specifically, if we have an omniscient algorithm which always produces an assignment which is a
β-approximation to the optimum, then using it we can create an ordinal algorithm which only
knows the ordinal preferences of the agents instead of their true distances to the facilities, but has
distortion of at most 1 + 2β.
Many well-known problems fall into our facility assignment model; Table 1 summarizes some
of our results. For example, classic facility location with facility costs, minimum weight bipartite
matching, egalitarian bipartite matching, k-center, and k-median are all special cases. In particular
our results show that if we are given unbounded computational resources, then it is always possible
to form an assignment with distortion of at most 3 for these problems, and no better bound is
possible simply due to the fact that we do not possess all the relevant information to compute the
3
Omniscient: Agents' ordinal prefs Only agents' ordinal
prefs (lower bounds)
full distances
and facility locations
Total (Sum) Social Choice
Median Social Choice
Min Weight Bipartite Matching
Egalitarian Bipartite Matching
Facility Location
k-center
k-median
1
1
1
1
1.488 [27]
2 [25]
2.675 [12]
3
3
3
3
3.976
5
6.35
5(3)
5(5)
n(3)
-(2)
∞ (∞)
- (-)
- (Ω(n))
Table 1: Best known distortion of polynomial-time algorithms in different settings. "Omniscient"
stands for the setting where all the distances between agents and facilities are known, and the
numbers represent the best-known approximation ratios. The second column represent our setting,
in which the ordinal preferences of the agents, and the numerical distances between facilities are
known. The last column represents the pure ordinal setting in which only the agent ordinal prefer-
ences are known, but the distances between facilities are unknown; this setting has been previously
studied, and we include the known lower bounds on the possible distortion in parentheses, including
some which we prove in the Appendix.
true optimum. This is a large improvement over previously known distortion bounds: for minimum
cost ordinal matching the best-known distortion bound is n using random serial dictatorship (RSD)
[13]; by using the knowledge of facility locations we are able to reduce this approximation ratio to
3.
1.2 Discussion and Related Work
Ordinal approximation [2] for the minimum social cost (or maximum social welfare) with underlying
utilities/distances between agents and alternatives has been studied in many settings including
social choice [3, 4, 11, 14, 16, 21, 23, 28, 29], matchings [5–7, 9, 13, 17, 22], secretary problems [26],
participatory budgeting [8], general graph problems [1, 5] and many other models in recent years.
The general assumption of the ordinal setting is that we only have the ordinal preferences of agents
over alternatives, and the goal is to form a solution that has close to optimal social cost. There
are different models: social choice, matching, facility location, etc.; different objectives: minimizing
social cost, maximizing social welfare, total cost objective, median objective, egalitarian objective,
etc.; different assumptions on utility or cost functions: unit-sum, unit-range, metric space, etc. In
this paper, we study general facility assignment problems in a metric space, and assume that the
ordinal preferences of agents over alternatives are given. Unlike previous work on this topic, we also
assume the locations of the alternatives are known; we show that this extra information enables us
to achieve much better approximation ratios than in the pure ordinal setting for many problems.
The distortion of social choice functions was first introduced in [28], to describe the ratio
between the total utility of the optimal candidate and the candidate selected by a mechanism
using only ordinal preferences. [3,23,29] studied the distortion of social choice functions in a metric
space; the assumption that the underlying numerical costs have this metric property allows for
much better results than more general costs.
In particular, for the objective of minimizing the
total distance from the agents to the chosen candidate, the above papers were able to show good
distortion bounds for many well-known mechanisms, in particular a bound of 5 for Copeland [3],
a bound of O(ln m) for Single Transferable Vote (STV) [29], and many others.
In addition, [3]
proved that no deterministic mechanism can have worst-case distortion better than 3, and [29]
4
√
ln m − 1. Goel
showed that all scoring rules for m-candidates have a distortion of at least 1 + 2
et al. [23] showed that Ranked Pairs, and the Schulze rule have a worst-case distortion of at least
5, and the expected worst-case distortion of any (weighted)-tournament rule is at least 3. They
also introduced the notion of "fairness" of social choice rules, and discussed the fairness ratio of
Copeland, Randomized Dictatorship, and a general class of cost functions. Finding a deterministic
mechanism with distortion less than 5 has been an open problem for several years. In this paper,
we show that if we know the exact locations of the candidates in addition to the ordinal ranking of
the agents, then there is a simple algorithm which achieves a distortion of 3, and no better bound
is possible.
While the above work, as well as our paper, only focuses on deterministic algorithms, the
distortion of randomized algorithms in social choice has also been considered, see for example
[4,19,21,24]. In a slightly different flavor of result, [15,16] consider the special case where candidates
are randomly and independently drawn from the set of voters. While we leave the analysis of
randomized algorithms which know the location of the facilities to future work, and consider the
worst-case candidate locations, it is worth pointing out that our deterministic algorithm achieves
a distortion of 3, which is also the best known distortion bound for any randomized mechanism
which only knows the ordinal preferences of the agents. Similarly, another common goal is to form
truthful mechanisms with small distortion for matching and social choice, as in [6, 13, 21]; we focus
on general mechanisms in this paper in order to understand the limitations of knowing only certain
kinds of ordinal information, and leave the goal of forming truthful mechanisms for future work.
For the median objective of social choice problems, [3] showed that Copeland gives a distortion
of at most 5, while no deterministic mechanism can achieve a distortion of better than 5 . [4] also
gave a randomized algorithm that has a distortion of at most 4.
In this paper, we are able to
improve this bound to a tight worst-case distortion of 3 by a deterministic mechanism, because we
also know how each candidate ranks all the other candidates, in addition to how each voter ranks
all the candidates.
The distortion of matching in a metric space has received far less attention than social choice
questions. [5–7] analyzed maximum-weight metric matching; the maximization objective makes this
problem far easier, and even choosing a uniformly random matching yields a distortion of a small
constant. This is very different from our goal of computing a minimum-cost matching, for which no
ordinal approximations better than O(n) are known. [13] studied facility assignment problems in a
metric space; they considered the problem with or without resource augmentation, and the cases
without augmentation are exactly the minimum weight bipartite matching problem. [13] showed
that the approximation ratio of random serial dictatorship (RSD) is at most n, and gave a lower
bound of 2n − 1 for the approximation ratio of serial dictatorship (SD), and a lower bound of n0.26
for RSD. Their results are the best known ordinal approximations for this problem. In this paper,
we are able to give a tight 3-approximation for the minimum weight matching problem, given the
locations of facilities in addition to the agents' ordinal preferences.
2 Model and Notation: Social Choice
For the social choice problems studied in this paper, we let A = {1, 2, . . . , n} be a set of agents,
and let F = {F1, F2, . . . , Fm} be a set of alternatives, which we will also refer to sometimes as
candidates or facilities. We will typically use i and j to refer to agents and W, X, Y, Z to refer
to alternatives. Let S be the set of total orders on the set of alternatives F. Every agent i ∈ A
has a preference ranking σ ∈ S; by X (cid:31)i Y we will mean that X is preferred over Y in ranking
σ. Although we assume that each agent has a total order of preference over the alternatives and
5
that this order is known to us, for many of our results it is only necessary that the top choice of
each agent is known. We say X is i's top choice if i prefers X to every other alternative in F. We
call the vector σ = (σ1, . . . , σn) ∈ S n a preference profile. We say that an alternative X pairwise
defeats Y if {i ∈ A : X (cid:31)i Y } > n
2 . The goal is to choose a single winning alternative.
Cardinal Metric Costs.
In this work we take the utilitarian view, and assume that the ordinal
preferences σ are derived from underlying (latent) cardinal agent costs. Formally, we assume that
there exists an arbitrary metric d : (A ∪ F)2 → R≥0 on the set of agents and alternatives. The
cost incurred by agent i of alternative X being selected is represented by d(i, X), which is the
distance between i and X. Such spatial preferences are relatively common and well-motivated, see
for example [3, 18] and the references therein. The underlying distances d(i, X) are unknown, but
unlike most previous work we do assume the distances between alternatives are given. For example,
when alternatives represent facilities or stores to be opened, it makes sense that their specific
locations would be known, while the distances from the customers to the stores may be private.
Similarly, when the alternatives represent political candidates, it may be easy to estimate their
locations in ideological space (for example based on their voting records and public statements),
but the ideology of the voters is much harder to estimate, with mechanism designers only knowing
which candidates the voters prefer but not how much they prefer them. The distance between two
alternatives X and Y is denoted by l(X, Y ). We say that d is consistent with l if ∀X, Y ∈ F,
d(X, Y ) = l(X, Y ).
The metric costs d naturally give rise to a preference profile. We say that d is consistent with
σ if ∀i ∈ A, ∀X, Y ∈ F, if d(i, X) < d(i, Y ), then X (cid:31)i Y . It means that the cost of X is less
than the cost of Y for agent i, so agent i prefers X over Y . As described above, we know exactly
the distances l and the preferences σ, but do not know the true costs d which give rise to σ. Let
D(σ, l) be the set of metrics that are consistent with σ and l; we know that one of the metrics from
this possibly infinite space captures the true costs, but do not know which one.
(cid:80)
the most common notion of social cost is the sum objective function, defined as SC(cid:80)(X,A) =
Social Cost Distortion We study several objective functions for social cost in this paper. First,
i∈A d(i, X). We also study the median objective function, SCmed(X,A) = medi∈A(d(i, X)), as
well as the egalitarian objective and many others (see Section 3.2). We use the notion of distortion
to quantify the quality of an alternative in the worst case, similar to the notation in [11, 28]. For
any alternative W , we define the distortion of W as the ratio between the social cost of W and the
optimal alternative:
dist(cid:80)(W, σ, l) = sup
d∈D(σ,l)
distmed(W, σ, l) = sup
d∈D(σ,l)
SC(cid:80)(W,A)
minX∈F SC(cid:80)(X,A)
SCmed(W,A)
minX∈F SCmed(X,A)
In other words, saying that the distortion of W is at most 3 means that, no matter what the
true costs d are (as long as they are consistent with the σ and l which we know), it must be that
the social cost of W is within a factor of 3 of the true optimum alternative, which is impossible to
compute without knowing the true costs. Because of this, a small distortion value means that there
is no need to obtain the true agent costs, and the ordinal information σ (together with information
l about the alternatives) is enough to form a good solution.
6
A social choice function f on A and F takes σ and l as input, and returns the winning alter-
native. We say the distortion of f is the same as the distortion of the winning alternative chosen
by f on σ and l. In other words, the distortion of a social choice mechanism f on a profile σ and
facility distances l is the worst-case ratio between the social cost of W = f (σ, l), and the social cost
of the true optimal alternative.
3 Distortion of Social Choice Mechanisms
3.1 Distortion of Total Social Cost
In this section, we study the sum objective and provide a deterministic algorithm that gives a
distortion of at most 3. According to [3], the lower bound on the distortion for deterministic social
choice functions with only ordinal preferences (without knowing l) is 3. This occurs in the simple
example with 2 alternatives which are tied with approximately half preferring each one. No matter
which one is chosen, the true optimum could be the other one, and its social cost can be as much
as 3 times better. Because the example in Theorem 3 from [3] only has two alternatives, knowing
l does not provide any extra information, and thus that example also provides a lower bound of
3 in our setting, although we assume the distances l between facilities are known in this paper.
Therefore, our mechanism achieves the best possible distortion in this setting. Note that if we only
have ordinal preferences of the agents without the distances between facilities, then the best known
approach so far is Copeland, which gives a distortion at most 5. Thus our results establish that
by knowing the distances l between alternatives, it is possible to reduce the distortion from 5 to 3,
and no better deterministic mechanism is possible.
Lemma 3.1. Let W, X be alternatives. If W (cid:31)i X, then d(i, X) ≥ d(X,W )
. [Lemma 5 in [3]]
In the following algorithm, we generate a set of projected agents as follows: Given agents A,
alternatives F, and the preference profile σ, for each agent i denote alternative Xi as i's top choice.
Then we create a new agent i at the location of Xi in the metric space, as shown in Figure 1 (a);
consequently, ∀ Y ∈ F, d(i, Y ) = d(Xi, Y ). Denote the set of the new agents as A = {1, 2, . . . , n}.
For any metric d consistent with l, d(i, Y ) = d(Xi, Y ) = l(Xi, Y ), so the distances between agents
in A and alternatives in F are known to us, unlike the true distances between A and F.
2
Algorithm 1: Algorithm for the minimum total social cost.
Input : Agents A = {1, 2, . . . , n},
Alternatives F = {F1, F2, . . . , Fm},
Each agent i's top choice alternative,
Distances between alternatives, i.e., l(Y, Z), ∀ Y, Z ∈ F
on A by choosing X, i.e., SC(cid:80)(X, A) =(cid:80)
Output: The winning alternative W .
Generate projected agent set A. For each alternative X ∈ F, calculate the total social cost
Final Output: Return the alternative W that has the minimum social cost SC(cid:80)(W, A) .
i∈ A d(i, X) =(cid:80)
i∈ A l(i, X) .
Theorem 3.2. The distortion of Algorithm 1 for minimum total social cost on A is at most 3.
Proof. Let W denote the winning alternative. W has the minimum social cost on the agent set A,
so for any alternative Y , it must be that
7
Figure 1: (a) For each agent, generate a projected agent at the location of its top choice alternative.
(b) A figure demonstrating agent i, i's top choice alternative Y , i's projected agent i located at Y ,
the winner W , and the optimal alternative X for the proof of Theorem 3.2.
SC(cid:80)(W, A)
SC(cid:80)(Y, A)
=
(cid:80)
(cid:80)
(cid:80)
(cid:80)
Let X denote the true optimal alternative for A. We want to get dist(cid:80)(W, σ, l) by upper
i∈ A d(i, W )
i∈ A d(i, Y )
=
i∈A d(i, W )
i∈A d(i, Y )
≤ 1
(1)
(2)
bounding the cost incurred by W compared to X:
SC(cid:80)(W,A)
SC(cid:80)(X,A)
=
≤
=
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
i∈A d(i, W )
i∈A d(i, X)
i∈A(d(i,i) + d(i, W ))
(cid:80)
i∈A d(i, X)
(cid:80)
(cid:80)
i∈A d(i,i)
i∈A d(i, X)
+
i∈A d(i, W )
i∈A d(i, X)
The inequality d(i, W ) ≤ d(i,i) + d(i, W ) is due to the triangle inequality since d is a metric,
as shown in Figure 1 (b). ∀i ∈ A, we know that i is located at i's top choice alternative, so the
(cid:80)
distance between i and i must be less than (or equal to) the distance between i and any alternative;
(cid:80)
i∈A d(i,X) ≤ 1. For any agent i such
thus d(i,i) ≤ d(i, X). Summing up for all i ∈ A, we get that
that X is not i's top choice, suppose alternative Y is i's top choice, then i has the same location as
Y and d(i, X) = d(X, Y ). By Lemma 3.1, d(i, X) ≥ d(X,Y )
, thus d(i, X) ≥ d(i,X)
. For all i that X
holds for all i ∈ A. Together with
is i's top choice, d(i, X) = 0, so the inequality d(i, X) ≥ d(i,X)
inequality 1 and 2,
i∈A d(i,i)
2
2
2
8
SC(cid:80)(W,A)
SC(cid:80)(X,A)
≤ 1 +
(cid:80)
(cid:80)
i∈A d(i, W )
i∈A d(i,X)
2
(cid:80)
(cid:80)
i∈A d(i, W )
i∈A d(i, X)
≤ 3
= 1 + 2
3.2 Distortion of Median Social Cost
In this section, we study the median objective function, and provide a deterministic mechanism
that gives a distortion of at most 3. Recall that we define the median social cost of an alternative
X as SCmed(X,A) = medi∈A(d(i, X)). We will refer to this as med(X) when d and A are fixed.
If n is even, we define median to be the ( n
2 + 1)th smallest value of the distances. Note that no
deterministic mechanism which only knows ordinal preferences can have worst-case distortion better
than 5 (Theorem 14 in [3]). With known distances between facilities, we are able to provide a natural
social choice function with distortion of 3, which is also provably the best possible distortion in our
setting (consider the example in Theorem 3 from [3] again). Moreover, our social choice function
only uses ordinal information about the alternatives, and not the full distances l; in particular as
long as we have ordinal preferences of each alternative for each other alternative (and thus a total
order of the distances from each alternative to the others), then our mechanism will work properly.
Such ordinal information may be easier to obtain than full distances l; for example candidates
can rank all the other candidates. In particular, given agents with ordinal preferences such that
the candidates are a subset of the agents, our mechanism will always form an outcome with small
distortion, even if we do not know the distances l.
Note that using only agents' top choices over alternatives and the distances between alterna-
tives, as Algorithm 1 does for the total social cost objective, is not enough to give a worst-case
distortion of 3 for the median objective. Consider the following example: there are 4 alternatives
W, X, Y, Z, the distances between them are: d(W, Y ) = d(Y, X) = d(X, Z) = d(Z, W ) = 2 and
d(W, X) = d(Y, Z) = 4. Suppose W is agents 1, 2's top choice, X is agent 3, 4's top choice, Y is
agent 5, 6's top choice, and Z is agent 7, 8's top choice. This graph is symmetric, so we choose
an arbitrary alternative as the winner. Suppose we choose W as the winner, and the distances
between agents and facilities are: the distances from agents 1, 2 to W are both 100, the distances
from agents 1, 2 to X, Y, Z are all 102. The distances from agents 5, 6 to Y, X are all 1, and the
distances from agents 5, 6 to W, Z are all 3. The distances from agents 7, 8 to Z, X are all 1, and
the distances from agents 7, 8 to Y, W are all 3. The distances from agents 3, 4 to X are both
1, the distances from 3, 4 to Y, Z are all 3, and the distances from 3, 4 to W are both 5. In this
example, the median is the distance from 5th closest agent to the winning alternative. X is the
optimal alternative with med(W ) = 1, while med(W ) = 5 has a distortion of 5.
We will use the following Lemmas from [3] in the proof of our algorithm:
Lemma 3.3. For any two alternatives W and Y , we have med(W ) ≤ med(Y ) + d(Y, W ). [Lemma
11 in [3]]
Lemma 3.4. For any two alternatives Y and P , if P pairwise defeats (or pairwise ties) Y , then
med(Y ) ≥ d(Y,P )
Lemma 3.5. Let W, Y be an alternatives ∈ F, if W pairwise defeats (or pairwise ties) Y , then
med(W ) ≤ 3med(Y ). [Proved in Theorem 8 in [3]]
. [Proved in Theorem 16 in [3]]
2
The main easy insight which we use in the formation of our algorithm comes from the following
lemma.
9
Lemma 3.6. For any three alternatives W , Y , and P , if P pairwise defeats (or pairwise ties) Y ,
and d(Y, W ) ≤ d(Y, P ), then med(W ) ≤ 3med(Y ).
Proof. By Lemma 3.3, med(W ) ≤ med(Y ) + d(Y, W ). By Lemma 3.4, med(Y ) ≥ d(Y,P )
know that d(Y, P ) ≥ d(Y, W ), thus
. And we
2
med(W ) ≤ med(Y ) + d(Y, W )
≤ med(Y ) + d(Y, P )
≤ med(Y ) + 2med(Y )
≤ 3med(Y )
We use a natural Condorcet-consistent algorithm to approximate the minimum median social
cost with the agents' preference rankings σ and the ordinal preferences of every alternative over
other alternatives. First, create the majority graph G = (F, E), i.e., a graph with alternatives as
vertices and an edge (Y, Z) ∈ E if Y pairwise defeats or pairwise ties Z. If a Condorcet winner (i.e.
an alternative which pairwise defeats all others) exists, then we return it immediately.
Otherwise, we consider each pair of alternatives. By Lemma 3.5, if the edge (W, Y ) ∈ E, then
med(W ) ≤ 3med(Y ). When considering an alternative pair W, Y , if (W, Y ) (cid:54)∈ E and we know that
there exists another alternative P which meets the conditions of Lemma 3.6, then we add an edge
(W, Y ) to G. It is not difficult to see that whenever the edge (W, Y ) is in our graph, this means
that med(W ) ≤ 3med(Y ). As we prove below, at the end of this process there always exists at
least one alternative which has edges to all the other alternatives, and thus the distortion obtained
from selecting it is at most 3, no matter which alternative is the true optimal one.
Note that from the ordinal preferences of alternatives over each other, we can get a partial order
of distances between the alternatives. Denote this partial order as (cid:22), i.e., we say that d(W, Y ) (cid:22)
d(W, Z) if we know that W prefers Y to Z (we do not have information about strict preference).
This is the information we have on hand: we only know the partial order of distances between pairs
of alternatives which share an alternative in common. Note, however, that if there exists a cycle
in this partial order, i.e., d(Y1, Y2) (cid:22) d(Y2, Y3) (cid:22) d(Y3, Y4) (cid:22) ··· (cid:22) d(Yk, Y1) (cid:22) d(Y1, Y2), then this
implies that all the distances in the cycle are actually equal, and thus we can also add the relations
d(Y1, Y2) (cid:23) d(Y2, Y3) (cid:23) d(Y3, Y4) (cid:23) ··· (cid:23) d(Yk, Y1) (cid:23) d(Y1, Y2). Such cycles are easy to detect (e.g.,
by forming a graph with a node for every alternative pair and then searching for cycles), and thus
we can assume that whenever a cycle exists in our partial order, then for every pair of distances
d(W, Y ) and d(W, Z) in the cycle, we have both d(W, Y ) (cid:22) d(W, Z) and d(W, Y ) (cid:23) d(W, Z).
Lemma 3.7. Consider the modified majority graph G = (F, E) at any point during Algorithm 2.
For any edge (W, Y ) ∈ E, we have that med(W ) ≤ 3med(Y ).
Proof. By Lemma 3.5, for any edge (W, Y ) in the original majority graph, med(W ) ≤ 3med(Y ).
Now consider an edge (W, Y ) added to E when processing the alternative pair W, Y . It must
be the case that there exists an alternative P , such that d(Y, W ) ≤ d(Y, P ) and P pairwise defeats
(or ties) Y . By Lemma 3.6, med(W ) ≤ 3med(Y ).
Lemma 3.8. At the end of Algorithm 2, there must exist an alternative W such that ∀Y ∈ F−{W},
(W, Y ) ∈ E.
10
Algorithm 2: Algorithm for the minimum median social cost.
Input : Agents A = {1, 2, . . . , n},
Alternatives F = {F1, F2, . . . , Fm},
The majority graph G = (F, E),
Ordinal preferences of each alternative over other alternatives,
Partial order of distances between alternatives.
Output: The winning alternative W .
If there is a Condorcet winner W , return W as the winner.
forall alternative pairs W, Y do
if (W, Y ) (cid:54)∈ E or (Y, W ) (cid:54)∈ E then
WLOG, suppose (Y, W ) exists, but (W, Y ) does not exist.
if there exists an alternative P , such that we have d(Y, W ) (cid:22) d(Y, P ) in our partial
order information, and P pairwise defeats (or ties) Y then
Add edge (W, Y ) to E;
continue;
end
end
end
There must exists an alternative W such that ∀Y ∈ F − {W}, (W, Y ) ∈ E. Return W as the
winner.
Proof. We prove this lemma by contradiction. Suppose no such alternative W exists. Then for each
alternative Y , there is at least one alternative Z, such that only (Z, Y ) ∈ E and (Y, Z) (cid:54)∈ E. This is
because we start with the majority graph, so at least one edge always exists between every pair. We
create another directed graph G(cid:48) = (F, E(cid:48)), with E(cid:48) being all the edges (Z, Y ) such that Y, Z (cid:54)∈ E.
Thus any pair of alternatives in G(cid:48) have at most one direction of edge between them. And by our
assumption, each alternative Y has at least one incoming edge in G(cid:48). Since the in-degree of each
node is at least 1 in G(cid:48), there must be at least one cycle in G(cid:48). To see this, one can for example
take the edge (Y2, Y1) coming into Y1, then the edge (Y3, Y2) coming into Y2, and proceed in this
way until a cycle is formed. Note that every edge in G(cid:48) must be in the original majority graph,
because if we add an edge when processing a pair of alternatives in our algorithm, that pair must
have edges in both directions.
Consider a cycle formed by edges (Y1, Y2), (Y2, Y3), . . . , (Yk−1, Yk), (Yk, Y1). When processing
the alternative pair Y2, Y3 in Algorithm 2, we did not add edge (Y3, Y2) to E, so it must be the
case that no alternative P exists such that d(Y2, Y3) (cid:22) d(Y2, P ) and P pairwise defeats (or ties) Y2.
But we know that Y1 pairwise defeats (or ties) Y2, because edge (Y1, Y2) is in the original majority
graph. Then the only possibility is we don't know if d(Y2, Y3) ≤ d(Y1, Y2), i.e., either d(Y2, Y3) and
d(Y1, Y2) are incomparable in our partial order, or we only know that d(Y2, Y3) (cid:23) d(Y1, Y2). They
cannot be incomparable, since we have the ordinal preferences of Y2 for Y1 and Y3, thus our partial
order must state that d(Y2, Y3) (cid:23) d(Y1, Y2), i.e., Y2 prefers Y1 to Y3. By the same reasoning, we
also get that Y3 prefers Y2 to Y4, and more generally that Yi prefers Yi−1 to Yi+1 for all i, where
Y0 = Yk and Yk+1 = Y1 since it is a cycle. This means that in our partial order, we have that
d(Y1, Y2) (cid:22) d(Y2, Y3) (cid:22) ··· (cid:22) d(Yk−1, Yk) (cid:22) d(Yk, Y1) (cid:22) d(Y1, Y2). Recall, however, that this means
we know d(Y1, Y2) = d(Y2, Y3) = ··· = d(Yk−1, Yk) = d(Yk, Y1), and before running Algorithm 2, we
detect cycles in the partial order of alternative distances, and add the equality information to the
partial order. This means that whenever d(Y1, Y2) (cid:22) d(Y2, Y3) (cid:22) ··· (cid:22) d(Yk−1, Yk) (cid:22) d(Yk, Y1) (cid:22)
11
d(Y1, Y2) exists in our partial order, we also have d(Y1, Y2) (cid:23) d(Y2, Y3) (cid:23) ··· (cid:23) d(Yk−1, Yk) (cid:23)
d(Yk, Y1) (cid:23) d(Y1, Y2) in the partial order as well. But this gives us a contradiction, since having
d(Y2, Y3) (cid:22) d(Y1, Y2) in the partial order, combined with the fact that Y1 pairwise defeats Y2, would
cause us to add the edge (Y3, Y2) in our algorithm, which contradicts the statement that only the
edge (Y2, Y3) is in the final graph produced by the algorithm, but not (Y3, Y2). Thus there must
exist at least one alternative with edges from it to all the others.
Theorem 3.9. The distortion of Algorithm 2 for minimum median social cost is at most 3.
Proof. If there is a Condorcet winner, by Lemma 3.5, the distortion is at most 3.
Otherwise, by Lemma 3.8, the algorithm always returns a winner. Suppose it returns alternative
W as the winner, by Lemma 3.7, W has a distortion at most 3 with any alternative X as the optimal
solution.
3.2.1 Generalizing Median: Percentile Distortion
Instead of just considering the median objective, we also consider a more general objective: the α-
percentile social cost. Let α-PC(Y ) denote the value from the set {d(i, Y ) : i ∈ A}, that α fraction
of the values lie below α-PC(Y ). Thus median is a special case when α = 1
2 , med(Y ) = 1
2 -PC(Y ).
It was shown in [3] Theorem 17 that the worst-case distortion when α ∈ [0, 1
2 ] in that setting (only
have agent's ordinal preferences over alternatives) is unbounded, and the same example shows
α ∈ [0, 1
2 ] in our setting is also unbounded. However, we are able to give a distortion of 3 for
α ∈ [ 1
3 ] is
5. The reason is that the ordinal preferences between alternatives are also available in our setting.
We will show that Algorithm 2 gives a distortion of at most 3 not only for the median objective, but
also for the general α-percentile objective, because all the lemmas we used to prove the conclusion
for the median objective could be generalized to α-percentile.
2 , 1] in this paper, while for the setting in [3], the lower bound for distortion when α ∈ [ 1
2 , 2
We use the following lemma from [3] in the proof of our algorithm:
Lemma 3.10. For any two alternatives W and Y , we have α-PC(W ) ≤ α-PC(Y ) + d(Y, W ).
[Lemma 18 in [3]]
We can generalize Lemma 3.6 to the following lemma, and the proof is by using Lemma 3.10
instead of Lemma 3.3 in the proof of Lemma 3.6,
Lemma 3.11. For any three alternatives W , Y , and P , if P pairwise defeats (or pairwise ties) Y ,
and d(Y, W ) ≤ d(Y, P ), then α-PC(W ) ≤ 3α-PC(Y ).
Theorem 3.12. The distortion of Algorithm 2 for the α-PC objective social cost with 1
is at most 3.
2 ≤ α ≤ 1
2 ≤ α ≤ 1, for any alternative Y , we know α-PC(Y ) ≥ med(Y ). Then Lemma 3.7
Proof. Note that Lemma 3.10 is actually a generalization of Lemma 3.3, and Lemma 3.11 is a
generalization of Lemma 3.6. Lemma 3.4 and Lemma 3.5 also generalize to the α-PC objective,
because when 1
also generalizes to the α-PC objective, because it only uses Lemma 3.5 and Lemma 3.6 in the
proof. And Lemma 3.8 still holds for the same algorithm. Thus all the lemmas and properties of
the median objective used in the proof of Theorem 3.9 could be generalized into the α-PC objective,
so the conclusion still holds for the α-PC objective when 1
2 ≤ α ≤ 1.
12
3.2.2 Algorithm 2 and the Total Social Cost
Although Algorithm 2 is designed for the median objective, it also performs quite well for the sum
objective. Interestingly, the distortion of this algorithm for the minimum total social cost is at
most 5, which is the same as Copeland (the best known deterministic algorithm with no knowledge
of candidate preferences). Thus this algorithm gives a distortion of 3 for median (and in fact for
all α-percentile objectives) and distortion of 5 for sum simultaneously. In settings where we are
not sure which objectives to optimize, or ones where we care both about the total social good, and
about fairness, this social choice mechanism provides the best of both worlds. The lemmas and
proofs for this result are similar to Theorem 3.9, as follows.
Lemma 3.13. Let W, Y be alternatives ∈ F. If W pairwise defeats (or pairwise ties) Y , then
SC(cid:80)(W,A) ≤ 3SC(cid:80)(Y,A). [Proved in Theorem 7 in [3]]
and d(Y, W ) ≤ d(Y, P ), then SC(cid:80)(W,A) ≤ 5SC(cid:80)(Y,A).
Proof. For all i ∈ A, we know d(i, W ) ≤ d(i, Y ) + d(Y, W ) by the triangle inequality. Summing up
for all i ∈ A, we get SC(cid:80)(W,A) ≤ SC(cid:80)(Y,A) + n · d(Y, W ).
we get SC(cid:80)(Y,A) ≥ n
P pairwise defeats (or pairwise ties) Y , so at least half of the agents prefer P to Y ; thus the
total social cost of Y is at least the sum of the social cost of these half of agents. By Lemma 3.1,
Lemma 3.14. For any three alternatives W , Y , and P , if P pairwise defeats (or pairwise ties) Y ,
d(Y,P )
2
4 d(Y, P ). Thus,
2 = n
SC(cid:80)(W,A) ≤ SC(cid:80)(Y,A) + n · d(Y, W )
≤ SC(cid:80)(Y,A) + n · d(Y, P )
≤ SC(cid:80)(Y,A) + 4SC(cid:80)(Y,A)
≤ 5SC(cid:80)(Y,A)
Lemma 3.15. Consider the modified majority graph G = (F, E) at any point during Algorithm 2.
For any edge (W, Y ) ∈ E, we have that SC(cid:80)(W,A) ≤ 5SC(cid:80)(Y,A).
Proof. By Lemma 3.13, for any edge (W, Y ) in the original majority graph, SC(cid:80)(W,A) ≤ 3SC(cid:80)(W,A).
Now consider an edge (W, Y ) added to E when processing the alternative pair W, Y . It must
be the case that there exists an alternative P , such that d(Y, W ) ≤ d(Y, P ) and P pairwise defeats
(or ties) Y . By Lemma 3.14, SC(cid:80)(W,A) ≤ 5SC(cid:80)(Y,A).
Theorem 3.16. The distortion of Algorithm 2 for minimum total social cost is at most 5, and this
bound is tight.
Proof. If there is a Condorcet winner, by Lemma 3.13, the distortion is at most 3. Otherwise,
suppose the algorithm returns alternative W as the winner; by Lemma 3.15 W has a distortion at
most 5 with any alternative X as the optimal solution.
To see that this bound is tight, consider the following example. There are three facilities W , Y ,
and P . There are q agents who prefer Y to W to P , q agents who prefer P to Y to W , and 1 agent
who prefers W to P to Y . We denote these three sets of agents as AY , AP and AW separately.
By the preferences of agents, we know that Y pairwise defeats W , W pairwise defeats P , and
P pairwise defeats Y . The distances between facilities are: d(Y, W ) = 2 − 2, d(W, P ) = 2 − ,
13
d(P, Y ) = 2, where is a very small positive number. AY is located at the same location as Y ,
so d(AY , Y ) = 0, d(AY , P ) = 2, and d(AY , W ) = 2 − 2. The distances between AP and the
alternatives are: d(AP , Y ) = d(AP , P ) = 1, d(AP , W ) = 3 − 2. AW has a distance of 1 to all
alternatives. Run Algorithm 2 on this example, and consider the alternative pair W , Y . Because
P pairwise defeats Y and d(Y, W ) (cid:22) d(Y, P ), we add edge (W, Y ) to the graph and make W the
winner. The total social cost of W is q ∗ (2 − 2) + q ∗ (3 − 2) + 1 = q(5 − 4) + 1. While the
optimal solution is to choose Y as the winner, and get a total social cost of q + 1. When q is very
large and is very small, the distortion in this example approaches 5.
4 Model and Notation: Facility Assignment Problems
The mechanism we used for approximation of total social cost in Theorem 3.2 can be applied to
far more general problems. In this section, we describe a set of facility assignment problems that
fit in this framework. As before, let A = {1, 2, . . . , n} be a set of agents, and F = {F1, F2, . . . , Fm}
be a set of facilities, with each agent i having a preference ranking σi over the facilities, and
σ = (σ1, . . . , σn).
As in the social choice model, we assume that there exists an arbitrary unknown metric d :
(A ∪ F)2 → R≥0 on the set of agents and facilities. The distances d(i, Fj) between agents and
facilities are unknown, but the ordinal preferences σ and the distances l between facilities are
given. Let D(σ, l) be the set of metrics consistent with σ and l, as defined previously in Section 2.
Unlike for social choice, our goal is now to choose which facilities to open, and which agents
should be assigned to which facilities. Formally, we must choose an assignment x : A → F, where
x(i) is the facility that i is assigned to. Every i ∈ A must be assigned to one (and only one)
facility in F; other than that, there could be arbitrary constraints on the assignment. Here are
some examples of constraints which fall into our framework: each facility Fi has a capacity ci,
which is the maximum number of agents that can be assigned to Fi; at least (or at most) p facilities
should have agents assigned to them; agents i and j must be (or must not be) assigned to the same
facility, etc. The social choice model is a special case of this one with the constraint that exactly
one facility must be opened, and all agents must be assigned to it. Note that the constraints are
only on the assignment, and independent of the metric space d. An assignment x is valid if it
satisfies all constraints. Let X be the set of all valid assignments.
The cost function of assignments. The cost of an assignment x consists of two parts. The first
part is the distance cost between agents and facilities. ∀i ∈ A, let si denote the distance between
i and the facility it is assigned to, i.e., si = d(i, x(i)). For a given metric d and assignment x, let
s(x, d) denote the vector of distances between each i ∈ A and x(i), i.e., s(x, d) = (s1, s2, . . . , sn).
Let cd : Rn≥0 → R≥0 be a cost function that takes a vector of distances as input. For example, this
could simply sum up all the distances, take the maximum distance for an egalitarian objective, etc.
To be as general as possible, instead of fixing a specific function cd we consider the set of distance
cost functions that are monotone nondecreasing and subadditive. Formally, cd is monotonically
nondecreasing means that for any vectors s and s(cid:48) such that s ≤ s(cid:48) componentwise, we have that
cd(s) ≤ cd(s(cid:48)). Any reasonable cost function should satisfy this property if agents desire to be
assigned to closer facilities. cd being subadditive means that for any vectors s and s(cid:48), we have that
cd(s + s(cid:48)) ≤ cd(s) + cd(s(cid:48)). While not all functions are subadditive, many important ones are, as
they represent the concept of "economies of scale", a common property of realistic costs.
The second part of the assignment cost is the facility cost. Let cf (x) denote the facility cost
for assignment x. cf can be an arbitrary function over the assignments, for example, the opening
14
cost of facilities, the penalty (or reward) for assigning certain agents to the same facility, etc. Our
framework includes all such functions, and thus is quite general, as we discuss below. The main
components needed for our framework to work is that the function cf does not depend on the
distances, only on x, and that the function cd is subadditive.
The total cost c(x, d) of an assignment x is the sum of the distance cost and the facility cost,
i.e. c(x, d) = cd(s(x, d)) + cf (x). We study algorithms to approximate the minimum cost assign-
ment given only agents' ordinal preferences over facilities, and the distances between facilities, as
described above.
Social Cost Distortion As for social choice, we use the notion of distortion to measure the
quality of an assignment in the worst case, similar to the notation in [11,28]. For any assignment x,
we define the distortion of x as the ratio between the social cost of x and the optimal assignment:
dist(x, σ, l) = sup
d∈D(σ,l)
c(x, d)
minx(cid:48)∈X c(x(cid:48), d)
A social choice function f on A and F takes σ and l as input, and returns a valid assignment
on A and F. We say the distortion of f on σ and l is the same as the distortion of the assignment
returned by f . In other words, the distortion of an assignment function f on a profile σ and facility
distances l is the worst-case ratio between the social cost of x = f (σ, l), and the social cost of the
true optimal assignment, to obtain which we would need the true distances d.
Approximation ratio of omniscient algorithms Consider omniscient algorithms which know
the true numerical distances between agents and facilities for the facility assignment problems, in
other words, the metric d. In some sense, the goal of our work is to determine when algorithms
with only limited information can compete with such omniscient algorithms. With the full dis-
tances information, we can of course obtain the optimal assignment using brute force, while for
our algorithms with limited knowledge this is impossible even given unlimited computational re-
sources. Nevertheless, we are also interested in what is possible to achieve if we restrict ourselves to
polynomial time. To differentiate traditional approximation algorithms from algorithms with small
distortion, suppose that an omniscient approximation algorithm f returns assignment x. Then we
denote the approximation ratio of a valid assignment x as:
ratio(x) =
c(x, d)
minx(cid:48)∈X c(x(cid:48), d)
Thus we say the approximation ratio of an omniscient algorithm f is β if for any input of the
problem, the assignment x returned by f has ratio(x) ≤ β.
4.1 Examples of Facility Assignment Problems
In this section we illustrate that our framework is quite general by giving various important ex-
amples which fit into our framework. In the section which follows, we prove a general black-box
reduction theorem for our framework, and thus immediately obtain mechanisms with small distor-
tion for all these examples simultaneously.
The total social cost problem we discussed in Section 3.1 is a special case of the facility assign-
ment problem such that the constraint is only one facility (alternative) is chosen, and all agents
are assigned to it. For any assignment x, the facility cost function cf (x) = 0, and the distance cost
function cd(s(x, d)) is the sum of distances from the winning alternative to all agents in the metric
15
d. cd is monotone and additive (thus subadditive). Here are some other examples that fit in our
framework:
Minimum weight metric bipartite matching. Given a set of agents A and a set of facilities
F such that A = F = n. G = (A,F, E) is an undirected complete bipartite graph. The facilities
and agents lie in a metric space d. The weight of each edge (i, F ) ∈ E is the distance between i
and F , w(i, F ) = d(i, F ). The goal is to find a minimum weight perfect matching of the bipartite
graph given only ordinal information. This setting has been studied before, and the best distortion
bound known is n [13] given by RSD for the case when only the ordinal preferences σ are known.
Our results show that if we also know the distances l between facilities, then even without knowing
the distances d between agents and facilities, it is possible to create simple mechanisms with dis-
tortion at most 3 (we can show that no better bound is possible for this setting). Thus having a bit
more information about the facilities immediately improves the distortion bound by a very large
amount. We show this result by using our facility assignment framework above: the constraint
here is that each facility has a capacity of 1, thus a valid assignment is a perfect matching of the
bipartite graph. For any assignment x, the facility cost function is cf (x) = 0, and the distance cost
function cd(s(x, d)) is the total edge weight in the assignment. cd is monotone and additive (thus
subadditive).
Egalitarian bipartite matching. With the same bipartite graph as in minimum weight
matching problems, the only difference is that the goal of egalitarian bipartite matching is to find
a perfect matching such that maximum edge weight (instead of the total weight) in the matching
is minimized [10].
The egalitarian bipartite matching problem is the same as minimum weight bipartite matching
except the distance cost function cd(s(x, d)) is the maximum edge weight in the assignment. This
function is also monotone and subadditive.
Metric Facility Location.
In this problem, one is given a set of agents A and a set of
facilities F such that A = n, F = m. The facilities and agents lie in a metric space d. Each
facility Fj ∈ F has an opening cost fj. Each agent is assigned to a facility; in different versions
there may be capacities on the number of agents assigned to a facility, lower bounds on the number
of agents assigned to a facility, or various other constraints [20]. The goal is to find a subset of
facilities F ⊆ F to open, so that the sum of opening costs for facilities in F and total distance of
the assignment is minimized.
Our framework allows arbitrary constraints on what constitutes a valid assignment, which cap-
tures facilities with capacities or lower bounds if needed. For any assignment x, the facility cost
function cf (x) is the sum of the opening costs fj for those facilities Fj that have at least one agent
assigned to it. The distance cost function cd(s(x, d)) is the total distances in the assignment, which
is monotone increasing and additive (thus subadditive).
k-center problem. The goal in this classic problem is to open a set of k facilities, with
each agent assigned to the closest one. The optimal solution is the subset of F which minimizes
maxi∈A d(i, x(i)). To express this in our framework, the constraint is that no more than k facilities
have agents assigned to them. For any assignment x, the facility cost function cf (x) = 0, and the
distance cost function cd(s(x, d)) is the maximum distance between any agent and facility in the
assignment.
k-median problem. This classic problem is the same as k-center, except the goal is to minimize
16
the sum of distances of agents to the facilities instead of the maximum distance.
5 Distortion of Facility Assignment Problems
In this section, we study general facility assignment problems, as described in Section 4, and form
mechanisms with small distortion. First, we construct a projected problem such that the distances
between agents and facilities are known, so it could be solved by an omniscient algorithm. Then
we map the result of the projected problem to the original problem and bound the distortion of
the original problem.
Given agents A = {1, 2, . . . , n} and facilities F = {F1, F2, . . . , Fm}, suppose facility F (cid:48) is i's
top choice in F. We create a new agent i at the location of F (cid:48) in the metric space. Consequently,
∀F ∈ F, d(i, F ) = d(F (cid:48), F ). Denote the set of the new agents as A = {1, 2, . . . , n}.
The original assignment problem is on agents A and facilities F, and only ordinal preferences
of agents in A over facilities are given. The projected problem is on agents A and facilities F, and
we know the actual distances between agents in A and facilities F, since we know the distances l
between facilities. The constraints and costs cd and cf remain the same for both the original and
the projected problem; the only difference is in the distances d. Our main result is that if we have
a β-approximation assignment to the minimum assignment cost on the projected problem, then
we can get an assignment that has a distortion of 2β +1 for the original problem in polynomial time.
Theorem 5.1. Given a valid assignment x for the projected problem on A and F, with ratio(x) ≤
β, the assignment x(i) = x(i) has distortion of at most (1 + 2β) for original assignment problem
on A and F.
Proof. First, x is a valid assignment for the projected problem on A and F, so x must also be a
valid assignment for the original problem on A and F. This is because the constraints are only on
the assignment, and are independent of the metric space d. For the same reason, the facility cost
of x equals the facility cost of x, cf (x) = cf (x).
Now consider the distance cost of x. Let x∗ denote the optimal assignment for the original
problem. ∀i ∈ A, let si = d(i, x(i)), ti = d(i,i), bi = d(i, x(i)). Similarly, let s∗
i = d(i, x∗(i)),
b∗
i = d(i, x∗(i)).
For any agent i and facility x(i), by triangle inequality,
si = d(i, x(i)) ≤ d(i,i) + d(i, x(i)) = ti + bi
Because cd is monotonically nondecreasing and subadditive,
cd(s1, s2, . . . , sn) ≤ cd(t1 + b1, t2 + b2, . . . , tn + bn)
≤ cd(t1, t2, . . . , tn) + cd(b1, b2, . . . , bn)
Therefore, the cost of our assignment x is bounded as follows:
cf (x) + cd(s(x, d)) = cf (x) + cd(s1, s2, . . . , sn)
= cf (x) + cd(s1, s2, . . . , sn)
≤ cf (x) + cd(t1, t2, . . . , tn) + cd(b1, b2, . . . , bn)
17
Because i is located at i's top choice facility, and x∗(i) is a facility, we thus know that ti ≤ s∗
i ,
and by monotonicity cd(t1, t2, . . . , tn) ≤ cd(s∗
1, s∗
2, . . . , s∗
n). Thus,
cf (x) + cd(s(x, d)) ≤ cf (x) + cd(t1, t2, . . . , tn) + cd(b1, b2, . . . , bn)
n) + cd(b1, b2, . . . , bn)
≤ cf (x) + cd(s∗
2, . . . , s∗
1, s∗
We know that x is a β-approximation to the optimum assignment for the projected problem.
Its total cost is exactly cf (x) + cd(b1, b2, . . . , bn), since the distance from i to x(i) = x(i) is exactly
bi. Now consider another assignment for the projected problem, in which i is assigned to x∗(i). The
cost of this assignment is cf (x∗)+cd(b∗
i . Since x is a β-approximation,
we therefore know that
n), by definition of b∗
2, . . . , b∗
1, b∗
cf (x) + cd(b1, b2, . . . , bn) ≤ βcf (x∗) + βcd(b∗
1, b∗
2, . . . , b∗
n),
and thus
cf (x) + cd(s(x, d)) ≤ cf (x) + cd(s∗
1, s∗
2, . . . , s∗
2, . . . , s∗
n) + βcf (x∗) + βcd(b∗
n) + cd(b1, b2, . . . , bn)
1, b∗
2, . . . , b∗
n)
≤ cd(s∗
1, s∗
For any agent i and facility x∗(i) in x∗, by triangle inequality,
i = d(i, x∗(i)) ≤ d(i, x∗(i)) + d(i,i) ≤ 2d(i, x∗(i)) = 2s∗
b∗
i
d(i,i) ≤ d(i, x∗(i)) above since i is located at the closest facility to i. Because cd is monotone and
subadditive, we also have that
cd(b∗
1, b∗
2, . . . , b∗
n) ≤ cd(2s∗
1, 2s∗
2, . . . , 2s∗
n) ≤ 2cd(s∗
1, s∗
2, . . . , s∗
n)
Putting everything together,
cf (x) + cd(s(x, d)) ≤ cd(s∗
1, s∗
2, . . . , b∗
1, b∗
n) + βcf (x∗) + βcd(b∗
2, . . . , s∗
n)
≤ βcf (x∗) + cd(s∗
n) + 2βcd(s∗
2, . . . , s∗
1, s∗
1, s∗
2, . . . , s∗
n)
1, s∗
= βcf (x∗) + (1 + 2β)cd(s∗
2, . . . , s∗
n)
≤ (1 + 2β)(cf (x∗) + cd(s(x∗, d)))
Note that the above theorem immediately implies that if we are only concerned with what is
possible to achieve given limited ordinal information in addition to distances between facilities, and
are not worried about our algorithms running in polynomial time, then we can always form an
assignment with distortion of at most 3 from knowing only σ and l. This is because we can solve
the projected problem with brute force, and then we have β = 1. This bound of 3 is tight for many
facility assignment problems: consider for example an instance of min-cost metric matching with
two agents and two facilities, with both preferring F1 to F2. One of the agents has distance to F1
18
of 0, and one is located halfway between F1 and F2, but since we only have ordinal information
we do not know which agent is which. If we assign the wrong agent to F1, then we end up with
distortion of 3, and it is impossible to do better for any deterministic mechanism.
If on the other hand we want to form poly-time algorithms with small distortion, the above
theorem gives a black-box reduction: if we have a β-approximation algorithm for the omniscient
case, then we can form a 1 + 2β-distortion algorithm for the ordinal case. Actually, we get a
1 + 2β-distortion for the distance cost, and a β-distortion for the facility cost, which is shown in
the second-to-last line of the proof for Theorem 5.1. This leads to the following corollaries:
Corollary 5.1.1. We can achieve the following distortion in polynomial time:
1. At most 3 for the minimum weight bipartite matching problem.
2. At most 3 for Egalitarian bipartite matching.
3. At most 3.976 for the facility location problem (1.488-approximation for the facility cost, and
3.976-approximation for the distance cost).
4. At most 5 for the k-center problem.
5. At most 6.35 for the k-median problem.
Proof. Min-cost matching and egalitarian matching are poly-time solvable, so β = 1. For the
latter, one can fix the threshold weight t such that every edge chosen should be at most t, and then
determine if such a matching exists. Performing a binary search on t gives an efficient algorithm.
For facility location, one can use the omniscient algorithm which is a 1.488-approximation in [27].
For the k-center problem, a greedy algorithm [25] gives a 2-approximation for the setting that agents
are a subset of facilities, which is the case in our projected problem. [12] gives a 2.675-approximation
omniscient algorithm for the k-median problem when agents are a subset of facilities, thus it also
gives a 2.675-approximation for our projected problem.
Note that the median function, unlike sum and maximum, is not subadditive, and thus does
not fit into our framework. In fact, while both min-cost and egalitarian matching problems have
algorithms with small distortion in our setting, the same is not possible for forming a matching
where the objective function is the cost of the median edge: see the Appendix for a lower bound.
6 Conclusion
In this paper, we provided two mechanisms to solve different social cost problems. The first one
makes use of the distances between facilities and an omniscient algorithm to get a low distortion for
general facility assignment problems. The second mechanism is a new voting rule for social choice
which simultaneously achieves a distortion of 3 for many objectives, including the cost of the median
voter, and a distortion of 5 for the total social cost at the same time. The first mechanism requires
the full distances l, but only needs the top choice from each agent. Thus, it puts only a small
load on the agents which submit their preferences, but requires the mechanism designer to collect
more information about the facilities and their distances to each other. The second mechanism, on
the other hand, only requires ordinal preference information from the facilities, but needs the full
preference ranking from the agents instead of just the top choice. It is especially appropriate for
settings in which the candidates or alternatives are agents themselves.
Many open questions remain for our setting. How well can facility location problems be ap-
proximated given information about facilities? While we established upper bounds on distortion,
19
we have no lower bounds besides the trivial bound of 3. What about randomized mechanisms, or
what if the mechanisms must be truthful? And more generally, exactly what information is enough
to guarantee mechanisms with small distortion? While our results show that knowing information
about facility locations is enough to result in small distortion, it may be possible that obtaining
even a bit of targeted information would result in powerful approximation algorithms. We look
forward to future work on this topic.
References
[1] Ben Abramowitz and Elliot Anshelevich. Utilitarians without utilities: Maximizing social
welfare for graph problems using only ordinal preferences. In AAAI 2018.
[2] Elliot Anshelevich. Ordinal approximation in matching and social choice. ACM SIGecom
Exchanges, 15(1):60–64, 2016.
[3] Elliot Anshelevich, Onkar Bhardwaj, and John Postl. Approximating optimal social choice
under metric preferences. In AAAI 2015.
[4] Elliot Anshelevich and John Postl. Randomized social choice functions under metric prefer-
ences. In IJCAI 2016.
[5] Elliot Anshelevich and Shreyas Sekar. Blind, greedy, and random: Algorithms for matching
and clustering using only ordinal information. In AAAI 2016.
[6] Elliot Anshelevich and Shreyas Sekar. Truthful mechanisms for matching and clustering in an
ordinal world. In WINE 2016.
[7] Elliot Anshelevich and Wennan Zhu. Tradeoffs between information and ordinal approximation
for bipartite matching. In SAGT 2017.
[8] Gerdus Benade, Swaprava Nath, Ariel D Procaccia, and Nisarg Shah. Preference elicitation
for participatory budgeting. In AAAI 2017.
[9] Anand Bhalgat, Deeparnab Chakrabarty, and Sanjeev Khanna. Social welfare in one-sided
matching markets without money. In APPROX 2011.
[10] Anna Bogomolnaia and Herv´e Moulin. Random matching under dichotomous preferences.
Econometrica, 72(1):257–279, 2004.
[11] Craig Boutilier, Ioannis Caragiannis, Simi Haber, Tyler Lu, Ariel D Procaccia, and Or Sheffet.
Optimal social choice functions: A utilitarian view. Artificial Intelligence, 227:190–213, 2015.
[12] Jaros(cid:32)law Byrka, Thomas Pensyl, Bartosz Rybicki, Aravind Srinivasan, and Khoa Trinh. An
improved approximation for k-median, and positive correlation in budgeted optimization. In
SODA 2014.
[13] Ioannis Caragiannis, Aris Filos-Ratsikas, Søren Kristoffer Stiil Frederiksen, Kristoffer Arnsfelt
Hansen, and Zihan Tan. Truthful facility assignment with resource augmentation: An exact
analysis of serial dictatorship. In WINE 2016.
[14] Ioannis Caragiannis, Swaprava Nath, Ariel D Procaccia, and Nisarg Shah. Subset selection
via implicit utilitarian voting. Journal of Artificial Intelligence Research, 58:123–152, 2017.
20
[15] Yu Cheng, Shaddin Dughmi, and David Kempe. Of the people: voting is more effective with
representative candidates. In EC 2017.
[16] Yu Cheng, Shaddin Dughmi, and David Kempe. On the distortion of voting with multiple
representative candidates. In AAAI 2018.
[17] George Christodoulou, Aris Filos-Ratsikas, Søren Kristoffer Stiil Frederiksen, Paul W Gold-
berg, Jie Zhang, and Jinshan Zhang. Social welfare in one-sided matching mechanisms. In
AAMAS 2016.
[18] James M. Enelow and Melvin J. Hinich. The Spatial Theory of Voting: An Introduction.
Cambridge University Press, New York, NY, 1984.
[19] Brandon Fain, Ashish Goel, Kamesh Munagala, and Sukolsak Sakshuwong. Sequential delib-
eration for social choice. In WINE 2017.
[20] Reza Zanjirani Farahani and Masoud Hekmatfar. Facility location: concepts, models, algo-
rithms and case studies. Springer, 2009.
[21] Michal Feldman, Amos Fiat, and Iddan Golomb. On voting and facility location. In EC 2016.
[22] Aris Filos-Ratsikas, Søren Kristoffer Stiil Frederiksen, and Jie Zhang. Social welfare in one-
sided matchings: Random priority and beyond. In SAGT 2014.
[23] Ashish Goel, Anilesh Kollagunta Krishnaswamy, and Kamesh Munagala. Metric distortion of
social choice rules: Lower bounds and fairness properties. In EC 2017.
[24] Stephen Gross, Elliot Anshelevich, and Lirong Xia. Vote until two of you agree: Mechanisms
with small distortion and sample complexity. In AAAI 2017.
[25] Dorit S Hochbaum and David B Shmoys. A best possible heuristic for the k-center problem.
Mathematics of operations research, 10(2):180–184, 1985.
[26] Martin Hoefer and Bojana Kodric. Combinatorial secretary problems with ordinal information.
In ICALP 2017.
[27] Shi Li. A 1.488 approximation algorithm for the uncapacitated facility location problem. In
ICALP 2011.
[28] Ariel D Procaccia and Jeffrey S Rosenschein. The distortion of cardinal preferences in voting.
In CIA 2006.
[29] Piotr Krzysztof Skowron and Edith Elkind. Social choice under metric preferences: Scoring
rules and stv. In AAAI 2017.
21
A Bad Examples and Lower Bounds
Note that our Algorithm 2 is only for social choice problems, and does not fit in the definition of
our general facility assignment problems. This is because the median cost function, unlike sum and
maximum, is not subadditive. In fact, while both min-cost and egalitarian matching problems have
algorithms with small distortion in our setting, the same is not possible for forming a matching
where the objective function is the cost of the median edge.
Theorem A.1. The worst-case distortion of the median-cost bipartite matching problem in a metric
space (given both agent preference profiles and distances between facilities) is unbounded.
Proof. Consider the following example: there are three agents a, b, c, and three facilities X, Y ,
Z. The preferences of agents are: a, b ∈ XY Z, while c ∈ ZXY . The distances between facilities
are: l(X, Y ) = 2, l(X, Z) = l(Y, Z) = 1000. The distances between the agents and facilities are,
of course, unknown. Consider the instance d(c, Z) = , d(a, X) = 2, and d(b, X) = d(b, Y ) = 1.
is a very small positive real number, and other distances not given obey triangle inequality. In
this instance, the optimal solution is x∗ = {(a, X), (b, Y ), (c, Z)}, which gives a median value of
2. But because a and b have the same preference profile, the instance could also be d(c, Z) = ,
d(b, X) = 2, and d(a, X) = d(a, Y ) = 1. If we still return the assignment x∗ for this instance, the
median would be 1. The distortion is arbitrarily bad when approaches 0.
The following Theorems show some of the lower bounds mentioned in Table 1.
Theorem A.2. The worst-case distortion for the facility location problem in a metric space (given
only agents' preference profiles) is unbounded.
Proof. Consider the following example: there are two agents 1, 2, and two facilities X, Y . Agent
1 prefers X to Y , while agent 2 prefers Y to X. The opening costs are: cf (X) = 1, cf (Y ) = 100.
We can choose to open one facility or both of them.
Case 1. Suppose we only open X. Consider the following distances between the agents and
facilities: d(1, X) = d(2, Y ) = 1, d(1, Y ) = d(2, X) = L, for some very large L. If we only open X,
then the total cost is > L. While the optimal solution is to open both X and Y , which has a total
cost of 103. The distortion is unbounded.
Case 2. Suppose we only open Y . Consider the same distances as in Case 1, then the total
cost is also L. And the optimal solution still has a total cost of 103. The distortion is unbounded.
Case 3. Suppose we open both facilities. Consider the following distances between the agents
and facilities: d(1, X) = d(1, Y ) = d(2, X) = d(2, Y ) = , where is a very small positive real
number. If we open both facilities, the total cost is 101 + 2. While the optimal solution is to
only open X , which has a total cost of 1 + 2. If we increase cf (Y ), the approximation ratio is
unbounded.
Theorem A.3. The worst-case distortion for the k-median problem in a metric space (given only
agents' preference profiles) is at least Ω(n).
Proof. Consider the following example: There are three facilities X, Y , and Z. There are q agents
who prefer X to Y to Z, q agents who prefer Y to X to Z, and 1 agent who prefers Z to X to Y .
We denote these three sets of agents as AX , AY and AZ separately. Suppose k = 2, then we have
three choices of the winners:
Case 1. Choose X, Y as the winners. Consider the following distances between agents and
facilities: d(X, Y ) = 1, d(Y, Z) = d(X, Z) = L for some very large L. AX is located at the same
location as X, AY is at the same location as Y , and AZ is at the same location as Z. The cost of
22
choosing X, Y as the winners is L because we need to assign the agent in AZ to X or Y . While
the optimal solution is to choose Y, Z as the winners, and get a total cost of q. So the distortion in
this case is unbounded.
Case 2. Choose X, Z as the winners. Consider the following distances between agents and
facilities: d(X, Y ) = d(Y, Z) = d(X, Z) = 1, and AX locate on top of X, AY locate on top of Y ,
and AZ locate on top of Z. The cost of choosing X, Z as the winner is q, while the optimal solution
is to choose X, Y as the winners, and get a total cost of 1. The distortion is q in this case.
Case 3. Choose Y, Z as the winners. Consider the same distances as in Case 2. If we choose
Y, Z as the winners, the total cost is still q, and the distortion of this case is also q.
The total number of agents is n = 2q + 1, so we can conclude that the distortions in all these
three cases are at least Ω(n).
Theorem A.4. The worst-case distortion of the egalitarian bipartite matching problem in a metric
space (given only agents' preference profiles) is at least 2.
Proof. Consider the following example: there are two agents 1, 2, and two facilities X, Y . Both
agents prefer X to Y . W.L.O.G., assume we match agent 1 to X, and agent 2 to Y . Suppose the
distances between agents and facilities are: d(1, X) = d(1, Y ) = 1, d(2, X) = , d(2, Y ) = 2, where
is a very small positive real number. The egalitarian cost of our matching is 2, while the optimal
solution is to match agent 1 to Y , and agent 2 to X, which has a cost of 1.
23
|
1308.1049 | 1 | 1308 | 2013-08-05T17:43:58 | Coevolutionary networks of reinforcement-learning agents | [
"cs.MA",
"cs.LG",
"nlin.AO"
] | This paper presents a model of network formation in repeated games where the players adapt their strategies and network ties simultaneously using a simple reinforcement-learning scheme. It is demonstrated that the coevolutionary dynamics of such systems can be described via coupled replicator equations. We provide a comprehensive analysis for three-player two-action games, which is the minimum system size with nontrivial structural dynamics. In particular, we characterize the Nash equilibria (NE) in such games and examine the local stability of the rest points corresponding to those equilibria. We also study general n-player networks via both simulations and analytical methods and find that in the absence of exploration, the stable equilibria consist of star motifs as the main building blocks of the network. Furthermore, in all stable equilibria the agents play pure strategies, even when the game allows mixed NE. Finally, we study the impact of exploration on learning outcomes, and observe that there is a critical exploration rate above which the symmetric and uniformly connected network topology becomes stable. | cs.MA | cs |
Coevolutionary networks of reinforcement-learning agents
Information Sciences Institute, University of Southern California, Marina del Rey, CA 90292, USA
Ardeshir Kianercy and Aram Galstyan
This paper presents a model of network formation in repeated games where the players adapt
their strategies and network ties simultaneously using a simple reinforcement learning scheme. It
is demonstrated that the co-evolutionary dynamics of such systems can be described via coupled
replicator equations. We provide a comprehensive analysis for three-player two-action games, which
is the minimum system size with nontrivial structural dynamics. In particular, we characterize the
Nash equilibria (NE) in such games and examine the local stability of the rest points corresponding
to those equilibria. We also study general n-player networks via both simulations and analytical
methods and find that in the absence of exploration, the stable equilibria consist of star motifs as
the main building blocks of the network. Furthermore, in all stable equilibria the agents play pure
strategies, even when the game allows mixed NE. Finally, we study the impact of exploration on
learning outcomes, and observe that there is a critical exploration rate above which the symmetric
and uniformly connected network topology becomes stable.
PACS numbers: 89.75.Fb,05.45.-a,02.50.Le,87.23.Ge
I.
INTRODUCTION
Networks depict complex systems where nodes corre-
spond to entities and links encode interdependencies be-
tween them. Generally, dynamics in networks is intro-
duced via two different approaches. In the first approach,
the links are assumed to be static, while the nodes are en-
dowed with internal dynamics (epidemic spreading, opin-
ion formation, signaling, synchronizing and so on). And
in the second approach, nodes are treated as passive ele-
ments, and the main focus is on the evolution of network
topology.
More recently, it has been suggested that separating
individual and network dynamics fails to capture real-
istic behavior of networks.
Indeed, in most real -- world
networks both the attributes of individuals (nodes) and
the topology of the network (links) evolve in tandem.
Models of such adaptive co-evolving networks have at-
tracted significant interest in recent years both in statis-
tical physics [1 -- 5] and game theory and behavioral eco-
nomics communities [6 -- 11].
To describe coupled dynamics of individual attributes
and network topology, here we suggest a simple model of
a coevolving network that is based on the notion of inter-
acting adaptive agents. Specifically, we propose network --
augmented multiagent systems where the agents play re-
peated games with their neighbors, and adapt both their
behaviors and the network ties depending on the out-
come of their interactions. To adapt, the agents use a
simple learning mechanism to reinforce (penalize) be-
haviors and network links that produce favorable (un-
favorable) outcomes. Furthermore, the agents use an
action selection mechanism that allows one to control
exploration/exploitation tradeoff via a temperature-like
parameter.
We have previously demonstrated [12] that the collec-
tive evolution of such a system can be described by ap-
propriately defined replicator dynamics equations. Orig-
inally suggested in the context of evolutionary game the-
ory (e.g., see Refs. [13, 14]), replicator equations have
been used to model collective learning in systems of in-
teracting self -- interested agents [15]. Refrence [12] pro-
vides a generalization to the scenario where the agents
adapt not only their strategies (probability of selecting a
certain action) but also their network structure (the set
of other agents that play against). This generalization
results in a system of coupled non-linear equations that
describe the simultaneous evolution of agent strategies
and network topology.
Here we use the framework suggested in Ref. [12] to
examine the learning outcomes in networked games. We
provide a comprehensive analysis of three-player two-
action games, which are the simplest systems that ex-
hibit non-trivial structural dynamics. We analytically
characterize the rest-points and their stability properties
in the absence of exploration. Our results indicate that
in the absence of exploration, the agents always play pure
strategies even when the game allows mixed NE. For the
general n-player case, we find that the stable outcomes
correspond to star-like motifs, and demonstrate analyti-
cally the stability of a star motif. We also demonstrate
the instability of the symmetric network configuration
where all the pairs are connected to each other with uni-
form weights.
We also study the the impact of exploration on the co-
evolutionary dynamics. In particular, our results indicate
that there is a critical exploration rate above which the
uniformly connected network is a globally stable outcome
of the learning dynamics.
The rest of the paper is organized as follows: we next
derive the replicator equations characterizing the coevo-
lution of the network structure and the strategies of the
agents. In Sec. III we focus on learning without explo-
ration, describe the NE of the game, and characterize the
restpoints of learning dynamics according to their stabil-
ity properties. We consider the the impact of exploration
on learning in Sec. IV and provide some concluding re-
marks in Sec. V.
II. CO-EVOLVING NETWORKS VIA
REINFORCEMENT LEARNING
Let us consider a set of agents that play repeated
games with each other. We differentiate agents by in-
dices x, y, z, . . .. At each round of the game, an agent
has to choose another agent to play with, and an action
from the pool of available actions. Thus, time -- dependent
mixed strategies of agents are characterized by a joint
probability distribution over the choice of the neighbors
and the actions.
We assume that the agents adapt to their environment
through a simple reinforcement mechanism. Among dif-
ferent reinforcement schemes, here we focus on (stateless)
Q-learning [16]. Within this scheme, the strategies of
the agents are parameterized through, so-called Q func-
tions that characterize the relative utility of a particular
strategy. After each round of game, the Q functions are
updated according to the following rule,
Qi
xy(t + 1) = Qi
xy(t) + α[Ri
x,y(t) − Qi
xy(t)]
(1)
x,y(Qi
where Ri
x,y) is the expected reward (Q value) of
agent x for playing action i with agent y, and α is a
parameter that determines the learning rate (which can
be set to α = 1 without a loss of generality).
Next, we have to specify how agents choose a neighbor
and an action based on their Q function. Here we use the
Boltzmann exploration mechanism where the probability
of a particular choice is given as [17]
pi
xy =
xy(cid:80)
eβQi
y,j eβQj
x y
(2)
where pi
xy is the probability that agent x will play with
agent y and choose action i. Here the inverse temperature
β ≡ 1/T > 0 controls the tradeoff between exploration
and exploitation; for T → 0 the agents always choose the
action corresponding to the maximum Q value, while for
T → ∞ the agents' choices are completely random.
x,y = (cid:80)
We now assume that the agents interact with each
other many times between two consecutive updates of
their strategies. In this case, the reward of the i th agent
in Eq. ( 1) should be understood in terms of the average
reward, where the average is taken over the strategies of
other agents, Ri
xy is the re-
ward (payoff) of agent x playing strategy i against agent
y who plays strategy j. Note that, generally speaking,
the payoff might be asymmetric.
We are interested in the continuous approximation to
the learning dynamics. Thus, we replace t + 1 → t + δt,
α → αδt, and take the limit δt → 0 in Eq. (1) to obtain
yx, where Aij
j Aij
xypj
Qi
xy = α[Ri
x,y − Qi
xy(t)]
(3)
Differentiating Eq. (2), using Eqs. (2) and (3), and scal-
ing the time t → αβt, we obtain the following replicator
2
(cid:88)
equation [15]:
(cid:88)
yx −(cid:88)
j
y,j
=
i,j,y
Aij
Aij
xypj
xypi
xypj
yx + T
pj
xy ln
pi
xy
pi
xy
pj
xy
pi
xy
(4)
Equations 4 describe the collective adaptation of the Q --
learning agents through repeated game -- dynamical inter-
actions. The first two terms indicate that the probabil-
ity of playing a particular pure strategy increases with a
rate proportional to the overall goodness of that strat-
egy, which mimics fitness-based selection mechanisms in
population biology [13]. The second term, which has an
entropic meaning, does not have a direct analog in pop-
ulation biology [15]. This term is due to the Boltzmann
selection mechanism that describes the agents' tendency
to randomize over their strategies. Note that for T = 0
this term disappears, so the equations reduce to the con-
ventional replicator system [13].
So far, we have discussed learning dynamics over a
general strategy space. We now make the assumption
that the agents' strategies factorize as follows,
(cid:88)
(cid:88)
pi
xy = cxypi
x ,
cxy = 1,
y
pi
x = 1.
i
(5)
Here cxy is the probability that the agent x will initiate
a game with the agent y, whereas pi
x is the probability
that he will choose action i. Thus, the assumption behind
this factorization is that the probability that the agent
will perform action i is independent of whom the game
is played against. Substituting Eq. (5) in Eq. (4) yields
(cid:20)(cid:88)
x −(cid:88)
j
y −(cid:88)
cxy ln cxy −(cid:88)
i,y,j
y
j
(cid:20)
−T
ln cxy + ln pi
cxypi
x + cxy pi
x = cxypi
x
aij
xycyxpj
aij
x,ycxycyxpi
xpj
y
(cid:21)(cid:21)
pj
x ln pj
x
(6)
Next, we sum both sides in Eq. (6), once over y and then
over i, and make use of the normalization conditions in
Eq. (5) to obtain the following coevolutionary dynamics
of action and connection probabilities:
(7)
(8)
+ T
cxy ln(cxy/cxy)
y
Equations (7) and
(8) are the replicator equations
that describe the collective evolution of both the agents'
strategies and the network structure.
The following remark is due: Generally, the replicator
dynamics in matrix games are invariant with respect to
xycxycyxpj
Aij
xycxycyxpi
xpj
y
+ T
pj
x ln(pj
x/pi
x)
= cyx
Aij
xypi
xpj
Aij
xycxycyxpi
xpj
y
y −(cid:88)
y −(cid:88)
i,j,y
i,j,y
pi
x
pi
x
=
cxy
cxy
y,j
Aij
(cid:88)
(cid:88)
(cid:88)
(cid:88)
i,j
j
adding any column vector to the payoff matrix. However,
this invariance does not hold in the present networked
game. The reason for this is the following:
if an agent
does not have any incoming links (i.e., no other agent
plays with him or her), then he always gets a zero re-
ward. Thus, the zero reward of an isolated agent serves
as a reference point. This poses a certain problem. For
instance, consider a trivial game with a constant reward
matrix aij = P . If P > 0, then the agents will tend to
play with each other, whereas for P < 0 they will try to
avoid the game by isolating themselves (i.e., linking to
agents that do not reciprocate).
To address this issue, we introduce an isolation payoff
CI that an isolated agent receives at each round of the
game.
It can be shown that the introduction of t his
payoff merely subtracts CI from the reward matrix in
the replicator learning dynamics. Thus, we paramet rize
the game matrix as follows:
aij = bij + CI
(9)
3
FIG. 1: (Color online) Categorization of two-action games
based on the reward matrix structure in the (a, b) plane.
where matrix B defines a specific game.
where
Although it is beyond the scope of the present paper,
an interesting question is what the reasonable values for
the parameter CI are. In fact, what is important is the
value of CI relative to the reward at the corresponding
Nash equilibria, i.e., whether not playing at all is better
than playing and receiving a potentially negative reward.
Different values of CI describe different situations.
In
particular, one can argue that certain social interactions
are apparently characterized by large CI , where not par-
ticipating in a game is seen as a worse outcome than par-
ticipating and getting negative rewards. In the following,
we treat CI as an additional parameter that changes in
a certain range, and examine its impact on the learning
dynamics.
A. Two-action games
We focus on symmetric games where the reward matrix
is the same for all pairs (x, y), Axy = A:
(cid:19)
(cid:18) a11 a12
a21 a22
A =
(10)
Let pα, α ∈ {x, y, . . . ,}, denote the probability for agent
α to play action 1 and cxy is the probability that agent
x will initiate a game with the agent y. For two action
games, the learning dynamics Eqs. (7) , and (8) becomes:
px
px(1 − px)
cxy
cxy
=
(apy + b)cxycyx + T log
y
= rxy − Rx + T
(cid:88)
y
1 − px
px
(11)
(cid:88)
(cid:88)
rxy = cyx(apxpy + bpx + dpy + a22)
Rx =
(apxpy + bpx + dpy + a22)cxycyx
y
Here we have defined the following parameters:
a = a11 − a21 − a12 + a22
b = a12 − a22
d = a21 − a22
(13)
(14)
(15)
(16)
(17)
The parameters a and b allow a categorization of two
action games as follows (Fig. 1):
• dominant action games: − b
a < 0
• coordination game: a > 0, b < 0 and 1 ≥ − b
• anti-coordination (Chicken) game: a < 0, b >
a > 1 or − b
a
0 and 1 ≥ − b
a
Before proceeding further, we elaborate on the con-
nection between the rest points of the replicator system
for T = 0 and the game-theoretic notion of NE (NE) 1.
For T = 0 (no exploration) in the conventional replicator
equations, all NE are necessarily the rest points of the
learning dynamic. The inverse is not true - not all rest
points correspond to NE - and only the stable ones do.
Note that in the present model the first statement does
not necessarily hold. This is because we have assumed
the strategy Eq.( 5), due to which equilibria where the
cxy ln
cxy
cxy
(12)
1 Recall that a joint strategy profile is called NE if no agent can
increase his expected reward by unilaterally deviating from the
equilibrium.
baDominant action Dominant action Coordination game Chicken game b = -a 4
FIG. 3: (Color online) Three-player network NE for prisoner's
dilemma and the coordination game; see the text for more
details.
operation and Defection, and the payoff matrix elements
satisfy b21 > b11 > b22 > b12; (see Fig. 2).
In a two-
player PD game, defection is a dominant strategy -- it al-
ways yields a better reward regardless of the other player
choice -- thus, the only NE is a mutual defection. And in
coordination game, the players have an incentive to se-
lect the same action. This game has two pure NE, where
the agents choose the same action, as well as a mixed
NE. In a general coordination game the reward elements
have the relationship b11 > b21, b22 > b12 (see Fig. 2).
In the three-agent scenario, a simple analysis yields
four possible network topologies corresponding to NE de-
picted in Fig. 3. In all of those configurations, the agents
that are not isolated select strategies that correspond
to two-agent NE. Thus, in the case of PD, non-isolated
agents always defect, whereas for the coordination game,
they can select one of three possible NE. We now examine
those configurations in more details.
Configuration I In this configuration, the agents x
and y play only with each other, whereas agent is z s
isolated: cxy = cyx = 1. Note that for this to be a NE,
agents x and y should not be "tempted" to switch and
play with the agent z. For instance, in the case of PD,
this yields pzb21 < b22, otherwise players x and y will be
better of linking with the isolated agent z and exploiting
his cooperative behavior. 3
Configuration II In the second configuration, there
is a central agent (z) who plays with the other two:
cxz = cyz = 1, czx + czy = 1. Note that this configuration
is continuously degenerate as the central agent can
FIG. 2: Examples of reward matrices for typical two-action
games.
agents adopt different strategies with different players are
not allowed. Thus, any NE that do not have the factor-
ized form simply cannot be described in this framework.
The second statement, however, remains true, and stable
rest points do correspond to NE.
III.
LEARNING WITHOUT EXPLORATION
For T = 0, the learning dynamics Equations (11), (12)
attain the following form:
(cid:88)
px
px(1 − px)
cxy
cxy
=
(apy + b)cxycyx
(18)
y
= rxy − Rx
(19)
Consider the dynamics of the strategies given by
Eq. 18. Clearly, the vertices of the simplex, px = {0, 1}
are the rest points of the dynamics. Furthermore, in
case the game allows a mixed NE, then the configuration
where all the agents play the mixed NE px = −b/a is
also a rest point of the dynamics. As is shown below,
however, this configuration is not stable, and for T = 0,
the only stable configurations correspond to the agents
playing pure strategies.
A. Three-player games
We now consider the case of three players in two-action
games. This scenario is simple enough for studying it
comprehensively, yet it still has non-trivial structural dy-
namics, as we demonstrate below.
1. Nash equilibria
We start by examining the NE for two classes of two-
action games, prisoner dilemma (PD) and a coordination
game. 2 In PD, the players have to choose between Co-
2 The behavior of the coordination and anti-coordination games
are qualitatively similar in the context of the present work, so
here we do not consider the latter.
3 Note that the dynamics will eventually lead to a different rest
point where z is now plays defect with both x and y.
XZ Y
XZ Y
XZ Y
XZ Y
Mixed strategy p∈{1,0,−ba} Nash strategy 1≥p≥0 d b c a distribute his link weight arbitrarily among the two
players. Additionally, the isolation payoff should be
smaller then than the reward at the equilibrium (e.g.,
if the latter condition is
b22 > CI for PD). Indeed,
reversed, then one of the agents, say x,
is better off
linking with y instead of z, thus "avoiding" the game
altogether.
Configuration III: The third configuration corre-
sponds to a uniformly connected networks where all
the links have the same weight cxy = cyz = ccx = 1
2 .
It is easy to see that when all three agents play NE
strategies, there is no incentive to deviate from the
uniform network structure.
in the last configuration
Configuration IV: Finally,
none of the links are reciprocated so that the players do
not play with each other: cxycyx = cxzczx = cyzczy = 0.
This cyclic network is a NE when the isolation payoff CI
is greater than the expected reward of playing NE in the
respective game.
2. Stable rest points of learning dynamics
The factorized NE discussed in the previous section
are the rest points of the replicator dynamics. However,
not all of those rest points are stable, so that not all the
equilibria can be achieved via learning. We now discuss
the stability property of the rest points.
One of the main outcomes of our stability analysis is
that at T = 0 the symmetric network configuration is not
stable. This is in fact a more general results that applies
to n-agent networks, as is shown in the next section. As
we will demonstrate later, the symmetric network can be
stabilized when one allows exploration.
The second important observation is that even when
the game allows mixed NE, such as in the coordination
game, any network configuration where the agents play
mixed strategy is unstable for T = 0 (see Appendix A).
Thus, the only outcome of the learning is a configuration
where the agents play pure strategies.
The surviving (stable) configurations are listed in
Fig. 4. Their stability can be established by analyzing
the eigenvalues of the corresponding Jacobian. Consider,
for instance, the configuration with one isolated player.
The corresponding eigenvalues are
λ1 = rxz − rxy , λ2 = ryz − ryx , λ3 = 0
λ4 = (1 − 2px)(r1
λ5 = (1 − 2py)(r1
x) < 0 ,
y) < 0 , λ6 = 0
x − r2
y − r2
For PD this configuration is marginally stable when
agents x and y play defect and rxy > 0 and ryx > 0.
It happens only when b22 ≥ −CI which means that
the isolation payoff should be less than the expected re-
ward for defection. Furthermore, one should also have
rxz < rxy , ryz < ryx, which indicates that the neither
5
FIG. 4:
(Color online) Stable rest points of the learning
dynamics for PD (upper panel) and the coordination game
(lower panel).
x nor y would get a better expected reward by switching
and playing with z (e.g., condition for NE). And for the
coordination game , assuming that b11 > b22 this config-
uration is stable when b11 ≥ −CI > b22 and b22 ≥ −CI .
Similar reasoning can be used for the other configu-
rations shown in Fig. 4. Note, finally, that there is a
coexistence of multiple equlibria for range of parameter,
except when the isolation payoff is sufficiently large, for
which the cyclic (non-reciprocated) network is the only
stable configuration.
B.
n-player games
In addition to the three agent scenario, we also ex-
amined the co-evolutionary dynamics of general n-agent
systems, using both simulations and analytical methods.
We observed in our simulations that the stable outcomes
of the learning dynamic consist of star motifs Sn (Fig. 5),
where a central node of degree n − 1 connects to n − 1
nodes of degree 1. 4 Furthermore, we observed that the
basin of attraction of motifs shrinks as motif size grows,
so that smaller motifs are more frequent.
4 This is true when the isolation payoff is smaller compared to
the NE payoff. In the opposite case the dynamics settles into a
configuration without reciprocated links.
b22≥−CIb22<−CI Mixed strategy Action 1 Action 2
−CI>b11b11≥−CI>b22−CI≤b22 6
IV. LEARNING WITH EXPLORATION
In this section we consider the replicator dynamics for
non-vanishing exploration rate T > 0. For two agent
games, the effect of the exploration has been previously
examined in Ref. [18], where it was established that for a
class of games with multiple Nash equilibria the asymp-
totic behavior of learning dynamics undergoes a drastic
changes at critical exploration rates and only one of those
equilibria survives. Below, we study the impact of the ex-
ploration in the current networked version of the learning
dynamics.
For 3-player, 2- action games we have six independent
variables px, py, pz, cxy, cyz, andczx. The strategy vari-
ables evolve according to the following equations:
FIG. 5: Observed stable configurations of co-evolutionary dy-
namics for T = 0.
px
py
pz
cyz
czx
We now demonstrate the stability of the star motif
Sn in n player two action games. Let player x be the
central player, so that all other players are only con-
nected to x, cαx = 1. Recall that the Jacobian of the
system is a block diagonal matrix with blocks J11 with
elements ∂ cij
( see
∂cmn
Appendix A). When all players play a pure strategy
pi = 0, 1 in a star shape motif, it can be shown that
J22 is diagonal matrix with diagonal elements of form
y(apy + b)cxycyx, whereas J11 is an upper tri-
angular matrix, and its diagonal elements are either zero
or have the form −(apxpy + bpx + dpy + a22)cxy where x
is the central player.
and J22 with has elements as ∂ pm
∂pn
(1− 2px)(cid:80)
For the Prisoner's Dilemma, the Nash Equilibrium cor-
responds to choosing the second action (defection) , i.e.
pα = 0. Then the diagonal elements of J22, and thus
its eigenvalues, equal bcxy. J11, on the other hand, has
n2− 2n eigenvalues , (n− 1) of them are zero and the rest
have the form of λ = −a22cxy. Since for the Prisoner's
Dilemma one has b < 0 then the start structure is stable
as long as b22 > CI .
A similar reasoning can be used for the Coordination
game, for which one has b < 0 and a + b > 0. In this
case, the star structure is stable when either b11 > −CI or
b22 > −CI , depending on whether the agents coordinate
on the first or second actions, respectively.
We conclude this section by elaborating on the
(in)stability of the n-agent symmetric network configu-
ration, where each agent is connected to all the other
agents with the same connectivity 1
n−1 . As shown in Ap-
pendix B, this configuration can be a rest point of the
learning dynamics Eq. (18) only when all agents play the
same strategy, which is either 0, 1 or −b/a. Consider
now the first block of the Jacobian in Eq. A1, i.e. J11.
It can be shown that the diagonal elements of J11 are
identically zero, so that T r(J11) = 0. Thus, either all
the eigenvalues of J11 are zero (in which case the con-
figuration is marginally stable), or there is at least one
eigenvalue that is positive, thus making the symmetric
network configuration unstable at T = 0.
1 − px
px
1 − py
py
1 − pz
pz
cxy
= (apy + b)wxy + (apz + b)wxz + T log
= (apz + b)wyz + (apx + b)wxy + T log
(1 − px)px
(1 − py)py
(1 − pz)pz
cxy(1 − cxy)
cyz(1 − cyz)
czx(1 − czx)
Here we have defined wxy = cxy(1 − cyz), wxz = (1 −
cxy)czx, and wyz = cyz(1−czx), and a, b, andd are defined
in Eqs. 15, 16 and 17.
= (apx + b)wxz + (apy + b)wyz + T log
= rxy − rxz + T log
= ryz − ryx + T log
= rzx − rzy + T log
1 − cxy
cxy
1 − cyz
cyz
1 − czx
czx
Figure 6(a) shows three possible network configura-
tions that correspond to the fixed points of the above
dynamics. The first two configurations are perturbed ver-
sion of a star motif ( stable solution for T = 0), whereas
the third one corresponds to a symmetric network where
all players connect to the other players with equal link
weights.
Furthermore, in Fig. 6(b) we show the behavior of the
learning outcomes for a PD game, as one varies the tem-
perature. For sufficiently small T , the only stable config-
urations are the perturbed star motifs, and the symmet-
ric network is unstable. However, there is a critical value
Tc above which the symmetric network becomes globally
stable.
Next, we consider the stability of the symmetric net-
works. As shown in Appendix B, the only possible solu-
tion in this configuration is when all the agents play the
same strategy, which can be found from the following
equation:
(ap + b) = 2T log
p
1 − p
(20)
The behavior of this equation (without the factor 2 in the
right-hand side) was analyzed in details in Ref. [18]. In
S2SnS4S3... 7
FIG. 7: (Color online) Impact of the exploration on the sta-
ble outcomes of a coordination game in Fig. (2). The top
panel shows the bifurcation of strategy p versus T , whereas
the bottom panel shows the stability region of the symmetric
network configuration in the CI − T plane. Here the critical
temperature is Tc = 0.36.
reward (Fig. 7).
V. DISCUSSION
We have
studied the
co-evolutionary dynamics
of
strategies and link structure in a network of
reinforcement-learning agents. By assuming that the
agents' strategies allow appropriate factorization, we de-
rived a system of a coupled replicator equations that de-
scribe the mutual evolution of agent behavior and net-
work topology. We used these equations to fully charac-
terize the stable learning outcomes in the case of three
agents and two action games. We also established some
analytical results for the more general case of n-player
two-action games.
We demonstrated that in the absence of any strat-
(a)
(b)
FIG. 6: a) (Color online) Possible network configurations for
three-player PD (Fig. 2)., (b) Bifurcation diagram for varying
temperature. Two blue solid lines correspond to the config-
urations with one isolated agent and one central agent. The
symmetric network configuration is unstable at low temper-
ature (red line), and becomes globally stable above a critical
temperature.
particular, for games with a single NE, this equation al-
lows a single solution that corresponds to the perturbed
NE. For games with multiple equilibria, on the other
hand, there is a critical exploration rate Tc: For T < Tc
there are two stable solutions and one unstable solution,
whereas for T ≥ Tc there is a single globally stable solu-
tion.
We use these insights to examine the stability of the
symmetric network configuration for the coordination
game, depending on the parameters T and CI ; see Ap-
pendix C. In this example a = 5 , b = −2 and d = 1
for all three agents. Figure 7 shows the bifurcation dia-
gram of p (probability of choosing the first action) plot-
ted versus T . Below the critical temperature, there are
three three solutions, two of which (that correspond to
the perturbed pure NE) are stable. And Fig. (7) shows
the domain of T and CI for stable homogenous equilib-
rium. When T → 0, the domain of CI shrinks until it
becomes a point at T = 0 where −CI is equal to the NE
Perturbed pure NE 0<p<1 Strong connection Weak connection Uniform connection CxyT0.0Tc−b22−b11 egy exploration (zero temperature limit) learning leads
to a network composed of star-like motifs. Furthermore,
the agents on those networks play only pure NE, even
when the game allows a mixed NE. Also, even though
the learning dynamics allows rest points with a uniform
network (e.g., an agent plays with all the other agents
with the same probability) , those equilibria are not sta-
ble at T = 0. The situation changes when the agents
explore their strategy space. In this case, the stable net-
work structures undergo bifurcation as one changes the
exploration rate. In particular, there is a critical explo-
ration rate above which the uniform network becomes a
globally stable outcome of the learning dynamics.
We note that the main premise behind the strategy fac-
torization use here is that the agents use the same strat-
egy profile irrespective of whom they play against. While
this assumption is perhaps valid under certain circum-
stances, it certainly has its limitations that need to be
studied further through analytical results and empirical
data. Furthermore, the other extreme where the agent
employs unique strategy profiles for each of his partners
does not seem very realistic either, as it would impose
considerable cognitive load on the agent. A more realis-
tic assumption is that the agent has a few strategy profile
that roughly corresponds to the type of agent he is in-
teracting with. The approach presented here can be, in
principle, generalized to the latter scenario.
VI. ACKNOWLEDGMENTS
We thank Armen Allahverdyan for his comments and
contributions during the initial phase of this work. This
research was supported in part by the National Sci-
ence Foundation under Grant No. 0916534 and the U.S.
AFOSR MURI under Grant No. FA9550-10-1-0569.
Appendix A: Local Stability Analysis of the Rest
Points
To study the local stability properties of the rest points
in the system given by Eqs.18 and 19 , we need to analyze
the eigenvalues of the corresponding Jacobian matrix.
For n-player two-action game, we have n action variables
and l = n(n − 2) link variables, so that the total number
of independent dynamical variables is n + l = n(n − 1).
We can represent the Jacobian as follows,
Here the diagonal blocks J11 and J22 are l × l and n × n
square matrices, respectively. Similarly, J12 and J21 are
l × n and n × l matrices, respectively.
In the most general case, the full analysis of the Jaco-
bian is intractable. However, the problem can be simpli-
fied for T = 0. Indeed, consider the lower off-diagonal
∂ cij
∂cmn
∂ pm
∂cij
∂ cij
∂pm
∂ pm
∂pn
=
(cid:18)J11 J12
(cid:19)
J21 J22
J =
(A1)
8
block of the Jacobian, J21, the elements of which have
the form
= pi(1 − pi)cji(api + b)
∂ pi
∂cij
(A2)
Consulting the rest point condition given by Eqs. 18, one
can see that J21 is identically zero. By using the block
matrix determinant identity, the characteristic polyno-
mial of the Jacobian assumes the following factorized
form
p(λ) = det(J11 − λI)det(J22 − λI) = 0
(A3)
The above factorization facilitates the stability analysis
for certain cases that we now focus on:
a.
(In)Stability of mixed strategies for T = 0
Let us
show that the configurations where the agents mix either
on their actions or links cannot be stable at T = 0. Here
we just need to consider the submatrix J22. We now
show that this matrix always has at least one positive
eigenvalues when players adopt the mixed NE p = −b/a.
Indeed, it can be shown that J22 is a non-zero matrix
with zero diagonal elements. Recall that for any square
matrix A the T r(A) =(cid:80) λi then T r(J11) = 0 means at
least one of its eigenvalues is always positive, so that the
mixed Nash configuration is unstable. The same line of
reasoning can be applied to the configuration where the
agents mix over the links.
Appendix B: Agent Strategies in Symmetric
Networks
Let us consider a two-action n-players game. Each
player i chooses action one with probability pi. Here we
prove that player n and player n−1 in a homogenous net-
work have the same strategy, i.e., pn = pn−1. Consider
Eq. (11) for players n, n − 1 and n − 2,
p1 + p2 + ··· + pn−2 + pn−1 = k log
− c (B1)
pn
1 − pn
p1 + p2 + ··· + pn−2 + pn = k log
pn−1
1 − pn−1
− c (B2)
where
K = − T (n − 1)2
a
, c =
b(n − 1)
a
Also, let us define a function g as
g(pn) = xn + k log
pn
(1 − pn)
(B3)
(B4)
Now , by subtracting the two Eqs.(B1) and (B2), we
have g(pn) = g(pn−1). Since 0 < pi < 1 , then function
g is a monotonic function, so g(pn) = g(pn−1) ↔ pn =
pn−1. By repeating the same reasoning for the remaining
pi one can prove that p1 = p2 = ··· = pn.
Appendix C: Stability of Symmetric three-player
network
For three-player two-action games, the Jacobian cor-
responding to the symmetric network configuration con-
sists of the following blocks:
−v −T −v
−v −v −T
,
−T −v −v
0 m −m
,
0 −g
,
,
−T
−m 0 m
m −m 0
g
0 −g
0
g
k
k
k −T
k
k −T
k
g
−g
J11 =
J12 =
J21 =
J22 =
9
and p is the probability of selecting the first action, which
is the same for all the agents in the symmetric network
configuration. The six eigenvalues that determine the
stability of the configuration can be calculated analyti-
cally and are as follows:
(C1)
(C2)
(C3)
λ1 = 2k − T,
λ2 = −T − 2v,
λ3,4 =
(−k − 2T + v −(cid:112)12gm + (k + v)2),
(−k − 2T + v +(cid:112)12gm + (k + v)2).
1
2
1
2
(C4)
λ5,6 =
where we have defined
v =
m =
g =
k =
ap2 + bp + dp + b22 + CI
4
,
,
8
ap + d
p(1 − p)(ap + b)
2
ap(1 − p)
,
,
4
(C5)
(C6)
(C7)
(C8)
[1] Jorge M Pacheco, Arne Traulsen, and Martin A Nowak.
Coevolution of strategy and structure in complex net-
works with dynamical linking. Physical Review Letters,
97(25):258103, 2006.
[2] T. Gross and B. Blasius. Adaptive coevolutionary net-
works: a review. Journal of the Royal Society Interface,
5(20):259, 2008.
[3] S.; Castellano, C.;Fortunato and V. Loreto. Statistical
physics of social dynamics. Reviews of Modern Physics,
81(2):591, 2009.
[4] M. Perc and A. Szolnoki. Coevolutionary games -- a mini
review. BioSystems, 99(2):109 -- 125, 2010.
[5] Gerd Zschaler. Adaptive-network models of collective dy-
namics. The European Physical Journal-Special Topics,
211(1):1 -- 101, 2012.
[6] David Lazer. The co-evolution of individual and network.
Journal of Mathematical Sociology, 25(1):69 -- 108, 2001.
[7] M.O. Jackson and A. Watts. On the formation of interac-
tion networks in social coordination games. Games and
Economic Behavior, 41(2):265 -- 291, 2002.
These expressions can be used to (numerically) identify
the stability region of the configuration in the parameter
space (T, CI ), as shown in Fig. 7.
[10] S. Goyal. Connections: an introduction to the economics
of networks. Princeton University Press, 2009.
[11] Mathias Staudigl.
bayesian interaction games.
Game Theory, 42(1):179 -- 210, 2013.
Co-evolutionary dynamics and
International Journal of
[12] Ardeshir Kianercy, Aram Galstyan, and Armen Al-
lahverdyan. Adaptive agents on evolving networks. In
Proceedings of the 11th International Conference on Au-
tonomous Agents and Multiagent Systems, 2012.
[13] J. Hofbauer and K. Sigmund. Evolutionary games and
Population dynamics. Cambridge University Press, 1998.
[14] J. Hofbauer and K. Sigmund. Evolutionary game dy-
namics. Bulletin of the American Mathematical Society,
40(4):479, 2003.
[15] Y. Sato and J.P. Crutchfield. Coupled replicator equa-
tions for the dynamics of learning in multiagent systems.
Physical Review E, 67(1), 2003.
[16] C.J.C.H. Watkins and P. Dayan. Technical note: Q-
learning. Machine learning, 8(3):279 -- 292, 1992.
[17] R.S. Sutton and A.G. Barto. Reinforcement learning: An
[8] G. Demange. Group formation in economics: networks,
introduction. The MIT press, 2000.
clubs and coalitions. Cambridge Univ. Press, 2005.
[9] Sanjeev Goyal and Fernando Vega-Redondo. Network
formation and social coordination. Games and Economic
Behavior, 50(2):178 -- 207, 2005.
[18] A. Kianercy and A. Galstyan. Dynamics of boltzmann q
learning in two-player two-action games. Physical Review
E, 85(4):041145, 2012.
|
1502.02076 | 1 | 1502 | 2015-02-06T23:50:28 | Can Other People Make You Less Creative? | [
"cs.MA"
] | This paper explains in layperson's terms how an agent-based model was used to investigate the hypothesis that culture evolves more effectively when novelty-generating creative processes are tempered by imitation processes that preserve proven successful ideas. Using EVOC, an agent-based model of cultural evolution we found that (1) the optimal ratio of inventing to imitating ranged from 1:1 to 2:1 depending on the fitness function, (2) there was a trade-off between the proportion of creators to conformers and how creative the creators were, and (3) when agents in increased or decreased their creativity depending on the success of their latest creative efforts, they segregated into creators and conformers, and the mean fitness of ideas across the society was higher. It is tentatively suggested that through the unconscious use of social cues, members of a society self-organizes to achieve a balanced mix of creators and conformers. | cs.MA | cs |
Gabora, L. (2015). Can Other People Make You Less Creative?
Psychology Today (online). https://www.psychologytoday.com/blog/mindbloggling/201502/can-
other-people-make-you-less-creative
Can Other People Make You Less Creative?
Liane Gabora
Abstract
Okanagan Campus, Arts Building, 333 University Way, Kelowna BC, V1V 1V7, CANADA
Department of Psychology, University of British Columbia
This paper explains in layperson’s terms how an agent-based model was
used to investigate the hypothesis that culture evolves more effectively when
novelty-generating creative processes are tempered by imitation processes
that preserve proven successful ideas. Using EVOC, an agent-based model
of cultural evolution we found that (1) the optimal ratio of inventing to
imitating ranged from 1:1 to 2:1 depending on the fitness function, (2) there
was a trade-off between the proportion of creators to conformers and how
creative the creators were, and (3) when agents in increased or decreased
their creativity depending on the success of their latest creative efforts, they
segregated into creators and conformers, and the mean fitness of ideas
across the society was higher. It is tentatively suggested that through the
unconscious use of social cues, members of a society self-organizes to
achieve a balanced mix of creators and conformers.
Do you feel more creative in some social environments than others, or even when you
are completely alone? Although it is widely believed that stimulating environments
enhance creativity there is evidence that this is not always the case; indeed it turns out
that creativity can be enhanced by spending time in an isolation tank blocked off from
sensory stimulation entirely (Forgays & Forgays, 1992; Norlander, Bergman, & Archer,
1998; Vartanian, & Suedfeld, 2011). This suggests that social stimulation may not
always enhance creativity, and may even decrease it.
The idea that social environments could interfere with creativity came from thinking
about creativity in the context of culture as an evolutionary process. Evolutionary
processes require a balance between exploratory processes that generate new variants
and conservative processes that perpetuate ‘tried and true’ variants. This is the case in
biological evolution, and it seemed reasonable that it would also be the case for cultural
evolution. In other words: people may send out social signals to each other that ensure
that creativity---the process that fuels cultural novelty---is properly balanced by
conformity,i.e., imitation of ideas that have already proven to be successful.
The notion that society might aim to temper novelty-generating creativity with novelty-
perpetuating imitation flies in the face of the widespread assumption that more creativity
is necessarily better. It goes without saying that creativity is a good thing, and that
everyone should be creative. Or does it? Sure, our capacity for self-expression, for
2
finding practical solutions to problems of survival, and coming up with aesthetically
pleasing objects that delight the senses, all stem from the creative power of the human
mind. But there are drawbacks to creativity.
First, creative people tend to be more emotionally unstable and prone to affective
disorders such as depression and bipolar disorder. They have a higher incidence of
‘schizotypal’ leanings than other segments of the population. They are also more prone
to abuse drugs and alcohol, and to commit suicide. So there is a ‘dark side’ to creativity
(Cropley, Cropley, Kaufman, & Runco, 2010).
Second, a creative solution to one problem often generates other problems, or
unexpected negative side effects that may only become apparent after much has been
invested in the creative solution. There is a cultural version of what in biology is referred
to as epistasis, where what is optimal with respect to one part depends on what is done
with respect to another part. Once both parts of a problem have been solved in a
mutually beneficial way, too much creativity can cause these ‘co-adapted’ partial
solutions to break down.
Third, in a group of interacting individuals, only a fraction of them need be creative for
the benefits of creativity to be felt throughout the group. Uncreative people can reap the
benefits of the ideas of ‘creative types’ without having to withstand the ‘dark side’ of
creativity by simply imitating, or admiring them. Few of us know how to build a computer,
or write a symphony, or a novel, but they are nonetheless ours to use and enjoy when
we please. An excess of creative types all completely absorbed in their own creative
process might effectively insulate themselves and block the rapid diffusion of the best
ideas.
This opens up some interesting questions. Would it be good for the society as a whole if
everyone were highly creative? In order for a culture to evolve optimally, what is the
ideal ratio of creators to imitators, and how creative should the ’creative types’ be? And
perhaps most interesting of all: do people upgrade or downgrade how creative they are
in response to social cues they receive from other people about the perceived value of
their creative outputs?
My colleagues and I are investigating these questions using a computer model of
cultural evolution. I'll begin by telling you a bit about the computer model itself. Then I’ll
explain the experiments that led up to the experiments that explored the hypothesis that
people socially regulate each others’ creativity. Finally, I’ll explain the social regulation
experiments themselves.
THE COMPUTER MODEL
The current model's predecessor was called Meme and Variations or MAV (Gabora,
1995). Its name is a pun on the musical form, ‘theme and variations'. MAV was the
earliest computer program to model culture as an evolutionary process in its own right.
MAV was inspired by the genetic algorithm (GA), a search technique that finds solutions
to complex problems by generating a 'population' of candidate solutions (through
processes akin to mutation and recombination), selecting the best, and repeating until a
satisfactory solution is found.
The computer model is composed of an artificial society of agents in a two-dimensional
grid-cell world. Agents consist of (1) a neural network, which encodes ideas for actions
2
3
and detects trends in what constitutes an effective action, and (2) a body, which
implements their ideas as actions. The agents can do two things: (1) invent ideas for
new actions, and (2) imitate their neighbors' actions. The computer model enables us to
investigate what happens to the diversity and effectiveness of actions in the artificial
society over successive rounds (called ‘iterations') of invention and imitation. Since the
ideas in the model are ideas for actions, diversity is measured by counting how many
different actions are being implemented by the agents. Evolution in the biological sense
is not taking place; the agents neither die nor have offspring. But evolution in the cultural
sense is taking place through the generating and sharing of ideas for actions amongst
agents, which over time leads to more effective actions.
In MAV, all agents were equally capable of both inventing and imitating. In the latest
version of the computer model called EVOC (for EVOlution of Culture), it is possible to
vary how likely an agent is to invent versus imitate.
A TYPICAL RUN
Each iteration, every agent has the opportunity to (1) acquire an idea for a new action,
either by imitation, copying a neighbor, or by invention, creating one anew, (2) update
their knowledge about what constitutes an effective action, and (3) implement a new
action. Effectiveness of actions starts out low because initially all agents are just
standing still doing nothing. Soon some agent invents an action that has a higher
effectiveness than doing nothing, and this action gets imitated, so effectiveness
increases. Effectiveness increases further as other ideas get invented, assessed,
implemented as actions, and spread through imitation. The diversity of actions initially
increases due to the proliferation of new ideas, and then decreases as agents hone in
on the fittest actions. Thus MAV successfully models how 'descent with modification' can
occur in a cultural context.
FIRST SET OF EXPERIMENTS:
WHAT IS THE OPTIMAL RATIO OF EFFORT INVESTED IN CREATIVITY TO EFFORT
INVESTED IN IMITATION?
In the earliest version of this computer model (MAV), all agents were equally capable of
both inventing and imitating (Gabora, 1995). It was possible to vary the probability that,
in a given iteration, they would invent versus imitate. The agents are too rudimentary to
suffer from depression or commit suicide, so it is just the other detrimental aspects of
creativity listed above that we thought might play a role in these experiments.
What the results showed was that if they all imitated each other all the time, nothing
happened at all: everyone watched everyone else and no one did anything. If they all
invented all the time, the progress of ideas was slow because they weren’t taking
advantage of each other’s hard work. The optimal ratio of inventing to imitating was
about 2:1. Of course, these results hold for just this little idealized world, and they may or
may not be generalizable to the world at large. But they do ring true for many people,
who say they spend about 2/3 of their time alone in their studio or office immersed in
their work, and about 1/3 of their time talking or reading about or studying things related
to their creative project. (Subsequent experiments showed that this ratio can be as low
as 1:1 depending on the ‘fitness function’, i.e., the kind of task the agents had to solve.)
SECOND SET OF EXPERIMENTS:
3
4
WHAT PROPORTION OF SOCIETY SHOULD BE CREATIVE, AND HOW CREATIVE
SHOULD THEY BE?
The finding that very high levels of creativity can be detrimental for society led to the
hypothesis that there is an adaptive value to society's ambivalent attitude toward
creativity; society as a whole may benefit from a distinction between the conventional
workforce and what has been called a “creative class” (Florida, 2002). Using the new
version of the computer model, EVOC we investigated how patterns of cultural evolution
are affected by how numerous and creative the creators are. We conducted experiment
that varied, not just the ratio of creators to imitators, but also the creativeness of the
creators (Leijnen & Gabora, 2009). Each agent could be a pure imitator, a pure creator,
or something in between. The pure imitators never invented; they simply copied the
successful innovations of the creative agents. The creators were able to invent as well
as well as imitate. The percentage of iterations in which they invented varied up to 100%.
The results were provocative. We found that cultural diversity – that is, the number of
different actions in the artificial society -- was positively correlated with both the
percentage of creators, and their level of creativity. However, for cultural fitness or
effectiveness, the situation was more complex. So long as the creative types weren’t
THAT creative, the more of them there are, the better. But when the creative types were
HIGHLY creative, the mean fitness of ideas across the society was higher if there were
fewer of them. The results seemed to show that the more creative the creators are, the
less numerous they should be.
We then conducted more extensive investigations of these questions employing more
detailed and sophisticated analytical methods. The amount of time it takes for the
effectiveness of ideas across the artificial society to reach a threshold level of
performance is affected by the creator to imitator ratio (C), and the creator innovation
probability (p).
The same general trends emerged. Cultural diversity was once again positively
correlated with both the percentage of creators, and their level of creativity (Gabora &
Leijnen, 2013). But when we looked at, not the variety of ideas, but how fit or effective
they were, the pattern of results was more complicated. What we found was that if C is
low the p should be high, but if C is high then p should be intermediate (Gabora &
Firouzi, 2012). In other words, once again there was a tradeoff between how creative the
creators were, and how many of them there should be.
These results supported the hypothesis that too much creativity causes ‘co-adapted’
partial solutions to problems to break down (the ‘don’t fix it if it ain’t broke’ phenomenon).
They also supported the hypothesis that creative types, while they are a necessary
source of novelty, constitute pinholes in the fabric of culture that block the spread of
ideas. An iteration spent inventing is an iteration not spent imitating, and imitation is
extremely valuable. It’s not just a form of free-riding, nor just ‘the greatest compliment’,
but an indispensable social mechanism that serves everyone. By simply copying the
successful innovations of the creative types, imitators serve as a ’memory’ for preserving
the fittest configurations. So, contrary to popular belief, it might not be best for the
society as a whole if everyone were creative.
THIRD SET OF EXPERIMENTS:
HOW IS SOCIETY AFFECTED IF PEOPLE CAN UPGRADE/DOWNGRADE THEIR
4
5
CREATIVITY IN RESPONSE TO SOCIAL SIGNALS?
We then hypothesized that society as a whole might perform even better with the ability
to adjust creativity in accordance with their perceived creative success, through
mechanisms such as selective ostracization of deviant behaviour unless accompanied
by the generation of valuable cultural novelty, and encouraging of successful creators. A
first step in investigating this hypothesis was to determine whether it is algorithmically
possible to increase the mean fitness of ideas in a society by enabling them to self-
regulate how creative they are. To test the hypothesis that the mean fitness of cultural
outputs across society increases faster with social regulation (SR) than without it, we
increased the relative frequency of invention for agents that generated superior ideas,
and decreased it for agents that generated inferior ideas (Gabora & Tseng, 2014a).
Each iteration, for each agent, the fitness of its current action relative to the mean fitness
of actions for all agents at the previous iteration was assessed. If its action was fitter
than the mean it created more, and if its action was less fit than the mean it imitated
more.
The typical pattern was observed with respect to the diversity, or number of different
ideas: an increase as the space of possibilities is explored followed by a decrease as
agents converge on fit actions. However, this pattern occurred earlier, and was more
pronounced, in societies with SR than societies without it. Interestingly, the mean fitness
of the cultural outputs in societies with SR was higher than that in societies without SR.
Even more interestingly, the societies with SR ended up separating into two distinct
groups: one that primarily invented, and one that primarily imitated. Thus the observed
increase in fitness could indeed be attributed to increasingly pronounced individual
differences in their degree of creative expression over the course of a run. Agents that
generated superior cultural outputs had more opportunity to do so, while agents that
generated inferior cultural outputs became more likely to propagate proven effective
ideas rather than reinvent the wheel.
We conducted another set of experiments in which agents were able to generate not just
simple single-step actions but complex multi-step actions that were more like real-world
actions such as dancing or tool-making (Gabora & Tseng, 2014b). The same general
patterns emerged: the agents segregated into creators and conformers, and the mean
fitness of actions across the society as a whole increased.
These results don't show that this kind of segregation into creators and
conformers actually takes place in real societies. What they suggest is that if this
did happen it could have an adaptive benefit for society as a whole.
CONCLUSIONS
The computer model differs substantially from the real world, so any attempt to
extrapolate these results should be taken with a grain of salt. The agents’ neural
networks are so small that creative novelty is generated does not involve noticing and
refining new kinds of connections the way it happens in real minds (Gabora, 2000).
Moreover, in these simulations, unlike the real world, agents had only one task to
accomplish. However, the results of these computer simulations are provocative, and
inspire new ways of thinking about creativity. They show that societies may benefit as a
5
6
whole by self-organizing into a balanced mix of novelty generating creators and
continuity perpetuating imitators, and lead to the speculation that this happens
spontaneously in real societies.
There are many speculative but fascinating implications of this research. It leads to the
suggestion that the reason creative individuals often isolate themselves is not just to
decrease disturbances so they can more fully concentrate on their art, but because
isolation safeguards you from social signals to downgrade your creativity, which can
negatively impact creative performance. It suggests that as societies become
increasingly denser it may become increasingly difficult to isolate oneself from the cues
by which people unconsciously socially regulate each other’s level of creativity.
Another possibility, suggested by my colleague Kiley Hamlin, is that girls are more
responsive to social cues to downgrade creativity than boys, which might prematurely
streamline them into an “imitator” track. The more you imitate, the more you rely on
imitation as a source of ideas and ways of doing things, so it becomes a vicious circle of
sorts, and vice versa: the more you create, the more you keep the creative juices flowing
and rely on your own creative processes as a source of not just new ideas, but pleasure.
ACKNOWLEDGMENTS
The author is grateful for funding from the Natural Sciences and Engineering Research
Council of Canada.
REFERENCES
Cropley, A., Cropley, D., Kaufman, J. & Runco, M. Eds. (2010). The Dark Side of
Creativity (pp. 277-296). Cambridge UK: Cambridge University Press.
Florida, R. (2002). The rise of the creative class. London: Basic Books.
Gabora, L. (2000). Toward a theory of creative inklings. In (R. Ascott, Ed.) Art,
Technology, and Consciousness (pp. 159-164). Intellect Press, Bristol, UK.
Gabora, L. (1995). Meme and variations: A computer model of cultural evolution. In (L.
Nadel & D. Stein, Eds.) 1993 Lectures in Complex Systems, Addison-Wesley, 471-486.
Gabora, L., & Firouzi, H. (2012). Society functions best with an intermediate level of
creativity.Proceedings of the 34th Annual Meeting of the Cognitive Science Society (pp.
1578-1583). Austin TX: Cognitive Science Society. [http://arxiv.org/abs/1310.4753]
Gabora, L., & Leijnen, S. (2013). The relationship between creativity, imitation, and
cultural diversity. International Journal of Softward and Informatics, 7(4), 615-627.
Gabora, L. & Tseng, S. (2014a). Computational evidence that self-regulation of creativity
is good for society. Proceedings of the 36th Annual Meeting of the Cognitive Science
Society (pp. 2240-2245). Austin TX: Cognitive Science Society.
[http://arxiv.org/abs/1408.2512]
Gabora, L. & Tseng, S. (2014b). The social impact of self-regulation on the evolution of
6
7
simple and complex creative ideas. Proceedings of the 5th International Conference on
Computational Creativity. Palo Alto: Association for the Advancement of Artificial
Intelligence.
Leijnen, S. & Gabora, L. (2009). How creative should creators be to optimize the
evolution of ideas? A computational model. Electronic Proceedings in Theoretical
Computer Science, 9, 108-119.
Forgays, D. G., & Forgays, D. K. (1992). Creativity enhancement through flotation
isolation. Journal of Environmental Psychology, 12, 329-335.
Norlander, T., Bergman, H., & Archer, T. (1998). Effects of flotation rest on creative
problem solving and originality. Journal of Environmental Psychology, 18, 399-408.
Vartanian, O., & Suedfeld, P. (2011). The effect of the flotation version of Restricted
Environmental Stimulation Technique (REST) on jazz improvisation. Music and Medicine,
3, 234-238.
7
|
1206.3793 | 1 | 1206 | 2012-06-17T21:19:45 | A distributed classification/estimation algorithm for sensor networks | [
"cs.MA",
"eess.SY",
"math.OC"
] | In this paper, we address the problem of simultaneous classification and estimation of hidden parameters in a sensor network with communications constraints. In particular, we consider a network of noisy sensors which measure a common scalar unknown parameter. We assume that a fraction of the nodes represent faulty sensors, whose measurements are poorly reliable. The goal for each node is to simultaneously identify its class (faulty or non-faulty) and estimate the common parameter.
We propose a novel cooperative iterative algorithm which copes with the communication constraints imposed by the network and shows remarkable performance. Our main result is a rigorous proof of the convergence of the algorithm and a characterization of the limit behavior. We also show that, in the limit when the number of sensors goes to infinity, the common unknown parameter is estimated with arbitrary small error, while the classification error converges to that of the optimal centralized maximum likelihood estimator. We also show numerical results that validate the theoretical analysis and support their possible generalization. We compare our strategy with the Expectation-Maximization algorithm and we discuss trade-offs in terms of robustness, speed of convergence and implementation simplicity. | cs.MA | cs |
A DISTRIBUTED CLASSIFICATION/ESTIMATION ALGORITHM
FOR SENSOR NETWORKS∗
FABIO FAGNANI†, SOPHIE M. FOSSON‡, AND CHIARA RAVAZZI§
Abstract. In this paper, we address the problem of simultaneous classification and estimation of
hidden parameters in a sensor network with communications constraints. In particular, we consider
a network of noisy sensors which measure a common scalar unknown parameter. We assume that
a fraction of the nodes represent faulty sensors, whose measurements are poorly reliable. The goal
for each node is to simultaneously identify its class (faulty or non-faulty) and estimate the common
parameter.
We propose a novel cooperative iterative algorithm which copes with the communication con-
straints imposed by the network and shows remarkable performance. Our main result is a rigorous
proof of the convergence of the algorithm and a characterization of the limit behavior. We also
show that, in the limit when the number of sensors goes to infinity, the common unknown param-
eter is estimated with arbitrary small error, while the classification error converges to that of the
optimal centralized maximum likelihood estimator. We also show numerical results that validate
the theoretical analysis and support their possible generalization. We compare our strategy with
the Expectation-Maximization algorithm and we discuss trade-offs in terms of robustness, speed of
convergence and implementation simplicity.
Key words. Classification, Consensus, Gaussian mixture models, Maximum-likelihood estima-
tion, Sensor networks, Switching systems.
1. Introduction. Sensor networks are one of the most important technologies
introduced in our century. Promoted by the advances in wireless communications
and by the pervasive diffusion of smart sensors, wireless sensor networks are largely
used nowadays for a variety of purposes, e.g., environmental and habitat surveillance,
health and security monitoring, localization, targeting, event detection.
A sensor network basically consists in the deployment of a large numbers of small
devices, called sensors, that have the ability to perform measurements and simple
computations, to store few amounts of data, and to communicate with other devices.
In this paper, we focus on ad hoc networks, in which communication is local: each
sensor is connected only with a restricted number of other sensors. This kind of
cooperation allows to perform elaborate operations in a self-organized way, with no
centralized supervision or data fusion center, with a substantial energy and economic
saving on processors and communication links. This allows to construct large sensor
networks at contained cost.
A problem that can be addressed through ad hoc sensor networking is the dis-
tributed estimation: given an unknown physical parameter (e.g., the temperature in
a room, the position of an object), one aims at estimating it using the sensing ca-
pabilities of a network. Each sensor performs a (not exact) measurement and shares
it with the sensors with which it can establish a communication; in turn, it receives
information and consequently updates its own estimate. If the network is connected,
∗A preliminary version of some of the results has appeared in the proceedings of the 50st IEEE
Conference on Decision and Control and European Control Conference, Orlando, Florida, 12-15
December 2011.
†DISMA (Dipartimento di Scienze Matematiche), Politecnico di Torino, Corso Duca degli Abruzzi,
24, I-10129 TO, (e-mail: [email protected])
‡DET (Dipartimento di Elettronica e Telecomunicazioni), Politecnico di Torino, Corso Duca degli
Abruzzi, 24, I-10129 TO (e-mail: [email protected])
§DET (Dipartimento di Elettronica e Telecomunicazioni), Politecnico di Torino, Corso Duca degli
Abruzzi, 24, I-10129 TO (e-mail: [email protected])
1
by iterating the sharing procedure, the information propagates and a consensus can
be reached. Neither centralized coordinator nor data fusion center is present. The
mathematical model of this problem must envisage the presence of noise in measure-
ments, which are naturally corrupted by inaccuracies, and possible constraints on the
network in terms of communication, energy or bandwidth limitations, and of necessity
of quantization or data compression.
Distributed estimation in ad hoc sensor networks has been widely studied in the
literature. For the problem of estimating an unknown common parameter, typical
approach is to consider distributed versions of classical maximum likelihood (ML)
or maximum-a-posteriori (MAP) estimators. Decentralization can be obtained, for
instance, through consensus type protocols (see [1], [2], [3]) adapted to the communi-
cation graph of the network, or by belief propagation methods [4] and [5].
A second important issue is sensors' classification, which we define as follows [6].
Let us imagine that sensors can be divided into different classes according to peculiar
properties, e.g., measurements' or processing capabilities, and that no sensor knows
to which class it belongs: by classification, we then intend the labeling procedure
that each sensor undertakes to determine its affiliation. This task is addressed to a
variety of clustering purposes, for example, to rebalance the computation load in a
network where sensors can be distinguished according to their processing power. On
most occasions, sensors' classification is faced through some distributed estimation,
the underlying idea being the following: each sensor performs its measurement of a
parameter, then iteratively modifies it on the basis of information it receives; during
this iterative procedure the sensor learns something about itself which makes it able
to estimate its own configuration.
i ηi, where θ⋆ ∈ R is the unknown global parameter, ω⋆
In this paper, we consider the following model: each sensor i performs a measure-
ment yi = θ⋆ + ω⋆
i > 0 is the
unknown status of the sensor, and ηi is a Gaussian random noise. The more ω⋆
is
i
large, the more the sensor i is malfunctioning, that is, the quality if its measurement
is low. The ω⋆
i parameter is supposed to belong to a discrete set, in particular in this
paper we consider the binary case.
The goal of each unit i is to estimate the parameter θ⋆ and the specific configu-
i . The presence of the common unknown parameter θ⋆ imposes a coupling
ration ω⋆
between the different nodes and makes the problem interesting.
An additive version of the aforementioned model has been studied in [7], where
measurement is given by yi = θ⋆ + ω⋆
i + ηi. Another related problem is the so-called
calibration problem [8,9]: sensor i performs a noisy linear measurement yi = Aiθ + ηi
where the unknown θ and Ai are a vector and a matrix, respectively, while ηi is a
noise; the goal consists in the estimation of θ and of Ai, the latter being known as
calibration problem.
All these are particular cases of the problem of the estimation of Gaussian mix-
tures' parameters [10, 11]. This perspective has been studied for sensor networks
in [12], [13], [14], and [15] where distributed versions of the Expectation-Maximization
(EM) algorithm have been proposed. A network is given where each node inde-
pendently performs the E-step through local observations.
In particular, in [14] a
consensus filter is used to propagate the local information. The tricky point of such
techniques is the choice of the number of averaging iterations between two consecutive
M-steps, which must be sufficient to reach consensus.
The aim of this paper is the development of a distributed, iterative procedure
which copes with the communication constraints imposed by the network and com-
2
putes an estimation (bθ,bω) approximating the maximum likelihood optimal solution
of the proposed problem. The core of our methodology is an Input Driven Consensus
Algorithm (IA for short), introduced in [16], which takes care of the estimation of
the parameter θ∗. IA is coupled with a classification step where nodes update the
estimation of their own type ω∗i by a simple threshold estimator based on the current
estimation of θ∗. The fact of using a consensus protocol working on inputs instead,
as more common, on initial conditions, is a key strategic fact: it serves the purpose of
using the innovation coming from the units who are modifying the estimation of their
status, as time passes by. Our main theoretical contribution is a complete analysis of
the algorithm in terms of convergence and of behaviour with respect to the size of the
network. With respect to other approaches like distributed EM for which convergence
results are missing, this makes an important difference. We also present a number of
numerical simulations showing the remarkable performance of the algorithm which,
in many situations, outperform classical choices like EM.
The outline of the paper is the following. In Section 2 we shortly present some
graph nomenclature needed in the paper. Section 3 is devoted to a formal description
of the problem and to a discussion of the classical centralized maximum likelihood
solution. In Section 4, we present the details and the analysis of our IA. Our main
results are Theorems 4.1 and 4.2: Theorem 4.1 ensures that, under suitable assump-
tions on the graph, the algorithm converges to a local maximum of the log-likelihood
function; Theorem 4.2 is a concentration result establishing that when the number
of nodes N → +∞, the estimate bθ converges to the true value θ∗ (a sort of asymp-
totic consistency). Finally, we also study the behavior of the discrete estimate bω
by analyzing the performance index the relative classification error over the network
when N → +∞ (see Corollary 4.4). Section 5 contains a set of numerical simulations
carried on different graph architectures: complete, circulant, grids, and random geo-
metric graphs. Comparisons are proposed with respect to the optimal centralized ML
solution and also with respect to the EM solution. Finally, a long Appendix contains
all the proofs.
2. General notation and graph theoretical preliminaries. Throughout
this paper, we use the following notational convention. We denote vectors with small
letters, and matrices with capital letters. Given a matrix M , M T denotes its trans-
pose. Given a vector v, v denotes its Euclidean norm. 1A is the indicator function
of set A. Given a finite set V, RV denotes the space of real vectors with components
labelled by elements of V. Given two vectors x, z ∈ RV , dH(x, z) = {i ∈ V : xi 6= zi}.
We use the convention that a summation over an empty set of indices is equal to zero,
while a product over an empty set gives one.
A symmetric graph is a pair G = (V,E) where V is a set, called the set of vertices,
and E ⊆ V × V is the set of edges with the property that (i, i) 6∈ E for all i ∈ V and
(i, j) ∈ E implies (j, i) ∈ E. G is strongly connected if, for all i, j ∈ V, there exist
vertices i1, . . . is such that (i, i1), (i1, i2), . . . , (is, j) ∈ E. To any symmetric matrix
P ∈ RV×V with non-negative elements, we can associate a graph GP = (V,EP ) by
putting (i, j) ∈ EP if and only if Pij > 0. P is said to be adapted to a graph
G if GP ⊆ G. A matrix with non-negative elements P is said to be stochastic if
Pij = 1 for every i ∈ V. Equivalently, denoting by 1 the vector of all 1 in RV ,
P is stochastic if P 1 = 1. P is said to be primitive if there exists n0 ∈ N such that
P n0
ij > 0 for every i, j ∈ V. A sufficient condition ensuring primitivity is that GP is
strongly connected and Pii > 0 for some i ∈ V.
Pj∈V
3. Bayesian modeling for estimation and classification.
3
3.1. The model. In our model, we consider a network, represented by a sym-
metric graph G = (V,E). G represents the system communication architecture. We
denote the number of nodes by N = V. We assume that each node i ∈ V measures
the observable
yi = θ⋆ + ω⋆
i ηi
(3.1)
where θ⋆ ∈ R is an unknown parameter, ηi's Gaussian noises N(0, 1), ω⋆
random variables taking values in {α, β} (with P(ω⋆
random variables ηi's and ω⋆
is a Gaussian mixture distributed according to the probability density function
i 's Bernoulli
i = β) = p). We assume all the
i 's to be mutually independent. Notice that each yi ∈ R
f (yi) = (1 − p)f (yiθ⋆, α) + pf (yiθ⋆, β)
f (yiθ⋆, x) =
x ∈ {α, β}.
e− (yi −θ⋆ )2
x√2π
2x2
1
(3.2)
(3.3)
The binary model of ω⋆ is motivated by different scenarios: as an example, if 0 < α <<
β, the nodes of type β may represent a subset of faulty sensors, whose measurements
are poorly reliable; the aim may be the detection of faulty sensors in order to switch
them off or neglect their measurements, or for other clustering purposes. It is also
realistic to assume that some a-priori information about the quantity of faulty sensors
is extracted, e.g., from experimental data on the network, and it is conceivable to
represent such information as an a-priori distribution. This is why we assume a
Bernoulli distribution on each ω⋆
i ; on the other hand, we suppose that no a-priori
information is available on the unknown parameter θ⋆. However, the addition of an a
priori probability distribution on θ∗ does not significantly alter our analysis and our
results.
3.2. The maximum likelihood solution. The goal is to estimate the param-
eter θ⋆ and the specific configuration ω⋆
i of each unit. Disregarding the network
constraints, a natural solution to our problem would be to consider a joint ML in θ⋆
and MAP in the ω⋆
i 's (see [17, 18]). Let f (y, ωθ) be the joint distribution of y and ω
(density in y and probability in ω) given the parameter θ, and consider the rescaled
log-likelihood function
LN (θ, ω) :=
1
N
log f (y, ωθ).
(3.4)
The hybrid ML/MAP solution, which for simplicity for now on we will refer to as the
ML solution, prescribes to choose θ and ω which maximize LN (θ, ω)
θ∈R, ω∈{α,β}V
(bθML,bωML) := argmax
2β2 + 1{ωj=α}(cid:18) (yj − θ)2
2
1
N Xj∈V(cid:18) (yj − θ)2
Standard calculations lead us to
LN (θ, ω) = −
LN (θ, ω).
(3.5)
(cid:18) 1
α2 −
1
β2(cid:19) + log
1 − p
p
β
α(cid:19)(cid:19)+c
(3.6)
where c is a constant. It can be noted that partial maximizations of LN (θ, ω) with
respect to just one of the two variables have simple representation. Let
bθ(ω) := argmax
θ
LN (θ, ω)
bω(θ) := argmax
ω
4
LN (θ, ω).
(3.7)
β otherwise
bω(θ)i =(α if yi − θ < δ
ln(cid:16) 1−p
α(cid:17)
α2 − 1
β2
β
p
.
j
j
bθ(ω) = Pj yj/ω2
Pj 1/ω2
δ =vuut2
bθML = argmax
L(θ,bω(θ)) ,
1
θ
bωML =bω(bθML).
(3.8)
(3.9)
The ML solution can then be obtained, for instance, by considering
Then
where
this can be done in a decentralized way. Moreover, further difficulties are caused by
It should be noted how the computation of the (bωML)i's becomes totally decentralized
oncebθML has been computed. For the computation ofbθML instead one needs to gather
information from all units to compute LN (θ,bω(θ)) and it is not at all evident how
the fact that LN (θ,bω(θ)) may contain many local maxima, as shown in Figure 3.1.
It should be noted that LN (θ,bω(θ)) is differentiable except at a finite number of
yi! .
LN (θ,bω(θ)) =(cid:18) 1
points, and between two successive non-differentiable points the function is concave.
Therefore, the local maxima of the function coincide with its critical points. On the
other hand, the derivative, where it exists, is given by
(θ − yi)1{yi−θ<δ} −
β2 θ −
α2(cid:19) 1
N Xi∈V
N Xi∈V
β2 −
(3.10)
d
dθ
1
1
1
Stationary points can therefore be represented by the relation
θ =
1
β2Pi yi +(cid:16) 1
β2 +Pi
N 1
β2(cid:17)Pi yi1{yi−θ<δ}
α2 − 1
1{yi−θ<δ}(cid:16) 1
β2(cid:17)
α2 − 1
.
(3.11)
A moment of thought shows us that (3.11) is equivalent to the relation θ =bθ(bω(θ)).
This representation will play a key role in the sequel of this paper.
3.3. Iterative centralized algorithms. The computational complexity of the
optimization problem (3.5) is practically unfeasible in most situations. However,
relations (3.8) suggest a simple way to construct an iterative approximation of the
ML solution (which we will denote IML). The formal pattern is the following: fixed
bω(0) = α1, for t = 0, 1, . . . , we consider the dynamical system
j i−2
j=1 yjhbω(t)
bθ(t+1) = PN
j i−2
j=1hbω(t)
PN
=(α if yi − θ < δ
β otherwise
5
i
bω(θ)(t+1)
for any i = 1, . . . , N.
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
)
)
θ
(
bω
,
θ
(
N
L
N = 50
N = 100
N = 400
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
)
)
θ
(
bω
,
θ
(
N
L
N = 500
N = 1000
N = 5000
−0.8
−10
−8
−6
−4
−2
0
θ
2
4
6
8
10
−0.8
−10
−8
−6
−4
−2
2
4
6
8
10
0
θ
Figure 3.1. α = 0.3, β = 10, p = 0.25: Plot of function LN (θ, bω(θ)) as a function of θ and
size N ∈ {50, 100, 400, 500, 1000, 5000}.
A more refined iterative solution is given by the so-called Expectation-Maximization
The algorithm stops whenever bθ(t+1) −bθ(t) < ε, for some fixed tolerance ε > 0.
(EM) algorithm [19]. The main idea is to introduce a hidden (say, unknown and un-
observed) random variable in the likelihood; then, at each step, one computes the
mean of the likelihood function with respect to the hidden variable and finds its max-
imum. Such a method seeks to find the maximum likelihood solution, which in many
cases cannot be formulated in a closed form. EM is widely and successfully used in
many frameworks and in principle it could also be applied to our problem. In our
context, making the variable ω to play the part of the hidden variable, equations for
EM become (see the tutorial [20] for their derivation)
q(t)
1. E-step: for all node i ∈ V,
Given bθ(0) ∈ R, for t = 0, 1, . . . ,
i = αy,bθ(t)(cid:17) =
i = P(cid:16)bω(t)
(1 − p)f(cid:16)ybω(t)
yj(cid:16)q(t)
bθ(t+1) = Pj∈V
j α−2 + (1 − q(t)
Pj∈V
j α−2 + (1 − q(t)
q(t)
(1 − p)f(cid:16)ybω(t)
i = α,bθ(t)(cid:17)
i = α,bθ(t)(cid:17) + pf(cid:16)ybω(t)
j )β−2(cid:17)
2. M-step:
j )β−2
i = β,bθ(t)(cid:17) .
.
{bω(t)
i =α}
computed in the E-step actually is the expectation of the
is worth to notice that q(t)
i
binary random variable 1
is the maximum of such expectation.
The algorithm stops whenever bθ(t+1) −bθ(t) < ε, for some fixed tolerance ε > 0. It
. On the other hand bθ(t+1) computed in the M-step
sequence {bθ(t)}∈N to a local maximum of the expected value of the log-likelihood with
respect to the unknown data ω, a result which is instead not directly available for IML.
Both algorithms however share the drawback of requiring centralization. Distributed
versions of the EM have been proposed (see, e.g., [12], [14]) but convergence is not
In Section 5 we will compare both these algorithms against
guaranteed for them.
An important feature of EM is that it is possible to prove the convergence of the
6
the distributed IA we are going to present in the next section. While it is true that
EM always outperforms IML, algorithm IA outperforms both of them for small size
algorithms, while shows comparable performance to EM for large networks.
4. Input driven consensus algorithm.
4.1. Description of the algorithm. In this section we propose a distributed
iterative algorithm approximating the centralized ML estimator. The algorithm is
suggested by the expressions in (3.8) and consists of the iteration of two steps: an
averaging step where all units aim at computing bθ through a sort of Input Driven
Consensus Algorithm (IA) followed by an update of the classification estimation per-
formed autonomously by all units.
Formally, IA is parametrized by a symmetric stochastic matrix P , adapted to the
communication graph G (Pij > 0 if and only if, (i, j) ∈ E), and by a real sequence
γ(t) → 0. Every node i has three messages stored in its memory at time t, denoted
with µ(t)
i = 0 and the initial
i
. Given the initial conditions µ(0)
i = 0, ν(0)
, ν(t)
i
i = α, the dynamics consists of the following steps.
i
, and bω(t)
1. Average step:
estimate bω(0)
(4.1a)
(4.1b)
(4.1c)
< δ
(4.2)
µ(t+1)
i
ν(t+1)
i
= (1 − γ(t))Xj
= (1 − γ(t))Xj
Pijµ(t)
Pijν(t)
i (cid:17)−2
j + γ(t)yi(cid:16)bω(t)
i (cid:17)−2
j + γ(t)(cid:16)bω(t)
i
i
i
.
/ν(t+1)
= µ(t+1)
bθ(t+1)
=bωi(bθ(t+1)) =(cid:26) α if yi −bθ(t+1)
bω(t+1)
i
β otherwise.
i
2. Classification step:
It should be noted that the algorithm provides a distributed protocol: each node only
needs to be aware of its neighbours and no further information about the network
topology is required.
4.2. Convergence. The following theorem ensures the convergence of IA. The
proof is rather technical and therefore deferred to Appendix A.
Theorem 4.1. Let
(a) γ(t) → 0, γ(t) ≥ 1/t, and γ(t) = γ(t+1) + o(γ(t+1)) for t → +∞;
(b) P ∈ RV×V+
eigenvalues.
be a stochastic, symmetric, and primitive matrix with positive
1.
Then, there exist bωIA ∈ {α, β}V and bθIA ∈ R such that
t→+∞bθ(t)
t→+∞bω(t) a.s.= bωIA ,
lim
lim
i
for all i ∈ V;
2. they satisfy the relations
a.s.= bθIA
A number of remarks are in order.
bθIA =bθ(bωIA) , bωIA =bω(bθIA).
7
• The assumption on the eigenvalues of P is essentially a technical one:
in
simulations it does not seem to have a crucial role, but we need it in our proof
of convergence. On the other hand, given any symmetric stochastic primitive
P , we cam consider a 'lazy' version of it Pτ = (1 − τ )I + τ P and notice that
for τ ∈ (0, 1) sufficiently small, indeed Pτ will satisfy the assumption on the
eigenvalues.
• The requirement γ(t) ≥ 1/t is not new in decentralized algorithms (see for
instance the Robbins-Monro algorithm, introduced in [21]) and serves the
need of maintaining 'active' the system input for sufficiently long time. Less
classical is the assumption γ(t) ∼ γ(t+1) which is essentially a request of
regularity in the decay of γ(t) to 0. Possible choices of γ(t) satisfying the
above conditions are γ(t) = t−ζ for ζ ∈ (0, 1), or γ(t) = t−1(ln t)α for any
α > 0.
• The proof (see Appendix A) will also give an estimation on the speed of
• Relations in item 2.
4.3. Limit behavior. In this section we present results on the behavior of our
algorithm for N → +∞. All quantities derived so far are indeed function of network
size N . In order to emphasize the role of N , we will add an index N when dealing
N ). Instead we will not add anything to expressions
convergence: indeed it will be shown that bθ(t) −bθIA = O(γ(t)) for t → ∞.
implies that bθIA is a local maximum of the function
LN (θ,bω(θ)) (see (3.11)).
with quantities like θ⋆ (e.g. bθML
Figure 3.1 shows a sort of concentration of the local maxima of LN (θ,bω(θ)) to
a global maximum for large N . Considering that IA converges to a local maximum,
this observation would lead to the conclusion that, for large N , the IA resembles the
optimal ML solution. This section provides some results which make rigorous these
considerations.
where there are vectors ω involved since their dimension is itself N .
Notice first that, applying the uniform law of large numbers [22] to the expression
(3.6), we obtain that, for any compact K ⊆ R, almost surely
lim
N→+∞
where
max
θ∈K(cid:12)(cid:12)(cid:12)(cid:12)LN (θ,bω(θ)) −ZR J (s, θ)f (s)ds(cid:12)(cid:12)(cid:12)(cid:12) = 0
β2(cid:19) + log
(cid:18) 1
α2 −
2
(4.3)
1
β
1 − p
p
2β2 + 1{ωj =α}(cid:18) (s − θ)2
out to be differentiable for every value of θ and to have a unique stationary point for
θ = θ∗ which turns out to be the global minimum. Unfortunately, this fact by itself
does not guarantee that global and local minima will indeed converge to θ∗. In our
J (s, θ) = −(cid:18) (s − θ)2
α(cid:19)(cid:19) + c (4.4)
where c is the same constant as in (3.6). The limit function RR J (s, θ)f (s)ds turns
derivations the properties of the function RR J (s, θ)f (s)ds will not play any direct
Theorem 4.2. Denote by SN the set of local maxima of L(θ,bω(θ)). Then,
role and therefore they will not be proven here. The main technical result which will
be proven in Appendix B is the following:
(4.5)
lim
max
ξ∈SN ξ − θ∗ = 0
N→+∞
almost surely and in mean square sense.
8
This has an immediate consequence,
Corollary 4.3.
almost surely and in mean square sense.
N = lim
N = θ⋆
lim
N→+∞bθIA
N→+∞bθML
Regarding the classification error, we have instead the following result:
Proposition 4.4.
lim
N→+∞
1
N
where
EdH (bωIA, ω⋆) = lim
N→+∞
1
N
= q(p, α, β)
EdH (bωML, ω⋆)
β√2(cid:19)(cid:21)
α√2(cid:19) + p(cid:20)1 − erfc(cid:18) δ
q(p, α, β) = (1 − p)erfc(cid:18) δ
(4.6)
(4.7)
and erfc(x) := 2√πR +∞
x
e−t2
dt is the complementary error function.
These results ensure that the IA performs, in the limit of large number of units
N , as the centralized optimal ML estimator. Moreover, they also show, consistency in
the estimation of the parameter θ⋆. As expected, for N → +∞ the classification error
does not go to 0 since the increase of measurements is exactly matched by the same
increase of variables to be estimated. Consistency however is obtained when p goes to
zero since we have that limp→0 q(p, α, β) = 0. Moreover, notice that the dependence
of function q on the parameters α and β is exclusively through their ratio β/α. In
particular, we have
lim
β/α→+∞
q(p, α, β) = 0
lim
β/α→1
q(p, α, β) = 1.
5. Simulations. In this section, we propose some numerical simulations. We
test our algorithm for different graph architectures and dimensions, and we compare
it with the IML and EM algorithms. Our goal is to give evidence of the theoretical
results' validity and also to evaluate cases that are not included in our analysis: the
good numerical outcomes we obtain suggest that convergence should hold in broader
frameworks. The numerical setting for our simulations is now presented.
Model: the sensors perform measurements according to the model (3.1) with
θ⋆ = 0, α = 0.3, β = 10; the prior probability P(ω⋆
i = β) is equal to p = 0.25.
Communication architectures: given a strongly connected symmetric graph
G = (V,E), we use the so-called Metropolis random walk construction for P (see [23])
which amounts to the following: if i 6= j
Pij =(cid:26) 0
(max{deg(i) + 1, deg(j) + 1})−1
if (i, j) 6∈ E
if (i, j) ∈ E
where deg(i) denotes the degree (the number of neighbors) of unit i in the graph G.
P constructed in this way is automatically irreducible and aperiodic.
We consider the following topologies:
9
1. Complete graph: Pij = 1
the centralized case.
N for every i, j = 1, . . . , N ; it actually corresponds to
2. Ring: N agents are disposed on a circle, and each agent communicates with
its first neighbor on each side (left and right). The corresponding circulant
symmetric matrix P is given by Pij = 1
3 for every i = 2, . . . , N − 1 and
3 ; PN 1 = PN N−1 = PN N = 1
j ∈ {i − 1, i, i + 1}; P11 = P12 = P1N = 1
3 ;
Pij = 0 elesewhere.
3. Torus-grid graph: sensors are deployed on a two dimensional grid and are
each connected with their four neighbors; the last node of each row of the
grid is connected with the first node of the same row, and analogously on
columns, so that a torus is obtained. The so-obtained graph is regular.
4. Random Geometric Graph with radius r = 0.3: sensors are deployed in the
square [0, 1]× [0, 1], their positions being randomly generated with a uniform
distribution; links are switched on between two sensors whenever the distance
is less than r. We only envisage connected realizations.
From Theorem 4.1, P is required to possess positive eigenvalues: our intuition
is that these hypotheses, that are useful to prove the convergence of the IA, are not
really necessary. We test this conjecture on the ring graph, whose eigenvalues are
known [24] to be λm = 1
necessarily positive.
N (cid:1)(cid:1) , m = 0, . . . , N − 1 and which are not
3(cid:0)1 + 2 cos(cid:0) 2πm
Algorithms: We implement and compare the following algorithms: IA with
γ(t) = 1/tζ for different choices of ζ ∈ {0.5, 0.7, 0.9}, and the two centralized iterative
algorithms IML, and EM described in Section 3.3.
Outcomes: we show the performance of the aforementioned algorithms in terms
of classification error and of mean square error on the global parameter, in function
of the number of sensors N . All the outcomes are obtained averaging over 400 Monte
Carlo runs.
We observe that the classification error (Figure 5.1) converges for N → ∞ for
all the considered algorithms. On the other hand, when N is small, IA performs
better than IML and EM, no matter which graph topology has been chosen: this
suggests that decentralization is then not a drawback for IA. Moreover, for smaller
γ(t) (i.e., slowing down the procedure), we obtain better IA performance in terms
of classification. Nevertheless, this is not universally true:
in other simulations, in
fact, we have noticed that if γ(t) is too small, the performance are worse. This is
not surprising, since γ(t) determines the weights assigned to the consensus and input
driven parts, whose contributions must be somehow balanced in order to obtain the
best solution. An important point that we will study in future is the optimization
of γ(t), whose choice may in turn depend on the graph topology and on the weights
assigned in the matrix P .
Analogous considerations can be done for the mean square error on θ: when N
increases, the mean square error decays to zero.
We remark that convergence is numerically shown also for the ring topology, which
is not envisaged by our theoretical analysis. Hence, our guess is that convergence
should be proved even under weaker hypotheses on matrix P .
For the interested reader, a graphical user interface of our algorithm is available
and downloadable on http://calvino.polito.it/∼fosson/software.html.
10
0.35
0.3
0.25
0.2
0.15
0.1
0.05
)
⋆
T
,
bT
(
H
d
E
1 N
IML
EM
IA ζ =0.5
IA ζ =0.7
IA ζ =0.9
0.35
0.3
0.25
0.2
0.15
0.1
0.05
)
⋆
T
,
bT
(
H
d
E
1 N
IML
EM
IA ζ =0.5
IA ζ =0.7
IA ζ =0.9
0
10
20
30
40
50
60
70
80
90
100
0
10
20
30
40
50
60
70
80
90
100
N
(a) Complete graph.
N
(b) Ring.
0.35
0.3
0.25
0.2
0.15
0.1
0.05
)
⋆
T
,
bT
(
H
d
E
1 N
IML
EM
IA ζ =0.5
IA ζ =0.7
IA ζ =0.9
0.35
0.3
0.25
0.2
0.15
0.1
0.05
)
⋆
T
,
bT
(
H
d
E
1 N
IML
EM
IA ζ =0.5
IA ζ =0.7
IA ζ =0.9
0
10
20
30
40
50
60
70
80
90
100
0
10
20
30
40
50
60
70
80
90
100
N
(c) Grid.
N
(d) Random geometric graph.
Figure 5.1. Relative classification error
6. Concluding remarks. In this paper, we have presented a fully distributed
algorithm for the simultaneous estimation and classification in a sensor network, given
from noisy measurements. The algorithm only requires the local cooperation among
units in the network. Numerical simulations show remarkable performance. The
main contribution includes the convergence of the algorithm to a local maximum of
the centralized ML estimator. The performance of the algorithm has been also studied
when the network size is large, proving that the solution of the proposed algorithm
concentrates around the classical ML solution.
Different variants are possible, for example the generalization to multiple classes
with unknown prior probabilities should be inferred. The choice of sequence {γ(t)}t∈N
is critical, since it influences both convergence time and final accuracy; the determi-
nation of a protocol for the adaptive search of sequence {γ(t)}t∈N is left for a future
work.
7. Acknowledgment. The authors wish to thank Sandro Zampieri for bringing
the problem to our attention and Luca Schenato for useful discussions. F. Fagnani
and C. Ravazzi further acknowledge the financial support provided by MIUR under
the PRIN project no. 20087W5P2K.
REFERENCES
[1] M. Huang and J. H. Manton, "Coordination and consensus of networked agents with noisy
11
measurements: Stochastic algorithms and asymptotic behavior," SIAM J. Control and
Optimization, vol. 48, no. 1, pp. 134 -- 161, 2009.
[2] R. Olfati-Saber, "Distributed kalman filter with embedded consensus filters," in Proceedings of
44th IEEE Conference on Decision and Control, pp. 8179 -- 8184, 2005.
[3] R. Olfati-Saber, "Kalman-consensus filter : Optimality, stability, and performance," in Proceed-
ings of 48th IEEE Conference on Decision and Control, pp. 7036 -- 7042, 2009.
[4] C. Moallemi and B. V. Roy, "Consensus propagation," IEEE Transactions on Information
Theory, vol. 52, no. 11, pp. 4753 -- 4766, 2006.
[5] V. Saligrama, M. Alanyali, and O. Savas, "Distributed detection in sensor networks with packet
losses and finite capacity links," IEEE Transactions on Signal Processing, vol. 54, no. 11,
pp. 4118 -- 4132, 2006.
[6] R. Duda, P. Hart, and D. Stork, Pattern Classification. Wiley, 2001.
[7] A. Chiuso, F. Fagnani, L. Schenato, and S. Zampieri, "Gossip algorithms for simultaneous
distributed estimation and classification in sensor networks," IEEE Journal of Selected
Topics in Signal Processing, pp. 691 -- 706, 2011.
[8] A. Dogandzic and B. Zhang, "Distributed estimation and detection for sensor networks using
hidden markov random field models," IEEE Transactions on Signal Processing, vol. 54,
no. 8, pp. 3200 -- 3215, 2006.
[9] N. Ramakrishnan, E. Ertin, and R. Moses, "Gossip-based algorithm for joint signature estima-
tion and node calibration in sensor networks," Selected Topics in Signal Processing, IEEE
Journal of, vol. 5, no. 4, pp. 665 -- 673, 2011.
[10] D. M. Titterington, A. F. M. Smith, and U. E. Makov, Statistical Analysis of Finite Mixture
Distributions. New York: John Wiley, 1985.
[11] R. A. Redner and H. F. Walker, "Mixture densities, maximum likelihood and the em algorithm,"
SIAM Review, pp. 195 -- 239, 1984.
[12] R. Nowak, "Distributed EM algorithms for density estimation and clustering in sensor net-
works," IEEE Transactions on Signal Processing, vol. 51, no. 8, pp. 2245 -- 2253, 2003.
[13] M. Rabbat and R. Nowak, "Distributed optimization in sensor networks," in ISPN '04: Pro-
ceedings of the 3rd international symposium on Information processing in sensor networks,
2004.
[14] D. Gu, "Distributed EM algorithms for gaussian mixtures in sensor networks," IEEE Transac-
tions on Neural Networks, vol. 19, no. 7, pp. 1154 -- 1166, 2008.
[15] A. Das and M. Mesbahi, "Distributed linear parameter estimation over wireless sensor net-
works," IEEE Transactions on Aerospace and Electronic Systems, vol. 45, no. 4, pp. 1293
-- 1306, 2009.
[16] F. Fagnani, S. M. Fosson, and C. Ravazzi, "Input driven consensus algorithm for distributed
estimation and classification in sensor networks," in Proceedings of 50th IEEE Conference
on Decision and Control, no. 20087, pp. 6654 -- 6659, 2011.
[17] A. Yeredor, "The joint MAP-ML criterion and its relation to ML and to extended least-squares,"
IEEE Transactions on Signal Processing, vol. 48, no. 12, pp. 3484 -- 3492, 2000.
[18] H. V. Trees and K. Bell, eds., Bayesian Bounds for parameter estimation and nonlinear filter-
ing/tracking. Wiley & Sons, 2007.
[19] A. P. Dempster, N. M. Laird, and D. B. Rubin, "Maximum likelihood from incomplete data via
the EM algorithm," J. Roy. Statist. Soc. Ser. B, vol. 39, no. 1, pp. 1 -- 38, 1977.
[20] J. Bilmes, "A gentle tutorial of the EM algorithm and its application to parameter estimation
for Gaussian mixture and hidden Markov models," Tech. Rep. TR-97-021, ICSI, 1997.
[21] H. Robbins and S. Monro, "A stochastic approximation method," Annals of Mathematical
Statistics, vol. 22, no. 3, pp. 400 -- 407, 1951.
[22] D. Andrews, "Generic uniform convergence," Econometric Theory, pp. 241 -- 257, 1992.
[23] L. Xiao, S. Boyd, and S. Lall, "Distributed average consensus with time-varying metropolis
weights," available on www.stanford.edu/∼boyd/papers/avg_metropolis.html, 2006.
[24] P. J. Davis, Circulant matrices. A Wiley-Interscience Publication, Pure and Applied Mathe-
maticsJohn Wiley & Sons, New York-Chichester-Brisbane, 1979.
[25] F. R. Gantmacher, The theory of matrices. New York: Chelsea publ., 1959.
[26] V. Borkar, Probability Theory. New York: Springer-Verlag, 1995.
Appendix A. Proof of Theorem 4.1. Consider the discrete-time dynamical
system defined by the update equations (4.1) and (4.2): the proof of its convergence
is obtained through intermediate steps.
1. First, we show that, for sufficiently large t, vectors µ(t), ν(t), and bθ(t) are
close to consensus vectors and we prove their convergence, assumingbω(t) has
12
already stabilized.
2. Second, we prove the stabilization of bω(t) in finite time, by modelling the
system in (4.1) and (4.2) as a switching dynamical system.
3. Finally, combining these facts together we conclude the proof.
A.1. Towards consensus. We start with some notation: let Ω := I − N−1
Given a bounded sequence u(t) ∈ RN , consider the dynamics
t ∈ N
x(t+1) =(cid:16)1 − γ(t)(cid:17) P x(t) + γ(t)u(t)
Tx so that x = x1 + Ωx.
(A.1)
T;
11
given x ∈ RV, let x := N−1
1
where x(0) is any fixed vector, and where, we recall the standing assumptions,
(a) γ(t) ∈ (0, 1), γ(t) ≥ 1/t, γ(t) ց 0 and γ(t) = γ(t+1) + o(γ(t+1)) for t → +∞;
(b) P ∈ RV×V+
is a stochastic, symmetric, primitive matrix with positive eigen-
A useful fact consequence of the assumptions on γ(t), is the following:
values.
(1 − γ(s)) ≤ e
t−1P
s=t0
−
1/s
t0
t ≤ t0γ(t)
≤
t−1Ys=t0
(A.2)
for any choice of t ≥ t0 > 0.
On the other hand, as a consequence of the assumptions of P (see [25]) we have
that P t → N−1
T , or equivalently that P tΩ → 0 for t → +∞. More precisely, we
can order the eigenvalues of P as 1 = µ1 > µ2 ≥ ··· ≥ µN ≥ 0, and we have that
P tΩ ≤ µt
2.
11
Lemma A.1. It holds
Ωx(t) = O(γ(t)) ,
for t → +∞.
Proof. From (A.1) and the fact that ΩP = P Ω we get, for any fixed t0 and t ≥ t0,
Ωx(t+1) =
tYs=t0(cid:16)1 − γ(s)(cid:17) P tΩx(t0) +
tXs=t0
tYk=s+1(cid:16)1 − γ(k)(cid:17) γ(s)P t−sΩu(s).
(A.3)
This yields
Ωx(t+1)2 ≤
≤
tYs=t0(cid:16)1 − γ(s)(cid:17) Ωx(t0)2 +
tYk=s+1(cid:16)1 − γ(k)(cid:17) γ(s)µ2t−su(s)2
tXs=t0
tYk=s+1(cid:16)1 − γ(k)(cid:17) γ(s)µ2t−s
tYs=t0(cid:16)1 − γ(s)(cid:17) Ωx(t0)2 + K
tXs=t0
(A.4)
with K := maxs u(s)2.
Fix now 0 < ε < 1 − µ2 and let t0 ∈ N be such that γ(t+1)
all t ≥ t0. Hence, for t ≥ s ≥ t0, we have that γ(s) < γ(t)
estimation (A.4) with this choice of t0. We get
γ(t) ∈ (1 − ε, 1) for
(1−ε)t−s . Consider now the
13
Ωx(t+1)2 ≤
≤
tYs=t0(cid:16)1 − γ(s)(cid:17) Ωx(t0)2 + Kγ(t)
tYs=t0(cid:16)1 − γ(s)(cid:17) Ωx(t0)2 +
Kγ(t)
1 − µ21−ε
.
tXs=t0(cid:18) µ2
1 − ε(cid:19)t−s
Using now (A.2) the proof is completed.
Proposition A.2. If ∃ t0 ∈ N s.t. u(t) = u ∀t ≥ t0 then
x(t) = u1.
lim
t→+∞
Proof. Write x(t) = x(t)
to prove that limt→+∞ x(t)
1 + Ωx(t) and notice that from Lemma A.1 it is sufficient
T , we obtain
1 = u1. From (A.1) and the fact that 1
T P = 1
x(t) − u =
t−1Ys=t0
(1 − γ(s))(x(s) − u)
which goes to zero from the non-summability of γ(t).
result.
We now apply these results to the analysis ofbθ(t). We start with a representation
Lemma A.3. It holds, for t → +∞,
¯µ(t)
¯ν(t) 1 +
1
¯ν(t) Ω(cid:18)µ(t) −
¯µ(t)
¯ν(t) ν(t)(cid:19) + o(cid:16)γ(t)(cid:17) .
(A.5)
bθ(t) =
Proof. For any i ∈ V,
µ(t)
i
ν(t)
i
¯µ(t)
¯ν(t) =
−
=
=
¯µ(t)
¯ν(t) +
−
µ(t)
i
ν(t)
i
µ(t)
i − ¯µ(t)
¯ν(t)
+ µ(t)
1
¯ν(t)(cid:16)Ωµ(t)(cid:17)i −
µ(t)
i
¯ν(t) −
µ(t)
i
¯ν(t)
1
−
i 1
¯ν(t)!
i ¯ν(t)(cid:16)Ων(t)(cid:17)i
ν(t)
i
µ(t)
i
ν(t)
.
It follows from Lemma A.1 that µ(t) = ¯µ(t)
1 + O(γ(t)) for
t → +∞. This and the fact that ¯ν(t) is bounded away from 0 (indeed ¯ν(t) ≥ α−2 for
all t > 0), yields
1 + O(γ(t)) and ν(t) = ¯ν(t)
µ(t)
i
ν(t)
i ¯ν(t)(cid:16)Ων(t)(cid:17)i
=
¯µ(t)
¯ν(t)"(cid:0)Ων(t)(cid:1)i
¯ν(t)
#(cid:16)1 + O(cid:16)γ(t)(cid:17)(cid:17)
from which thesis follows.
We can now present our first convergence result.
Corollary A.4. It holds, for t → +∞,
=
(t)
¯bθ
¯µ(t)
¯ν(t) + o(γ(t)) , Ωbθ(t) = O(γ(t)).
14
i
lim
]−2
Proof. Both relations are obtained from (A.5). The first one is immediate. The
second one follows from Lemma A.1 and the fact that ¯ν(t) stays bounded away from
0.
finite time as explained in the next result.
Corollary A.4 says that the estimate bθ(t) is close to a consensus for sufficiently
large t. Something more precise can be stated if we know that if bω(t) stabilizes at
Corollary A.5. If ∃ t0 ∈ N s.t. bω(t) =bωIA ∀t ≥ t0 then
i (cid:3)−2
yi(cid:2)bωIA
i (cid:3)−2 1.
Pi∈V(cid:2)bωIA
t→+∞bθ(t) =bθ(bωIA) = Pi∈V
Proof. Proposition A.2 guarantees that µ(t) and ν(t) converge to 1
and 1
[bωIA
1, respectively. This yields the thesis.
stabilizes in finite time: this, by virtue of previous considerations will complete our
proof. To prove this fact will take lots of effort and will be achieved through several
intermediate steps.
yi[bωIA
NPi∈V
A.2. Stabilization ofbω(t). We are going to prove that vectorbω(t) almost surely
We start observing that, sincebω(t) can only assume values in a finite set, equations
in (4.1) and (4.2) can be conveniently modeled by a switching system as shown below.
For reasons which will be clear below, in this subsection we will replace the
configuration space {α, β}V with the augmented state space {α, β+, β−}V. If ω ∈
{α, β+, β−}V, define
Θω = {x ∈ RV :xi − yi < δ, if ωi = α, xi ≥ yi + δ, if ωi = β+, xi ≤ yi − δ, if ωi = β−}.
We clearly have RV =Sω∈{α,β+,β−}V Θω.
RV and gω : R × RV → RV by
On each Θω the dynamical system is linear. Indeed, define the maps fω : R×RV →
NPi∈V
]−2
i
1
ω2
i
[fω(t, x)]i = (1 − γ(t))[P x(t)]i + γ(t) yi
[gω(t, x)]i = (1 − γ(t))[P x(t)]i + γ(t) 1
i = β2 if ωi = β+, β−. Then, ifbθ(t) ∈ Θω, (4.1a), (4.1b), and
ω2
i
µ(t+1) = fω(t, µ(t))
ν(t+1) = gω(t, ν(t))
where, conventionally, ω2
(4.1c) can be written as
= µ(t+1)
i
/ν(t+1)
i
.
i
bθ(t+1)
Notice that this is a closed-loop switching system, since the switching policy is deter-
mined by bθ(t). It is clear that the stabilization of bω(t) is equivalent to the fact that
there exist an ω ∈ {α, β+, β−}N and a timeet such that bθ(t) ∈ Θω for all t ≥et.
From Corollary A.5 candidate limit points for bθ(t) are
ω ∈ {α, β+, β−}N .
bθ(ω)1 = Pi∈V
Pi∈V
yiω−2
i
ω−2
i
1
15
Also, from Proposition A.4, the dynamics can be conveniently analysed by study-
We now make an assumption which holds almost everywhere with respect to the
ing it in a neighborhood of the line Λ = {λ1λ ∈ R}.
choice of yi's and, consequently, does not entail any loss of generality in our proof.
ASSUMPTION:
• yi − yj 6∈ {0,±δ,±2δ} for all i 6= j;
This assumption has a number of consequences which will be used later on:
• bθ(ω) − yi 6∈ {±δ} for all ω ∈ {α, β+, β−}V and for all i.
(C1) bθ(ω)1, yi1 ∈ Sω∈{α,β+,β−}V int(Θω) for all ω ∈ {α, β+, β−}V and for all
i ∈ V;
(C2) Λ ∩ ¯Θω ∩ ¯Θω′ ∩ ¯Θω′′ = ∅ for any triple of distinguished ω, ω′, ω′′. In other
terms, Λ always crosses boundaries among regions Θω at internal point of
faces.
We now introduce some further notation, which will be useful in the rest of the
paper.
Θǫ := {x ∈ RV : Ωx2 < ǫ}, Θǫ
ω := Θǫ ∩ Θω
Γ := {ω ∈ {α, β+, β−}V : Θω ∩ Λ 6= ∅}.
For any ω ∈ Γ consider
Πω = {π = ¯Θω ∩ ¯Θω′ : dH(ω, ω′) = 1, π ∩ Λ = ∅}
and define σω := minπ∈Πω d(Θω ∩ Λ, π) > 01.
X < Y means that x < y for all x ∈ X and y ∈ Y .
follower of ω (or also that ω is the past-follower of ω′) if the following happens:
In the sequel, we will use the natural ordering on Λ: given the sets X, Y ⊆ Λ,
Definition A.6. Given two elements ω, ω′ ∈ Γ, we say that ω′ is the future-
(A) There exists i0 such that ωi = ω′i for all i 6= i0 and ωi0 6= ω′i0;
(B) Θω ∩ Λ < Θω′ ∩ Λ.
Notice that, in order for ω and ω′ to satisfy definition above, it must necessarily
happen that either ωi0 = α and ω′i0 = β+, or ωi0 = β− and ω′i0 = α. Given ω ∈ Γ,
its future-follower (if it exists) will be denoted by ω+. It is clear that (because of
property (C2) described above) that we can order elements in Γ as ω1, ω2, . . . , ωM in
such a way that ωr+1 = (ωr)+ for every r = 1, . . . , M − 1.
Given ω ∈ Γ, consider the following subsets of RN (see Fig. A.1):
ω := {x ∈ Θǫ
Mǫ
Lǫ
ω,ω+ := {x ∈ Θǫ : Mǫ
ω : x1 + Ωz ∈ Θǫ
ω, ∀z : z2 < ǫ}
ω ∩ Λ < ¯x < Mǫ
ω+ ∩ Λ} .
(with the implicit assumption that Lǫ
ω,ω+ = ∅ if ω+ does not exist.) We clearly have
Θǫ =Sω∈Γ Mǫ
such that
ω ∪ Lǫ
ω,ω+ .
Notice that, because of property (C1), we can always choose ǫ0 ∈ (0, minω∈Γ σω)
bθ(ω)1, yi1 ∈ [ω′∈Γ
Mǫ0
ω′
∀ω ∈ Γ,∀i ∈ V.
1 d(Θω ∩ Λ, π) denotes the distance between the two sets Θω ∩ Λ and the set π
16
Θω
Mǫ
ω
Θ
+
ω
Lǫ
ω,ω
+
Λ
ǫ
Figure A.1. Given the couple (ω, ω′) the sets Lǫ
ω,ω′ and Mǫ
ω are visualized.
This implies that there exists c > 0 such that
d [ω′∈Γ
∂Λ (Mǫ
ω′ ∩ Λ) ,{bθ(ω), yi}! ≥ c,
∀ǫ ≤ ǫ0
(A.6)
where ∂Λ(·) denotes the boundary of a set in the relative topology of Λ.
ω then Mǫ
Corollary A.4). From now on we consider times t ≥ tǫ. Our aim is to prove through
intermediate steps the following facts
Fix now ǫ ≤ ǫ0 and choose tǫ such that bθ(t) ∈ Θǫ for all t ≥ tǫ (it exists by
F1) if bθ(ω) ∈ Mǫ
ω is an asymptotically invariant set for bθ(t), namely,
when t is sufficiently large, if bθ(t) ∈ Mǫ
F2) if bθ(ω) /∈ Mǫ
F3) bθ(t) /∈Sω∈{α,β+,β−}V Lǫ
Lemma A.7. Ifbθ(t) ∈ Θω then there exists c(t) ∈ [α2/β2, β2/α2] and r(t) = o(γ(t))
ω then bθ(t+1) ∈ Mǫ
ω when bθ(ω)1 ∈ Mǫ
(t)(cid:19) + r(t)
F1) Asymptotic invariance of Mǫ
ω then bθ(t) /∈ Mǫ
for t → +∞ such that
ω,ω+ for t sufficiently large.
ω for t sufficiently large;
(A.7)
(t+1)
ω.
ω;
bθ
=bθ
Proof. If bθ(t) ∈ Θω then
µ(t+1)
ν(t+1) −
µ(t)
ν(t) =
(t)
+ c(t)γ(t)(cid:18)bθ(ω) −bθ
(1 − γ(t))µ(t) + γ(t)N−1PN
(1 − γ(t))µ(t) + γ(t)N−1PN
i − µ(t)γ(t)N−1PN
ν(t)γ(t)N−1PN
bθ(ω) −
= γ(t) N−1PN
i=1 yiω−2
i=1 ω−2
ν(t)!
i=1 yiω−2
ν (t+1)ν(t)
i=1 ω−2
ν(t+1)
µ(t)
ν(t)
µ(t)
−
=
i
i
i
17
i=1 ω−2
i
Choosing c(t) = N −1 PN
i=1 ω−2
i
ν(t+1)
easily follows.
∈ [α2/β2, β2/α2] and using Corollary A.4 thesis
then
Proposition A.8 (Proof of F1)). There exists t′ ≥ tǫ such that, if bθ(ω)1 ∈ Θω,
ω and if t is large enough so that
ω ∀t ≥ t′ .
c(t)γ(t) < 1 , we have, by convexity, that
bθ(t) ∈ Mǫ
ω ⇒ bθ(t+1) ∈ Mǫ
Proof. Consider the relation (A.7). If bθ(t) ∈ Mǫ
+ c(t)γ(t)(cid:18)bθ(ω) −bθ
(t)
(t)(cid:19) ∈ Mǫ
ω.
z := bθ
Moreover, because of (A.6) and the fact that c(t) is bounded away from 0, there exists
c′ > 0 such that d(z, ∂(Mǫ
ω∩ Λ)) ≥ c′γ(t). Proof is then completed by selecting t′ ≥ tǫ
such that c(t)γ(t) < 1 and r(t) < c′γ(t)/2 for all t ≥ t′.
F2) Transitivity of Mǫ
if bθ(ω)1 /∈ Mǫ
technical lemma based on convexity arguments is required.
ω, then, at a certain time t, bθ(t) will definitively be outside Mǫ
Lemma A.9. Let ω ∈ Γ be such that there exists its future-follower ω+. Then,
ω. Our next goal is to prove that
ω. A
ω when bθ(ω)1 /∈ Mǫ
bθ(ω)1 > Θω ∩ Λ
⇒ bθ(ω+)1 > Θω ∩ Λ
bθ(ω+)1 < Θω+ ∩ Λ ⇒ bθ(ω)1 < Θω+ ∩ Λ.
Proof. Suppose ωi = ω+
i0 = α (the other case can
be treated in an analogous way). Pick x′ ∈ Θω ∩ Λ and x′′ ∈ Θω+ ∩ Λ. From
x′′ − yi0 < δ, and x′ − yi0 > δ it immediately follows that x′′ > yi0 − δ, x′ < yi0 − δ
and, in particular, the fact
i ,∀i 6= i0 and ωi0 = β−, ω+
Notice now that
yi0 1 > Θω ∩ Λ .
(A.8)
bθ(ω+) =
=
=
2 + yi0
β2
2
1
ω+
i
yi0(cid:16) 1
β2(cid:17)
α2 − 1
Pi∈V
β2(cid:17)
yi0(cid:16) 1
α2 − 1
Pi∈V
yi0(cid:16) 1
β2(cid:17)
α2 − 1
Pi∈V
1
ω+
i
1
ω+
i
2
2
1
ω+
i
1
ωi
2
2
yi
ω+
i
+ Pi∈V\i0
Pi∈V
+ bθ(ω)Pi∈V
Pi∈V
+ bθ(ω)(cid:20)Pi∈V
1
ω+
i
2
α2 − 1
β2(cid:17)(cid:21)
.
1
ω+
i
2 −(cid:16) 1
Pi∈V
1
ω+
i
2
A convexity argument and the use of (A.8) now allow to conclude.
In Figures A.2 and A.3 a picture of the various points is depicted whenbθ(ω) > Θω∩Λ.
Proposition A.10 (Proof of F2)). If bθ(ω)1 /∈ Θω, then there exists t′′ such that
bθ(t) /∈ Θǫ
ω ∀t > t′′.
18
Θω
Θω+
Θω
Θω+
Λ
yi0 1 bθ(ω+)1 bθ(ω)1
(a) yi0 < yω+ < yω.
¯yi0 1
Λ
bθ(ω)1 bθ(ω+)1
(b) yi0 < bθ(ω+) < bθ(ω).
Figure A.2. ωi0 = β, bθ(ω)1 > Θω ∩ Λ
Θω
Θω+
Λ
¯yi0 1
bθ(ω)1 bθ(ω+)1
Figure A.3. bθ(ω)1 > Θω ∩ Λ
Proof. Supposebθ(ω)1 > Θω ∩ Λ (the case when is < can be treated analogously).
Lemma A.9 implies that bθ(ω+)1 > Θω ∩ Λ. Let c be the constant given in (A.6) and
put
Consider the relation (A.7) and choose t1 in such a way that
A := {x ∈ Θǫ
(t+1)
(t)
ω ∪ Θǫ
ω+ x ≤ α := min{bθ(ω),bθ(ω+)} − c/2}.
≤ c2(max{yi} − min{yi})γ(t) + r(t) < c/2
(t+1)
(t)
−
¯bθ
≥bθ
+ α2c γ(t)/4β2.
(A.9)
and r(t) < α2cγ(t)/4β2 for all t ≥ t1. It also follows from (A.7) that, if for some
¯bθ
t ≥ t1 bθ(t) ∈ A, then,
Owing to the non-summability of γ(t) it follows that if bθ(t) enters in Θǫ
t ≥ t1, then, in finite time it will enter into A \ Θǫ
A. In particular there must exist t2 ≥ t1 such that bθ
bθ
that bθ
< α − c. Because of (A.9), it must be that bθ
contradicts the fact that on A, bθ
F3) Transitivity ofSω,ω+∈{α,β+β−}N Lω,ω+. We start with the following tech-
exists a strictly increasing sequence of switching times {τk}+∞k=0 such that
> α. We now prove that
1 > Θω for every t ≥ t2. If not there must exist a first time index t3 > t2 such
< α − c/2 but this
ω for some
ω and then it will finally exit
(t2)
Lemma A.11. Let x(t) be the sequence defined in (A.1) and suppose that there
nical result concerning the general system (A.1).
is increasing (A.10).
(t2)
(t3)
(A.10)
bθ
(t)
(t3−1)
u(s+1)
i
= u(s)
i
∀i 6= i0
and ∀s ∈ [τ0, +∞[
i0 =(v′
u(s)
∀s ∈ I′ :=S+∞k=0[τ2k, τ2k+1)
v′′ ∀s ∈ I′′ :=S+∞k=0[τ2k+1, τ2k+2).
19
b
b
b
b
b
b
b
b
b
Then, for every δ > 0, there exists ¯tδ and two sequences a(t)
such that
δ ≥ 0 and b(t)
δ ≤ δγ(t),
(cid:16)Ω(cid:16)x(t+1) − x(t)(cid:17)(cid:17)i0
= A(t)
δ γ(t) (v′ − v′′) + b(t)
δ
for t ∈ I′ with t ≥ ¯tδ.
eigenvalues 1 = λ1 > λ2 ≥ ··· ≥ λN ≥ 0. Also assume we have chosen φ1 = N−1/2
Proof. Let φj ∈ RV be an orthonormal basis of eigenvectors for P relative to the
1.
We put
F (t) := Qt
k=0(cid:0)1 − γ(k)(cid:1)
γ(t)
and we notice that
F (s+1)
F (s) = (1 − γ(s+1))
γ(s)
γ(s+1) → 1 , for s → +∞.
Fix ǫ in such a way that λ2(1 + ǫ) < 1 and choose s0 such that
F (s+1)
F (s) ≤ 1 + ǫ , ∀s ≥ s0.
Let t0 ≥ s0 to be fixed later. From (A.3) we can write
Ω(x(t+1) − x(t)) =
=
t−1Ys=t0(cid:16)1 − γ(s)(cid:17)h(cid:16)1 − γ(t)(cid:17) P − Ii P t−t0Ωx(t0)
tYk=s+1(cid:16)1 − γ(k)(cid:17) γ(s)P t−sΩu(s) −
tXs=t0
t−1Xs=t0
t−1Ys=t0(cid:16)1 − γ(s)(cid:17)h(cid:16)1 − γ(t)(cid:17) P − Ii P t−t0Ωx(t0)
t−1Xs=t0
t−1Xs=t0−1
t−1Ys=t0(cid:16)1 − γ(s)(cid:17)h(cid:16)1 − γ(t)(cid:17) P − Ii P tΩx(t0)
F (s+1) Ωu(s+1) − γ(t−1)
P t−s−1 F (t)
+
=
=
+ γ(t)
+ (γ(t) − γ(t−1))
F (t0) Ωu(t0)
t−1Xs=t0
F (t−1)
P t−s−1 F (t−1)
F (s) Ωu(s) + γ(t)P t−t0 F (t−1)
F (s) (cid:19) Ωu(s+1)
F (s) Ω(cid:16)u(s+1) − u(s)(cid:17) .
F (s+1) −
P t−s−1(cid:18) F (t)
P t−s−1 F (t−1)
+ γ(t)
+ γ(t)
t−1Xs=t0
t−1Xs=t0
20
t−1Yk=s+1(cid:16)1 − γ(k)(cid:17) γ(s)P t−s−1Ωu(s)v
P t−s−1 F (t−1)
F (s) Ωu(s)
(A.11)
(A.12)
(A.13)
(A.14)
It follows from the assumptions on P , the assumptions on γ(t) and relation (A.2)
that the terms (A.11) and (A.12) are both o(γ(t)) for t → +∞. We now estimate
(A.13):
≤
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2
(A.15)
(A.16)
(A.17)
where
F (t−1)
F (s+1) −
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
P t−s−1(cid:18) F (t)
t−1Xs=t0
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
F (s) (cid:18) F (t)
t−1Xs=t0
[λ2(1 + ǫ)]t−s−1(cid:12)(cid:12)(cid:12)(cid:12)(cid:18) F (s)
t−1Xs=t0
F (t−1)
F (s)
K
βt0
1 − λ2(1 + ǫ)
=
F (t−1)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2
F (s) (cid:19) Ωu(s+1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
F (s+1) − 1(cid:19) P t−s−1Ωu(s+1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
F (s+1) − 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) K ≤
t≥s≥t0(cid:12)(cid:12)(cid:12)(cid:12)(cid:18) F (t)
F (s) Ω(cid:16)u(s+1) − u(s)(cid:17)#i0
βt0 := sup
F (t−1)
=
K = maxu(s)2 ,
F (s)
F (s+1) − 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) .
λt−τh
j
F (t − 1)
F (τh − 1)
(−1)h(v′ − v′′).
(φj )2
P t−s−1 F (t−1)
" t−1Xs=t0
Xj≥2
i0 Xh:t0≤τh≤t−1
i0 Xk:t0≤τ2k≤t−1(cid:20)λt−τ2k
j
(φj)2
Xj≥2
If t ∈ I′, the above expression can be rewritten as
We now concentrate on the component i0 of the term (A.14). Using the spectral
decomposition of P and the assumptions on u(t), we can write,
Notice that
λt−τ2k
j
F (t − 1)
F (τ2k − 1)−λt−τ2k−1
j
F (t − 1)
F (τ2k−1 − 1)
= λt−τ2k
j
F (t − 1)
F (τ2k − 1) − λt−τ2k−1
j
F (t − 1)
F (τ2k−1 − 1)(cid:21) (v′ − v′′).
F (τ2k − 1)(cid:18)1 − λτ2k−τ2k−1
F (t − 1)
j
F (τ2k − 1)
F (τ2k−1 − 1)(cid:19) > 0
(we have used the fact that 0 ≤ λj (1 + ǫ) < 1 for all j ≥ 2). To complete the proof
now proceed as follows. For a fixed δ > 0, choose t0 ≥ s0 in such a way that (A.15) is
below δ/2. Then, fix ¯tδ ≥ t0 in such a way that the summation of (A.11) and (A.12)
is below δγ(t)/2 for t ≥ ¯tδ. It is now sufficient to define
a(t)
δ
:=Xj
(φj )2
i0 Xk:t0≤τ2k≤t−1(cid:20)λt−τ2k−1
j
F (t − 1)
F (τ2k − 1) − λt−τ2k−1
j
F (t − 1)
F (τ2k−1 − 1)(cid:21)
and b(t)
δ
equal to the sum of the terms (A.11), (A.12), and (A.13),
Proposition A.12 (Proof of F3)). There exists t′′′ ∈ N such that
bθ(t) 6∈ [ω,ω′∈{α,β}N Lω,ω+
21
(A.18)
for all t > t′′′.
Proof.
In view of the results in Propositions A.8 and A.10, and the fact that
From Lemma A.3 and Corollary A.4 we can write
bθ(t+1) −bθ(t) goes to 0 for t → +∞, if (A.18) negation of (A.18) yields that there
exists ω ∈ Γ such that bθ(t) ∈ Lω,ω+ for t large enough. Now, if bθ(t) ∈ Lω,ω+ ∩ Θω (or
if bθ(t) ∈ Lω,ω+ ∩ Θω+ ) for t sufficiently large, a straightforward application of (A.7)
would imply that bθ(t) would necessarily exit Lω,ω+ in finite time. Therefore, it must
hold that bθ(t) keeps switching, for large t, between Lω,ω+ ∩ Θω and Lω,ω+ ∩ Θω+ .
bθ(t+1) −bθ(t) =
¯ν(t) 1 +
¯ν(t) Ω(cid:16)ν(t+1) − ν(t)(cid:17)(cid:21) + o(γ(t))
¯bθ
¯bθ
I′ := {tbθ(t) ∈ Θω} ,
¯ν(t)(cid:20)Ω(cid:16)µ(t+1) − µ(t)(cid:17) −
I′′ := {tbθ(t) ∈ Θω+}
i0 . From Lemma A.7, and applying Lemma A.11
and put v′ = 1/ω2
to µ(t) and ν(t), we get that for t ∈ I′ sufficiently large, it holds
i0 and v′′ = 1/ω+2
¯ν(t+1) −
Define now
¯µ(t)
(t+1)
(t)
1
(t)
(t)
1
¯ν(t) γ(t)a(t)
δ
(t)
) +
) + a(t)
δ + r(t). (A.19)
(v′ − v′′) (yi0 −bθ
i0 = c(t)γ(t)(¯yω −bθ
i0 −bθ(t)
bθ(t+1)
If bθ(ω) > Θω ∩ Λ, then also, by Lemma A.9, ¯yω+ > Θω ∩ Λ. This, using (A.7), would
imply that bθ(t) would necessarily exit Lω,ω+ in finite time. Therefore, we must have
bθ(ω) < Θω ∩ Λ. Hence, ¯yω −bθ
(v′ − v′′) (yi0 −bθ
notice that, since bθ(t) ∈ Lω,ω+,
Choose now δ such that δ < α2c/16β2 and ¯t ≥ ¯tδ such that r(t) < δγ(t). It then
follows from (A.19) that for t ∈ I′ and t ≥ ¯t, it holds
) < 0. Recall now the definition of the constant c in (A.6) and
c(t)γ(t)(¯yω −bθ
< 0. Moreover, it is easy to check that in any case
) ≤ −α2c/4β2γ(t).
(t)
(t)
i0
bθ(t+1)
−bθ(t)
i0 ≤ −α2c/8β2γ(t) < 0.
thus implies the thesis.
A.3. Proof of Theorem 4.1. Propositions A.8, A.10, and A.12 imply that
This says that as long as bθ(t) ∈ Θω, its i0-th component decreases. But this entails
thatbθ(t) can never leave Θω, which contradicts the infinite switching assumption and
there exists bωIA ∈ {α, β}V such that bθ(t) ∈ ΘbωIA for t sufficiently large. This im-
mediately implies that bω(t) = bωIA for t sufficiently large. Corollary A.5 implies
that bθIA = limt→+∞bθ(t) = θ(bωIA) Finally, since θ(bωIA) ∈ ΘbωIA, we also have that
bωIA = ω(bθIA).
Appendix B. Proof of concentration results.
22
B.1. Preliminaries. For a more efficient parametrization of the stationary points,
we introduce the notation:
ω ∈ {α, β}V Θω := {x ∈ Rx − yi < δ ⇔ ωi = α}
(B.1)
It is then straightforward to check from (3.11) that the set of local maxima SN can
be represented as
Since,
SN := {θ =bθ(ω) ω ∈ {α, β}V , bθ(ω) ∈ Θω}.
Θω 6= ∅ ⇔ ω =bω(x) for some x ∈ R
(B.2)
(B.3)
1 )
1 )
E(y1bω(x)−2
E(bω(x)−2
for analysing the set SN we can restrict to consider ω of type ω =bω(x). Consider the
sequence of random functions γN (x) :=bθ(bω(x)).
From (3.8), applying the strong law of large numbers, we immediately get that
lim
N→+∞
γN (x) a.s.= γ∞(x) :=
.
(B.4)
Something stronger can indeed be said by a standard use of Chernoff bound [26]:
Lemma B.1. For every ǫ > 0, there exists q < 1 such that, for any x ∈ R,
P (γN (x) − γ∞(x) > ǫ) ≤ 2qN .
Proof. Let ai = yiωN (x)−2
i
and bi = ωN (x)−2
b denote the corresponding expected values.
i with i ∈ {1, . . . , N} and let a and
By Chernoff's bound and by Hoeffding's inequality we have, respectively, that
1
N
NXi=1
P (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ai − a(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ ǫ1! ≤ qN
1
1
N
NXi=1
P (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
bi − b(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ ǫ2! ≤ 2qN
2
with
q1 = e−
α2ǫ2
1
4
q2 = e−2ǫ2
2(α−2−β−2)−2
.
(B.5)
Fix ǫ1 < ǫ
2bβ4 and ǫ2 < ǫ
2aβ4 , then
P(cid:0)(cid:12)(cid:12)¯yωN (x) − y∞(x)(cid:12)(cid:12) > ǫ(cid:1) ≤ P
≤ qN
= qN
1 + qN
1 + qN
2
(cid:12)(cid:12)(cid:12) 1
i=1 bi(cid:12)(cid:12)(cid:12)
i=1 ai − a(cid:12)(cid:12)(cid:12) b + a(cid:12)(cid:12)(cid:12)b − 1
NPN
NPN
i=1 bi
2 + 1{β4(ǫ1b+aǫ2)>ǫ}
NPN
b 1
> ǫ
where the last step follows by the way ǫ1 and ǫ2 have been chosen.
There is still a point to be understood: in our derivation q1 and q2 depend on the
choice of x through a and b. However, it is immediate to check that a and b are both
bounded in x. This allows to conclude.
23
has an important property which will be useful later on.
From (B.4) is immediate to see that γ∞ is a bounded function of class C1 and it
Lemma B.2. There exists a constant C > 0 such that
x − γ∞(x) ≥ C(x − θ∗)
γ∞(x) − x ≥ C(θ∗ − x)
γ∞(θ∗) = θ∗
if x ∈ (θ⋆, +∞)
if x ∈ (−∞, θ⋆)
Proof.
If x ∈ (θ⋆, +∞) and f is the density of each yi (a mixture of two Gaussians) then
x − y∞(x) =
β2RR\(x−δ,x+δ) (x − t)f (t)dt
β2RR\(x−δ,x+δ) f (t)dt
1
1
α2R x+δ
x−δ (x − t)f (t)dt + 1
α2R x+δ
x−δ f (t)dt + 1
β2RR (x − t)f (t)dt
α2R x+δ
x−δ f (t)dt + 1
1
1
≥
β2RR\(x−δ,x+δ) f (t)dt
where the last inequality follows from the fact that R x+δ
conclude that
x−δ (x − t)f (t)dt ≥ 0. We
x − y∞(x) ≥
1
β2 (x − θ⋆)
1
α2R x+δ
x−δ f (t)dt + 1
β2RR\(x−δ,x+δ) f (t)dt
> 0.
Second statement if x ∈ (−∞, θ⋆) can be verified in a completely analogous way. The
third statement then simply follows by continuity.
We now come to a key result.
Lemma B.3. For any fixed ǫ > 0, there exist q ∈ (0, 1) and χ > 0 such that
P(cid:0)γN (x) ∈ Θbω(x)(cid:1) ≤ χqN
(B.6)
(B.7)
for all x such that x − θ⋆ > ǫ.
Proof. We assume x > θ⋆ + ǫ (the other case x < θ⋆ − ǫ being completely
equivalent). Fix ǫ′ ∈ (0, Cǫ) where C was defined in Lemma B.2 and estimate as
follows
P(cid:0)γN (x) ∈ Θbω(x)(cid:1) ≤ P(cid:0)γN (x) ∈ Θbω(x) , γN (x) − γ∞(x) ≤ ǫ′(cid:1)
+ P (γN (x) − γ∞(x) > ǫ′) .
Using Lemma B.2 we get
Thus
{γN (x) − γ∞(x) ≤ ǫ′} ⊆ {γN (x) ≤ x − (Cǫ − ǫ′)} .
(cid:8)γN (x) ∈ Θbω(x),γN (x) − γ∞(x) ≤ ǫ′(cid:9)
⊆ {∄i : yi ∈ (γN (x) − δ, γN (x) − δ + min{Cǫ − ǫ′, δ})}
and, consequently, the first term in (B.7) can be estimated as
P(cid:0)γN (x) ∈ ΘbωN (x) , γN (x) − γ∞(x) ≤ ǫ′(cid:1) ≤ 1 −Z γN (x)−δ+min{Cǫ−ǫ′,δ}
γN (x)−δ
f (y)dy!N
(B.8)
24
where f (y) is the density of each yi. Considering now that f (y) is bounded away from
0 on any bounded interval, that γN (x) − γ∞(x) ≤ ǫ′ and that γ∞(x) is a bounded
function, we deduce that the right hand side of (B.8) can be uniformly bounded as
qN for some q ∈ (0, 1). Substituting in (B.7), and using Lemma B.1 we finally obtain
the thesis.
B.2. Proof of Theorem 4.2. Define
for any ǫ > 0 and
AN (ǫ) :=n∃ω ∈ {α, β}V :bθ(ω) ∈ Θω,bθ(ω) − θ⋆ > ǫo
B1 := {∃i ∈ V : yi − θ⋆ > N}
B2 :=(cid:8)∃(i, j) ∈ V × V : yi − yj < N−4(cid:9)
B3 := {∃(i, j) ∈ V × V : yi − yj ∈(cid:0)2δ, 2δ + N−4(cid:9)
2 ∩ Bc
1 ∩ Bc
and estimate P (AN (ǫ)) ≤ P (AN (ǫ),Bc
3) + P(B1) + P(B2) + P(B3). Standard
considerations allow to upper bound the probability of each event Bi by a common
term K/N 2. We now focus on the estimation of the first term. The crucial point is
that, the condition Bc
3 allow us to reinforce condition (B.3) in the sense
cardinality is polynomial in N . Specifically, define
that all ω for which Θω 6= ∅ can be obtained as ω =bω(x) as x varies in a set whose
2 ∩ Bc
1 ∩ Bc
Z = {ζj = θ⋆ − N − δ + jN−4 : j ∈ N, j ≤ jmax}
where jmax := ⌈N 4(2N + 2δ)⌉ and notice that, assuming that the yi's satisfy Bc
we have that bω(ζj ) and bω(ζj+1) differ in at most one component and that bω(x) ∈
{bω(ζj ),bω(ζj+1)} for every x ∈ [ζj , ζj+1]. Moreover, because of Bc
1 we have thatbω(x)i =
bω(ζ0)i = β for all x ≤ θ⋆
N − δ and for all i. Similarly, bω(x)i = bω(ζjmax )i = β for all
1 ∩ Bc
x ≥ θ⋆ + N + δ and for sll i. In other terms, under the assumption that the yi's satisfy
1 ∩ Bc
Bc
3, it holds {ω ∈ {α, β}V Θω 6= ∅} = {bω(x) x ∈ Z}. Hence,
P (AN (ǫ),Bc
2 ∩Bc
3,
2 ∩ Bc
2 ∩ Bc
3) ≤
≤ P[ζ∈Z(cid:8)γN (ζ) ∈ Θbω(ζ),γN (ζ) − θ⋆ > ǫ(cid:9)
≤ P[ζ∈Z(cid:8)γN (ζ) ∈ Θbω(ζ),γ∞(ζ) − θ⋆ > ǫ/2(cid:9) +
+ P (γN (ζ) − γ∞(ζ) ≤ ǫ/2) .
Notice that, because of the continuity of γ∞, there exists ǫ > 0 such that γ∞(ζ)−θ⋆ >
ǫ/2 ⇒ ζ − θ > ǫ. We can then use Lemma B.3,
P[ζ∈Z(cid:8)γN (ζ) ∈ Θbω(ζ),γ∞(ζ) − θ⋆ > ǫ/2(cid:9)
≤ ZP(cid:0)γN (ζ) ∈ Θbω(ζ),γN (ζ) − θ⋆ > ǫ(cid:1) ≤ cN 5 qN
where c and q are those coming from Lemma B.3 relatively to ǫ. Putting together all
the estimations we have obtained and using Lemma B.1, we finally obtain that there
25
(notice that f does not depend on i). We can compute
i = σ)
2σ2 ds
if σ = β
2σ2 ds
if σ = α
θ−δ e− (s−θ⋆ )2
θ−δ e− (s−θ⋆ )2
=
EdH (bωIA, ω⋆) =
f (θ, σ) = P(bω(θ)i 6= σ ω⋆
1√2πσ2R θ+δ
1 − 1√2πσ2R θ+δ
N Xi
= pEf (bθIA, α) + (1 − p)Ef (bθIA, β).
i
P(bωIA
6= ω⋆
i )
1
N
1
exists χ > 0 such that P (AN (ǫ)) ≤ χ/N 2. Using Borel-Cantelli Lemma and standard
arguments, it follows now that the relation (4.5) hold in an almost surely sense.
It remains to be shown convergence in mean square sense. For this we need to go
the second additive term in the right hand side of (3.10) can be bounded uniformly
back to the form (3.10) of the derivative of L(θ,bω(θ)). The key observation is that
in modulus by some constant C. If we denote ¯γN = N−1Pi yi, this implies that the
function is increasing for θ > ¯γN + β2C and decreasing for θ < ¯γN − β2C. Hence,
necessarily,
ξ − ¯γN ≤ β2C ∀ξ ∈ SN .
(B.9)
0. This, together with (B.9), yields E max
ergodic theorem also E¯γN − θ⋆2 → 0 for N → +∞, the proof is complete.
On the other hand, by the law of large numbers, ¯γN almost surely converges to θ⋆
ξ∈SN ξ − ¯γN converges to
and this implies, by the previous part of the theorem that max
ξ∈SN ξ − ¯γN2 → 0 for N → +∞. Since by the
B.3. Proof of Proposition 4.4. We prove it for bωIA, the other verification
being completely equivalent). If σ ∈ {α, β}, we define
Since f (θ, σ) is a C1 function of θ, we immediately obtain that
and, by Corollary 4.3, this last expression converges to 0, for N → +∞. Hence,
Ef (bθIA, σ) − Ef (θ⋆, σ) ≤ CEbθIA − θ⋆
EdH (bωIA, ω⋆) = pEf (θ⋆, α) + (1 − p)Ef (θ⋆, β).
1
N
Straightforward computation now proves the thesis.
26
|
1710.11460 | 1 | 1710 | 2017-10-31T13:32:29 | Investigating the effect of social groups in uni-directional pedestrian flow | [
"cs.MA",
"physics.soc-ph"
] | The influence of cohesion among members of dyads is investigated in scenarios characterized by uni-directional flow by means of a discrete model: a corridor and the egress from a room with a bottleneck of varying width are simulated. The model manages the dynamics of simulated group members with an adaptive mechanism, balancing the probability of movement according to the dispersion of the group; the cohesion mechanism is calibrated through the parameters $\kappa_c$ and $\delta$. All scenarios are simulated with two procedures: (Proc. 1) population composed of individual pedestrians, in order to validate the simulation model and to provide baseline data; (Proc. 2) population including dyads (50% of the simulated pedestrians), in order to verify their impact. In the corridor scenario, the presence of dyads causes a reduction of the velocities and specific flow at medium-high densities. Egress from a square room with a unique central exit produces results in line with recent studies in the literature, but also shows that the dyads negatively affect the dynamics, leading generally to a slower walking speed and a lower pedestrian flow. Ignoring the presence of dyads would lead to an overestimation of egress flows. | cs.MA | cs |
Investigating the effect of social groups in
uni-directional pedestrian flow
Luca Crociani, Yiping Zeng, Andrea Gorrini, Giuseppe Vizzari, Weiguo Song
Abstract The influence of cohesion among members of dyads is investigated in sce-
narios characterized by uni-directional flow by means of a discrete model: a corridor
and the egress from a room with a bottleneck of varying width are simulated. The
model manages the dynamics of simulated group members with an adaptive mech-
anism, balancing the probability of movement according to the dispersion of the
group; the cohesion mechanism is calibrated through the parameters κc and δ . All
scenarios are simulated with two procedures: (Proc. 1) population composed of indi-
vidual pedestrians, in order to validate the simulation model and to provide baseline
data; (Proc. 2) population including dyads (50% of the simulated pedestrians), in
order to verify their impact. In the corridor scenario, the presence of dyads causes a
reduction of the velocities and specific flow at medium-high densities. Egress from
a square room with a unique central exit produces results in line with recent studies
in the literature, but also shows that the dyads negatively affect the dynamics, lead-
ing generally to a slower walking speed and a lower pedestrian flow. Ignoring the
presence of dyads would lead to an overestimation of egress flows.
1 Introduction
Recently, the topic of social groups has gathered a growing interest from re-
searchers studying the dynamics of pedestrians. One one side, field observations
Luca Crociani · Andrea Gorrini · Giuseppe Vizzari
CSAI research center, University of Milano - Bicocca, Milano, Italy
e-mail: {name.surname}@disco.unimib.it
Yiping Zeng ( ) · Weiguo Song
Sate Key Laboratory of Fire Science, University of Science and Technology of China, Hefei, China
e-mail: [email protected]
Weiguo Song
e-mail: [email protected]
1
2
Crociani, L., Zeng, Y., Gorrini, A., Vizzari, G., Song, W.
(e.g. [4]) have highlighted that a crowd of pedestrians is mostly composed of groups
of up to 4 members (dyads are typically the most frequent) and that people walking
in a group are generally slower than individuals. Controlled experiments involving
groups have been designed to investigate their aggregated effect, and a relevant im-
pact is observed in a corridor setting with bi-directional flow (see, e.g. [3, 2]) and a
bottleneck scenario(see, e.g. [5]). On the modelling side, many works are being pro-
posed for the simulation of group behaviour: for example, the work in [9] proposes
a continuous model for the behavior of groups of 2 or 3 members based on field ob-
servations, while [7] design leader-follower structure of dyads using dynamic floor
fields. In this paper, we present a discrete model considering groups, based on the
work in [1], which is extended and calibrated to fit the data from a controlled ex-
periment [3]. The calibrated model is then used to investigated the aggregated effect
of dyads in uni-directional pedestrian flows at high densities (Sec. 4) and passing
through a bottleneck (Sec. 5).
2 A Model for Group Cohesion
The simulation model here presented is designed to achieve a more realistic
simulation of pedestrian and group dynamics, with particular attention to the shape
of the group. The model is discrete in space and time, and at each time-step of
the simulation agents evaluate cells c of the Moore neighbourhood with the utility
function U(c). This aggregates the components associated to the reproduction of a
particular behaviour by means of a weighted sum:
U(c) =
κgG(c) + κobOb(c) + κsS(c) + κcC(c) + κdD(c) + κovOv(c)
d
(1)
Individual functions model respectively: (i) goal attraction; (ii) obstacle repul-
sion; (iii) keeping distance from other pedestrians; (iv) cohesion with other group
members; (v) direction inertia; (vi) overlapping to avoid gridlock in counter-flow
situations. The first three elements are modelled with the usage of the well-known
floor-field approach to model the base behaviour of pedestrians: movement towards
a target, obstacle avoidance, proxemics with other pedestrians in a repulsive sense.
The function C(c) is introduced to manage the cohesion among group members
and its strength is mainly adjustable by the parameter κc. The function calculates the
level of attractiveness of each neighbour cell, according to the position and velocity
of the other agents of the group. These information are used to estimate their next
positions and then evaluate the attractiveness of each cell according to the cohesion,
as the following:
(cid:34)(cid:32)
η · ∑
a∈G\{a}
C(c) =
(cid:33)
(cid:35)
(cid:107)xa − x•
a(cid:107)−(cid:107)xc − x•
a(cid:107)
G− 1
· 2
− 1
(2)
Investigating the effect of social groups in uni-directional pedestrian flow
3
(a)
(b)
Fig. 1 (a) Example of calculation of C(c) for a 2-members group. Cells ahead both pedestrians
will be the most desirable according to the function. (b) Example of values of the function Balance
for the different calibration weights, configuring different values of δ .
Where η is a normalization factor that, along with numerical values, allows
to translate the cohesion value into the range [−1,1]. G defines the size of the
group and (cid:107)x(cid:107) denotes the 2-norm of a vector x. x•
a describes a prediction of the
next position of the agent a and it represents a small advancement to the model that
allowed to improve the plausibility of the simulated behaviour. The estimation is
calculated using the velocity vector v a: x•
a = x a +v a. The functioning of the function
is exemplified in Fig. 1(a).
By acting on the parameters of U(c) during the simulation it is possible to
describe different states of the same pedestrian in different moments of a single
simulated scenario. This strategy is applied to allow group members to adapt their
behaviour in dense situations or in presence of obstacles. According to this method,
weights κg,κc,κi are varied and possibly inhibited according to an index that de-
scribes the instantaneous dispersion of the group, as shown in Fig. 1(b). This adap-
tation of the behaviour of group members can be calibrated according to the param-
eter δ that, together with κc, will be investigated for the calibration of the model in
the next Section.
After the utility evaluation for all the cells of the neighbourhood, the choice of
action is decided by the probability to move in each cell c as (N is the normalization
factor): P(c) = N · eU(c).
3 Calibration of the Group Behaviour
In order to calibrate the model and reproduce plausible movement of pedes-
trians and groups, data from the first procedure of a controlled experiment study-
ing the movement of dyads [3] have been used. In the experimental procedure,
4
Crociani, L., Zeng, Y., Gorrini, A., Vizzari, G., Song, W.
(a)
(b)
Fig. 2 Example configuration of the first procedure of the controlled experiment in [3] (a) and its
realization with the simulation model (b).
Individual Dyads Population
Experiment
Simulations
1.32
1.308
1.30
1.305
1.31
1.3067
Table 1 Comparison of average speeds [m/s] of pedestrians in the experiment and in the simula-
tions.
a uni-directional pedestrian flow composed of 30 individuals and 24 dyad mem-
bers1 crossed a corridor-like setting of 3m width. Trajectories were automatically
extracted in a measurement area of 10m length located at the center of the corridor,
after a buffer zone of 2m that allowed participants to reach a stable speed.
As shown in Fig. 2, a corridor-like setting similar to the experiment has been
designed with the simulation model, with an analogous initial configuration of the
pedestrian flow: starting positions are given by 9 start areas which generates 6 pedes-
trians each. Members of dyads are generated close to each other to avoid bias given
by the initial configuration. Moreover, only information about pedestrians inside the
measurement area are analysed. Given that all participants were young male student,
we configured a desired speed of simulated pedestrians of 1.6 m/s.
The aim of the simulation campaign was to calibrate parameters δ and κc,
managing the group behaviour, to fit the results about average speeds of all par-
ticipants and the distribution of relative positions of dyads, which describes their
spatial behaviour during the experimental iteration. To gather stable results, a set of
100 simulation iterations is configured for each configuration of the two parameters
that, as pure assumption, is explored in the range [0 : 30] with an increase rate of 1.
The calibration phase led to the optimal configuration of parameters (δ ,κc) =
(7,12), which generates average speeds of singles, dyads and of the overall popu-
lation as shown in Tab. 1. Results are in line with the observation. By looking at
the distribution of relative positions (Fig. 3(a) and 3(b)), the difference is a bit more
marked than with the speed, but overall the results are close and they highlights that
also in the simulations the most frequent pattern was the line-abreast one.
1 Dyads were artificially and randomly formed at the beginning of the iteration by asking partici-
pants to try to walk close to their companion as they were friends.
Investigating the effect of social groups in uni-directional pedestrian flow
5
(a)
(b)
Fig. 3 Distribution of relative positions of dyads (direction of movement towards right) of the
experiment (a) and simulation (b). The same color-scale is applied in both pictures.
4 Validation at basic movement
In this section, we will study the effects of the presence of dyads on the spe-
cific flow in a corridor environment with uni-directional movement. A graphic de-
scription of the setting is shown in Fig. 4. Two case-studies have been configured: a
baseline with only singles and one with 50% presence of dyads. To compute the fun-
damental diagram, simulation campaigns of 10 iterations of 5000 steps each have
been run per each investigated average density of pedestrians in the whole environ-
ment. The environment was configured as toroidal, so that the number of pedestrians
is constant during each iteration. To gain precision with the analysis, data are gen-
erated within a measurement area of 8 m length, where entrance/exit events are
recorded for each agent. This allowed to compute the number and the average speed
of pedestrian inside the measurement area for a given time window and to achieve
larger datasets with every simulation. Finally, at each run we considered only data
generated from step 2000, where a steady state was generally reached.
Results in the form density–speed and density–flow are shown in Fig. 5. Both
datasets are in agreement with empirical guidelines and datasets from the literature.
At the same time, the specific flow of dyads is smaller than that of individuals for
densities higher than 1.5 p/m2. On one hand, dyads walk slower than individuals in
low density situations, but the difference is quite limited with a similar configura-
tion of their desired speed. The increase of density, on the other hand, has a stronger
effects on the average speed of the population with the presence of dyads, empha-
sizing that dyads had a negative effect on the observed dynamics. This is due to the
cohesion mechanism configured for the walking behaviour of dyads, which makes
them to keep short distances to each other and to slow down while congested sit-
uations leads to a fragmentation of the group. While this behaviour is assumed for
this model and it could be considered rather strong compared to the reality, a similar
6
Crociani, L., Zeng, Y., Gorrini, A., Vizzari, G., Song, W.
Fig. 4 The corridor environment used for the test of the fundamental diagram.
Fig. 5 Results of the fundamental diagram in the form density–speed (left) and density–flow
(right).
difference among fundamental diagrams observed with only individuals and with
about 40% of dyads was observed in an experiment also presented in this book [2].
5 Analysis of group influence on a bottleneck scenario
Up to now, the model is calibrated to reproduce a plausible behaviour of dyads
and it is validated for the simulation of uni-directional flow in a corridor. In this
section the model will be used to study the effect of two-members groups in a pe-
riodical scenario representing a room with a bottleneck, as in Fig. 6(a). Again, we
configure two simulation campaigns to study the phenomena: (a) a scenario with
only individuals and (b) one with 50% presence of dyads. 400 simulated pedestri-
ans are generated in the scenario for each configuration of the bottleneck width and
5 iterations of 5000 steps are run to get a sufficient dataset. Again, output data are
gathered only in a steady state of the system (time-step from 2000 to 5000). Fig. 7
exemplifies two density maps for one iteration with bottleneck width of 4.0 m and
Investigating the effect of social groups in uni-directional pedestrian flow
7
(a)
(b)
Fig. 6 (a) Simulated scenario to test the outflow from the bottleneck. (b) Total flow achieved in the
two case studies and two empirical datasets from the literature [6, 8] for comparison.
highlights a slightly higher congestion in the scenario with dyads. This provides an
explanation of the results describing the outflow of pedestrians from the bottleneck,
which is calculated as J = N
t . The corresponding results are shown in Fig. 6(b).
While the baseline scenario generates a specific flow at the bottleneck of about 2.1
ped/m*s, the scenario with dyads provides a sensibly lower result following the
trend of about 1.8 ped/m*s observed in the dataset from Rupprecht et Al. [8].
6 Conclusions
The influence of cohesion among members of dyads and its aggregated effect in
scenarios characterized by uni-directional flow has been investigated by applying a
discrete model to two benchmark scenarios: a one way corridor and the egress from
a room with a central bottleneck. The simulation model represents the dynamics of
simulated group members with an adaptive mechanism, balancing the probability
of movement according to the dispersion of the group; the cohesion mechanism is
calibrated through two parameters κc and δ , acting on the utility of a movement,
and according to data coming from a controlled experiments involving dyads.
The investigated scenarios are simulated with two procedures, configuring a
population composed of only individual pedestrians and one half composed of dyad
members. In accordance with recent studies in the literature, simulated results show
that the dyads negatively affects the dynamics, leading generally to a lower velocity
and a lower pedestrian flow. The difference in the fundamental diagram observed
in the scenario of the corridor is sensible and for the range of densities lower than
1.5 ped/m2 is similar with the empirical data from the experiment used for the cali-
bration (see [2] in this book). Results coming from the simulation of the bottleneck
show again a negative effect of the presence of dyads, which in this case is even
more apparent. The achieved trend and the difference between the two simulation
8
Crociani, L., Zeng, Y., Gorrini, A., Vizzari, G., Song, W.
(a)
(b)
Fig. 7 Cumulative mean density at a steady state for the scenario with only individuals (a) and the
one with 50% presence of dyads (b).
scenarios are in line with other studies from the literature and, overall, suggests fur-
ther research to investigate the microscopic behaviour of pedestrian groups and its
aggregated effects on pedestrian dynamics.
Acknowledgements This research has been supported by the Key Research and Development
Program (2016YFC0802508), the Program of Shanghai Science and Technology Committee
(16DZ1200106), the Specialized Research Fund for the Doctoral Program of Higher Education
of China (20133402110009) the China Scholarship Council (CSC) and Fundamental Research
Funds for the Central Universities (WK2320000035).
References
1. Bandini, S., Crociani, L., Vizzari, G.: An approach for managing heterogeneous speed profiles
in cellular automata pedestrian models. J. Cellular Automata 12(5), 401–421 (2017)
2. Crociani, L., Gorrini, A., Feliciani, C., Vizzari, G., Nishinari, K., Bandini, S.: Micro and macro
pedestrian dynamics in counterflow: the impact of social groups. In: this volume (2018)
3. Gorrini, A., Crociani, L., Feliciani, C., Zhao, P., Nishinari, K., Bandini, S.: Social groups and
pedestrian crowds: Experiment on dyads in a counter flow scenario. In: Proceedings of the 8th
International Conference on Pedestrian and Evacuation Dynamics - PED2016, Hefei, China,
October 17-21, 2016, pp. 179–184 (2016)
4. Gorrini, A., Vizzari, G., Bandini, S.: Age and group-driven pedestrian behaviour: from obser-
vations to simulations. Collective Dynamics 1, 1–16 (2016)
5. von Kruchten, C., Schadschneider, A.: Empirical study on social groups in pedestrian evacua-
tion dynamics. Physica A: Statistical Mechanics and its Applications 475, 129–141 (2017)
6. Liao, W., Seyfried, A., Zhang, J., Boltes, M., Zheng, X., Zhao, Y.: Experimental study on pedes-
trian flow through wide bottleneck. Transportation Research Procedia 2(Supplement C), 26–33
(2014). The Conference on Pedestrian and Evacuation Dynamics 2014 (PED 2014)
7. Muller, F., Schadschneider, A.: Evacuation dynamics of asymmetrically coupled pedestrian
pairs. In: Traffic and Granular Flow'15, pp. 265–272. Springer (2016)
8. Rupprecht, T., Klingsch, W., Seyfried, A.: Influence of Geometry Parameters on Pedestrian
Flow through Bottleneck, pp. 71–80. Springer US (2011)
9. Zanlungo, F., Ikeda, T., Kanda, T.: Potential for the dynamics of pedestrians in a socially inter-
acting group. Phys. Rev. E 89, 012,811 (2014)
|
1907.09597 | 1 | 1907 | 2019-07-22T21:54:44 | Agent Modeling as Auxiliary Task for Deep Reinforcement Learning | [
"cs.MA",
"cs.LG"
] | In this paper we explore how actor-critic methods in deep reinforcement learning, in particular Asynchronous Advantage Actor-Critic (A3C), can be extended with agent modeling. Inspired by recent works on representation learning and multiagent deep reinforcement learning, we propose two architectures to perform agent modeling: the first one based on parameter sharing, and the second one based on agent policy features. Both architectures aim to learn other agents' policies as auxiliary tasks, besides the standard actor (policy) and critic (values). We performed experiments in both cooperative and competitive domains. The former is a problem of coordinated multiagent object transportation and the latter is a two-player mini version of the Pommerman game. Our results show that the proposed architectures stabilize learning and outperform the standard A3C architecture when learning a best response in terms of expected rewards. | cs.MA | cs | Agent Modeling as Auxiliary Task for Deep Reinforcement Learning
Pablo Hernandez-Leal,∗ Bilal Kartal∗ and Matthew E. Taylor
{pablo.hernandez, bilal.kartal, matthew.taylor}@borealisai.com
Borealis AI
Edmonoton, Canada
9
1
0
2
l
u
J
2
2
]
A
M
.
s
c
[
1
v
7
9
5
9
0
.
7
0
9
1
:
v
i
X
r
a
Abstract
In this paper we explore how actor-critic methods in deep
reinforcement learning, in particular Asynchronous Advan-
tage Actor-Critic (A3C), can be extended with agent model-
ing. Inspired by recent works on representation learning and
multiagent deep reinforcement learning, we propose two ar-
chitectures to perform agent modeling: the first one based on
parameter sharing, and the second one based on agent policy
features. Both architectures aim to learn other agents' poli-
cies as auxiliary tasks, besides the standard actor (policy) and
critic (values). We performed experiments in both coopera-
tive and competitive domains. The former is a problem of
coordinated multiagent object transportation and the latter is
a two-player mini version of the Pommerman game. Our re-
sults show that the proposed architectures stabilize learning
and outperform the standard A3C architecture when learning
a best response in terms of expected rewards.
Introduction
An important ability for agents to have is to reason about the
behaviors of other agents by constructing models that make
predictions about the modeled agents (Albrecht and Stone
2018). This agent modeling (Schadd, Bakkes, and Spronck
2007)1 area usually takes concepts and algorithms from mul-
tiagent systems (since the environment includes at least two
agents), game theory (which studies the strategic interac-
tions among agents), and reinforcement learning (since the
model may be based on information observed from interac-
tions).
Agent modeling usually serves two purposes in multia-
gent settings: it improves the coordination efficiency in co-
operative scenarios (Chalkiadakis and Boutilier 2003) and,
in competitive scenarios, it helps the agent to better optimize
(best respond) its actions against the predicted opponent pol-
icy (Carmel and Markovitch 1995), e.g., by exploiting oppo-
nent mistakes.
Early algorithms for agent modeling came from game the-
ory literature, e.g., fictitious play (Brown 1951). Later, many
works adapted reinforcement learning algorithms for this
∗Equal contribution
1Sometimes referred as opponent modelling since "opponent"
is used to refer to another agent in the environment.
task (Banerjee and Peng 2005). Recently, agent modeling
has been also considered in the context of deep reinforce-
ment learning (DRL).
DRL has shown outstanding results in Atari games, Go,
Poker and recently in strategy video games (Mnih et al.
2013; Torrado et al. 2018). Due to these successes, it is
natural that DRL is now being tested in multiagent envi-
ronments (Hernandez-Leal, Kartal, and Taylor 2018). Some
works have explored using DRL to evaluate emergent behav-
iors in multiagent environments (Tampuu et al. 2017), and
others have proposed algorithms for multiagent DRL (Fo-
erster et al. 2017). In contrast, our goal is to estimate other
agents' (opponent or teammate) policies by means of an aux-
iliary task at the same time that the agent is learning its re-
spective (best response) policy. In general, (self-supervised)
auxiliary tasks are not used for anything other than shaping
the features of the agent, i.e., facilitating the representation
learning process (Bellemare et al. 2019), improving learn-
ing stability (Jaderberg et al. 2017), and have broadened the
horizons of RL to learn from all experience, whether re-
warded or not. Self-supervision defines losses via surrogate
annotations that are synthesized from unlabeled inputs. Ex-
amples are reward prediction which can be cast into a regres-
sion task (Jaderberg et al. 2017) and dynamics prediction
that captures state, action, and successor relationships. Since
the purpose is representation learning and not full modeling
of the dynamics and reward, the losses need not form a tran-
sition model and proxies can suffice to help tune the repre-
sentation, i.e., these losses are expected to give gradients not
necessarily a generative model (Shelhamer et al. 2017).
In this work we take advantage of auxiliary tasks when
learning a best response and the opponent/teammate model.
Since these two elements are linked to each other, we pro-
pose two architectures that take advantage of this realization.
Recently, asynchronous actor-critic methods have become
widely used in DRL; Asynchronous Advantage Actor-Critic
(A3C) (Mnih et al. 2016) is a major representative of this
category, which does not use an experience buffer and learns
completely on-policy. Thus, we take A3C as baseline and set
off to evaluate and better understand the use of agent mod-
eling as auxiliary task with on-policy actor-critic methods in
DRL with the following contributions:
• Agent modeling in DRL is still an open research area with
opportunities in video games (Zhao and Szafron 2009;
Torrado et al. 2018; Borovikov et al. 2019). Our exper-
iments are performed in two recent multiagent environ-
ments, one cooperative and one competitive domain (mini
version of the Pommerman game).
• We propose two new architectures that take inspiration
from multiagent DRL and representation learning to do
agent modeling. The first architecture, Agent Modeling
by parameter Sharing (AMS-A3C), takes inspiration from
the concept of parameter sharing to learn the oppo-
nent/teammate policy as an auxiliary task as well as the
standard actor and critic.
• The second architecture, Agent Modeling by policy Fea-
tures (AMF-A3C), leverages the concept of policy fea-
tures to learn latent space features that are used as input
when computing the actor and critic of the learning agent.
Our results show that modeling the opponent/teammate
increases the expected rewards and improves the stability of
the learning process. In particular, in this work we show the
benefits of using opponent/teammate policy prediction as an
auxiliary task with respect to non-learning stochastic agents
in both cooperative and competitive scenarios.
Related Work
In this section, we describe related work on agent modeling
in DRL, multiagent DRL and auxiliary tasks.
Deep Reinforcement Opponent Network (DRON) (He
et al. 2016) was the first DRL work that performed op-
ponent modeling. DRON's idea is to have two networks:
one learns Q values (similar to DQN (Mnih et al. 2013))
and a second learns a representation of the opponent's pol-
icy. DRON used hand-crafted features to define the oppo-
nent network. In contrast, Deep Policy Inference Q-Network
(DPIQN) and Deep Policy Inference Recurrent Q-Network
(DPIRQN) (Hong et al. 2018) learned opponent policy fea-
tures directly from raw observations of the other agents. The
way to learn these policy features is by means of auxiliary
tasks (Jaderberg et al. 2017) that provide additional learn-
ing goals; in this case, the auxiliary task is to learn the op-
ponent's policy. Then, the Q value function of the learning
agent is conditioned on the policy features, which aim to
reduce the non-stationarity of the multiagent environment.
In contrast, our proposals do not need an experience replay
buffer, learn completely on-policy and we make use of full
parameter sharing (Foerster et al. 2017).
Deep Cognitive Hierarchies (Lanctot et al. 2017) is an al-
gorithm that aims to avoid overfitting in two-player games. It
uses deep reinforcement learning to compute best responses
to a distribution over policies and empirical game-theoretic
analysis to compute new meta-strategy distributions. The-
ory of Mind Network (Rabinowitz et al. 2018) tackles the
problem of meta-learning, i.e., the proposed network should
acquire a strong prior model for agents' behaviour to boot-
strap to richer predictions. DeepBPR+ studies the problem
of efficient policy detection and reuse when playing against
non-stationary agents in Markov games (Zheng et al. 2018).
In contrast, our goal is to estimate the opponent/teammate's
policy at the same time that the agent is learning its re-
spective (best response) policy; since these two elements are
linked to each other our proposals improve the stability of
the learning process as well as increase the obtained rewards.
Self Other Modeling (SOM) (Raileanu et al. 2018) is a
recently proposed algorithm that uses the agent's own pol-
icy as a means to predict the opponent's goal (and actions).
SOM is based on the assumption that the agents are identi-
cal, which is more suitable when agents share a fixed set of
goals and have similar abilities.
Auxiliary tasks were originally presented as hints that im-
proved the network performance and learning time. Sud-
darth and Kergosien (1990) presented a minimal example
of a small neural network where it was shown that adding an
auxiliary task effectively removed local minima. Recently,
some works have used them in single-agent RL problems,
for example, Mirowski et al. (2017) studied self-supervised
tasks (like depth prediction) in a navigation problem. Their
results show that augmenting an RL agent with auxiliary
tasks supports representation learning, which provides richer
training signals that enhance data efficiency. Another inter-
esting result is that using the auxiliary task as a loss was
better than using the value as input. Another example was
presented by Lample and Chaplot (2017) who added an aux-
iliary task (to predict game feature information such as the
presence of enemies or items) to a DQN network to improve
learning in First-Person-Shooting games. These ideas also
relate to Multi-Task Learning where by learning tasks in par-
allel using a shared representation, what is learned for each
task can help the learning of the others (Caruana 1997).
Preliminaries
We start with the standard reinforcement learning setting of
an agent interacting in an environment over a discrete num-
ber of steps. At time t the agent in state st takes an action
at and receives a reward rt. The state-value function is the
expected return (sum of discounted rewards) from state s
following a policy π(as):
V π(s) = E[Rt:∞st = s, π],
and the action-value function is the expected return follow-
ing policy π after taking action a from state s:
Qπ(s, a) = E[Rt:∞st = s, at = a, π].
Algorithms, such as Q-learning, or its (deep) neural net-
work variant, DQN, approximate the action-value function
Q(s, a; θ) using parameters θ, and then update parameters
to minimize the mean-squared error, using the loss function:
i ) − Q(s, a; θi))2]
LQ(θi) = E[(r + γmaxa(cid:48)Q(s(cid:48), a(cid:48); θ−
where θ− represents the parameters of the target network
that is held constant, but synchronized to the behaviour net-
work θ− = θ, at certain periods to stabilize learning.
A3C (Asynchronous Advantage Actor-Critic) is an al-
gorithm that employs a parallelized asynchronous train-
ing scheme (e.g., using multiple CPUs) for efficiency; it
is an on-policy RL method that does not need an ex-
perience replay buffer. A3C allows multiple workers to
n−1(cid:88)
simultaneously interact with the environment and com-
pute gradients locally. All the workers pass their com-
puted local gradients to a global network that performs
the optimization and synchronizes the updated actor-critic
NN parameters with the workers asynchronously. A3C
maintains a parameterized policy (actor) π(as; θ) and
value function (critic) V (s; θv), which are updated as fol-
lows: (cid:52)θ = ∇θ log π(atst; θ)A(st, at; θv) and (cid:52)θv =
A(st, at; θv)∇θv V (st) where
A(st, at; θv) =
γkrt+k + γnV (st+n) − V (st),
k=0
with A(s, a) = Q(s, a) − V (s) representing the advantage
function, commonly used to reduce variance.
The policy and the value function are updated after ev-
ery tmax actions or when a terminal state is reached. It
is common to use one softmax output for the policy head
π(atst; θ) and one linear output for the value function head
V (st; θv), with all non-output layers shared, see Figure 1 (a).
The loss function for A3C is composed of two terms:
policy loss (actor), Lπ, and value loss (critic), Lv. An en-
tropy loss for the policy, H(π), is also commonly added to
help to improve exploration by discouraging premature con-
vergence to suboptimal deterministic policies (Mnih et al.
2016). Thus, the loss function is given by:
LA3C = λvLv + λπLπ − λHEs∼π[H(π(s,·, θ)]
with λv = 0.5, λπ = 1.0, and λH = 0.01, being standard
weighting terms on the individual loss components.
The UNsupervised REinforcement and Auxiliary Learn-
ing (UNREAL) framework (Jaderberg et al. 2017) is built
on top of A3C. UNREAL proposes unsupervised auxiliary
tasks to speed up the learning process that requires no ad-
ditional feedback from the environment. The idea of ad-
ditional auxiliary predictions is to help with the represen-
tational learning problem (Bengio, Courville, and Vincent
2013) and had also been useful to improve the robustness
and stability of the learning process (Jaderberg et al. 2017).
UNREAL proposes two auxiliary tasks: auxiliary control
and auxiliary prediction that share the previous layers that
the base agent uses to act. By using this jointly learned repre-
sentation, the base agent learns to optimize extrinsic reward
much faster and, in many cases, achieves better policies at
the end of training. The UNREAL algorithm optimizes a
single combined loss function with respect to the joint pa-
rameters of the agent that combines the A3C loss, LA3C, to-
gether with an auxiliary control loss LP C, an auxiliary re-
ward prediction loss LRP and a replayed value loss LV R. In
contrast to A3C, UNREAL uses an experience replay buffer
that is sampled with more priority given to interactions with
positive rewards to improve the critic network.
Agent Modeling with A3C
In this section we first describe the challenges of opponent
modeling in the context of reinforcement learning and mul-
tiagent systems, then we present our two main contributions:
the AMS-A3C and AMF-A3C architectures.
Opponent modeling and multiagent systems
In a multiagent environment, agents interact at the same time
with the environment and potentially with each other (Tuyls
and Weiss 2012). These environments are commonly for-
malized as a Markov game (cid:104)S,N , A, T, R(cid:105), which can be
seen as an extension of an MDP to multiple agents (Littman
1994). One key distinction is that the transition, T , and re-
ward function, R, depend on the actions of all, N , agents.
Given a learning agent i and using the common shorthand
notation −i = N \ {i} for the set of opponents, the value
function now depends on the joint action a = (ai, a−i), and
the joint policy π(s, a) =(cid:81)
j πj(s, aj):
(cid:88)
a∈A
V π
i (s) =
(cid:88)
T (s, ai, a−i, s(cid:48))
π(s, a)
[R(s, ai, a−i, s(cid:48)) + γVi(s(cid:48))].
s(cid:48)∈S
(1)
The optimal policy is dependant on the other agents' poli-
cies:
π∗
i (s, ai, π−i) = arg max
πi
V (πi,π−i)
i
(s).
However, if the other agents' policies are stationary (can
still be stochastic) then the problem can be reduced to a stan-
dard MDP where RL algorithms can be used to effectively
learn a best response to those other agents, irrespective if the
domain is cooperative or competitive. Our goal therefore is
to accurately estimate the opponent/teammate policy at the
same time that the agent is learning its respective (best re-
sponse) policy. Since these two elements are linked to each
other, below we propose two architectures that take advan-
tage of this realization. In this work we show advantage of
agent policy prediction with respect to non-learning agents.
We leave as future work how to deal with learning agents.
AMS-A3C: Agent Modeling by parameter Sharing
This architecture builds on the concepts of parameter shar-
ing and auxiliary tasks. Parameter sharing has been pro-
posed in multiagent DRL as a way to reduce the number
of parameters to learn and improve the performance. The
idea is to perform centralized learning where agents share
the same network (i.e., parameters) but the outputs represent
different agent actions (Foerster et al. 2017).
Building on the same principle, in our architecture we
want to also predict the opponent/teammate policies as well
as the standard actor and critic, with the key characteris-
tic that the previous layers will share all their parameters,
see Figure 1 (b). The change in the architecture is accompa-
nied by a modification in the loss function. In this case, we
treat the learning of the other agents' policies as auxiliary
tasks (Jaderberg et al. 2017) by refining the loss function as:
LAMS-A3C = LA3C +
1
N
λAMiLAMi
where and λAMi is weight term and LAMi is an auxiliary
loss for opponent/teammate i:
LAMi = − 1
M
i log(aj
aj
i )
M(cid:88)
j
N(cid:88)
i
Figure 1: CNN represents convolutional layers, FC represents fully connected layers, and ⊗ represents an element-wise vector
multiplication. (a) A3C outputs values and the agent's policy. (b) AMS-A3C is similar to A3C but adds a head that predicts the
other agents' policies. (c) AMF-A3C aims to learn other agents' policy features in the latent space, hoppi, which are then used
to compute the value and policy of the learning agent. Both approaches can be generalized to N opponents/teammates.
which is the supervised cross entropy loss between the ob-
served one-hot encoded opponent/teammate action (ground
truth), aj
i , for a trajectory of length
M.
i , and the prediction, aj
AMF-A3C: Agent Modeling by policy Features
The second architecture uses the concepts of policy features
and auxiliary tasks. Hong et al. (2018) proposed a modified
DQN architecture that conditions Q-values of the learning
agent on features in the latent space2 that also predict the
opponent/teammate policy, i.e., policy features.
We take a similar approach but in the context of on-policy
actor-critic methods, which means that policy features con-
dition on both actor and critic. In this case, after the convo-
lutional layers, the fully connected layers are divided in two
sections, one specialized in the opponent/teammate policy
and the other in the actor and critic (of the learning agent).
Then, we directly use opponent/teammate policy features,
hoppi vector, to be conditioned (via an element-wise mul-
tiplication) when computing the actor and critic, see Fig-
ure 1 (c). The loss function is similarly refined as follows:
LAMF-A3C = LA3C + 1N
Note that we described AMS-A3C and AMF-A3C in the
general case with N agents in the environment. In the ex-
periments we evaluated scenarios with one opponent or one
teammate.
(cid:80)N
i λAMiLAMi
Implementation details
For A3C, AMS-A3C, and AMF-A3C we used 3 or 4 con-
volutional layers (depending on the domain), with 32 fil-
ters, 3 × 3 kernels, stride and padding of 1. For A3C and
AMS-A3C the convolutional layers are followed with 2 fully
connected layers with 128 hidden units each, followed by
2-heads: the critic has a single output for state-value esti-
mate for the observation, and the actor has A outputs for
the policy probabilities for the given observation. For AMF-
A3C, the fully connected layers have 64 units (to keep the
same number of weights as AMS-A3C). For AMS-A3C and
AMF-A3C, the opponent/teammate policy head has Aopp
outputs corresponding to the opponent/teammate policy. We
used ELU activation functions. The parameters of all archi-
tectures have entropy weight of 0.01, a value loss weight of
2Not to be confused with latent variables.
(a)
(b)
Figure 2: (a) The coordinated multiagent object transport
moving problem (Palmer et al. 2018). Two agents need to
coordinate to pick up an object and delivery it to the goal
zone. Our experiments use a stochastic teammate that moves
with higher probability towards the object and then to the
goal. (b) An example of the mini Pommerman with board
size 8×8. The board is randomly generated varying the num-
ber of wood (light brown cell), walls (dark brown cell) and
power-ups (yellow circle). Initial positions of the agents are
randomized close to any of the 4 corners of the board.
0.5, a policy loss weight of 1, and a discount factor of 0.99.
The parameters of the learning agent's policy are optimized
using Adam with lr = 0.0001, β1 = 0.9, β2 = 0.999, =
1× 10−8, and weight decay 1× 10−5. In the next section we
compare different settings for λAM .
Experiments
This section describes the two experimental domains: a co-
operative multiagent transport moving problem and a com-
petitive mini version of Pommerman. We then present the
experimental results in terms of sensitivity of the loss weight
parameter λAM for AMS-A3C and AMF-A3C in the coor-
dination domain, and then we compare with A3C in terms
of rewards for the two domains.
Domains and setup
Coordination This domain is
nated Multi-Agent Object Transportation
(CMOTPs) (Palmer et al. 2018),
inspired by Coordi-
Problems
in which two agents
are tasked with delivering one object to a goal within a
grid-world. The agents must locate and pick up the object
by standing in the grid cells on the left and right hand
side. The task is fully cooperative, i.e., objects can only
be transported upon both agents grasping the item (this
happens automatically when situated next to the object) and
choosing to move in the same direction. Agents only receive
a positive reward after placing the object in the goal, see
Figure 2(a). Agents have 1900 time steps to complete this
task, otherwise the object is reset to the starting position.
The actions available to each agent are to either stay in
place or move left, right, up, or down. We tested two
teammates: one hesitant agent which moves randomly but
with higher probability towards the object and once it has
grasped it then moves with higher probability towards the
goal; and a stubborn agent which prefers to follow a certain
path after grasping the object (i.e., some actions are fully
deterministic). Agents receive one observation per time step
from the environment as a 16 × 16 pixel representation. We
used 12 CPU workers in these experiments.
Competition The Pommerman environment (Resnick et
al. 2018) is based on the classic console game Bomber-
man. Our experiments use the simulator in a mode with two
agents, see Figure 2(b). Each agent can execute one of 6
actions at every time step: move in any of four directions,
stay put, or place a bomb. Each cell on the board can be a
passage, a rigid wall, or wood. The maps are generated ran-
domly, albeit there is always a guaranteed path between any
two agents. The winner of the game is the last agent stand-
ing and receives a reward of 1. Whenever an agent places a
bomb it explodes after 10 time steps, producing flames that
have a lifetime of 2 time steps. Flames destroy wood and kill
any agents within their blast radius. When wood is destroyed
either a passage or a power-up is revealed. Power-ups can be
of three types: increase the blast radius of bombs, increase
the number of bombs the agent can place, or give the ability
to kick bombs. A single game of two-player Pommerman is
finished when an agent dies or when reaching 800 timesteps.
We considered the rule-based opponent baseline that
comes within the simulator (a.k.a. SimpleAgent). Its be-
haviour is stochastic since it collects power-ups and places
bombs when it is near an opponent. It is skilled in avoid-
ing blasts from bombs. It uses Dijkstra's algorithm on each
time-step, resulting in longer training times.
We evaluated our two proposed architectures and the stan-
dard A3C against the opponents mentioned above. In all
cases we provided learning agents with dense rewards and
we did not tune those reward terms. In our setting the entire
board is visible and agents receive one observation per time
step from the environment as a 18× 8× 8 matrix which con-
tains the current time step board description of the board for
the current time step, similar to Resnick et al. (2019).
Results
Sensitivity of λAM In the first set of experiments
we used the coordination domain to evaluate different
weights for the opponent modeling loss value: annealing
(a)
(c)
(b)
(d)
Figure 3: Comparison for the weight for the opponent mod-
eling loss value, λAM , annealing 1.0 → 0.0 with varying
discount rates (exponentially) or fixing the value. Learning
curves in the coordination domain with the hesitant team-
mate for (a) AMS-A3C and (b) AMF-A3C: no significant
variation; with the stubborn teammate for (c) AMS-A3C and
(d) AMF-A3C: best results were obtained with λAM = 0.1
λAM = 1.0 → 0.0 varying discount rates exponentially
{0.999, 0.9999, 0.99999} or keeping the value fixed with
λAM = {0.1, 0.5}. With the hesitant teammate both AMS-
A3C and AMF-A3C show similar behavior for all the eval-
uated parameters (better than A3C), see Figures 3(a)-(b).
When testing with the stubborn teammate we observed more
variation among parameters, for both AMS-A3C and AMS-
A3C using a fixed λAM = 0.1 or quickly annealing with
0.999 gave the best results, see Figures 3(c)-(d). Our hypoth-
esis is that this teammate is easier to learn and the network
does not need too much weight on their modeling; instead it
can focus on policy learning.
Coordination Using the best parameters for AMS-A3C
and AMF-A3C we compare to A3C. Figure 4 depicts learn-
ing curves3 (average with standard deviations of 10 runs)
where it can be seen that in the first part of the learning
(30k episodes), all learning agents behave similarly, how-
ever, in the long run AMS-A3C and AMF-A3C obtained
higher rewards than A3C (AMS-A3C was statistically sig-
nificantly better than A3C from episode 60k, α = 0.05).
We noted that against the hesitant teammate A3C decreases
its rewards, likely because of its on-policy nature, see Fig-
ure 4(a). In contrast, AMS-A3C and AMF-A3C show stabil-
ity and start increasing their rewards. When facing the stub-
born teammate, AMS-A3C and AMF-A3C show less vari-
3Because of the stochasticity of the opponent actions an upper
bound of the expected reward is ≈ 0.7 (experimentally computed)
with the selected parameters.
(a)
(b)
(a)
(b)
Figure 4: Coordination domain: Learning curves with two
different teammates (a) hesitant and (b) stubborn in the co-
ordination problem. Vanilla A3C shows instability and even
reduces its rewards after some episodes, in contrast, AMS-
A3C and AMF-A3C are more stable, with lower variance
and higher rewards.
Figure 6: T-SNE analysis from trained (a) A3C and (b)
AMS-A3C agents from 100 episodes (colors are obtained
from the value head). AMS-A3C t-SNE shows many more
clusters, in particular, the cluster highlighted on the left cor-
responds to states when AMS-A3C is about to win the game
(value close to 1).
the opponent commit suicide by blocking its moves (in Pom-
merman, if two agents simultaneously want to move to the
same cell, they both stay in their current locations) and make
it stand on the path of the flames, in contrast to A3C which
was unable to learn this strategy and obtained lower rewards.
Lastly, we performed a visual analysis similar to Zahavy,
Ben-Zrihem and Mannor (2016). We took trained agents of
A3C and AMS-A3C and for 100 episodes we recorded both
the activations of the last layer and the value output. We ap-
plied t-SNE (Maaten and Hinton 2008) on the activations
data (as input) and the value outputs (as labels). Figure 6
depicts the t-SNEs where it can be seen that AMS-A3C
has more well-defined clusters than A3C's, in particular the
highlighted cluster on the left represents states when AMS-
A3C is about to win the game because it can accurately pre-
dict the opponent's moves, which implies values close to 1.
Conclusions
Deep reinforcement learning has shown outstanding results
in recent years. However, there are still many open ques-
tions regarding different recent learning algorithms. We take
as base a major representative of actor-critic methods, i.e.,
A3C, and propose two architectures that are designed to do
agent modelling as an auxiliary task. This means that at the
same time the network improves the representation learn-
ing, it will also aim to learn other agents policies. Even
though auxiliary tasks are not new, their use in deep RL
and opponent modeling is still not well studied. Our work
serves as an important stepping stone in this direction by
proposing two architectures that improve learning when do-
ing opponent/teammate modeling in deep RL. Our archi-
tectures AMS-A3C and AMF-A3C are inspired by multi-
agent DRL concepts: parameter sharing and opponent pol-
icy features. We experimented in both cooperative and com-
petitive domains. In the former, our proposals were able to
learn coordination faster and more robustly compared to the
vanilla A3C. In the latter, our agents were able to predict
opponent moves in complex simultaneous move, Pommer-
man, and successfully obtain a best response that resulted in
higher scores in terms of rewards. As future work, we are
Figure 5: Competition domain: Moving average over 10k
games of the rewards (shaded lines are standard deviation
over 5 runs) obtained by the two proposed architectures and
A3C against the rule-based opponent: AMS-A3C and AMF-
A3C obtained significantly higher scores than A3C.
ance than A3C due to their accurate agent modeling (AMS-
A3C is statistically significant over A3C from episode 90k
with α = 0.05), see Figure 4(b). Examining the trained
agents, AMS-A3C and AMF-A3C show better coordination
skills once the object is grasped compared to vanilla A3C,
i.e., agents reached the goal faster once grabbing the object.
Competition One clear distinction from the previous do-
main is that it is more elaborate and stochastic (board is
randomized and changes depending on the agents' actions).
In this experiments we set λAM = 0.01 and we evaluate
against the rule-based opponent. In this case, we let the
learning agents train for 6 million episodes to guarantee con-
vergence (≈ 3 days of training with 50 workers). Results are
depicted in Figure 5 (with standard deviations over 5 runs),
where it can be seen that AMS-A3C and AMF-A3C both
clearly outperform A3C in terms of rewards (AMS-A3C is
statistically significant over A3C from episode 3.5M and
AMF-A3C from 5.5M, α = 0.05). When observing the poli-
cies generated we noted that during game play the agents
trained with AMS-A3C and AMF-A3C were able to make
interested in exploring self-play, learning agents, and mixed
(coordination-competition) environments.
References
2013.
[Albrecht and Stone 2018] Albrecht, S. V., and Stone, P.
2018. Autonomous agents modelling other agents: A com-
prehensive survey and open problems. Artificial Intelligence
258:66 -- 95.
[Banerjee and Peng 2005] Banerjee, B., and Peng, J. 2005.
In AAMAS,
Efficient learning of multi-step best response.
60 -- 66.
[Bellemare et al. 2019] Bellemare, M. G.; Dabney, W.;
Dadashi, R.; Taiga, A. A.; Castro, P. S.; Roux, N. L.; Schu-
urmans, D.; Lattimore, T.; and Lyle, C. 2019. A geomet-
ric perspective on optimal representations for reinforcement
learning. arXiv preprint arXiv:1901.11530.
Y.;
[Bengio, Courville, and Vincent 2013] Bengio,
Representation
Courville, A.; and Vincent, P.
IEEE trans-
learning: A review and new perspectives.
actions on pattern analysis and machine intelligence
35(8):1798 -- 1828.
[Borovikov et al. 2019] Borovikov, I.; Zhao, Y.; Beirami, A.;
Harder, J.; Kolen, J.; Pestrak, J.; Pinto, J.; Pourabolghasem,
R.; Chaput, H.; Sardari, M.; et al. 2019. Winning isn't ev-
erything: Training agents to playtest modern games. In AAAI
Workshop on Reinforcement Learning in Games.
[Brown 1951] Brown, G. W. 1951.
Iterative solution of
games by fictitious play. Activity analysis of production and
allocation 13(1):374 -- 376.
[Carmel and Markovitch 1995] Carmel, D., and Markovitch,
S. 1995. Opponent Modeling in Multi-Agent Systems. In
IJCAI. Springer-Verlag.
[Caruana 1997] Caruana, R. 1997. Multitask learning. Ma-
chine learning 28(1):41 -- 75.
[Chalkiadakis and Boutilier 2003] Chalkiadakis, G.,
and
Boutilier, C. 2003. Coordination in Multiagent Reinforce-
ment Learning: A Bayesian Approach. In AAMAS.
[Foerster et al. 2017] Foerster, J. N.; Nardelli, N.; Farquhar,
G.; Afouras, T.; Torr, P. H. S.; Kohli, P.; and Whiteson, S.
2017. Stabilising Experience Replay for Deep Multi-Agent
Reinforcement Learning. In ICML.
[He et al. 2016] He, H.; Boyd-Graber, J.; Kwok, K.; and
Daume, H. 2016. Opponent modeling in deep reinforce-
ment learning. In ICML, 2675 -- 2684.
[Hernandez-Leal, Kartal, and Taylor 2018] Hernandez-Leal,
P.; Kartal, B.; and Taylor, M. E. 2018. Is multiagent deep
reinforcement learning the answer or the question? A brief
survey. arXiv preprint arXiv:1810.05587.
[Hong et al. 2018] Hong, Z.-W.; Su, S.-Y.; Shann, T.-Y.;
Chang, Y.-H.; and Lee, C.-Y. 2018. A Deep Policy Infer-
ence Q-Network for Multi-Agent Systems. In AAMAS.
[Jaderberg et al. 2017] Jaderberg, M.; Mnih, V.; Czarnecki,
W. M.; Schaul, T.; Leibo, J. Z.; Silver, D.; and Kavukcuoglu,
K. 2017. Reinforcement Learning with Unsupervised Aux-
iliary Tasks. In ICLR.
[Lample and Chaplot 2017] Lample, G., and Chaplot, D. S.
2017. Playing FPS Games with Deep Reinforcement Learn-
ing. In AAAI, 2140 -- 2146.
[Lanctot et al. 2017] Lanctot, M.; Zambaldi, V. F.; Gruslys,
A.; Lazaridou, A.; Tuyls, K.; P´erolat, J.; Silver, D.; and
Graepel, T. 2017. A Unified Game-Theoretic Approach to
Multiagent Reinforcement Learning. In NIPS.
[Littman 1994] Littman, M. L. 1994. Markov games as a
framework for multi-agent reinforcement learning. In ICML,
157 -- 163.
[Maaten and Hinton 2008] Maaten, L. v. d., and Hinton, G.
2008. Visualizing data using t-SNE. Journal of Machine
Learning Research 9(Nov).
[Mirowski et al. 2017] Mirowski, P.; Pascanu, R.; Viola, F.;
Soyer, H.; Ballard, A. J.; Banino, A.; Denil, M.; Goroshin,
R.; Sifre, L.; Kavukcuoglu, K.; et al. 2017. Learning to
navigate in complex environments. ICLR.
[Mnih et al. 2013] Mnih, V.; Kavukcuoglu, K.; Silver, D.;
Graves, A.; Antonoglou, I.; Wierstra, D.; and Riedmiller,
M. 2013. Playing Atari with Deep Reinforcement Learn-
ing. arXiv preprint arXiv:1312.5602v1.
[Mnih et al. 2016] Mnih, V.; Badia, A. P.; Mirza, M.; Graves,
A.; Lillicrap, T.; Harley, T.; Silver, D.; and Kavukcuoglu,
K. 2016. Asynchronous methods for deep reinforcement
learning. In ICML, 1928 -- 1937.
[Palmer et al. 2018] Palmer, G.; Tuyls, K.; Bloembergen, D.;
and Savani, R. 2018. Lenient Multi-Agent Deep Reinforce-
ment Learning. In AAMAS.
[Rabinowitz et al. 2018] Rabinowitz, N. C.; Perbet, F.; Song,
H. F.; Zhang, C.; Eslami, S. M. A.; and Botvinick, M. 2018.
Machine Theory of Mind. In ICML.
[Raileanu et al. 2018] Raileanu, R.; Denton, E.; Szlam, A.;
and Fergus, R. 2018. Modeling Others using Oneself in
Multi-Agent Reinforcement Learning. In ICML.
[Resnick et al. 2018] Resnick, C.; Eldridge, W.; Ha, D.;
Britz, D.; Foerster, J.; Togelius, J.; Cho, K.; and Bruna,
J. 2018. Pommerman: A multi-agent playground. AIIDE
Multi-Agent Workshop.
[Resnick et al. 2019] Resnick, C.; Raileanu, R.; Kapoor, S.;
Peysakhovich, A.; Cho, K.; and Bruna, J. 2019. Backplay:"
man muss immer umkehren". AAAI-19 Workshop on Rein-
forcement Learning in Games.
[Schadd, Bakkes, and Spronck 2007] Schadd, F.; Bakkes, S.;
and Spronck, P. 2007. Opponent modeling in real-time strat-
egy games. In GAMEON, 61 -- 70.
[Shelhamer et al. 2017] Shelhamer, E.; Mahmoudieh, P.; Ar-
gus, M.; and Darrell, T. 2017. Loss is its own reward: Self-
supervision for reinforcement learning. ICLR workshops.
[Suddarth and Kergosien 1990] Suddarth, S. C., and Ker-
gosien, Y. 1990. Rule-injection hints as a means of im-
proving network performance and learning time. In Neural
Networks. Springer. 120 -- 129.
[Tampuu et al. 2017] Tampuu, A.; Matiisen, T.; Kodelja, D.;
Kuzovkin, I.; Korjus, K.; Aru, J.; Aru, J.; and Vicente, R.
2017. Multiagent cooperation and competition with deep
reinforcement learning. PLOS ONE 12(4):e0172395.
[Torrado et al. 2018] Torrado, R. R.; Bontrager, P.; Togelius,
J.; Liu, J.; and Perez-Liebana, D. 2018. Deep Reinforce-
ment Learning for General Video Game AI. arXiv preprint
arXiv:1806.02448.
[Tuyls and Weiss 2012] Tuyls, K., and Weiss, G. 2012. Mul-
tiagent learning: Basics, challenges, and prospects. AI Mag-
azine 33(3):41 -- 52.
[Zahavy, Ben-Zrihem, and Mannor 2016] Zahavy, T.; Ben-
Zrihem, N.; and Mannor, S. 2016. Graying the black box:
Understanding DQNs. In ICML.
[Zhao and Szafron 2009] Zhao, R., and Szafron, D. 2009.
Learning character behaviors using agent modeling in
games. In AIIDE.
[Zheng et al. 2018] Zheng, Y.; Meng, Z.; Hao, J.; Zhang, Z.;
Yang, T.; and Fan, C. 2018. A deep bayesian policy reuse
approach against non-stationary agents. In NeurIPS. 962 --
972.
|
1804.04202 | 2 | 1804 | 2018-06-18T13:52:12 | Controlling Swarms: A Programming Paradigm with Minimalistic Communication | [
"cs.MA"
] | Inspired by natural swarms, numerous control schemes enabling robotic swarms, mobile sensor networks and other multi-agent systems to exhibit various self-organized behaviors have been suggested. In this work, we present a Wave Oriented Swarm Programming Paradigm (WOSPP) enabling the control of swarms with minimalistic communication bandwidth in a simple manner, yet allowing the emergence of diverse complex behaviors and autonomy of the swarm. Communi cation in the proposed paradigm is based on "ping"-signals inspired by strategies for communication and self organization of slime mold (dictyostelium discoideum) and fireflies (lampyridae). Signals propagate as information-waves throughout the swarm. We show that even with 1-bit bandwidth communication between agents suffices for the design of a substantial set of behaviors in the domain of essential behaviors of a collective. Ultimately, the reader will be enabled to develop and design a control scheme for individual swarms. | cs.MA | cs |
Controlling Swarms: A Programming Paradigm with
Minimalistic Communication
Joshua Cherian Varughese* · Hannes
Hornischer* · Payam Zahadat · Ronald
Thenius · Franz Wotawa · Thomas
Schmickl
Abstract Inspired by natural swarms, numerous control schemes enabling robotic
swarms, mobile sensor networks and other multi-agent systems to exhibit various
self-organized behaviors have been suggested. In this work, we present a Wave Ori-
ented Swarm Programming Paradigm (WOSPP) enabling the control of swarms
with minimalistic communication bandwidth in a simple manner, yet allowing the
emergence of diverse complex behaviors and autonomy of the swarm. Communi-
cation in the proposed paradigm is based on "ping"-signals inspired by strategies
for communication and self organization of slime mold (dictyostelium discoideum)
and fireflies (lampyridae). Signals propagate as information-waves throughout the
swarm. We show that even with 1-bit bandwidth communication between agents
suffices for the design of a substantial set of behaviors in the domain of essential
behaviors of a collective. Ultimately, the reader will be enabled to develop and
design a control scheme for individual swarms.
Keywords Swarm control · Programming paradigm · Bio-inspired · Self-
organization · Agent based modeling · Mobile sensor network
1 Introduction
Nature teems with various kinds of life forms with varying individual capabilities.
Many of these lifeforms have been found to depend not only on individual capa-
bilities but especially on emergent group dynamics. Moving around as a swarm
of individuals permits hunting together, foraging more efficiently, sharing food or
* these authors contributed equally to this work; order of their listed names was chosen ran-
domly
Joshua Cherian Varughese, Hannes Hornischer, Ronald Thenius and Thomas Schmickl
Institut fur Biologie
Karl Franzens Universitat Graz
E-mail: [email protected]
Joshua Cherian Varughese and Franz Wowtawa
Institut fur Softwaretechnologie
Technische Universitat Graz
E-mail: [email protected]
2
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
collaborative defense against predators to increase their collective probability of
survival and reproduction [14]. Fish increases their individual probability of sur-
vival by moving in schools and collectively performing escape maneuvers when a
predator is detected by one of the fish [5, 16]. The foraging efficiency of a hive
increases dramatically when bees perform waggle dances [43, 35] to inform other
bees about food sources in the vicinity. The synchronized blinking of fireflies [6, 7]
and the aggregation of slime mold cells to form a slug [13, 36] are other examples
of lifeforms relying on collective abilities for foraging and reproduction.
Various kinds of lifeforms with varying physical and cognitive complexity evolved
to perform decentralized behaviors in order to ensure greater probability of sur-
vival. The part each individual has to play in such emergent behaviors is often
simple, yet the result that emerges on a group level is resilient to failures of indi-
viduals or other disruptive events [23]. Numerous natural swarms following simple
behaviors of its individuals are scalable and can therefore consist of hundreds,
thousands or even millions of individual entities [20, 37].
The simplicity and decentralized nature of the individual's behaviors produc-
ing resilient collective phenomena have attracted interest from engineers seeking
to inherit properties of simplicity, resilience, scalability in engineered systems. Sev-
eral optimization algorithms relying on emergent behaviors have been proposed
and successfully implemented[34, 19, 44]. "Particle swarm optimization" [19] is an
optimization algorithm, inspired by flocks of birds, which is being used for multi
objective optimization. BEECLUST [34] is an optima finding algorithm, inspired
by a swarm of newly hatched bees, that can be used to find global optima. Nu-
merous algorithms [32, 21, 17, 45] enabling simple agents to accomplish complex
tasks such as source localization [17], task allocation [21, 45], collective mapping
et cetera have been developed. In [32, 12], swarm intelligent algorithms are pro-
posed that enable robots with very limited individual abilities to transport large
objects in a collective manner. In [17], a distributed algorithm for localizing the
source of an odor in an environment is proposed and tested on a swarm of robots.
Apart from enabling a group of simple robots to perform complex tasks, vari-
ous algorithms and methods have been suggested to control a swarm of robots
of varying sizes to perform specific actions and tasks in a coordinated manner
such as arranging themselves in a particular shape or responding collectively to
external cues or stimuli [9, 22, 33]. In this paper a paradigm for controlling a
group of agents is introduced and presented, using communication based on waves
of single bit signals, or "pings" propagating through the swarm. The single bit
pings propagating through the swarm is analogous to scroll waves of cAMP sig-
nals propagating through a swarm of dictyostelium discoideum while the periodic
initiation of ping signals is inspired by lampyridae. WOSPP enables a swarm of
agents with directional communication and locomotive capabilities to synchronize,
elect a leader, estimate the number of members in the swarm, localize the center
of the swarm, aggregate etc. In addition, a meta control scheme that enables a
user to combine individual behaviors to produce more complex swarm behaviors
is presented. In order to put this work into context of the current state of the art,
related algorithms and approaches will be discussed briefly in Section 2.
Generally, in contrast to existing approaches for swarm control, the paradigm
presented in this work allows a swarm to inherit rich self organized collective
behaviors. Instead of complex messages or encoded signals solely one-bit commu-
nication suffices for the presented behaviors and ultimately allows the design of
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
3
both top-down control interfaces and autonomous swarms. The basic communica-
tion behavior of the paradigm presented in this paper has already been used for
producing gradient taxis[42]. Its basic concept as well as several behaviors have
been explored in [39].
In this paper, a substantial set of behaviors in the domain of essential behaviors
of a collective is presented and an extensive literature review is conducted in order
to place the presented paradigm into perspective of existing work. For all presented
behaviors the detailed structure and its design is described giving the reader an
intuitive understanding of how to develop and design basic behavioral building
blocks, which will be referred to as "primitives".
Since each individual application of swarm control requires a specific set of be-
havioral abilities, depending on the swarm members' abilities, their environmental
conditions and their tasks the primitives have to be constructed or adjusted in
a fitting manner. Primitives serve as a basis for a meta control scheme and can
be combined in various ways in order to produce complex collective behaviors as
shown in Figure 1. Ultimately, the reader will be enabled to design primitives and
thus construct a meta control scheme for a swarm.
In Section 2, a literature review of relevant research is presented. Subsequently,
in Section 3 the bioinspiration for the communication mechanism from slime
mold and fireflies is presented. In Section 4, the communication mechanism of
the paradigm presented in this paper is introduced as well as the fundamental
concept of the paradigm. In Section 5 a set of primitives is introduced, classified
into three categories: internal organization, swarm awareness and locomotion. In
Section 6 possible methods for combining primitives and resulting complex behav-
iors are presented1.
Fig. 1 Schematic illustration of a meta control scheme. Every agent in WOSPP follows a fixed
basic operational structure. Behavioral changes are only introduced through minor changes in
two designated "codeBlocks", as presented later. This structure allows the coherent design of
fundamental behaviors or "primitives" and thus the development of a meta control interface
for controlling the swarm.
1 All primitives presented in this paper have been simulated using Python 3.6.3
4
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
2 Related Research
In the field of controlling groups of entities, there is already a variety of approaches
and previous work with diverse foci.
In [22], a classical approach to controlling a swarm is employed where users can
interact with and control a swarm of certain robots, referred to as "zooids", using
gestures. The authors achieved a responsive swarm using an external projector for
tracking the robots position, assigning a goal position for each individual robot
and then utilizing classical motion control strategies like Proportional-Integral-
Derivative (PID) control to move the agents precisely from the start position to
the goal position. Where as the programmer has control of the precise movements
of the swarms members, the presence of a higher organizational entity is inherently
necessary and thus depicts a classical example of top down control. There is a sub-
stantial amount of work related to controlling groups of entities. In the pioneering
work of Craig Reynolds, he introduced self propelled particles known as boids [30]
which exhibit self organized flocking and collision avoidance. Reynold's boids are
able to mimic a flock of birds whose individuals follow simple behavioral rules.
Due to its simplicity and decentralized structure it is applicable to large swarms.
Its focus is on the generation of realistic behaviors as found in natural flocks and
hence limited in its versatility.
In [26] a concept of self propelled particles with internal oscillators, or "swar-
mallators" is introduced. Attractive and repulsive forces are then used in relation
to the relative phase shift of the oscillators for generating a range of collective
behaviors. The internal processes and states of the swarms' entities substantially
influence and determine the interplay between individuals producing a small set
of collective phenomena.
In [24], the authors introduce an algorithm for self assembly of identical agents
on a surface into a predetermined global shape. Multiple gradients are developed
by propagating messages starting from the agents at the edges in order to develop a
relative positioning system among the agents. Subsequently, various shapes can be
generated by manipulating the behavior of agents with particular gradient values.
More complex shapes are achieved by repeating the process of generating gradients
and folding along the specific areas of interest. A variation of the aforementioned
idea was used to assemble various shapes in a self organized manner in a thousand
robot swarm [33]. In [1], programmable self assembly and other similar research
done by various researchers were unified as amorphous computing.
Among existing work, the paradigm presented in this paper exhibits most par-
allel characteristics with the approaches presented under amorphous computing,
however significantly differs in several points. Instead of multi-bit signals encoding
information, as used for communication in amorphous computing, the presented
paradigm already produces rich behavioral diversity with single bit communica-
tion. Another key difference between is that the presented paradigm refrains from
using "seed" agents[25] or global knowledge regarding edges and vertices in case of
programmable self assembly for origami generation. Instead agents initiate com-
munication with the rest of the swarm decentralized and randomly. In amorphous
computing, languages such as Origami Shape Language (OSL) [24], Growing Point
Language (GPL) [10] are utilized in order to enable a user to program a swarm.
In this paper, the proposed paradigm will be used to program a swarm to perform
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
5
collective behaviors and additionally, a meta control scheme can be designed for
the swarm to perform these behaviors autonomously.
3 Bioinspiration
The programming paradigm presented in this paper takes inspiration from the
communication mechanisms used by slime mold and fireflies. In the following the
relevant aspects of these lifeforms are briefly discussed.
3.1 Slime mold
Slime mold (dictyostelium discoideum), is a free living diploid life form which takes
advantage of swarming behavior to survive challenging environments. Slime mold
has a widely varying cooperation with other cells during its life cycle depending
on the environment. When there are ample food sources, cells grow and divide
individually without cooperation. In case of scarcity of food or other threats, sig-
nificant cooperation between the cells begin. During its cooperative phases, the
individual cells aggregate to form a multicellular organism. And the collective be-
gins to act similar to a single organism. The paradigm presented in this paper
will chiefly consider the signaling behavior of slime mold during the aggregation
of slime mold cells. For aggregating, some cells (centers) release Cyclic Adeno-
sine Monophosphate (cAMP) into the environment to recruit surrounding cells to
join the aggregate. This signal induces a short-lived chemical concentration spike
around the recruiting cells [36]. Other slime mold cells that perceive the chemical
signal will produce the emit the chemical themselves to relay the signal. Since
all slime mold cells relay the chemical signals they receive, the the original signal
produced by the recruiting center rapidly travels through the swarm. Additionally,
each cell needs around 12-15 seconds [2] in between two cAMP signals. During this
insensitive period, individual cells are insensitive to any cAMP pulses. This inter-
mediate insensitive time acts as a "refractory" period that prevents any echoing
between two amoeba cells. The signal relaying mechanism described above forms
the basis for spatio-temporal patterns known as scroll waves. Since the origin of
the waves are at the recruiting cells, the amoeba can move towards the incoming
signal to reach the recruiting center [36].
3.2 Fireflies
Fireflies (lampyridae) are a family of insects that are capable of producing bio-
luminescence to attract a mate or a prey [6]. Bio-luminescence of various families of
fireflies has been a subject of elaborate study in the past [6]. Apart from being able
to blink, fireflies are known to behave in cooperation with other fireflies in order to
attract mates or prey [6]. Such synchronicity is a result of a simple mechanism by
which initially the individual fireflies blink periodically. When a firefly perceives a
blink in its surrounding, it blinks again and then resets its own blinking frequency
to match the received blink [8]. This is analogous to a phase coupled oscillator
which adjusts its phase to match it to that of the faster one in the vicinity. This
6
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
trait emerges into a quasi synchronized blinking pattern while the frequency of
blinking will be influenced by the fastest blinking insect.
4 WOSPP - Wave Oriented Swarm Programming Paradigm
The wave oriented swarm programming paradigm WOSPP is strongly inspired by
the two aforementioned organisms slime mold and fireflies.
In particular, communication within the paradigm is based on waves propagat-
ing through a swarm. Every agent within a swarm has the ability of sending and
receiving information signals, which we will refer to as "pings", to nearby agents.
Connected to this, all agents can enter three different states, parallel to the be-
havior of slime mold: An inactive state in which agents are receptive to incoming
signals (responsive to cAMP or "pings"), an active state where they send or relay
a signal (release cAMP or send "ping") and optionally perform an action, which is
followed by a refractory state where agents are temporarily insensitive to incoming
signals. This is schematically shown in Figure 2(a).
(a)
(b)
Fig. 2 (a): Three states of agents in WOSPP: From inactive state, through external trigger
(e.g. incoming ping) or internal trigger (timer) the agent transitions into the active state where
it sends a ping and optionally performs an action. Afterwards, it enters the refractory state,
being insensitive to incoming signal until transitioning into the inactive state again. (b):The
conceptual operating structure of an individual agent. If an agent in the inactive state receives
a ping, it relays the signal by entering the active state and sending a ping itself. Specific to
the primitive a certain code will be executed and the refractory state is entered. If an agent's
internal time is up it will initiate a ping following the same structure, however executing a
different code (specific to the primitive) in the active state.
This operational structure results in a wave like propagation of signals through-
out the swarm. Agents in the inactive state get triggered to relay a signal, while the
refractory state prevents the system from continuously signaling and thus flooding
the system. In Figure 4 the propagation of waves is shown for a swarm of agents,
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
7
each agent represented by a dot with the color denoting their state. The inactive
state is denoted in black, the active state in red and the refractory state in green.
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 3 Illustration of wave based communication. In (a) almost all agents are in the inactive
state, shown in black, except one agent which broadcasts a ping, i.e. enters the active state,
shown in red. It afterwards transitions into the refractory state, shown in green. Neighboring
agents receive the signal and themselves transition into the active state as shown in (b) and
(c). The ping signal spreads in a wave like manner. In (d) the initiating agent transitions from
the refractory state into the inactive state again. Due to a fixed duration of the refractory
state, the transition to the inactive state as well spreads in a wave like manner, shown in (e)
and (f). Times [s]: (a) 0, (b) 2, (c) 4, (d) 11, (e) 16, (f) 19. Parameters (as defined in more
detail at the end of Section 4): number of agents N = 80, physical size of the swarm in units
perception range Rs = 5 r, refractory time in units timesteps tref = 5 s.
The ability of fireflies to adjust and reset their individual frequency of "blink-
ing" is ground to a concept used in this paradigm: internal timers. Every agent
has an internal timer which, when running out, will trigger the agent to enter the
"active" state where it broadcasts a ping, thus initiating a ping wave through the
swarm. For most primitives, this timer is reset right after running out, causing an
agent to repeatedly count down and, subsequently, ping.
This communication structure and the internal timers constitute the basic and
fixed structure of an agent in WOSPP and is shown as pseudo code in Algorithm 1.
This basic structure suffices for the behavior shown in Figure 4 and is fixed for
all agents. However, as will be presented later in this work, complex behavior can
be induced by adding simple command in the codeBlocks. When agents perform
simple actions when relaying or initiating pings, complex behavior can emerge.
This structure is conceptually shown in a flowchart in Figure 2(b). The three
8
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
states are marked, as well as two separate operational chains, one for relaying, one
for initiating pings. Both incorporate an optional field for executing the respective
codeBlocks.
For the behaviors, or primitives, presented in this work, agents are not only
able to send and receive pings, but also optionally have a heading and sense of
directionality. Agents are enabled to determine the direction from which they
receive a ping. Furthermore, for some tasks agents have the ability to move in
direction of their heading. Primitives in which the agents move requires agents to
move sufficiently slow in order to avoid single agents being left behind or splitting
the swarm up into smaller groups. The agents are considered as point particles
and thus collision detection is ignored within this work since it is highly specific
for each individual swarm and its environmental conditions. Aside from primitives
shown in this paper a large variety of behaviors can be developed, incorporating
abilities and environmental sensors of agents not discussed here. This work gives
a fundamental introduction into the concept for enabling a modular design and
development of primitives for swarm control tailored to specific conditions and
tasks of a swarm.
Regarding communication bandwidth, for the individual primitives presented
in this work exclusively single-bit communication is considered in order to show
that even for the minimum communication case remarkable behaviors can be pro-
duced. The use of multi-bit communication instead of pings would further increase
the possibilities for swarm control. Multi-bit communication is discussed in Sec-
tion 6 and in the discussion.
For the simulations presented in this work only swarms with agents spatially
distributed in two dimensions in an approximately circular shape were used. This
is chosen as a general shape without loss of generality, it does not limit or affect
functionality of WOSPP regarding shape or spatial dimension of swarms.
In the following parameters and quantities used in this work are introduced
and defined.
– In the numerical simulations presented here, time is measured in timesteps s.
An agent receiving a ping will be activated and itself send a ping one timestep
tactivate = 1 s later.
p
– Every agent has an internal timer tp which usually periodically resets to a
. However, for some primitives however the timer can be reset
maximum tmax
to a random number between tp ∈ (0, tmax
– The number of agents constituting a swarm is defined as N . The minimum N
necessary for the presented primitives to function is N = 2. The maximum can
theoretically be arbitrarily large, as further elaborated in the discussion.
p
].
– The refractory time tref denotes the time an agent remains in the refractory
state, i.e. insensitive to incoming pings. For the presented simulations tref is
set to be larger than the time a signal would take to propagate at the edge
of the swarm in a circular manner. This is for preventing a ping wave from
continuously propagating through the swarm.
– For distances the basic unit is perception range of agents r, the distance up to
which an agent will perceive the pinging of a nearby agent.
– For primitives including locomotion, agents take discrete spatial steps within
a timestep, where the length of their step d is set to one-sixth of a perception
range d = r/6.
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
9
– The physical size of the swarm is defined as Rs and will be given in units
perception range r.
Data: Paradigm parameters
Result: -
state ← inactive;
timer(tp) ← random integer ∈ (0, tmax
while primitive do
p
];
decrement timer(tp);
if agent in refractory state then
wait for refractory time;
if refractory time is over then
state ← inactive
end
end
if agent in active state then
broadcast ping;
state ← refractory
end
if agent in inactive state then
listen for incoming pings;
if ping received then
state ← active;
execute Relay-CodeBlock;
end
end
if timer(tp) ≤ 0 then
state ← active;
execute Initiate-CodeBlock;
end
end
Function Initiate-CodeBlock
-
Function Relay-CodeBlock
-
Algorithm 1: Basic pseudo code for every individual agent within WOSPP.
This structure is fixed, behavioral changes are only introduced through adding
commands to the initiate- and relay-codeBlocks, here highlighted. The timer tp
is initially set to a random value with upper limit tmax
.
p
5 Primitives
In this section a set of primitives is presented where small changes in the code-
Blocks produce large scale complex behavior. For every primitive plots of results
for intuitively visualizing the behavior are presented, as well as the code-block.
The presented primitives are divided into three categories:
10
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
1. Internal organization is about self organization of the swarm on an internal
level of each agent, including the primitives "leader election", "synchroniza-
tion" and "localize object".
2. Swarm awareness includes the individuals awareness about properties of the
swarm or properties of itself within the swarm. Presented are the primitives
"localize swarm center", "estimating number of swarm members" and "esti-
mate individual position within swarm".
3. The category locomotion is about physically self organizing or restructuring
the swarm, including the primitives "aggregation", "moving collectively" and
"gas expansion".
In the following, if not stated otherwise, every swarm is initially randomly
distributed within a circular area of radius Rs such that every agent is connected
to at least one neighbor.
5.1 Internal organization: Leader election
For various tasks it can be beneficial or even necessary for a swarm to have a cer-
tain agent "leading" a swarm. Having a certain agent assigned as a special entity
brings the risk of having this agent removed from a swarm and thus disabling the
entire swarm. Instead the swarm can collectively elect a leader thus eliminating
such risks. For deciding for a leader, initially all agents consider themselves po-
tential leaders, shown in Figure 4(a) in black. An agent pinging is illustrated in
red and an agent not considering itself a leader anymore is green. Every agent
sets its timer to a random number within tp ∈ (0, tmax
]. As soon as an agent
receives a ping before its own internal timer ended it will not consider itself a
candidate anymore and also deactivate its internal timer, i.e. not initiate ping-
ing. After an agent initiated a ping it will again randomly choose a time tp for
initiating pinging another time. This is shown as pseudo code in Algorithm 2.
p
Data: Paradigm parameters
Result: Leader election
unction Initiate-CodeBlock
. . F
Function Relay-CodeBlock
candidate ← true;
timer(tp) ← random integer ∈ (0, tmax
deactivate internal timer;
candidate ← false;
p
];
Algorithm 2: Code block for primitive "Leader election"
Figure 4(b) and (c) show agents initiating ping waves and immediately out-
rivaling their surrounding agents. The refractory mode prevents two initiating
agents from outrivaling each other, however more than one can be left as potential
leaders, as shown in Figure 4(d). Since every remaining candidate again chooses
a random time to ping, after few "negotiation cycles" a single candidate, which
then can be considered the leader, will remain, as shown in Figure 4 (e) and (f).
Alternative approaches for leader election in groups e.g. based on voting [18] as
well as decentralized probabilistic methods [4] can be found in the literature.
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
11
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 4 Leader election in a swarm. Candidates are shown in black, pinging agents in red and
agents not considering themselves candidates in green. Initially all agents consider themselves
candidates or potential leaders (a). After receiving and relaying a ping an agent will not
consider itself a candidate anymore. Agents initiating pinging thus outrival agents around
them. In (b)-(d) it is illustrated how several agents initiate pinging and not outrival each other
(due to refractory time). Since only candidates will initiate pinging, the remaining candidates
then repeat the process, indicated in (e) until only a single candidate remains, as shown in (f).
Times [s]: (a) 0, (b) 5, (c) 8, (d) 30, (e) 41, (f) 189. Parameters: N = 200, Rs = 6 r, tref = 10 s,
tmax
p
= 100 s.
5.2 Internal organization: Synchronization
For a swarm being able to perform coordinated actions it often is essential to syn-
chronize regarding their respective tasks. This primitive allows the swarm mem-
bers to synchronize regarding e.g. the sending of pings, allowing to perform actions
quasi-simultaneously.
p
Every agent sets its internal counter to a random value between tp ∈ (0, tmax
].
If an agent receives a ping, it resets its internal counter to tmax
. This is shown as
pseudo code in Algorithm 3. As a result, the first agent sending a ping (which is
then being relayed and propagates through the system wave-like) resets the timers
of all relaying agents to the maximum tmax
. Hence, the entire swarm will ping
quasi-simultaneously within a time period smaller or equal to the duration of a
ping propagating from one end of the swarm to the other. In Figure 5(a) the syn-
chronization process for a swarm of N = 15 agents is shown via an order parameter
∆φmax which decreases with increasing synchronicity within the swarm. ∆φmax
is calculated by determining the smallest phase interval containing the timers of
all agents and then taking the maximum phase difference of all timer pairs. At
p
p
12
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
t = 30 s the onset of synchronization is indicated with a gray vertical line. In
Figure 5(b) the corresponding internal timers of all agents are shown, incremen-
tally decreasing with time. Every line of points represents the internal timer of
one agent. At t = 30 s an agent initiates pinging and thus resets the timers of all
other agents such that at t = 35 s all agents reset and thus synchronized. Alter-
native approaches to synchronization in swarms employs pulse coupled oscillators
inspired by fireflies [27, 28].
Data: Paradigm parameters
Result: Synchronized Swarm
unction Initiate-CodeBlock
timer(tp)← tmax
timer(tp)← tmax
Function Relay-CodeBlock
p
p
Algorithm 3: Code block for primitive "Synchronization". The entire pseudo
code for each agent is shown. The highlighted part is the pseudo code which
differs for every primitive. For all other primitives presented here only the latter
part will be shown.
;
;
. . F
(a)
(b)
Fig. 5 (a): Onset of synchronized internal timers at t ≈ 30 of a swarm of N = 15 agents,
indicated with a gray vertical line. The order parameter ∆φmax is plotted against simulation
time t. ∆φmax is calculated by determining the smallest phase interval containing the timers
of all agents and then taking the maximum phase difference between two timers. After fully
synchronizing at t ≈ 36 s the maximum phase difference decreased from ∆φmax ≈ 4.4 rad
to ∆φmax ≈ 0.25 rad which corresponds to a time interval of ∆t ≈ 4 s. This interval can be
identified in (b), where the internal timers of the agents versus simulation time is shown. Every
line of data points corresponds to the internal timer of a single agent, which incrementally
counts down. All timers gradually decrease until at t = 30 s an agent's timer reaches tp = 0
and thus initiates pinging. This marks the onset of the synchronization process and is marked
with a gray vertical line. All agents relaying the ping then reset their timers. The reset signal
propagates through the swarm and at t = 35 s all agents reset and collectively count down in
a quasi-synchronous manner, that is, within a time interval of ∆t = 4 s. Parameters: N = 15,
Rs = 1.67 r, tref = 20 s, tmax
= 100 s.
p
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
13
5.3 Internal organization: Localize object
For distributing information about spatial structure of the surrounding, a swarm
needs to be able to communicate the location of nearby objects or events among
its members. This primitive enables a swarm to collectively localize the direction
of a direct path towards an object which one or few members of the swarm detect.
Each agent refrains from initiating pinging unless it itself detects an object. Every
agent receiving a ping, records the direction of the incoming ping. An estimate of
the direction towards the object is then obtained by taking a running average of
the directions of incoming pings. The pseudo code is shown in Algorithm 4.
Figure 6 shows the agents estimate of the rough location of the object as arrow,
placed at the position of the agent within the swarm. The red square represents an
object which can only be detected by agents in its vicinity. Figure 6 (a) shows the
initial (random) orientation of the agents. With increasing number of perceived
pings, the estimate of direction towards the object increases in accuracy until
agents accurately point towards the position of the object as shown in Figure 6
(b).
Data: Paradigm parameters
Result: Agent knows rough direction of an object
state ← inactive;
end
Function Relay-CodeBlock
record ping direction;
current estimate ← average ping directions
Algorithm 4: Code block for primitive "Localize object"
(a)
(b)
];
unction Initiate-CodeBlock
timer(tp) ← random integer ∈ (0, tmax
if no object detected then
p
. . F
Fig. 6 The arrows representing the agents estimation of the location of the object, the begin-
ning of the arrow denotes the position af an agent. (a): The estimate is initialized to random
direction at the start of the simulation. (b): The converged estimation of the location of
the object after one tmax
. All agents now point towards the object. Parameters: N = 100,
Rs = 3.34 r, tref = 5 s, tmax
= 2500 s.
p
p
14
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
5.4 Swarm awareness: Localize swarm center
For a swarm being able to execute spatially coordinated actions the knowledge
of the individual about the location of the center of the swarm can be of great
advantage. This primitive enables each swarm member to identify the direction
from where most signals originate from, which will be referred to as Average Origin
of Pings, or AOP. For a swarm of the presented type (circular, approximately
homogeneously distributed, agents have several communication neighbors) this
direction coincides with the direction towards the physical center of the swarm.
Each agent sets its internal counter tp to a random value between tp ∈ (0, tmax
],
as soon as a counter reaches tp = 0 an agent will send a ping. When an agent
receives a ping it stores the direction of the incoming ping and averages over all
stored directions. This is shown as pseudo code in Algorithm 5.
p
Data: Paradigm parameters
Result: Agent knows rough direction of swarm center
unction Initiate-CodeBlock
. . F
Function Relay-CodeBlock
center estimate ← mean of previous estimates;
timer(tp) ← random integer ∈ (0, tmax
record ping direction;
current estimate ← average ping directions
Algorithm 5: Code block for primitive "Localize swarm center"
];
p
After one cycle every agent once initiated a ping wave, unless a ping was
initiated while all surrounding agents were in refractory state and thus the ping
not relayed. Figure 7 shows a swarm in its initial state and after it equilibrated
where every agent's orientation is denoted by an arrow at the position of the agent
in the swarm. Initially the heading is random. After equilibrating, the agents on
the outside accurately point towards the center of the swarm.
(a)
(b)
Fig. 7 Agents estimation of the direction towards the center of the swarm. The beginning of
an arrow denotes the position of an agent. (a) shows the initial estimates of each agent as arrow
at its position in the swarm. (b) shows the converged estimates after t = 4 tmax
. Parameters:
N = 100, Rs = 3.34 r, tref = 5 s, tmax
= 2500 s
p
p
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
15
5.5 Swarm awareness: Estimating number of swarm members
For some tasks a swarm may needs to be constituted of a certain number of agents
in order to effectively operate. Or a swarm may needs check if the number of its
members substantially changed, due to loss of members or merging with another
swarm. This problem can be solved without an external observer as the swarm can
estimate its number of swarm members autonomously. Each agent sets its internal
counter tp to a random value between tp ∈ (0, tmax
]. Whenever an internal timer
is up, an agent will initiate pinging and randomly reset its timer to tp ∈ (0, tmax
].
Each time an agent receives a ping it relays the signal and increments a counter
Ncount.
p
p
Furthermore, every time an agent initiates pinging, i.e. one internal cycle has
passed, it will store its counter Ncount as its estimate of the number of swarm
members for the past cycle. The average over those estimates will be the agent's
opinion of the number of members in the swarm Nest. This is shown as pseudo
code in Algorithm 6.
Data: Paradigm parameters
Result: Agent knows approximate number of members in the swarm
unction Initiate-CodeBlock
. . F
estimate(Nest) ← mean of previous swarm size counters;
counter(Ncount) ← 0;
timer(tp) ← random integer ∈ (0, tmax
increment counter Ncount;
Function Relay-CodeBlock
];
p
Algorithm 6: Code block for primitive "Estimating number of swarm members"
In Figure 8 the estimate Nest averaged over all members of a swarm versus
simulation time is shown. The estimate quickly increases before slowly converging
to Nest ≈ 34. The error bars represent the standard deviation thus indicating
that the estimates of all agents are closely distributed around the mean. For the
data in Figure 8, a swarm of 50 agents was building estimates over a time of
20 · tmax
= 20 · 1000 s, so 2 · 104 timesteps. The estimate converges to a value
significantly lower than the actual number of swarm members however, for the
same swarm the estimate consistently converges to the same (lower) estimate.
p
16
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
Fig. 8 Estimated number of swarm members averaged over all agents in the swarm versus
time. The error bars represent the standard deviation. The estimate steeply increases from
Nest = 0 to Nest = 30 at t = 3 tmax
before it gradually converges to its final estimate of
Nest ≈ 34. Parameters: N = 50, Rs = 3.34 r, tref = 5 s, tmax
= 1000 s.
p
p
Figure 9 shows the percentage deviation Nerr from the actual number of swarm
. For every
, sufficiently long for the estimate to
members N depending on the maximum possible cycle length tmax
data point the simulation was run for 25tmax
converge.
p
p
Fig. 9 Percentage deviation of estimated number of agents in the swarm versus maximum
cycle length and the actual number of agents in the swarm. Agents consistently underestimate
the number of members of the swarm. The deviation decreases for decreasing N and increasing
tmax
∈ 100, 200, .., 1500 s.
p
. Parameters used: N ∈ {10, 20, .., 100}, Rs = 1 r, tref = 5 s, tmax
p
With increasing tmax
the deviation from the actual number of members of the
swarm decreases. Knowing the order of magnitude of N of a swarm, the tmax
can
be chosen sufficiently large such that the deviation sufficiently small. For instance,
p
p
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
17
considering a swarm of a maximum of 30 agents of the presented kind, a maximum
cycle length of tmax
> 1500 would ensure a deviation of around Ndev = 10 %.
p
However, a general corrective function for determining the approximate sys-
tematic deviation from the actual number of agents in the swarm can be developed,
though this will not be further discussed in this work.
5.6 Swarm awareness: Estimate individual position within swarm
Some actions require the agents in a swarm to determine their relative position
within the swarm. Considering for example a swarm protecting itself from environ-
mental hazards, requiring only the agents at the edge or outer shell of the swarm
to take measures, it is necessary for each individual agent to learn about their
approximate relative position. For this, each agent sets its internal timer tp to a
random value between tp ∈ (0, tmax
]. As previously explained, this will result in
agents randomly pinging at random time slots and each agent relays the received
pings to nearby agents. The agents then bin each of the pings received into four
directions of α = 90◦ each. If there is at least one empty bin with no pings re-
ceived, then the agents perceives itself as being on the periphery of the swarm.
Pseudo code is shown in Algorithm 7.
p
Data: Paradigm parameters
Result: Agent knows if it is at the periphery
if Is at least one bin empty? then
unction Initiate-CodeBlock
. . F
else
periphery ← true;
periphery ← false;
Function Relay-CodeBlock
end
timer(tp) ← random integer ∈ (0, tmax
record ping direction;
bin incoming ping directions into bins of 90◦;
p
];
Algorithm 7: Code block for primitive "Localize object"
Figure 10 shows the perception of agents regarding their position in the swarm.
Initially, no agents perceive if they are at the periphery of the swarm, denoted
through the black color of agents in Figure 10(a). As agents receive more pings
from the surrounding agents, they are able to have more accurate estimate of its
own position within the swarm as shown in Figure 10(b) where red colored agents
perceive that they are at the periphery of the swarm.
18
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
(a)
(b)
Fig. 10 Figures show the agents perception of their location within the swarm. Red colored
agents perceive that they are on the periphery and the black colored agents perceive them not
being on the periphery. (a), shows the initialization at the start of the simulation with all agents
perceiving themselves as "not being at the periphery". (b) shows the converged perception of
the agents after 10 tmax
. Parameters: N = 100, Rs = 4 r, tref = 5 s, tmax
p
p
= 2500 s
Another approach, enabling agents to estimate their distance from the average
origin of pings is to let agents collect the number of incoming pings per direc-
tional bin and then merge the four bins into two in a way, such that two bins of
α = 180◦ arise, one with the minimum count of received pings, the other half with
the maximum. With increasing number of members in a swarm the accuracy of this
approach increases. A large variety of approaches for localization within a swarm
can be found in the literature. Some of the established methods use Kalman Filter-
ing [31], Monte Carlo Localization [15] or localization based on local information
[11].
5.7 Locomotion: Aggregation
Considering a swarm of agents with the ability to move and spatially arrange itself,
for regrouping it needs to be able to gather or aggregate. For that every agent
randomly sets its internal counter tp to a random value between tp ∈ (0, tmax
].
An agent receiving a ping will, after relaying it, move a small distance towards the
incoming ping. This way, gradually all agents move towards each other. Pseudo
code is shown in Algorithm 8.
p
Figure 11 shows a swarm aggregating in such manner. From its initial state the
swarm steadily moves towards its average origin of pings, causing it to aggregate
at the center of the swarm. Figure 11(d) shows the aggregated state of the agents
as well as each agents trajectory as blue line. This illustrates how agents tend
to follow the paths of their fellow members of the swarm, producing a root-like
trajectory structure.
For illustrating the aggregation process, Figure 12 shows in blue the average
root mean square distance of all agents to the center, i.e. the average position of
all agents at that time. In red the same quantity averaged over 20 independent
simulations is shown.
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
19
Data: Paradigm parameters
Result: Aggregated swarm
Algorithm 8: Code block for primitive "Aggregation"
(a)
(b)
(c)
(d)
Fig. 11 Aggregation of the swarm. Initial state of the swarm is shown in (a), in (b) and (c)
it steadily aggregates. The final state and trajectories of each agent for the entire simulation
as blue lines are shown in (d). Times [tmax
]: (a) 0, (b) 0.7, (c) 1.4, (d) 2. Parameters: N = 80,
Rs = 2 r, tref = 10 s, tmax
= 500 s.
p
p
Fig. 12 Average root mean square distance of all agents from the center of the swarm plot-
ted against time. The blue graph shows the Rrms of the swarm in the simulation shown in
Figure 11. The Rrms linearly decreases until tmax
≈ 1 when the swarm almost fully aggre-
gated. If looked at in close, the linear decrease occurs in quasi-discrete steps, corresponding to
ping waves causing all agents to move towards each other quasi-simultaneously. Parameters:
N = 80, Rs = 2 r, tref = 10 s, tmax
= 500 s.
p
p
timer(tp) ← random integer ∈ (0, tmax
timer(tp) ← tmax
record ping direction;
Calculate average of incoming pings;
move towards incoming ping;
p
p
];
. . F
unction Initiate-CodeBlock
Function Relay-CodeBlock
;
20
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
Considering, a swarm needs to aggregate at a specific location, e.g. for measur-
ing or exploring a certain area or object, the primitive can be changed such that
only certain agents, which for example perceive stimuli such as the presence of an
object, are able to initiate pinging. This is shown in Figure 13. The stimulus can
also be an event or can be connected with a gradient. Considering agents with the
ability to perceive e.g. light intensity, the agents will be able to aggregate at the
brightest spot if every agent sets its internal counter to a value proportional to its
perceived brightness. The agents at the brightest spots will statistically ping first.
Furthermore, every agent receiving a ping will reset its counter thus allowing the
agents at the brightest spot to hijack the swarm. This process executed repeatedly
will result in a gradient taxis behavior as presented in [42]. It is worth noting that
various approaches for aggregation in swarms have been developed and presented
based on differing mechanisms and varying levels of complexity [40, 3, 34].
(a)
(b)
(c)
(d)
Fig. 13 A swarm aggregating at an object, marked as red square on right hand side of the
shown system. Initial state of the swarm is shown in (a). The swarm gradually aggregates at
the object in (b) and (c) until every agent directly perceives the object in (d). Also shown
are the trajectories for the entire simulation as blue lines. Times [tmax
]: (a) 0, (b) 10, (c) 20,
(d) 64. make the object square and bigger to make it more different from agents.
Parameters: N = 80, Rs = 3.34 r, tref = 15 s, tmax
= 50 s.
p
p
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
21
5.8 Locomotion: Moving collectively
For the mobility of a swarm the ability to collectively move to a specific location
can be crucial. For letting the entire swarm move towards a certain direction, a
single agent serves as leader. Exclusively this leader initiates pings and gradually
moves along a trajectory leading to the target location. All agents receiving pings
will move towards the direction of it and thus follow the leader. The pseudo code
is shown in Algorithm 9.
Data: Paradigm parameters
Result: Swarm follows a leader
unction Initiate-CodeBlock
. . F
Function Relay-CodeBlock
p
leader ← true;
timer(tp) ← random integer ∈ (0, tmax
deactivate timer;
leader ← false;
record ping direction;
calculate average of incoming pings;
move towards incoming ping;
];
Algorithm 9: Code block for primitive "Moving collectively"
Figure 14 shows a swarm aggregating towards a leader located at the far right
end of the swarm,which steadily moves towards the right. While following the
leader, the remaining swarm forms a line behind it, being lead away. This prim-
itive can be viewed as "aggregation at a specific, moving agent". For choosing a
leading agent, the primitive "leader election", which was earlier introduced, can
be executed prior to this primitive.
22
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
(a)
(b)
(c)
(d)
Fig. 14 This figure shows a swarm being lead by a single agent towards the right. From the
initial state in (a), the swarm aggregates towards the leading agent in (b) and (c). In (d)
the swarm formed a line following the leader. Also shown in blue are the trajectories of each
agent for the entire simulation. Times [tmax
: (a) 1, (b) 4, (c) 12, (d) 40. Parameters: N = 80,
Rs = 3.34 r, tref = 5 s, tmax
= 50.
p
p
5.9 Locomotion: Gas expansion
Considering a swarm exploring its environment, it can maximize its covered sur-
rounding area by physically expanding. The primitive "gas expansion" enables a
swarm to uniformly expand. Each agent sets its internal counter tp to a random
value between tp ∈ (0, tmax
]. As soon as the internal counter reaches tp = 0 an
agent sends a ping. Each agent moves away a small step from incoming pings. As
soon as an agent receives no pings anymore, it does not move further away. See
Algorithm 10 for pseudo code.
p
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
23
Data: Paradigm parameters
Result: Expanded swarm
Algorithm 10: Code block for primitive "Gas expansion"
In Figure 15 (a) an initially densely packed swarm is shown, which then expands
in Figure 15 (b) and (c) until it is fully expanded in Figure 15 (d). Agents can then
reconnect with its swarm members by moving back, in the opposite direction of
the previous step, or by integrating its entire trajectory and thus finding their way
back until they again perceive signals. Depending on the communication abilities
of the swarm, the perception range or sensitivity can be temporarily decreased
during the expansion such that afterwards the agents will again be connected.
(a)
(b)
(c)
(d)
Fig. 15 A swarm performing the primitive "gas expansion". The initial state of the swarm is
shown in (a), where it is fully aggregated. In (b) and (c) it gradually expands. The final state
is shown in (d) with the trajectories of each swarm as blue lines. Times [timesteps]: (a) 0, (b)
20, (c) 60, (d) 280. Parameters: N = 80, Rs = 0.67 r, tref = 5 s, tmax
= 50 s.
p
Function Relay-CodeBlock
unction Initiate-CodeBlock
timer(tp) ← random integer ∈ (0, tmax
timer(tp) ← tmax
record ping direction;
calculate average of incoming pings;
move away from incoming ping;
p
];
p
;
. . F
24
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
6 Combining primitives
All previously presented primitives are based on single-bit communication for
demonstrating that even with minimum communication bandwidth complex be-
haviors can be produced. Using multi-bit signals the spectrum of possible primi-
tives can be extended significantly. Examples and possibilities for multi-bit com-
munication based systems and behaviors are presented in the discussion.
However, let alone by combining primitives not only more elaborate behaviors
can be produced, by enabling a swarm to switch between a set of primitives it
can operate autonomously. The most intuitive way of combining primitives is to
execute primitives one after another in a sequential manner. This allows the design
of complex tasks which can be executed by the swarm autonomously. This is
schematically shown in Figure 16 (a).
(a)
(b)
Fig. 16 Schematic illustration of (a): sequential execution and (b): interleaved execution of
three primitives A, B and C.
An example for sequential execution of primitives producing an autonomously
acting swarm is a collective exploration procedure, shown in Figure 17. The fol-
lowing sequence of primitives is executed periodically: aggregation, leader election,
moving collectively, gas expansion. In Figure 17(a)-(b) the swarm aggregates and
then determines a leader in Figure 17(c)). This leader will choose a random di-
rection and lead the swarm to a new location, as shown in Figure 17(c) to (d).
Then the entire swarm expands and explores the area and for example, collects
data before again aggregating and restarting this procedure. Due to the limited
abilities of the individual members of the swarm, they have no awareness of the
collective state or if the execution of a primitve was completed. For this example,
the execution times of all primitives were fixed or "hard-coded".
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
25
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 17 Consecutive execution of the primitives Aggregation, Leader election, Moving collec-
tively and Gas expansion as example for an exploring routine of an autonomous swarm. The
swarm prepares for changing its location and thus aggregates from (a) to (c). It then decides
on a leading agent (marked in red) which then leads the swarm towards the top right of the
system, a target area, shown in (c) and (d). The swarm expands again for exploring the new
environment, shown in (e). In (f) the final state along with the trajectories of all agents over
the entire simulation are shown. Parameters: N = 50, tref and tmax
vary for each primitive.
p
Another approach to combining primitives is to execute several primitives in
an interleaved manner. This allows the emergence of a larger variety of complex
behaviors. This is schematically shown in Figure 16 (b). For executing several
primitives in a quasi-simultaneous manner, the previously presented single bit
communication can be extended. A simple option is to introduce several individual
layers of single bit communication, one for each primitive. Alternatively, multi bit
signals could be used, every signal encoding the primitive it is associated with.
The demonstrations shown here are only exemplary and many more combina-
tions of primitives are possible. In case an observer controls a swarm, autonomy
of the swarm is not necessary and full control through manual choice of primitives
can be exerted.
7 Discussion
In Section 5, a set of primitives which can be utilized and combined as basic
building blocks for a meta control scheme for a swarm is presented, covering the
categories "internal organization", "swarm awareness" and "locomotion". Two ex-
emplary realizations of combination of previously presented primitives for complex
26
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
collective behaviors are presented in Section 6. Through this paper, it is demon-
strated that WOSPP enables swarms consisting of agents with limited abilities to
collectively perform a large variety of complex behaviors using "scroll wave" based
communication. Due to its simple and flexible fundamental concept, this program-
ming paradigm is applicable to a large spectrum of different types of swarms and
environments while requiring minimal communication abilities. A cross section of
possible abilities of WOSPP is presented in Section 5 where both primitives and
basic analysis of their behaviors is shown. In this section WOSPP as a whole
including its advantages, scope for further work and application in robotics is
discussed.
In the introduction, several areas dealing with swarm control, such as amor-
phous computing were briefly examined. As pointed out, there are significant dif-
ferences between amorphous computing and WOSPP one of which being that
amorphous computing requires multi-bit communication. In amorphous comput-
ing, multiple gradients are propagated starting from a seed agent. The gradient
is essentially a hop count which enables the internal positioning of agents with
respect to the seed agent which in turn allows a group of robots to organize them-
selves globally. The paradigm presented in this paper functions even using single
bit communication between the agents. Therefore, the extreme case i.e, the re-
striction to single-bit communication, is focused on in this paper. Though many
systems do not have such constraint and could thus make use of complex signals
being transmitted yielding even greater versatility. One such possibility is the use
of so called "hop counts", encoding within the signal how often it was relayed, as
done in [24, 1]. It allows for instance the limitation of the range up to how many
agents or nodes a signal is relayed.
An analysis has been conducted on the resilience of "scroll wave" based com-
munication[41] where its robustness against signal loss was examined. It was shown
that due to redundancy in signal pathways, a system using slime mold based com-
munication, as utilized in WOSPP, can compensate up to 70% individual prob-
ability of signal loss without significant decrease in performance. The ability of
this basic behavior to cope with high amounts of signal loss endows WOSPP with
resilient functioning when pings fail to be sent or received.
As opposed to approaches such as [22], decentralized control in WOSPP allows
scalability limited primarily by the communication abilities relative to operational
time scales. For the class of swarms presented in this paper the main constraint
is constituted by the condition to choose the maximum internal cycle length tmax
significantly larger than the time for a ping wave to propagate from one end
of the swarm to the other. It ensures that ping waves likely propagate through
the entire system without colliding with other waves, thus enabling swarm-wide
communication. By sufficient choice of parameters, the number of swarm members
can be increased almost arbitrarily without loss of functionality.
p
However, considering a swarm encountering frequent collisions of ping waves,
for instance while aggregating a swarm could ultimately split into sub-swarms,
especially at sparsely connected regions in the swarm. Each group that splits out
of the parent swarm retains the properties of the original swarm. Thus, in case
of splitting, each sub-swarm immediately adapts to the new situation, being fully
functional. Such properties of WOSPP can even be used to add to the richness
in collective behavioral diversity. For example, swarms performing a search or
exploration task can deliberately split into sub-swarms and proceed separately.
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
27
One of the prerequisites for a swarm to be able to implement WOSPP is di-
rectional communication similar to most animals in nature which exhibit swarm
behavior. In this paper, most simulations conducted follow the assumption that
agents have the ability to precise detect the direction of incoming pings. In prac-
tice, this requirement can be substantially loosened for the agents to have a lower
angular resolution without significant loss of functionality. In other words, a rough
perception of the direction of the incoming ping is sufficient for the basic func-
tionality of WOSPP. Preliminary work was done but is not shown in this paper
and will be subject of future work and further analysis on WOSPP. Only a selec-
tion of primitives is shown in this work, more ideas for primitives are for instance
the incorporation of statistical regularities in very large swarms. The spectrum of
collective behaviors can be varied and extended much further and applied a vari-
ety of systems, naturally depending on the requirements, tasks and creativity of
the user. Consequently, the paradigm is applicable not only to swarms, but yields
great potential in network system such as sensor networks, internet of things et
cetera due to structural similarities in signal/information propagation throughout
the system. In those examples presumably the main focus will be on "internal
organization" and "swarm awareness" primitives.
Other future work connected to subCULTron[38], a project aiming to deploy
a heterogeneous swarm of underwater robots to monitor environmental param-
eters in the lagoon of Venice. Within the framework of this project, individual
primitives of WOSPP are already being used for swarm control. Robotic systems
such as subCULTron, which employ a large number of individual agents in noisy
environment aiming for autonomous operation, can benefit from WOSPP. In the
future, a WOSPP language will be developed enabling users to combine primitives
in a convenient manner and apply them as control scheme to a swarm of robots.
Alternatively, a programming language for robotic swarms, called "buzz" [29] can
be used for implementing the WOSPP. This will further facilitate the usage and
application of WOSPP and the development of increasingly elaborate primitives,
e.g. involving complex collective decision making, allowing in the future an easier
designing of fully autonomous swarms with the ability to flexibly adapt to varying
environmental conditions.
Acknowledgements This work was supported by EU-H2020 Project no. 640967, subCUL-
Tron, funded by the European Unions Horizon 2020 research and innovation programme.
References
1. Abelson, H., Allen, D., Coore, D., Hanson, C., Homsy, G., Knight Jr, T.F., Nagpal, R.,
Rauch, E., Sussman, G.J., Weiss, R.: Amorphous computing. Communications of the
ACM 43(5), 74–82 (2000)
2. Alcantara, F., Monk, M.: Signal propagation during aggregation in the slime mould dic-
tyostelium discoideum. Microbiology 85(2), 321–334 (1974)
3. Bahge¸ci, E., Sahin, E.: Evolving aggregation behaviors for swarm robotic systems: A sys-
tematic case study. In: Swarm Intelligence Symposium, 2005. SIS 2005. Proceedings 2005
IEEE, pp. 333–340. IEEE (2005)
4. Ben-Shahar, O., Dolev, S., Dolgin, A., Segal, M.: Direction election in flocking swarms.
Ad Hoc Networks 12, 250–258 (2014)
5. Brock, V.E., Riffenburgh, R.H.: Fish schooling: a possible factor in reducing predation.
ICES Journal of Marine Science 25(3), 307–317 (1960)
28
Varughese, Hornischer, Zahadat, Thenius, Wotawa, Schmickl
6. Buck, J., Buck, E.: Biology of synchronous flashing of fireflies. Nature 211, 562–564 (1966)
7. Buck, J., Buck, E.: Mechanism of rhythmic synchronous flashing of fireflies: Fireflies of
southeast asia may use anticipatory time-measuring in synchronizing their flashing. Science
159(3821), 1319–1327 (1968)
8. Camazine, S., Denenbourg, J.L., Franks, N.R., Sneyd, J., Theraulaz, G., Bonabeau, E.:
Synchronized flashing among fireflies. pp. 143–166. Princton University Press, Princton
(2001)
9. Christensen, A.L., Rehan, O., Dorigo, M., et al.: Morphology control in a multirobot
system. IEEE Robotics & Automation Magazine 14(4), 18–25 (2007)
10. Coore, D.: Towards a universal language for amorphous computing.
In: International
Conference on Complex Systems (ICCS2004) (2004)
11. Cornejo, A., Nagpal, R.: Distributed range-based relative localization of robot swarms. In:
Algorithmic Foundations of Robotics XI, pp. 91–107. Springer (2015)
12. Decugniere, A., Poulain, B., Campo, A., Pinciroli, C., Tartini, B., Os´ee, M., Dorigo, M.,
Birattari, M.: Enhancing the cooperative transport of multiple objects. In: International
Conference on Ant Colony Optimization and Swarm Intelligence, pp. 307–314. Springer
(2008)
13. Devreotes, P.: Dictyostelium discoideum: a model system for cell-cell interactions in de-
velopment. Science 245(4922), 1054–1058 (1989)
14. Eberhart, R.C., Shi, Y., Kennedy, J.: Swarm intelligence. Elsevier (2001)
15. Fox, D., Burgard, W., Kruppa, H., Thrun, S.: A probabilistic approach to collaborative
multi-robot localization. Autonomous robots 8(3), 325–344 (2000)
16. Hall, S., Wardle, C., MacLennan, D.: Predator evasion in a fish school: test of a model for
the fountain effect. Marine biology 91(1), 143–148 (1986)
17. Hayes, A.T., Martinoli, A., Goodman, R.M.: Swarm robotic odor localization: Off-line
optimization and validation with real robots. Robotica 21(4), 427–441 (2003)
18. Karpov, V., Karpova, I.: Leader election algorithms for static swarms. Biologically Inspired
Cognitive Architectures 12, 54–64 (2015)
19. Kennedy, J., Eberhart, R.C.: Particle swarm optimization. In: IEEE International Con-
ference on Neural Networks. IEEE Press, Los Alamitos, CA (1995)
20. Kennedy, J.S.: The migration of the desert locust (schistocerca gregaria forsk.) i. the
behaviour of swarms. ii. a theory of long-range migrations. Phil. Trans. R. Soc. Lond. B
235(625), 163–290 (1951)
21. Labella, T.H., Dorigo, M., Deneubourg, J.L.: Division of labor in a group of robots inspired
by ants' foraging behavior. ACM Transactions on Autonomous and Adaptive Systems
(TAAS) 1(1), 4–25 (2006)
22. Le Goc, M., Kim, L.H., Parsaei, A., Fekete, J.D., Dragicevic, P., Follmer, S.: Zooids:
Building blocks for swarm user interfaces. In: Proceedings of the 29th Annual Symposium
on User Interface Software and Technology, pp. 97–109. ACM (2016)
23. Middleton, E.J., Latty, T.: Resilience in social insect infrastructure systems. Journal of
The Royal Society Interface 13(116), 20151022 (2016)
24. Nagpal, R.: Programmable self-assembly using biologically-inspired multiagent control. In:
Proceedings of the first international joint conference on Autonomous agents and multia-
gent systems: part 1, pp. 418–425. ACM (2002)
25. Nagpal, R., Kondacs, A., Chang, C.: Programming methodology for biologically-inspired
self-assembling systems. In: AAAI Spring Symposium on Computational Synthesis, pp.
173–180 (2003)
26. O'Keeffe, K.P., Hong, H., Strogatz, S.H.: Oscillators that sync and swarm. Nature Com-
munications 8(1), 1504 (2017)
27. Perez-Diaz, F., Trenkwalder, S.M., Zillmer, R., Gross, R.: Emergence and inhibition of
synchronization in robot swarms. In: Distributed Autonomous Robotic Systems, pp. 475–
486. Springer (2018)
28. Perez-Diaz, F., Zillmer, R., Gross, R.: Control of synchronization regimes in networks of
mobile interacting agents. Physical Review Applied 7(5), 054002 (2017)
29. Pinciroli, C., Lee-Brown, A., Beltrame, G.: Buzz: An extensible programming language
for self-organizing heterogeneous robot swarms. arXiv preprint arXiv:1507.05946 (2015)
30. Reynolds, C.W.: Steering behaviors for autonomous characters. In: Game developers con-
ference, vol. 1999, pp. 763–782 (1999)
31. Roumeliotis, S.I., Bekey, G.A.: Collective localization: A distributed kalman filter approach
to localization of groups of mobile robots. In: Robotics and Automation, 2000. Proceedings.
ICRA'00. IEEE International Conference on, vol. 3, pp. 2958–2965. IEEE (2000)
Controlling Swarms: A Programming Paradigm with Minimalistic Communication
29
32. Rubenstein, M., Cabrera, A., Werfel, J., Habibi, G., McLurkin, J., Nagpal, R.: Collective
transport of complex objects by simple robots: theory and experiments. In: Proceedings
of the 2013 international conference on Autonomous agents and multi-agent systems, pp.
47–54. International Foundation for Autonomous Agents and Multiagent Systems (2013)
33. Rubenstein, M., Cornejo, A., Nagpal, R.: Programmable self-assembly in a thousand-robot
swarm. Science 345(6198), 795–799 (2014)
34. Schmickl, T., Hamann, H.: BEECLUST: A swarm algorithm derived from honeybees. In:
Y. Xiao (ed.) Bio-inspired Computing and Communication Networks. CRC Press (2011)
35. Seeley, T.D.: The tremble dance of the honey bee: message and meanings. Behavioral
Ecology and Sociobiology 31, 375–383 (1992)
36. Siegert, F., Weijer, C.J.: Three-dimensional scroll waves organize dictyostelium slugs.
PNAS 89(14), 6433–6437 (1992)
37. Simpson, S.J., McCaffery, A.R., Haegele, B.F.: A behavioural analysis of phase change in
the desert locust. Biological Reviews 74(4), 461–480 (1999)
38. subCULTron: Submarine cultures perform long-term robotic exploration of unconventional
environmental niches (2015). http://www.subcultron.eu/
39. Thenius, R., Varughese, J.C., Moser, D., Schmickl, T.: Wospp-a wave oriented swarm
programming paradigm. IFAC-PapersOnLine 51(2), 379–384 (2018)
40. Trianni, V., Gross, R., Labella, T.H., S¸ahin, E., Dorigo, M.: Evolving aggregation behaviors
in a swarm of robots. In: European Conference on Artificial Life, pp. 865–874. Springer
(2003)
41. Varughese, J.C., Thenius, R., Schmickl, T., Wotawa, F.: Quantification and analysis of the
resilience of two swarm intelligent algorithms. In: C. Benzmuller, C. Lisetti, M. Theobald
(eds.) GCAI 2017. 3rd Global Conference on Artificial Intelligence, EPiC Series in Com-
puting, vol. 50, pp. 148–161. EasyChair (2017)
42. Varughese, J.C., Thenius, R., Wotawa, F., Schmickl, T.: Fstaxis algorithm: Bio-inspired
In: Proceedings of the 15th International Conference on the
emergent gradient taxis.
Synthesis and Simulation of Living Systems. MIT Press (2016)
43. Von Frisch, K.: The dance language and orientation of bees. (1967)
44. Yang, X.S.: Firefly algorithms for multimodal optimization.
In: Stochastic algorithms:
foundations and applications, pp. 169–178. Springer (2009)
45. Zahadat, P., Hahshold, S., Thenius, R., Crailsheim, K., Schmickl, T.: From honeybees to
robots and back: division of labour based on partitioning social inhibition. Bioinspiration
& biomimetics 10(6), 066005 (2015)
|
1905.11838 | 1 | 1905 | 2019-05-28T14:20:33 | A Parameterized Perspective on Protecting Elections | [
"cs.MA",
"cs.AI",
"cs.CY",
"cs.DS",
"cs.GT"
] | We study the parameterized complexity of the optimal defense and optimal attack problems in voting. In both the problems, the input is a set of voter groups (every voter group is a set of votes) and two integers $k_a$ and $k_d$ corresponding to respectively the number of voter groups the attacker can attack and the number of voter groups the defender can defend. A voter group gets removed from the election if it is attacked but not defended. In the optimal defense problem, we want to know if it is possible for the defender to commit to a strategy of defending at most $k_d$ voter groups such that, no matter which $k_a$ voter groups the attacker attacks, the outcome of the election does not change. In the optimal attack problem, we want to know if it is possible for the attacker to commit to a strategy of attacking $k_a$ voter groups such that, no matter which $k_d$ voter groups the defender defends, the outcome of the election is always different from the original (without any attack) one. | cs.MA | cs |
A Parameterized Perspective on Protecting
Elections
Palash Dey
Indian Institute of Technology, Kharagpur, India
[email protected]
Neeldhara Misra
Indian Institute of Technology, Gandhinagar, India
[email protected]
Swaprava Nath
Indian Institute of Technology, Kanpur, India
[email protected]
Garima Shakya
Indian Institute of Technology, Kanpur, India
[email protected]
Abstract
We study the parameterized complexity of the optimal defense and optimal attack problems in voting.
In both the problems, the input is a set of voter groups (every voter group is a set of votes) and two
integers ka and kd corresponding to respectively the number of voter groups the attacker can attack
and the number of voter groups the defender can defend. A voter group gets removed from the election
if it is attacked but not defended. In the optimal defense problem, we want to know if it is possible for
the defender to commit to a strategy of defending at most kd voter groups such that, no matter which
ka voter groups the attacker attacks, the outcome of the election does not change.
In the optimal
attack problem, we want to know if it is possible for the attacker to commit to a strategy of attacking
ka voter groups such that, no matter which kd voter groups the defender defends, the outcome of
the election is always different from the original (without any attack) one. We show that both the
optimal defense problem and the optimal attack problem are computationally intractable for every
scoring rule and the Condorcet voting rule even when we have only 3 candidates. We also show that
the optimal defense problem for every scoring rule and the Condorcet voting rule is W[2]-hard for both
the parameters ka and kd, while it admits a fixed parameter tractable algorithm parameterized by the
combined parameter (ka, kd). The optimal attack problem for every scoring rule and the Condorcet
voting rule turns out to be much harder -- it is W[1]-hard even for the combined parameter (ka, kd).
We propose two greedy algorithms for the OPTIMAL DEFENSE problem and empirically show that they
perform effectively on reasonable voting profiles.
Keywords and phrases parameterized complexity, election control, optimal attack, optimal defense
1
Introduction
The problem of election control asks if it is possible for an external agent, usually with a
fixed set of resources, to influence the outcome of the election by altering its structure in
some limited way. There are several specific manifestations of this problem: for instance,
one may ask if it is possible to change the winner by deleting k voter groups, presumably
by destroying ballot boxes or rigging electronically submitted votes. Indeed, several cases
of violence at the ballot boxes have been placed on record [7, 2], and in 2010, Halderman
2
A Parameterized Perspective on Protecting Elections
and his students exposed serious vulnerabilities in the electronic voting systems that are
in widespread use in several states [1]. A substantial amount of the debates around the
recently concluded presidential elections in the United States revolved around issues of
potential fraud, with people voting multiple times, stuffing ballot boxes, etc. all of which
are well recognized forms of election control. For example, Wolchok et al. [54] studied
security aspects on Internet voting systems.
Parameters
kd
ka
(ka, kd)
m
OPTIMAL DEFENSE
OPTIMAL ATTACK
Scoring rules
Condorcet
Scoring rules
Condorcet
W[2]-hard [Theorem 15] W[2]-hard [Theorem 18] W[2]-hard [Theorem 16] W[2]-hard [Theorem 19]
W[2]-hard [Theorem 22] W[2]-hard [Theorem 23]
O∗(kkd
a ) [Theorem 28]
No poly kernel [Theorem 27]
para-NP-hard [Theorem 13]
W[1]-hard [Theorem 25] W[1]-hard [Theorem 26]
para-coNP-hard [Theorem 13]
Table 1 Summary of parameterized complexity results. kd : the maximum number of voter groups
that the defender can defend. ka : the maximum number of voter groups that the attacker can attack.
m : the number of candidates.
The study of controlling elections is fundamental to computational social choice: it is widely
studied from a theoretical perspective, and has deep practical impact. Bartholdi et al [4]
initiated the study of these problems from a computational perspective, hoping that compu-
tational hardness of these problems may suggest a substantial barrier to the phenomena of
control: if it is, say NP-hard to control an election, then the manipulative agent may not
be able to compute an optimal control strategy in a reasonable amount of time. This basic
approach has been intensely studied in various other scenarios. For instance, Faliszewski
et al. [27] studied the problem of control where different types of attacks are combined
(multimode control), Mattei et al [44] showed hardness of a variant of control which just
exercises different tie-breaking rules, Bulteau et al. [10] studied voter control in a combin-
atorial setting, etc [49, 52, 28, 11, 43, 31, 30, 29, 26, 45, 25, 24, 24, 34, 37, 33, 36, 32, 47,
48, 51, 14, 21, 20, 16, 17, 15].
Exploring parameterized complexity of various control problems has also gained a lot of
interest. For example, Betzler and Uhlmann [6] studied parameterized complexity of can-
didate control in elections and showed interesting connection with digraph problems, Liu
and Zhu [41, 42] studied parameterized complexity of control problem by deleting voters
for many common voting rules, and so on [40, 53, 38, 18, 22]. Studying election control
from a game theoretic approach using security games is also an active area of research. See,
for example, the works of An et al. and Letchford et al. [3, 39].
The broad theme of using computational hardness as a barrier to control has two distinct lim-
itations: one is, of course, that some voting rules simply remain computationally vulnerable
to many forms of control, in the sense that optimal strategies can be found in polynomial
time. The other is that even NP-hard control problems often admit reasonable heuristics,
can be approximated well, or even admit efficient exact algorithms in realistic scenarios.
Therefore, relying on NP-hardness alone is arguably not a robust strategy against control.
To address this issue, the work of Yin et al. [56] explicitly defined the problem of protect-
ing an election from control, where in addition to the manipulative agent, we also have a
"defender", who can also deploy some resources to spoil a planned attack. In this setting,
elections are defined with respect to voter groups rather than voters, which is a small dif-
ference from the traditional control setting. The voter groups model allows us to consider
P. Dey, N. Misra, S. Nath, and G. Shakya
3
attacks on sets of voters, which is a more accurate model of realistic control scenarios.
In Yin et al. [56], the defense problem is modeled as a Stackelberg game in which limited
protection resources (say kd) are deployed to protect a collection of voter groups and the
adversary responds by attempting to subvert the election (by attacking, say, at most ka
groups). They consider the plurality voting rule, and show that the problem of choosing the
minimal set of resources that guarantee that an election cannot be controlled is NP-hard.
They further suggest a Mixed-Integer Program formulation that can usually be efficiently
tackled by solvers. Our main contribution is to study this problem in a parameterized setting
and provide a refined complexity landscape for it. We also introduce the complementary
attack problem, and extend the study to voting rules beyond plurality. We now turn to a
summary of our contributions.
Contribution:
We refer the reader to Section 2 for the relevant formal definitions, while focusing here on a
high-level overview of our results. Recall that the OPTIMAL DEFENSE problem asks for a set
of at most kd voter groups which, when protected, render any attack on at most ka voter
groups unsuccessful. In this paper, we study the parameterized complexity of OPTIMAL DE-
FENSE for all scoring rules and the Condorcet voting rule (these are natural choices because
they are computationally vulnerable to control - - the underlying "attack problem" can be
resolved in polynomial time). We show that the problem of finding an optimal defense is
tractable when both the attacker and the defender have limited resources. Specifically, we
show that the problem is fixed-parameter tractable with the combined parameter (ka, kd)
by a natural bounded-depth search tree approach. We also show that the OPTIMAL DEFENSE
problem is unlikely to admit a polynomial kernel under plausible complexity theoretic as-
sumption. We observe that both these parameters are needed for fixed parameter tractab-
ility, as we show W[2]-hardness when OPTIMAL DEFENSE is parameterized by either ka or
kd.
Another popular parameter considered for voting problems is m, the number of candidates
-- as this is usually small compared to the size of the election in traditional application
scenarios. Unfortunately, we show that OPTIMAL DEFENSE is NP-hard even when the election
has only 3 candidates, eliminating the possibility of fixed-parameter algorithms (and even
XP algorithms). This strengthens a hardness result shown in Yin et al. [56]. Our hardness
results on a constant number of candidates rely on a succinct encoding of the information
about the scores of the candidates from each voter group. We also observe that the problem
is polynomially solvable when only two candidates are involved.
We introduce the complementary problem of attacking an election: here the attacker plays
her strategy first, and the defender is free to defend any of the attacked groups within the
budget. The attacker wins if she is successful in subverting the election no matter which
defense is played out. This problem turns out to be harder: it is already W[1]-hard when
parameterized by both ka and kd, which is in sharp contrast to the OPTIMAL DEFENSE prob-
lem. This problem is also hard in the setting of a constant number of candidates -- specific-
ally, it is coNP-hard for the plurality voting rule [Theorem 10] and the Condorcet voting
rule [Theorem 12] even when we have only three candidates if every voter group is en-
coded as the number of plurality votes every candidate receives from that voter group. Our
demonstration of the hardness of the attack problem is another step in the program of using
4
A Parameterized Perspective on Protecting Elections
computational intractability as a barrier to undesirable phenomenon, which, in this context,
is the act of planning a systematic attack on voter groups with limited resources.
We finally propose two simple greedy algorithms for the OPTIMAL DEFENSE problem and
empirically show that it may be able to solve many instances of practical interest.
2
Preliminaries
Let C = {c1, c2, . . . , cm} be a set of candidates and V = {v1, v2, . . . , vn} a set of voters. If
not mentioned otherwise, we denote the set of candidates by C, the set of voters by V, the
number of candidates by m, and the number of voters by n. Every voter vi has a preference
or vote ≻i which is a complete order over C. We denote the set of all complete orders over
C by L(C). We call a tuple of n preferences (≻1, ≻2, · · · , ≻n) ∈ L(C)n an n-voter preference
profile. Often it is convenient to view a preference profile as a multi-set consisting of its
votes. The view we are taking will be clear from the context. A voting rule (often called
voting correspondence) is a function r : ∪n∈NL(C)n −→ 2C \ {∅} which selects, from a
preference profile, a nonempty set of candidates as the winners. We refer the reader to [9]
for a comprehensive introduction to computational social choice. In this paper we will be
focusing on two voting rules -- the scoring rules and the Condorcet voting rule which are
defined as follows.
Scoring Rule: A collection of m-dimensional vectors −→
sm = (α1, α2, . . . , αm) ∈ Rm with
α1 > α2 > . . . > αm and α1 > αm for every m ∈ N naturally defines a voting rule --
a candidate gets score αi from a vote if it is placed at the ith position, and the score of
a candidate is the sum of the scores it receives from all the votes. The winners are the
candidates with the highest score. Given a set of candidates C, a score vector −→
α of length
−→
α
C, a candidate x ∈ C, and a profile P, we denote the score of x in P by s
P (x). When
the score vector −→
α from the superscript. A straight
forward observation is that the scoring rules remain unchanged if we multiply every αi by
any constant λ > 0 and/or add any constant µ. Hence, we assume without loss of generality
that for any score vector −→
sm, there exists a j such that αj − αj+1 = 1 and αk = 0 for all k > j.
We call such a score vector a normalized score vector.
α is clear from the context, we omit −→
Weighted Majority Graph and Condorcet Voting Rule: Given an election E = (C, (≻1, ≻2
, . . . , ≻n)) and two candidates x, y ∈ C, let us define NE(x, y) to be the number of votes
where the candidate x is preferred over y. We say that a candidate x defeats another can-
didate y in pairwise election if NE(x, y) > NE(y, x). Using the election E, we can construct
a weighted directed graph GE = (U = C, E) as follows. The vertex set U of the graph GE
is the set of candidates C. For any two candidates x, y ∈ C with x 6= y, let us define the
margin DE(x, y) of x from y to be NE(x, y) − NE(y, x). We have an edge from x to y in GE if
DE(x, y) > 0. Moreover, in that case, the weight w(x, y) of the edge from x to y is DE(x, y).
A candidate c is called the Condorcet winner of an election E if there is an edge from c to
every other vertices in the weighted majority graph GE. The Condorcet voting rule outputs
the Condorcet winner if it exists and outputs the set C of all candidates otherwise.
Let r be a voting rule. We study the r-OPTIMAL DEFENSE problem which was defined by
Yin et al. [56]. It is defined as follows. Intuitively, the r-OPTIMAL DEFENSE problem asks if
there is a way to defend kd voter groups such that, irrespective of which ka voter groups the
P. Dey, N. Misra, S. Nath, and G. Shakya
5
attacker attacks, the output of the election (that is the winning set of candidates) is always
same as the original one. A voter group gets deleted if only if it is attacked but not defended.
◮ Definition 1 (r-OPTIMAL DEFENSE). Given n voter groups Gi, i ∈ [n], two integers ka and
kd, does there exist an index set I ⊆ [n] with I 6 kd such that, for every I′ ⊂ [n] \ I with
I′ 6 ka, we have r((Gi)i∈[n]\I′ ) = r((Gi)i∈[n])? The integers ka and kd are called respectively
attacker's resource and defender's resource. We denote an arbitrary instance of the r-OPTIMAL
DEFENSE problem by (C, {Gi : i ∈ [n]}, ka, kd).
We also study the r-OPTIMAL ATTACK problem which is defined as follows. Intuitively, in the
r-OPTIMAL ATTACK problem the attacker is interested to know if it is possible to attack ka
voter groups such that, no matter which kd voter groups the defender defends, the outcome
of the election is never same as the original (that is the attack is successful).
◮ Definition 2 (r-OPTIMAL ATTACK). Given n voter groups Gi, i ∈ [n], two integers ka and kd,
does there exist an index set I ⊆ [n] with I 6 ka such that, for every I′ ⊆ [n] with I′ 6 kd,
we have r((Gi)i∈[n]\(I\I′)) 6= r((Gi)i∈[n])? We denote an arbitrary instance of the r-OPTIMAL
ATTACK problem by (C, {Gi : i ∈ [n]}, ka, kd).
Encoding of the Input Instance: In both the r-OPTIMAL DEFENSE and r-OPTIMAL ATTACK
problems, we assume that every input voter group G is encoded as follows. The encoding
lists all the different votes ≻ that appear in the voter group G along with the number of
times the vote ≻ appear in G. Hence, if a voter group G contains only k different votes over
m candidates and consists of n voters, then the encoding of G takes O(km log m log n) bits
of memory.
In parameterized complexity, each problem instance comes
Parameterized complexity:
with a parameter k. Formally, a parameterized problem Π is a subset of Γ ∗ × N, where Γ is
a finite alphabet. An instance of a parameterized problem is a tuple (x, k), where k is the
parameter. A central notion is fixed parameter tractability (FPT) which means, for a given
instance (x, k), solvability in time f(k) · p(x), where f is an arbitrary function of k and p is
a polynomial in the input size x. Just as NP-hardness is used as evidence that a problem
probably is not polynomial time solvable, there exists a hierarchy of complexity classes above
FPT, and showing that a parameterized problem is hard for one of these classes is considered
evidence that the problem is unlikely to be fixed-parameter tractable. The main classes in
this hierarchy are: FPT ⊆ W[1] ⊆ W[2] ⊆ · · · ⊆ W[P] ⊆ XP. We now define the notion of
parameterized reduction [13].
◮ Definition 3. Let A, B be parameterized problems. We say that A is fpt-reducible to B if
there exist functions f, g : N → N, a constant α ∈ N and an algorithm Φ which transforms an
instance (x, k) of A into an instance (x ′, g(k)) of B in time f(k)xα so that (x, k) ∈ A if and
only if (x ′, g(k)) ∈ B.
To show W-hardness in the parameterized setting, it is enough to give a parameterized
reduction from a known hard problem. For a more detailed and formal introduction to
parameterized complexity, we refer the reader to [13] for a detailed introduction to this
paradigm.
◮ Definition 4. [Kernelization] [50, 35] A kernelization algorithm for a parameterized prob-
lem Π ⊆ Γ ∗ × N is an algorithm that, given (x, k) ∈ Γ ∗ × N, outputs, in time polynomial in
x + k, a pair (x ′, k ′) ∈ Γ ∗ × N such that (a) (x, k) ∈ Π if and only if (x ′, k ′) ∈ Π and (b)
x ′, k ′ 6 g(k), where g is some computable function. The output instance x ′ is called the kernel,
6
A Parameterized Perspective on Protecting Elections
and the function g is referred to as the size of the kernel. If g(k) = kO(1), then we say that Π
admits a polynomial kernel.
For many parameterized problems, it is well established that the existence of a polynomial
kernel would imply the collapse of the polynomial hierarchy to the third level (or more
precisely, CoNP ⊆ NP/Poly). Therefore, it is considered unlikely that these problems would
admit polynomial-sized kernels. For showing kernel lower bounds, we simply establish
reductions from these problems.
◮ Definition 5. [Polynomial Parameter Transformation] [8] Let Γ1 and Γ2 be paramet-
erized problems. We say that Γ1 is polynomial time and parameter reducible to Γ2, written
Γ1 6Ptp Γ2, if there exists a polynomial time computable function f : Σ∗ × N → Σ∗ × N, and a
polynomial p : N → N, and for all x ∈ Σ∗ and k ∈ N, if f ((x, k)) = (x ′, k ′), then (x, k) ∈ Γ1 if
and only if (x ′, k ′) ∈ Γ2, and k ′ 6 p (k). We call f a polynomial parameter transformation (or
a PPT) from Γ1 to Γ2.
This notion of a reduction is useful in showing kernel lower bounds because of the following
theorem.
◮ Theorem 6. [8, Theorem 3] Let P and Q be parameterized problems whose derived classical
problems are Pc, Qc, respectively. Let Pc be NP−complete, and Qc ∈ NP. Suppose there exists
a PPT from P to Q. Then, if Q has a polynomial kernel, then P also has a polynomial kernel.
3
Classical Complexity Results
Yin et al. [56] showed that the OPTIMAL DEFENSE problem is polynomial time solvable for
the plurality voting rule when we have only 2 candidates. On the other hand, they also
showed that the OPTIMAL DEFENSE problem is NP-complete when we have an unbounded
number of candidates. We begin with improving their NP-completeness result by showing
that the OPTIMAL DEFENSE problem becomes NP-complete even when we have only 3 can-
didates and the attacker can attack any number of voter groups. Towards that, we reduce
the k-SUM problem to the OPTIMAL DEFENSE problem. The k-SUM problem is defined as
follows.
◮ Definition 7 (k-SUM). Given a set of n positive integers W = {wi, i ∈ [n]}, and two positive
integers k 6 n and M, does there exist an index set I ⊂ [n] with I = k such that Pi∈I wi =
M?
The k-SUM problem can be easily proved to be NP-complete by modifying the NP-
completeness proof of the Subset Sum problem in Cormen et al. [12]. We also need the
following structural result for normalized scoring rules which has been used before [5, 19].
◮ Lemma 8. Let C = {c1, . . . , cm} be a set of candidates and −→
α a normalized score vector of
length C. Let x, y ∈ C, x 6= y, be any two arbitrary candidates. Then there exists a profile Py
x
consisting of m votes such that we have the following.
sPy
(a) for every a ∈ C \ {x, y}
(x) + 1 = sPy
(y) − 1 = sPy
x
x
x
For any two candidates x, y ∈ C, x 6= y, we use Py
x to denote the profile as defined in
Theorem 8. We are now ready to present our NP-completeness result for the OPTIMAL
DEFENSE problem for the scoring rules even in the presence of 3 candidates only.
In the
interest of space, we will provide only a sketch of a proof for a several results.
P. Dey, N. Misra, S. Nath, and G. Shakya
7
◮ Theorem 9. The OPTIMAL DEFENSE problem is NP-complete for every scoring rule even if
the number of candidates is 3 and the attacker can attack any number of the voter groups.
Proof. The OPTIMAL DEFENSE problem for every scoring rule can be shown to belong to NP
by using a defense strategy S (a subset of at most kd voter groups) as a certificate. The fact
that the certificate can be validated in polynomial time involves checking if there exists a
successful attack despite protecting all groups in S. This can be done in polynomial time, but
due to space constraints, we defer a detailed argument to a full version of this manuscript.
We now turn to the reduction from k-SUM.
Let −→
α be any normalized score vector of length 3. The OPTIMAL DEFENSE problem for
the scoring rule based on −→
α belongs to NP. Let (W = {w1, . . . , wn}, k, M) be an arbitrary
instance of the k-SUM problem. We can assume, without loss of generality, that 8 divides M
and wi for every i ∈ [n]; if this is not the case, we replace M and wi by respectively 8M and
8wi for every i ∈ [n] which clearly is an equivalent instance of the original instance. Let us
also assume, without loss of generality, that 2k < n (if not then add enough copies of M + 1
to W) and M < Pn
i=1 wi (since otherwise, it is a trivial NO instance). We construct the
following instance of the OPTIMAL DEFENSE problem for the scoring rule based on −→
α . Let
M′ be an integer such that M′ > Pn
i=1 wi and 8 divides M′. We have 3 candidates, namely
a, b, and c. We have the following voter groups.
-- For every i ∈ [n], we have a voter group Gi consisting of wi copies of Pc
a (as defined in
Theorem 8) and M′ − wi copies of Pc
sGi (c) = sGi (a) + M′ + wi = sGi (b) + 2M′ − wi
b. Hence, we have the following.
-- We have one voter group G consisting of (kM′+M)/2 − 3 copies of Pa
c , (kM′−M)/2 − 1 copies
c , and (kM′−M)/2 − 1 copies of Pb
of Pb
s G(c) = s G(a) − (kM′ + M − 6) = s G(b) − (2kM′ − M − 6)
a. We have the following.
i=1 wi −M+6 = sQ(b)+(n−2k)M′ +M−Pn
i=1Gi ∪ G. We have sQ(c) = sQ(a) + (n − k)M′ +
Let Q be the resulting profile; that is Q = ∪n
Pn
i=1 wi,
we have sQ(c) > sQ(a) and sQ(c) > sQ(b). Thus the candidate c wins the election uniquely.
We define kd, the maximum number of voter groups that the defender can defend, to be
k. We define ka, the maximum number of voter groups that the attacker can attack, to be
n + 1. This finishes the description of the OPTIMAL DEFENSE instance. We claim that the
two instances are equivalent.
i=1 wi +6. Since n > 2k and M′ > Pn
In the forward direction, let the k-SUM instance be a YES instance and I ⊂ [n] with I =
k be an index set such that Pi∈I wi = M. Let us consider the defense strategy where
the defender protects the voter groups Gi for every i ∈ I. Since Pi∈I wi = M, we have
Pi∈I(M′ − wi) = kM′ − M. Let H be the profile of voter groups corresponding to the index
set I; that is, H = ∪i∈IGi. Let H′ be the profile remaining after the attacker attacks some
voter groups. Without loss of generality, we can assume that the attacker does not attack the
voter group G since otherwise the candidate c continues to win uniquely. We thus obviously
have H ∪ G ⊆ H′. We have sH∪ G(c) = sH∪ G(a) + kM′ + Pi∈I wi − (kM′ + M − 6) =
sH∪ G(a) + 6 and sH∪ G(c) = sH∪ G(b) + 2kM′ − Pi∈I wi − (2kM′ − M − 6) = sH∪ G(b) + 6.
Since the candidate c receives as much score as any other candidate in the voter group Gi
for every i ∈ [n], we have sH′∪ G(c) > sH′∪ G(a)+6 and sH′∪ G(c) > sH′∪ G(b)+6. Hence, the
candidate c wins uniquely in the resulting profile H′ after the attack and thus the defense is
successful.
8
A Parameterized Perspective on Protecting Elections
In the other direction, let the OPTIMAL DEFENSE instance be a YES instance. Without loss of
generality, we can assume that the attacker does not attack the voter group G and thus the
defender does not defend the voter group G. We can also assume, without loss of generality,
that the defender defends exactly k voter groups since the candidate c receives as much score
as any other candidate in the voter group Gi for every i ∈ [n]. Let I ⊂ [n] with I = k such
that defending all the voter groups Gi, i ∈ I is a successful defense strategy. We claim that
Pi∈I wi > M. Suppose not, then let us assume that Pi∈I wi < M. Since, wi is divisible
by 8 and positive for every i ∈ [n] and m is divisible by 8, we have Pi∈I wi 6 M − 8. Let
H be the profile of voter groups corresponding to the index set I; that is, H = ∪i∈IGi. We
have sH∪ G(c) = sH∪ G(a) + kM′ + Pi∈I wi − (kM′ + M − 6) 6 sH∪ G(a) + M − 8 − M + 6 =
sH∪ G(a) − 2. Hence attacking the voter groups Gi, i ∈ [n] \ I makes the score of c strictly
less than the score of a. This contradicts our assumption that defending all the voter groups
Gi, i ∈ I is a successful defense strategy. Hence we have Pi∈I wi > M. We now claim that
Pi∈I wi 6 M. Suppose not, then let us assume that Pi∈I wi > M. Since, wi is divisible by
8 and positive for every i ∈ [n] and m is divisible by 8, we have Pi∈I wi > M + 8. Let H′ be
the profile of voter groups corresponding to the index set I; that is, H′ = ∪i∈IGi. We have
sH′∪ G(c) = sH′∪ G(b) + 2kM′ − Pi∈I wi − (2kM′ − M − 6) 6 sH′∪ G(b) − (M + 8) + M + 6 =
sH′∪ G(b) − 2. Hence attacking the voter groups Gi, i ∈ [n] \ I makes the score of c strictly
less than the score of b. This contradicts our assumption that defending all the voter groups
Gi, i ∈ I is a successful defense strategy. Hence we have Pi∈I wi 6 M. Therefore we have
Pi∈I wi = M and thus the k-SUM instance is a YES instance.
◭
In the proof of Theorem 9, we observe that the reduced instance of the OPTIMAL DEFENSE
problem viewed as an instance of the OPTIMAL ATTACK problem is a NO instance if and
only if the k-SUM instance is a YES instance. Hence, the same reduction as in the proof of
Theorem 9 gives us the following result for the OPTIMAL ATTACK problem.
◮ Corollary 10. The OPTIMAL ATTACK problem is coNP-hard for every scoring rule even if the
number of candidates is 3 and the attacker can attack any number of voter groups.
We now prove a similar hardness result as of Theorem 9 for the Condorcet voting rule.
◮ Theorem 11. The OPTIMAL DEFENSE problem is NP-complete for the Condorcet voting rule
even if the number of candidates is 3 and the attacker can attack any number of voter groups.
Proof. The OPTIMAL DEFENSE problem for the Condorcet voting rule clearly belongs to
NP. To show NP-hardness, we reduce an arbitrary instance of the k-SUM problem to the
OPTIMAL DEFENSE problem for the Condorcet voting rule. Let ({w1, . . . , wn}, k, M) be an
arbitrary instance of the k-SUM problem. We construct the following instance of the OPTIMAL
DEFENSE problem for the Condorcet voting rule. Let M′ = max{wi : i ∈ [n]}. We have 3
candidates, namely a, b, and c. We have the following voter groups.
-- For every i ∈ [n], we have a voter group Gi where DGi (a, b) = 2wi, DGi (a, c) = 2(M′ −
wi), and DGi (b, c) = 0.
-- We have one voter group G where the candidates b and c receive respectively D G(b, a) =
2M − 1, D G(c, a) = 2(kM′ − M) − 1, and D G(b, c) = 1.
We define kd, the maximum number of voter groups that the defender can defend, to be
k. We define ka, the maximum number of voter groups that the attacker can attack, to
be n + 1. We observe that the candidate a is the Condorcet winner of the election. This
P. Dey, N. Misra, S. Nath, and G. Shakya
9
finishes the description of the OPTIMAL DEFENSE instance. We claim that the two instances
are equivalent.
In the forward direction, let the k-SUM instance be a YES instance and I ⊂ [n] with I =
k be an index set such that Pi∈I wi = M. Let us consider the defense strategy where
the defender protects the voter groups Gi for every i ∈ I. Since Pi∈I wi = M, we have
Pi∈I(M′ − wi) = kM′ − M. Without loss of generality, we can assume that the attacker
does not attack the voter group G. We observe that the candidate a is the Condorcet winner
of the election even when the attacker attacks all the voter groups Gj, j ∈ [n] \ I. Hence the
OPTIMAL DEFENSE instance is a YES instance.
In the other direction, let the OPTIMAL DEFENSE instance be a YES instance. Without loss of
generality, we can assume that the attacker does not attack the voter group G and thus the
defender does not defend the voter group G. We can also assume, without loss of generality,
that the defender defends exactly k voter groups since the candidate a continues to be the
Condorcet winner if the attacker attacks at most k − 1 voter groups. Let I ⊂ [n] with
I = k such that defending all the voter groups Gi, i ∈ I is a successful defense strategy.
We claim that Pi∈I wi > M. Suppose not, then let us assume that Pi∈I wi < M. Then
attacking the voter groups Gi, i ∈ [n] \ I makes the candidate b defeat the candidate a
in pairwise election. This contradicts or assumption that defending all the voter groups
Gi, i ∈ I is a successful defense strategy. Hence we have Pi∈I wi > M. We now claim that
Pi∈I wi 6 M. Suppose not, then let us assume that Pi∈I wi > M. Then attacking the
voter groups Gi, i ∈ [n] \ I makes the candidate c defeat the candidate a in pairwise election.
This contradicts or assumption that defending all the voter groups Gi, i ∈ I is a successful
defense strategy. Hence we have Pi∈I wi 6 M. Therefore we have Pi∈I wi = M and thus
the k-SUM instance is a YES instance.
◭
In the proof of Theorem 11, we observe that the reduced instance of OPTIMAL DEFENSE
viewed as an instance of the OPTIMAL ATTACK problem is a NO instance if and only if the
k-SUM instance is a YES instance. Hence, the same reduction as in the proof of Theorem 11
gives us the following result for the OPTIMAL ATTACK problem.
◮ Corollary 12. The OPTIMAL ATTACK problem is coNP-hard for the Condorcet voting rule
even if the number of candidates is 3 and the attacker can attack any number of voter groups.
4 W-Hardness Results
In this section, we present our hardness results for the OPTIMAL DEFENSE and the OPTIMAL
ATTACK problems in the parameterized complexity framework. We consider the following
parameters for both the problems -- number of candidate (m), defender's resource (kd), and
attacker's resource (ka). From Theorems 9 to 12 we immediately have the following result
for the OPTIMAL DEFENSE and OPTIMAL ATTACK problems parameterized by the number of
candidates for both the scoring rules and the Condorcet voting rule.
◮ Corollary 13. The OPTIMAL DEFENSE problem is para-NP-hard parameterized by the num-
ber of candidates for both the scoring rules and the Condorcet voting rule. The OPTIMAL ATTACK
problem is para-coNP-hard parameterized by the number of candidates for both the scoring
rules and the Condorcet voting rule.
10
A Parameterized Perspective on Protecting Elections
The NP-completeness proof for the OPTIMAL DEFENSE problem for the plurality voting rule
by Yin et al. [56] is actually a parameter preserving reduction from the HITTING SET problem
parameterized by the solution size. The HITTING SET problem is defined as follows.
◮ Definition 14 (HITTING SET). Given a universe U, a set S = {Si : i ∈ [t]} of subsets of U,
and a positive integer k which is at most U, does there exist a subset W ⊆ U with W = k such
that W ∩ Si 6= ∅ for every i ∈ [t]. We denote an arbitrary instance of HITTING SET by (U, S, k).
Since the HITTING SET problem parameterized by the solution size k is known to be W[2]-
complete [23], the following result immediately follows from Theorem 2 of Yin et al. [56].
⊲ Observation 1 ([56]). The OPTIMAL DEFENSE problem for the plurality voting rule is
W[2]-hard parameterized by kd.
We now generalize Observation 1 to any scoring rule by exhibiting a polynomial parameter
transform from the HITTING SET problem parameterized by the solution size.
◮ Theorem 15. The OPTIMAL DEFENSE problem for every scoring rule is W[2]-hard paramet-
erized by kd.
Proof. Let (U, S = {Sj : j ∈ [t]}, k) be an arbitrary instance of HITTING SET. Let U = {zi : i ∈
[n]}. Without loss of generality, we assume that Sj 6= ∅ for every j ∈ [t] since otherwise the
instance is a NO instance. Let −→
α be a normalized score vector of length t + 2. We construct
the following instance of the OPTIMAL DEFENSE problem for the scoring rule based on −→
α .
The set of candidates C = {xj : j ∈ [t]} ∪ {y, d}. We have the following voter groups.
-- For every i ∈ [n], we have a voter group Gi. For every j ∈ [t] with zi ∈ Sj we have 2
copies of Pd
xj
in Gi.
-- We have one group G where we have 2tn copies of Pxj
d for every j ∈ [n] and 2tn − 1
copies of Py
d.
i=1Gi ∪ G. We define the defender's resource
Let Q be the resulting profile; that is Q = ∪n
kd to be k + 1 and attacker's resource to be n. This finishes the description of the OPTIMAL
DEFENSE instance. Since Sj 6= ∅ for every j ∈ [t], we have sQ(y) > sQ(xj) for every j ∈ [t].
We also have sQ(y) > sQ(d). Hence the candidate y is the unique winner of the profile Q. We
now prove that the OPTIMAL DEFENSE instance (C, Q, ka, kd) is equivalent to the HITTING
SET instance (U, S, k).
In the forward direction, let us suppose that the HITTING SET instance is a YES instance. Let
I ⊂ [n] be such that I = k and {zi : i ∈ I} ∩ Sj 6= ∅. We claim that the defender's strategy of
defending the voter groups Gj for every j ∈ [t] \ I and G results in a successful defense. Let H
be the profile of voter groups corresponding to the index set I; that is, H = ∪i∈IGi. Let H′ be
the profile remaining after the attacker attacks some voter groups. We thus obviously have
H ∪ G ⊆ H′. Since {zi : i ∈ I} forms a hitting set, we have sH′(y) > sH′ (xj) for every j ∈ [t].
Also since the voter group G is defended, we have sH′ (y) > sH′ (d). Hence the candidate y
continues to win uniquely even after the attack and hence the OPTIMAL DEFENSE instance is
a YES instance.
In the other direction, let the OPTIMAL DEFENSE instance be a YES instance. Without loss
of generality, we can assume that the defender defends the voter group G since otherwise
the attacker can attack the voter group G which makes the score of the candidate d more
than the score of the candidate y and thus defense would fail. We can also assume, without
P. Dey, N. Misra, S. Nath, and G. Shakya
11
loss of generality, that the defender defends exactly k voter groups. Let I ⊂ [n] with I = k
such that defending all the voter groups Gi, i ∈ I and G is a successful defense strategy.
Let us consider Z = {zi : i ∈ I} ⊆ U. We claim that Z must form a hitting set.
Indeed,
otherwise let us assume that there exists a j ∈ [t] such that Z ∩ Sj = ∅. Consider the
situation where the attacker attacks voter groups Gi for every i ∈ [n] \ I. We observe that
s∪i∈IGi ∪ G(xj) > s∪i∈IGi∪ G(y). This contradicts our assumption that defending all the voter
groups Gi, i ∈ I and G is a successful defense strategy. Hence Z forms a hitting set and thus
the HITTING SET instance is a YES instance.
◭
In the proof of Theorem 15, we observe that the reduced instance of OPTIMAL DEFENSE
viewed as an instance of the OPTIMAL ATTACK problem is a NO instance if and only if the
k-SUM instance is a YES instance. Hence, the same reduction as in the proof of Theorem 15
gives us the following result for the OPTIMAL ATTACK problem.
◮ Corollary 16. The OPTIMAL ATTACK problem for every scoring rule is W[2]-hard parameter-
ized by kd.
We now show W[2]-hardness of the OPTIMAL DEFENSE problem for the Condorcet voting
rule parameterized by kd. Towards that, we need the following lemma which has been used
before [46, 55].
◮ Lemma 17. For any function f : C × C −→ Z, such that
1. ∀a, b ∈ C, f(a, b) = −f(b, a).
2. ∀a, b, c, d ∈ C, f(a, b) + f(c, d) is even,
there exists a n voters' profile such that for all a, b ∈ C, a defeats b with a margin of f(a, b).
Moreover,
n is even and n = O
X
{a,b}∈C×C
f(a, b)
Next, we show the W[2]-hardness of the OPTIMAL DEFENSE problem for the Condorcet voting
rule parameterized by kd. This is also a parameter-preserving reduction from the HITTING
SET problem.
◮ Theorem 18. The OPTIMAL DEFENSE problem for the Condorcet voting rule is W[2]-hard
parameterized by kd.
Proof. Let (U, S = {Sj : j ∈ [t]}, k) be an arbitrary instance of HITTING SET. Let U = {zi :
i ∈ [n]}. Without loss of generality, we assume that Sj 6= ∅ for every j ∈ [t] since otherwise
the instance is a NO instance. We construct the following instance of the OPTIMAL DEFENSE
problem for the Condorcet voting rule. The set of candidates C = {xj : j ∈ [t]} ∪ {y}. For every
i ∈ [n], we have a voter group Gi. For every j ∈ [t] with zi ∈ Sj we have DGi (y, xj) = 2.
Let Q be the resulting profile; that is Q = ∪n
i=1Gi. We define the defender's resource kd to
be k and attacker's resource to be n. This finishes the description of the OPTIMAL DEFENSE
instance. Since Sj 6= ∅ for every j ∈ [t], we have DQ(y, xj) > 2 for every j ∈ [t]. Hence
the candidate y is the Condorcet winner of the profile Q. We now prove that the OPTIMAL
DEFENSE instance (C, Q, ka, kd) is equivalent to the HITTING SET instance (U, S, k).
In the forward direction, let us suppose that the HITTING SET instance is a YES instance. Let
I ⊂ [n] be such that I = k and {zi : i ∈ I} ∩ Sj 6= ∅. We claim that the defender's strategy
12
A Parameterized Perspective on Protecting Elections
of defending the voter groups Gj for every j ∈ [t] \ I results in a successful defense. Let H
be the profile of voter groups corresponding to the index set I; that is, H = ∪i∈IGi. Let
H′ be the profile remaining after the attacker attacks some voter groups. We thus obviously
have H ⊆ H′. Since {zi : i ∈ I} forms a hitting set, we have DH′ (y, xj) > 2 for every
j ∈ [t]. Hence the candidate y continues to win uniquely even after the attack and hence the
OPTIMAL DEFENSE instance is a YES instance.
In the other direction, let the OPTIMAL DEFENSE instance be a YES instance. We can also
assume, without loss of generality, that the defender defends exactly k voter groups. Let
I ⊂ [n] with I = k such that defending all the voter groups Gi, i ∈ I is a successful defense
strategy. Let us consider Z = {zi : i ∈ I} ⊆ U. We claim that Z must form a hitting set.
Indeed, otherwise let us assume that there exists a j ∈ [t] such that Z ∩ Sj = ∅. Consider the
situation where the attacker attacks voter groups Gi for every i ∈ [n] \ I. We observe that
D∪i∈IGi (y, xj) = 0 and hence the candidate y is not the Condorcet winner. This contradicts
our assumption that defending all the voter groups Gi, i ∈ I is a successful defense strategy.
Hence Z forms a hitting set and thus the HITTING SET instance is a YES instance.
◭
In the proof of Theorem 18, we observe that the reduced instance of OPTIMAL DEFENSE
viewed as an instance of the OPTIMAL ATTACK problem is a NO instance if and only if the
k-SUM instance is a YES instance. Hence, the same reduction as in the proof of Theorem 18
gives us the following result for the OPTIMAL ATTACK problem.
◮ Corollary 19. The OPTIMAL ATTACK problem for the Condorcet voting rule is W[2]-hard
parameterized by kd.
We now show that the OPTIMAL DEFENSE problem for scoring rules is W[2]-hard paramet-
erized by ka also by exhibiting a parameter preserving reduction from a problem closely
related to HITTING SET, which is SET COVER problem parameterized by the solution size.
The SET COVER problem is defined as follows. This is a W[2]-complete problem [23]. We
now present our W[2]-hardness proof for the OPTIMAL DEFENSE problem for scoring rules
parameterized by ka, by a reduction from SET COVER.
◮ Definition 20 (SET COVER). Given an universe U, a set S = {Si : i ∈ [t]} of subsets of U, and
a non-negative integer k which is at most t, does there exists an index set I ⊂ [t] with I = k
such that Si∈I Si = U. We denote an arbitrary instance of SET COVER by (U, S, k).
◮ Theorem 21. The OPTIMAL DEFENSE problem for every scoring rule and Condorcet rule is
W[2]-hard parameterized by ka.
◮ Theorem 22. The OPTIMAL DEFENSE problem for every scoring rule is W[2]-hard paramet-
erized by ka.
Proof. Let (U, S = {Sj : j ∈ [t]}, k) be an arbitrary instance of SET COVER. Let U = {zi :
i ∈ [n]}. We assume that k > 3 since otherwise the SET COVER instance is polynomial
time solvable. For i ∈ [n], let fi be the number of j ∈ [t] such that zi ∈ Sj; that is, fi =
{j ∈ [t] : zi ∈ Sj}. We assume, without loss of generality, that for every i ∈ [n], t − fi −
k > 3k by adding at most 9t empty sets in S. We construct the following instance of the
OPTIMAL DEFENSE problem for the scoring rule induced by the score vector −→
α rule. The set
of candidates C = {xi : i ∈ [n]} ∪ {y, d}. Let −→
α be any normalized score vector of length n + 2.
We have the following voter groups.
P. Dey, N. Misra, S. Nath, and G. Shakya
13
-- For every j ∈ [t], we have a voter group Gj. For every i ∈ [n] and j ∈ [t] with zi /∈ Sj, we
have 2 copies of Pd
xi
.
-- We have another voter group H where, for every i ∈ [n], we have 2tn + (2(t − fi − k) + 1)
copies of Pxi
d and 2tn copies of Py
d.
We define attacker resource ka to be k and the defender's resource kd to be t − k. This
finishes the description of the OPTIMAL DEFENSE instance. We first observe that the score
of the candidate d is strictly less than the score of every other candidate. We now observe
that the candidate y is the unique winner of the election since the score of the candidate y
is 2k − 1 more than the score of the candidate xi for every i ∈ [n]. We now prove that the
OPTIMAL DEFENSE instance (C, ∪j∈[t]Gj ∪ H, ka, kd) is equivalent to the SET COVER instance
(U, S, k).
In the forward direction, let us suppose that the SET COVER instance is a YES instance. Let
I ⊂ [t] be such that I = k and Sj∈I Sj = U. We claim that the defender's strategy of
defending the voter groups Gj for every j ∈ [t] \ I results in a successful defense. To see
this, we first observe that, if the attacker attacks the voter group H, then the candidate y
continues to uniquely win the election irrespective of what other voter groups the attacker
attacks. Indeed, since t − fi − k > 3k for every i ∈ [n], the score of the candidate xi is strictly
less than the score of the candidate y irrespective of what other voter groups the attacker
attacks. Since, for every i ∈ [n] and j ∈ [t], the score of the candidate xi is not more than
the score of the candidate y in the voter group Gj, we may assume that the attacker attacks
the voter group Gj for every j ∈ I (since they are the only voter groups unprotected except
H). Now, since Sj, j ∈ I forms a set cover of U, after deleting the voter groups Gj, j ∈ I, the
score of the candidate xi increases by at most 2(k − 1) from the original election for every
i ∈ [n]. Hence, after deleting the voter groups Gj, j ∈ I, the score of the candidate xi is still
strictly less than the score of the candidate y. Hence the candidate y continues to win and
thus the defense is successful. Hence the OPTIMAL DEFENSE instance is a YES instance.
In the other direction, let us suppose that the OPTIMAL DEFENSE instance is a YES instance.
We assume, without loss of generality, that the defender protects exactly t − k voter groups.
We argued in the forward direction that we can assume, without loss of generality, that
the attacker never attacks the voter group H. Hence, we can also assume, without loss of
generality, that the defender also does not defend the voter group H. Let I ⊂ [t] be such
that I = k and the defender defends the voter group Gj for every j ∈ [t] \ I. We claim
that the sets Sj, j ∈ I forms a set cover of U. Suppose not, then let zi be an element in U
which is not covered by Sj, j ∈ I. We observe that attacking the voter groups Gj for every
j ∈ I increases the score of the candidate xi by 2k which makes the candidate y lose in the
resulting election (after deleting the voter groups Gj for every j ∈ I) since the score of xi is
strictly more than the score of y. This contradicts our assumption that defending the voter
group Gj for every j ∈ [t] \ I is a successful defense strategy. Hence Sj, j ∈ I forms a set cover
of U and thus the SET COVER instance is a YES instance.
◭
We now present our W[2]-hardness proof for the OPTIMAL DEFENSE problem for the Con-
dorcet voting rule parameterized by ka.
◮ Theorem 23. The OPTIMAL DEFENSE problem for the Condorcet voting rule is W[2]-hard
parameterized by ka.
Proof. Let (U, S = {Sj : j ∈ [t]}, k) be an arbitrary instance of SET COVER. Let U = {zi :
14
A Parameterized Perspective on Protecting Elections
i ∈ [n]}. We assume that k > 3 since otherwise the SET COVER instance is polynomial time
solvable. For i ∈ [n], let fi be the number of j ∈ [t] such that zi ∈ Sj; that is, fi = {j ∈
[t] : zi ∈ Sj}. We assume, without loss of generality, that for every i ∈ [n], t − fi − k > 3k
by adding at most 9t empty sets in S. We construct the following instance of the OPTIMAL
DEFENSE problem for the Condorcet voting rule. The set of candidates C = {xi : i ∈ [n]} ∪ {y}.
We have the following voter groups.
-- For every j ∈ [t], we have a voter group Gj. For every i ∈ [n] and j ∈ [t], we have
DGj (y, xi) = 2 if zi /∈ Sj and DGj (y, xi) = 0 otherwise. We also have DGj (xi, xℓ) = 0 for
every j ∈ [t], i, ℓ ∈ [n] with i 6= ℓ.
-- We have another voter group H where, for every i ∈ [n], we have DH(xi, y) = 2(t − fi −
k). We also have DH(xi, xℓ) = 0 for every i, ℓ ∈ [n] with i 6= ℓ.
We define attacker resource ka to be k and the defender's resource kd to be t − k. This fin-
ishes the description of the OPTIMAL DEFENSE instance. We first observe that the candidate
y is a Condorcet winner of the resulting election. We now prove that the OPTIMAL DEFENSE
instance (C, ∪j∈[t]Gj ∪ H, ka, kd) is equivalent to the SET COVER instance (U, S, k).
In the forward direction, let us suppose that the SET COVER is a YES instance. Let I ⊂ [t]
be such that I = k and Sj∈I Sj = U. We claim that the defender's strategy of defending
the voter groups Gj for every j ∈ [t] \ I results in a successful defense. To see this, we first
observe that, we can assume without loss of generality that the attacker does not attack the
voter group H since the candidate y loses every pairwise election in H. Since, for every
i ∈ [n] and j ∈ [t], the candidate y does not lose any pairwise election in the voter group
Gj, we may assume that the attacker attacks the voter group Gj for every j ∈ I (since they
are the only voter groups unprotected except H). Now, since Sj, j ∈ I forms a set cover of
U, after deleting the voter groups Gj, j ∈ I, we have D∪j∈[t]\IGi∪H(y, xi) > 2(t − fi − k +
1) − 2(t − fi − k) = 2 for every i ∈ [n]. Hence, after deleting the voter groups Gj, j ∈ I,
the candidate y continues to be the Condorcet winner of the remaining profile. Hence the
OPTIMAL DEFENSE instance is a YES instance.
In the other direction, let us suppose that the OPTIMAL DEFENSE instance is a YES instance.
We assume, without loss of generality, that the defender protects exactly t − k voter groups.
We argued in the forward direction that we can assume, without loss of generality, that
the attacker never attacks the voter group H. Hence, we can also assume, without loss of
generality, that the defender also does not defend the voter group H. Let I ⊂ [t] be such that
I = k and the defender defends the voter group Gj for every j ∈ [t] \ I. We claim that the
sets Sj, j ∈ I forms a set cover of U. Suppose not, then let zi be an element in U which is not
covered by Sj, j ∈ I. We observe that D∪j∈[t]\IGi ∪H(y, xi) = 2(t−fi−k)−2(t−fi −k) = 0 and
thus attacking the voter groups Gj for every j ∈ I makes the candidate y not the Condorcet
winner. This contradicts our assumption that defending the voter group Gj for every j ∈ [t]\I
is a successful defense strategy. Hence Sj, j ∈ I forms a set cover of U and thus the SET COVER
instance is a YES instance.
◭
We now show that the OPTIMAL ATTACK problem for the scoring rules is W[1]-hard even
parameterized by the combined parameter ka and kd. Towards that, we exhibit a polyno-
mial parameter transform from the CLIQUE problem parameterized by the size of the clique
we are looking for which is known to be W[1]-complete. The CLIQUE problem is defined as
follows.
P. Dey, N. Misra, S. Nath, and G. Shakya
15
◮ Definition 24 (CLIQUE). Given a graph G and an integer k, does there exist a clique in G of
size k? We denote an arbitrary instance of CLIQUE by (G, k).
◮ Theorem 25. The OPTIMAL ATTACK problem for every scoring rule is W[1]-hard parameter-
ized by (ka, kd).
Proof. Let (G = (V, E), k) be an arbitrary instance of the CLIQUE problem. Let V = {vi :
i ∈ [n]} and E = {ej : j ∈ [m]}. Let −→
α be any arbitrary normalized score vector of length
m + 2. We construct the following instance of the OPTIMAL ATTACK problem for the scoring
rule induced by the score vector −→
α . The set of candidates C = {xj : j ∈ [m]} ∪ {y, d}. We have
the following voter groups.
-- For every i ∈ [n], we have a voter group Gi. For every i ∈ [n], we have 10m copies of Px
d
in the voter group Gi if the
for every x ∈ C \ {d} in Gi. We also have two copies of Pd
xj
edge ej is incident on the vertex vi, for every i ∈ [m] and j ∈ [m].
-- We have another voter group H. We have one copy of Pxj
d for every j ∈ [m] in H.
We define attacker resource ka to be k and the defender's resource kd to be k − 2. This
finishes the description of the OPTIMAL ATTACK instance. Let Q be the resulting profile; that
it Q = ∪i∈[n]Gi ∪ H. We first observe that the candidate y is the winner of the resulting
election since sQ(y) = sQ(xj) + 3 and sQ(y) > sQ(d). This completes a description of the
construction. Due to lack of space, we defer the proof of equivalence to a longer version of
this manuscript. We now prove that the OPTIMAL ATTACK instance (C, Q, ka, kd) is equivalent
to the CLIQUE instance (G, k).
In the forward direction, let us assume that U = {vi : i ∈ I} ⊂ V with I = k forms a clique
in G. We claim that attacking all the voter groups Gi, i ∈ I forms a successful attack. Indeed,
suppose the defender defends all the voter groups Gi, i ∈ I except Gℓ and Gℓ′ . Let ej⋆ be
the edge between the vertices vℓ and vℓ′ in G. Let the profile after the attack be G; that is,
G = ∪i∈[n]\IGi ∪ Gℓ ∪ Gℓ′ ∪ H. Then we have s G(y) = s G(xj⋆ ) − 1 and thus the candidate y
does not win after the attack. Hence the OPTIMAL ATTACK instance is YES instance.
In the other direction, let the OPTIMAL ATTACK instance be a YES instance. We first observe
that the candidate d performs worse than everyone else in every voter group and thus d can
never win. Now we can assume, without loss of generality, that the attacker does not attack
the voter group H since the candidate y is not receiving more score than any other candidate
except d in H. Let attacking all the voter groups Gi, i ∈ I with I 6 k is a successful attack.
We observe that if I < k, then defending any k − 2 of the groups that are attacked foils the
attack -- since the candidate y continues to win even after deleting any one group. Hence we
have I = k. Let us consider the subset of vertices U = {vi : i ∈ I}. We claim that U forms a
clique in G. Indeed, if not, then let us assume that there exists two indices ℓ, ℓ′ ∈ I such that
there is no edge between the vertices vℓ and vℓ′ in G. Let us consider the defender strategy
of defending all the voter groups Gi, i ∈ I \ {ℓ, ℓ′}. We observe that the candidate y continues
to uniquely receive the highest score among all the candidates and thus y wins uniquely in
the resulting election. This contradicts our assumption that attacking all the voter groups
Gi, i ∈ I with I 6 k is a successful attack. Hence U forms a clique in G and thus the CLIQUE
instance is a YES instance.
◭
We now show similar result as of Theorem 25 for the Condorcet voting rule.
16
A Parameterized Perspective on Protecting Elections
◮ Theorem 26. The OPTIMAL ATTACK problem for the Condorcet voting rule is W[1]-hard
parameterized by (ka, kd).
Proof. Let (G = (V, E), k) be an arbitrary instance of the CLIQUE problem. Let V = {vi :
i ∈ [n]} and E = {ej : j ∈ [m]}. We construct the following instance of the OPTIMAL ATTACK
problem for the Condorcet voting rule. The set of candidates C = {xj : j ∈ [m]} ∪ {y}. We
have the following voter groups.
-- For every i ∈ [n], we have a voter group Gi. We have DGi (y, xj) = 2 if the edge ej is
incident on the vertex vi and DGi (y, xj) = 0 if the edge ej is not incident on the vertex
vi, for every i ∈ [n] and j ∈ [m]. We also have DGi (xℓ, xj) = 0 for every i ∈ [n], j, ℓ ∈ [m],
and j 6= ℓ.
-- We have another voter group H where we have DH(xj, y) = 2 for every j ∈ [m] and
DH(xℓ, xj) = 0 for every j, ℓ ∈ [m] and j 6= ℓ.
We define attacker resource ka to be k and the defender's resource kd to be k − 2. This
finishes the description of the OPTIMAL ATTACK instance. Let Q be the resulting profile;
that it Q = ∪i∈[n]Gi ∪ H. We first observe that the candidate y is the Condorcet winner
of the resulting election. We now prove that the OPTIMAL ATTACK instance (C, Q, ka, kd) is
equivalent to the CLIQUE instance (G, k).
In the forward direction, let us assume that U = {vi : i ∈ I} ⊂ V with I = k forms a clique
in G. We claim that attacking all the voter groups Gi, i ∈ I forms a successful attack. Indeed,
suppose the defender defends all the voter groups Gi, i ∈ I except Gℓ and Gℓ′ . Let ej⋆ be
the edge between the vertices vℓ and vℓ′ in G. Let the profile after the attack be G; that is,
G = ∪i∈[n]\IGi ∪ Gℓ ∪ Gℓ′ ∪ H. Then we have D G(y, xj⋆ ) = 0 and thus the candidate y is not
the unique winner after the attack. Hence the OPTIMAL ATTACK instance is YES instance.
In the other direction, let the OPTIMAL ATTACK instance be a YES instance. We can assume,
without loss of generality, that the attacker does not attack the voter group H since the
candidate y loses every pairwise election in H. Let attacking all the voter groups Gi, i ∈ I
with I 6 k is a successful attack. We observe that if I < k, then defending any k − 2
of the groups that are attacked foils the attack -- since the candidate y continues to be the
Condorcet winner even after deleting any one group. Hence we have I = k. Let us consider
the subset of vertices U = {vi : i ∈ I}. We claim that U forms a clique in G. Indeed, if not,
then let us assume that there exists two indices ℓ, ℓ′ ∈ I such that there is no edge between
the vertices vℓ and vℓ′ in G. Let us consider the defender strategy of defending all the voter
groups Gi, i ∈ I \ {ℓ, ℓ′}. We observe that the candidate y continues to be the Condorcet
winner in the resulting election. This contradicts our assumption that attacking all the voter
groups Gi, i ∈ I with I 6 k is a successful attack. Hence U forms a clique in G and thus the
CLIQUE instance is a YES instance.
◭
Once we have a parameterized algorithm for the OPTIMAL DEFENSE problem for the para-
meter (ka, kd), an immediate question is whether there exists a kernel for the OPTIMAL
DEFENSE problem of size polynomial in (ka, kd). We know that the HITTING SET problem
does not admit polynomial kernel parameterized by the universe size [23]. We observe that
the reductions from the HITTING SET problem to the OPTIMAL DEFENSE problem in The-
orem 15 and ?? are polynomial parameter transformations. Hence we immediately have the
following corollary.
P. Dey, N. Misra, S. Nath, and G. Shakya
17
◮ Corollary 27. The OPTIMAL DEFENSE and OPTIMAL ATTACK problems for the scoring rules
and the Condorcet rule do not admit a polynomial kernel parameterized by (ka, kd).
5
The FPT Algorithm
We complement the negative results of Observation 1 and Theorem 22 by presenting an FPT
algorithm for the OPTIMAL DEFENSE problem parameterized by (ka, kd). In the absence of
a defender, that is when kd = 0, Yin et al. [56] showed that the OPTIMAL DEFENSE problem
is polynomial time solvable for the plurality voting rule. Their polynomial time algorithm
for the OPTIMAL DEFENSE problem can easily be extended to any scoring rule. Using this
polynomial time algorithm, we design the following O∗(kkd
a ) time algorithm for the OPTIMAL
DEFENSE problem for scoring rules. This result shows that the OPTIMAL DEFENSE problem
is fixed parameter tractable with (ka, kd) as the parameter.
◮ Theorem 28. There is an algorithm for the OPTIMAL DEFENSE problem for every scoring
rule and the Condorcet voting rule which runs in time O∗(kkd
a ).
Proof. Let us prove the result for any scoring rule. The proof for the Condorcet voting rule
is exactly similar. Initially we run the attacking algorithm over the n voter groups without
any group being protected. If a successful attack exists, the algorithm outputs the ka groups
to be deleted. We recursively branch on ka cases by protecting one of these ka groups in
each branch and running the attacking algorithm again. In addition, the parameter kd is
also reduced by 1 each time a group is protected. When kd=0, the attacking algorithm is
run on all the leaves of the tree and a valid protection strategy exists as long as for at least
one of the leaves the attack outputs no i.e. after deploying resources to protect kd groups
the attacker is unable to change the outcome of the election with any strategy. The groups
to be protected is determined by traversing the tree that leads to the particular leaf which
did not output an attack. Clearly the number of nodes in this tree is bounded by kkd
a . The
amount of time taken to find an attack at each node is bounded by poly(n). Hence the
running time of this algorithm is bounded by kkd
◭
a .poly(n).
6
Experiments
Though the previous sections show that the optimal defending problem is computationally
intractable, it is a worst-case result.
In practice, elections have voting profiles that are
generated from some (possibly known) distribution. In this section, we conduct an empirical
study to understand how simple defending strategies perform for two such statistical voter
generation models. The defending strategies we consider are variants of a simple greedy
policy.
Defending strategy: For a given voting profile and a voting rule, the defending strategy finds
the winner. Suppose the winner is a. The strategy considers a with every other candidate,
and for each such pair it creates a sorted list of classes based on the winning margin of
votes for a in those classes, and picks the top kd classes to form a sub-list. Now, among
all these (m − 1) sorted sub-lists, the strategy picks the most frequent kd classes to protect.
We call this version of the strategy GREEDY 1. For certain profiles an optimal attacker (a)
may change the outcome by attacking some of the unprotected classes or (b) is unable to
18
A Parameterized Perspective on Protecting Elections
2
Y
D
E
E
R
G
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0.0
1000
1
Y
D
E
E
R
G
800
600
400
200
0
Plurality
Veto
Borda
P
V
B
P
V
B
P
V
B
P
V
B
P
V
B
P
V
B
P
V
B
P
V
B
P
V
B
optimal and defended
optimal but not defended
not optimal
7
8
9
10
2
3
4
5
6
kd
Figure 1 Performances of GREEDY 1 and GREEDY 2 for uniform voting profile generation model.
change the outcome. If (a) occurs, then there is a possibility that for the value of kd there
does not exist any defense strategy which can guard the election from all possible strategies
of the attacker. In that case, GREEDY 1 is optimal and is not optimal otherwise. It is always
optimal for case (b). Note that, given a profile and kd protected classes, it is easy to find
if there exists an optimal attack strategy, while it is not so easy to identify whether there
does not exist any defending strategy if the GREEDY 1 fails to defend. We find the latter
with a brute-force search for this experiment. A small variant of GREEDY 1 is the following:
when GREEDY 1 is unable to defend (which is possible to find out in poly-time), the strategy
chooses to protect kd classes uniformly at random. Call this strategy GREEDY 2.
2
Y
D
E
E
R
G
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0.0
1000
1
Y
D
E
E
R
G
800
600
400
200
0
Plurality
Veto
Borda
P
V
B
V
P
B
P
V
B
P
V
B
P
V
B
P
V
B
P
V
B
P
V
B
P
V
B
optimal and defended
optimal but not defended
not optimal
7
8
9
10
2
3
4
5
6
kd
Figure 2 Performances of GREEDY 1 and GREEDY 2 for voting profile generation model with two
major contesting candidates.
Voting profile generation: Fix m = 5. We generate 1000 preference profiles over these
alternatives for n = 12000, where each vote is picked uniformly at random from the set of
all possible strict preference orders over m alternatives. The voters are partitioned into 12
classes containing equal number of voters. We consider three voting rules: plurality, veto,
and Borda. The lower plot in Figure 1 shows the number of profiles which belongs to the
three categories: (i) GREEDY 1 defends (is optimal), (ii) GREEDY 1 cannot defend but no
defending strategy exists (is optimal), (iii) GREEDY 1 cannot defend but defending strategy
P. Dey, N. Misra, S. Nath, and G. Shakya
19
exists (not optimal). The x-axis shows different values of kd and we fix ka = 12 − kd.
The upper plot of Figure 1 shows the fraction of the profiles successfully defended by GREEDY
2 where GREEDY 1 is not optimal (i.e., cannot defend but defending strategy exists) when
GREEDY 2 uniformly at random picks kd classes 100 times. These fractions therefore serves
as an empirical probability of successful defense of GREEDY 2 given GREEDY 1 is not optimal.
In an election where the primary contest happens between two major candidates, even
though there are more candidates present, the generation model may be a little different.
We also consider another generation model that generates 40% profiles having a fixed al-
ternative a on top and the strict order of the (m − 1) alternatives is picked uniformly at
random, a similar 40% profiles with some other alternative b on top, and the remaining
20% preferences are picked uniformly at random from the set of all possible strict prefer-
ence orders. Similar experiments are run on this generation model and results are shown in
Figure 2.
The results show that even though optimal defense is a hard problem, a simple strategy
like greedy achieves more than 70% optimality. From the rest 30% non-optimal cases, the
variant GREEDY 2 is capable of salvaging it into optimal with probability almost 5% for
uniform generation model and above 5% for two-major contestant generation model for
kd = ka = 6. This empirically hints at a possibility that defending real elections may not be
too difficult.
7
Conclusion
We have considered the OPTIMAL DEFENSE problem from a primarily parameterized per-
spective for scoring rules and the Condorcet voting rule. We showed hardness in the number
of candidates, the number of resources for the defender or the attacker. On the other hand,
we show tractability for the combined parameter (ka, kd). We also introduced the OPTIMAL
ATTACK problem, which is hard even for the combined parameter (ka, kd), and also showed
the hardness for a constant number of candidates. Even though the OPTIMAL DEFENSE prob-
lem is hard, empirically we show that relatively simple mechanisms ensure good defending
performance for reasonable voting profiles.
References
1
2
3
4
halderman
Alex
http://www.popsci.com/brilliant-10-alex-halderman-strengthens-democracy-using-software,
2010.
strengthens
using
software,
democracy
Popular
Science,
Election day bombings
https://www.rt.com/news/pakistan-election-day-bombing-136, 2013.
sweep pakistan:
Over 30 killed, more than 200 injured.
Bo An, Matthew Brown, Yevgeniy Vorobeychik, and Milind Tambe. Security games with surveil-
lance cost and optimal timing of attack execution. In International conference on Autonomous
Agents and Multi-Agent Systems, AAMAS '13, Saint Paul, MN, USA, May 6-10, 2013, pages 223 --
230, 2013.
John J. Bartholdi, Craig A. Tovey, and Michael A. Trick. How hard is it to control an election?
Mathematical and Computer Modelling, 16(8):27 -- 40, 1992.
20
A Parameterized Perspective on Protecting Elections
5
6
7
Dorothea Baumeister, Magnus Roos, and Jörg Rothe. Computational complexity of two variants
of the possible winner problem. In The 10th International Conference on Autonomous Agents and
Multiagent Systems (AAMAS), pages 853 -- 860, 2011.
Nadja Betzler and Johannes Uhlmann. Parameterized complexity of candidate control in elec-
tions and related digraph problems. Theor. Comput. Sci., 410(52):5425 -- 5442, 2009.
Satarupa Bhattacharjya.
http://www.sundaytimes.lk/100411/News/nws_16.html, 2010.
Low turnout and invalid votes mark first post war general polls.
8 Hans L. Bodlaender, Stéphan Thomassé, and Anders Yeo. Kernel Bounds for Disjoint Cycles
and Disjoint Paths. In Amos Fiat and Peter Sanders, editors, Proc. 17th Annual European Sym-
posium,on Algorithms (ESA 2009), Copenhagen, Denmark, September 7-9, 2009., volume 5757 of
Lecture Notes in Computer Science, pages 635 -- 646. Springer, 2009.
9
10
11
12
Felix Brandt, Vincent Conitzer, Ulle Endriss, Jérôme Lang, and Ariel Procaccia. Handbook of
computational social choice, 2016.
Laurent Bulteau, Jiehua Chen, Piotr Faliszewski, Rolf Niedermeier, and Nimrod Talmon. Com-
binatorial voter control in elections. Theor. Comput. Sci., 589:99 -- 120, 2015.
Jiehua Chen, Piotr Faliszewski, Rolf Niedermeier, and Nimrod Talmon. Elections with few voters:
Candidate control can be easy. In Proc. Twenty-Ninth AAAI Conference on Artificial Intelligence,
January 25-30, 2015, Austin, Texas, USA., pages 2045 -- 2051, 2015.
Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction to
Algorithms, Third Edition. The MIT Press, 3rd edition, 2009.
13 Marek Cygan, Fedor V. Fomin, Lukasz Kowalik, Daniel Lokshtanov, Dániel Marx, Marcin Pilip-
czuk, Michal Pilipczuk, and Saket Saurabh. Parameterized Algorithms. Springer, 2015.
14
15
16
17
18
19
20
21
22
Palash Dey. Manipulative elicitation - A new attack on elections with incomplete preferences.
Theor. Comput. Sci., 731:36 -- 49, 2018.
Palash Dey. Optimal bribery in voting. In Proc. International Conference on Autonomous Agents
and Multiagent Systems, AAMAS, 2019.
Palash Dey and Neeldhara Misra. On the exact amount of missing information that makes finding
possible winners hard. In Proc. 42nd International Symposium on Mathematical Foundations of
Computer Science, MFCS, pages 57:1 -- 57:14, 2017.
Palash Dey, Neeldhara Misra, and Y. Narahari. Detecting possible manipulators in elections.
In Proc. International Conference on Autonomous Agents and Multiagent Systems, AAMAS, pages
1441 -- 1450, 2015.
Palash Dey, Neeldhara Misra, and Y. Narahari. Kernelization complexity of possible winner and
coalitional manipulation problems in voting. Theor. Comput. Sci., 616:111 -- 125, 2016.
Palash Dey, Neeldhara Misra, and Y. Narahari. Kernelization complexity of possible winner and
coalitional manipulation problems in voting. Theor. Comput. Sci., 616:111 -- 125, 2016.
Palash Dey, Neeldhara Misra, and Y. Narahari. Frugal bribery in voting. Theor. Comput. Sci.,
676:15 -- 32, 2017.
Palash Dey, Neeldhara Misra, and Y. Narahari. Complexity of manipulation with partial inform-
ation in voting. Theor. Comput. Sci., 726:78 -- 99, 2018.
Palash Dey, Neeldhara Misra, and Y. Narahari. Parameterized dichotomy of choosing committees
based on approval votes in the presence of outliers. Theor. Comput. Sci., 2019.
P. Dey, N. Misra, S. Nath, and G. Shakya
21
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
Rod G Downey and Michael Ralph Fellows. Parameterized Complexity, volume 3. springer Heidel-
berg, 1999.
Gábor Erdélyi, Edith Hemaspaandra, and Lane A. Hemaspaandra. More natural models of elect-
oral control by partition. In Algorithmic Decision Theory - 4th International Conference, ADT 2015,
Lexington, KY, USA, September 27-30, 2015, Proceedings, pages 396 -- 413, 2015.
Gábor Erdélyi, Markus Nowak, and Jörg Rothe. Sincere-strategy preference-based approval
voting fully resists constructive control and broadly resists destructive control. Math. Log. Q.,
55(4):425 -- 443, 2009.
Gábor Erdélyi and Jörg Rothe. Control complexity in fallback voting. In Theory of Computing
2010, CATS 2010, Brisbane, Australia, January 2010, pages 39 -- 48, 2010.
Piotr Faliszewski, Edith Hemaspaandra, and Lane A. Hemaspaandra. Multimode control attacks
on elections. J. Artif. Intell. Res. (JAIR), 40:305 -- 351, 2011.
Piotr Faliszewski, Edith Hemaspaandra, and Lane A. Hemaspaandra. Weighted electoral control.
J. Artif. Intell. Res. (JAIR), 52:507 -- 542, 2015.
Piotr Faliszewski, Edith Hemaspaandra, Lane A. Hemaspaandra, and Jörg Rothe. Llull and
copeland voting broadly resist bribery and control. In Proc. Twenty-Second AAAI Conference on
Artificial Intelligence, July 22-26, 2007, Vancouver, British Columbia, Canada, pages 724 -- 730,
2007.
Piotr Faliszewski, Edith Hemaspaandra, Lane A. Hemaspaandra, and Jörg Rothe. Copeland
voting fully resists constructive control. In Algorithmic Aspects in Information and Management,
4th International Conference, AAIM 2008, Shanghai, China, June 23-25, 2008. Proceedings, pages
165 -- 176, 2008.
Piotr Faliszewski, Edith Hemaspaandra, Lane A. Hemaspaandra, and Jörg Rothe. Llull and cope-
land voting computationally resist bribery and constructive control. J. Artif. Intell. Res. (JAIR),
35:275 -- 341, 2009.
Piotr Faliszewski, Edith Hemaspaandra, Lane A. Hemaspaandra, and Jörg Rothe. The shield
that never was: societies with single-peaked preferences are more open to manipulation and
control.
In Proc. 12th Conference on Theoretical Aspects of Rationality and Knowledge (TARK-
2009), Stanford, CA, USA, July 6-8, 2009, pages 118 -- 127, 2009.
Piotr Faliszewski, Edith Hemaspaandra, Lane A. Hemaspaandra, and Jörg Rothe. The shield that
never was: Societies with single-peaked preferences are more open to manipulation and control.
Inf. Comput., 209(2):89 -- 107, 2011.
Zack Fitzsimmons, Edith Hemaspaandra, and Lane A. Hemaspaandra. Control in the presence of
manipulators: Cooperative and competitive cases. In IJCAI 2013, Proc. 23rd International Joint
Conference on Artificial Intelligence, Beijing, China, August 3-9, 2013, pages 113 -- 119, 2013.
Jörg Flum and Martin Grohe. Parameterized Complexity Theory, volume 3. Springer, 2006.
Edith Hemaspaandra, Lane A. Hemaspaandra, and Jörg Rothe. Hybrid elections broaden
complexity-theoretic resistance to control. Math. Log. Q., 55(4):397 -- 424, 2009.
Edith Hemaspaandra, Lane A. Hemaspaandra, and Jörg Rothe. Controlling candidate-sequential
elections. In ECAI 2012 - 20th European Conference on Artificial Intelligence. Including Prestigi-
ous Applications of Artificial Intelligence (PAIS-2012) System Demonstrations Track, Montpellier,
France, August 27-31 , 2012, pages 905 -- 906, 2012.
38
Lane A. Hemaspaandra, Rahman Lavaee, and Curtis Menton. Schulze and ranked-pairs voting
are fixed-parameter tractable to bribe, manipulate, and control. In International conference on
22
A Parameterized Perspective on Protecting Elections
Autonomous Agents and Multi-Agent Systems, AAMAS '13, Saint Paul, MN, USA, May 6-10, 2013,
pages 1345 -- 1346, 2013.
39
Joshua Letchford, Vincent Conitzer, and Kamesh Munagala. Learning and approximating the
optimal strategy to commit to.
In Algorithmic Game Theory, Second International Symposium,
SAGT 2009, Paphos, Cyprus, October 18-20, 2009. Proceedings, pages 250 -- 262, 2009.
40 Hong Liu, Haodi Feng, Daming Zhu, and Junfeng Luan. Parameterized computational complex-
ity of control problems in voting systems. Theor. Comput. Sci., 410(27-29):2746 -- 2753, 2009.
41 Hong Liu and Daming Zhu. Parameterized complexity of control problems in maximin election.
Inf. Process. Lett., 110(10):383 -- 388, 2010.
42 Hong Liu and Daming Zhu. Parameterized complexity of control by voter selection in maximin,
copeland, borda, bucklin, and approval election systems. Theor. Comput. Sci., 498:115 -- 123,
2013.
43
44
45
46
47
48
49
50
51
52
53
Krzysztof Magiera and Piotr Faliszewski. How hard is control in single-crossing elections? In
ECAI 2014 - 21st European Conference on Artificial Intelligence, 18-22 August 2014, Prague, Czech
Republic - Including Prestigious Applications of Intelligent Systems (PAIS 2014), pages 579 -- 584,
2014.
Nicholas Mattei, Nina Narodytska, and Toby Walsh. How hard is it to control an election by
breaking ties? In Torsten Schaub, Gerhard Friedrich, and Barry O'Sullivan, editors, ECAI, volume
263 of Frontiers in Artificial Intelligence and Applications, pages 1067 -- 1068. IOS Press, 2014.
Cynthia Maushagen and Jörg Rothe. Complexity of control by partitioning veto and maximin
elections and of control by adding candidates to plurality elections.
In ECAI 2016 - 22nd
European Conference on Artificial Intelligence, 29 August-2 September 2016, The Hague, The Neth-
erlands - Including Prestigious Applications of Artificial Intelligence (PAIS 2016), pages 277 -- 285,
2016.
David C McGarvey. A theorem on the construction of voting paradoxes. Econometrica, pages
608 -- 610, 1953.
Curtis Menton. Normalized range voting broadly resists control.
53(4):507 -- 531, 2013.
Theory Comput. Syst.,
Curtis Glen Menton and Preetjot Singh. Control complexity of schulze voting. In IJCAI 2013,
Proc. 23rd International Joint Conference on Artificial Intelligence, Beijing, China, August 3-9,
2013, pages 286 -- 292, 2013.
Tomasz Miasko and Piotr Faliszewski. The complexity of priced control in elections. Ann. Math.
Artif. Intell., 77(3-4):225 -- 250, 2016.
Rolf Niedermeier.
Tübingen, 2002.
Invitation to fixed-parameter algorithms. Habilitationschrift, University of
David C. Parkes and Lirong Xia. A complexity-of-strategic-behavior comparison between
schulze's rule and ranked pairs. In Proc. Twenty-Sixth AAAI Conference on Artificial Intelligence,
July 22-26, 2012, Toronto, Ontario, Canada., 2012.
Tomasz Put and Piotr Faliszewski. The complexity of voter control and shift bribery under
parliament choosing rules. Trans. Computational Collective Intelligence, 23:29 -- 50, 2016.
Jianxin Wang, Weimin Su, Min Yang, Jiong Guo, Qilong Feng, Feng Shi, and Jianer Chen.
Parameterized complexity of control and bribery for d-approval elections. Theor. Comput. Sci.,
595:82 -- 91, 2015.
P. Dey, N. Misra, S. Nath, and G. Shakya
23
54
55
56
Scott Wolchok, Eric Wustrow, Dawn Isabel, and J. Alex Halderman. Attacking the washington,
D.C. internet voting system.
In Financial Cryptography and Data Security - 16th International
Conference, FC 2012, Kralendijk, Bonaire, Februray 27-March 2, 2012, Revised Selected Papers,
pages 114 -- 128, 2012.
Lirong Xia and Vincent Conitzer. Determining possible and necessary winners under common
voting rules given partial orders. J. Artif. Intell. Res., 41(2):25 -- 67, 2011.
Yue Yin, Yevgeniy Vorobeychik, Bo An, and Noam Hazon. Optimally protecting elections.
In
Proc. Twenty-Fifth International Joint Conference on Artificial Intelligence, IJCAI 2016, New York,
NY, USA, 9-15 July 2016, pages 538 -- 545, 2016.
|
1112.1333 | 2 | 1112 | 2012-03-18T18:11:34 | Reaching an Optimal Consensus: Dynamical Systems that Compute Intersections of Convex Sets | [
"cs.MA"
] | In this paper, multi-agent systems minimizing a sum of objective functions, where each component is only known to a particular node, is considered for continuous-time dynamics with time-varying interconnection topologies. Assuming that each node can observe a convex solution set of its optimization component, and the intersection of all such sets is nonempty, the considered optimization problem is converted to an intersection computation problem. By a simple distributed control rule, the considered multi-agent system with continuous-time dynamics achieves not only a consensus, but also an optimal agreement within the optimal solution set of the overall optimization objective. Directed and bidirectional communications are studied, respectively, and connectivity conditions are given to ensure a global optimal consensus. In this way, the corresponding intersection computation problem is solved by the proposed decentralized continuous-time algorithm. We establish several important properties of the distance functions with respect to the global optimal solution set and a class of invariant sets with the help of convex and non-smooth analysis. | cs.MA | cs |
Reaching an Optimal Consensus: Dynamical
Systems that Compute Intersections of Convex Sets∗
Guodong Shi, Karl Henrik Johansson†and Yiguang Hong‡
Abstract
In this paper, multi-agent systems minimizing a sum of objective functions, where each
component is only known to a particular node, is considered for continuous-time dynamics
with time-varying interconnection topologies. Assuming that each node can observe a convex
solution set of its optimization component, and the intersection of all such sets is nonempty,
the considered optimization problem is converted to an intersection computation problem.
By a simple distributed control rule, the considered multi-agent system with continuous-time
dynamics achieves not only a consensus, but also an optimal agreement within the optimal
solution set of the overall optimization objective. Directed and bidirectional communications
are studied, respectively, and connectivity conditions are given to ensure a global optimal
consensus. In this way, the corresponding intersection computation problem is solved by the
proposed decentralized continuous-time algorithm. We establish several important properties
of the distance functions with respect to the global optimal solution set and a class of invariant
sets with the help of convex and non-smooth analysis.
Keywords: Multi-agent systems, Optimal consensus, Connectivity Conditions, Distributed
optimization, Intersection computation
1
Introduction
In recent years, multi-agent dynamics has been intensively investigated in various areas including
engineering, natural science, and social science. Cooperative control of multi-agent systems is an
∗This work has been supported in part by the Knut and Alice Wallenberg Foundation, the Swedish Research
Council, KTH SRA TNG, and the NNSF of China under Grant 61174071.
†G. Shi and K. H. Johansson are with ACCESS Linnaeus Centre, School of Electrical Engineering, Royal
Institute of Technology, Stockholm 10044, Sweden. Email: [email protected], [email protected]
‡Y. Hong is with Key Laboratory of Systems and Control, Institute of Systems Science, Chinese Academy of
Sciences, Beijing 100190, China. Email: [email protected]
1
active research topic, and rapid developments of distributed control protocols via interconnected
communication have been made to achieve the collective tasks, e.g., [16, 15, 12, 25, 10, 9, 20,
22, 17, 18]. However, fundamental challenges still lie in finding suitable tools to describe and
design the dynamical behavior of these systems and thus providing insights in their functioning
principles. Different from the classical control design, the multi-agent studies aim at fully
exploiting, rather than avoiding, interconnection between agents in analysis and synthesis in
order to deal with distributed design and large-scale information process.
Consensus is a basic problem of the study of multi-agent coordination, which usually requires
that all the agents achieve the same state, e.g., a certain relative position or velocity. To achieve
collective behavior, connectivity plays a key role in the coordination of multi-agent network,
and various connectivity conditions have been used to describe frequently switching topologies
in different cases. The "joint connection" or similar concepts are important in the analysis of
stability and convergence to guarantee a suitable convergence. Uniform joint-connection, i.e.,
the joint graph is connected during all intervals which are longer than a constant, has been
employed for different consensus problems [16, 15, 24, 19, 7]. On the other hand, [t,∞)-joint
connectedness, i.e., the joint graph is connected in the time intervals [t,∞), is necessary [22, 25],
and therefore the most general form to secure the global coordination.
Moreover, distributed optimization of a sum of convex objective functions,(cid:80)N
i=1 fi(z), where
each component fi is known only to node i, has attracted much attention in recent years, due
to its wide application in multi-agent systems and wireless networks [29, 30, 32, 31, 33]. A class
of subgradient-based incremental when some estimate of the optimal solution can be passed
over the network via deterministic or randomized iteration were studied in [29, 30, 34]. Then a
non-gradient-based algorithm was proposed in [33], where each node starts at its own optimal
solution and updates using a pairwise equalizing protocol. In view of multi-agent systems, the
local information transmitted over the neighborhood is usually limited to a convex combina-
tion of its neighbors [16, 15, 25]. Combining the ideas of consensus algorithms and subgradient
methods, a number of significant results were obtained. A subgradient method in combination
with consensus steps was given for solving coupled optimization problems with fixed undirected
topology in [31].
Then, an important work on multi-agent optimization was [27], where a
decentralized algorithm was proposed as a simple sum of an averaging (consensus) part and a
subgradient part, and convergence bounds for a distributed multi-agent model under various
connectivity conditions were shown. Constrained consensus and optimization were further stud-
2
ied in [28], where each agent was always restricted in its own convex set. A "projected consensus
algorithm" was presented to solve the constrained consensus problem in which each agent takes
averaging and projection steps alternatively, and it was generalized to "projected subgradient
algorithm" with optimization goal also took into consideration [28].
Most of the literature on optimization and consensus algorithms is in discrete time, and it is
usually hard for the considered agents to reach both consensus and optimum unless the weights
rule of the links, the step size in the iteration and the connectedness of the communication
graph are properly selected [27, 28, 30]. Few researchers have considered continuous-time agent
dynamics that solves a distributed optimization problem. However, dynamical system solution
to optimization problem is of great interest since a simple vector-field solution may provide
important geometrical insights. The classical Arrow-Hurwicz-Uzawa flow was shown to converge
to the set of saddle points for a constrained convex optimization problem [41]. Then in [42],
a simple and elegant continuous-time protocol was presented which solves linear programming
The goal of this paper is to establish a simple distributed continuous-time control law which
problems.
can ensure consensus and minimize(cid:80)N
that the intersection set,(cid:84)N
i=1 fi(z) asymptotically. Each optimal solution set, Xi of
optimization objective fi(z), is assumed to be a convex set observed only by node i. Assuming
i=1 Xi, is nonempty, the optimal solution set of the group objective
becomes this intersection set, and the considered optimization problem is then converted to a
distributed intersection computation problem.
In fact, computing several convex sets' inter-
section is a classical problem, and "alternating projection algorithm" was a standard solution,
in which the algorithm is carried out by iteratively projecting onto each set [38, 39, 40]. The
"projected consensus algorithm" presented in [28] can be viewed as its generalized version. The
intersection computation problem is also of interest in the study of computational geometry, a
branch of computer science [44, 45]. Hence, an important motivation for our work is to provide
a system-theoretic insight into the convergence properties of certain distributed optimization
problems. Similar to the continuous-time approximation of recursive algorithms [43] and con-
strained optimizations [41, 42], we establish a suitable dynamical model for such analysis. Also
by itself, the considered continuous-time distributed optimization problem has many applica-
tions, e.g., wireless resource allocation [29, 30], formation control [9, 22, 14], and mobile sensing
[10, 37].
In this paper, we present a simple dynamical system solution to this convex intersection
3
computation problem, as the sum of a consensus part and a projection part. Since this pro-
jection part can be viewed as a special subgradient information, this protocol is actually a
continuous-time version of the algorithm proposed in [27]. We show that an optimal consen-
sus (i.e., consensus within the global optimal solution set), can be achieved under time-varying
communications. Both directed and bidirectional cases are investigated, and sharp connectivity
conditions are obtained in the sense that a general optimal consensus will no longer hold for a
general model with weaker connectedness. Additionally, we use quite general weights rule which
allow the weight of each arc in the communication graph to depend on time or system state.
The rest of the paper is organized as follows. In Section 2, some preliminary concepts are
introduced. In Section 3, we formulate the considered optimal consensus problem, and the main
results are shown. Then, in Sections 4 and 5, convergence to the optimal solution set and global
consensus are analyzed, respectively, based on which the proofs of the main results are obtained.
Finally, in Section 6 concluding remarks are given.
2 Preliminaries
In this section, we introduce some notations and theories on graph theory [4], convex analysis
[1, 3] and nonsmooth analysis [5].
A directed graph (digraph) G = (V,E) consists of a finite set V of nodes and an arc set
E,
in which an arc is an ordered pair of distinct nodes of V. An element (i, j) ∈ E de-
scribes an arc which leaves i and enters j. A walk in digraph G is an alternating sequence
W : i1e1i2e2 . . . em−1im of nodes iκ and arcs eκ = (iκ, iκ+1) ∈ E for κ = 1, 2, . . . , m− 1. A walk is
called a path if the nodes of this walk are distinct, and a path from i to j is denoted as i → j. G
is said to be strongly connected if it contains path i → j and j → i for every pair of nodes i and
j. A digraph G is called to be bidirectional when for any two nodes i and j, (i, j) ∈ E if and only
if (j, i) ∈ E. Ignoring the direction of the arcs, the connectedness of a bidirectional digraph will
be transformed to that of the corresponding undirected graph. A time-varying graph is defined
as Gσ(t) = (V,Eσ(t)) with σ : [0, +∞) → Q as a piecewise constant function, where Q is a finite
set indicating all possible graphs. Moreover, the joint graph of Gσ(t) in time interval [t1, t2) with
t1 < t2 ≤ +∞ is denoted as G([t1, t2)) = ∪t∈[t1,t2)G(t) = (V,∪t∈[t1,t2)Eσ(t)).
A set K ⊂ Rm is said to be convex if (1 − λ)x + λy ∈ K whenever x ∈ K, y ∈ K and
0 ≤ λ ≤ 1. For any set S ⊂ Rm, the intersection of all convex sets containing S is called the
4
convex hull of S, denoted by co(S). The next lemma can be found in [2].
Lemma 2.1 Let K be a subset of Rm. The convex hull co(K) of K is the set of elements of
the form
m+1(cid:88)
i=1
x =
where λi ≥ 0, i = 1, . . . , m + 1 with(cid:80)m+1
i=1 λi = 1 and xi ∈ K.
Let K be a closed convex subset in Rm and denote xK
λixi,
= inf y∈K x − y as the distance
.
between x ∈ Rm and K, where · denotes the Euclidean norm. There is a unique element
PK(x) ∈ K satisfying x − PK(x) = xK associated to any x ∈ Rm [2]. The map PK is called
the projector onto K. We also have
(cid:104)PK(x) − x, PK(x) − y(cid:105) ≤ 0,
∀y ∈ K.
Moreover, PK has the following non-expansiveness property:
PK(x) − PK(y) ≤ x − y, x, y ∈ Rm.
Clearly, x2
K is continuously differentiable at point x, and (see [2])
∇x2
K = 2(x − PK(x)).
The following lemma was obtained in [22], which is useful in what follows.
Lemma 2.2 Suppose K ⊂ Rm is a convex set and xa, xb ∈ Rm. Then
(cid:104)xa − PK(xa), xb − xa(cid:105) ≤ xaK · xaK − xbK .
Particularly, if xaK > xbK, then
(cid:104)xa − PK(xa), xb − xa(cid:105) ≤ −xaK · (xaK − xbK).
(1)
(2)
(3)
(4)
(5)
Next, the upper Dini derivative of a continuous function h : (a, b) → R (−∞ ≤ a < b ≤ ∞)
at t is defined as
D+h(t) = lim sup
s→0+
h(t + s) − h(t)
s
.
When h is continuous on (a, b), h is non-increasing on (a, b) if and only if D+h(t) ≤ 0 for any
t ∈ (a, b). The next result is given for the calculation of Dini derivative (see [6, 24]).
5
Lemma 2.3 Let Vi(t, x) : R × Rd → R (i = 1, . . . , n) be C1 and V (t, x) = maxi=1,...,n Vi(t, x).
If I(t) = {i ∈ {1, 2, . . . , n} : V (t, x(t)) = Vi(t, x(t))} is the set of indices where the maximum is
reached at t, then D+V (t, x(t)) = maxi∈I(t)
Vi(t, x(t)).
Finally, consider a system
x = f (t, x),
(6)
where f : R × Rd → Rd is piecewise continuous in t and continuous in x. Let x(t) = x(t, t0, x0)
be a solution of (6) with initial condition x(t0) = x0. Then Ω0 ⊂ Rd is called a positively
invariant set of (6) if, for any t0 ∈ R and any x0 ∈ Ω0, x(t, t0, x0) ∈ Ω0 when t ≥ t0.
3 Problem Formulation and Main Results
In this section, we first define the considered optimal consensus problem. We propose a multi-
agent optimization model and a distributed control law to solve this optimization problem. Then
the main results are presented on connectivity conditions which can ensure an optimal consensus
globally.
3.1 Multi-agent Model
Consider a multi-agent system with agent set V = {1, 2, . . . , N}, for which the dynamics of each
agent is a first-order integrator:
xi = ui,
i = 1, . . . , N
(7)
where xi ∈ Rm represents the state of agent i, and ui is the control input.
The communication in the multi-agent network is modeled as a time-varying graph Gσ(t) =
(V,Eσ(t)). Moreover, node j is said to be a neighbor of i at time t when there is an arc (j, i) ∈ Eσ(t),
and Ni(σ(t)) represents the set of agent i's neighbors at time t. As usual in the literature
[15, 24, 22], an assumption is given to the variation of Gσ(t).
A1 (Dwell Time) There is a lower bound constant τD > 0 between two consecutive switching
time instants of σ(t).
We have the following definition.
Definition 3.1 (i) Gσ(t) is said to be uniformly jointly strongly connected (UJSC) if there exists
a constant T > 0 such that G([t, t + T )) is strongly connected for any t ≥ 0.
6
(ii) Assume that Gσ(t), t ≥ 0 is bidirectional. Gσ(t) is said to be infinitely jointly connected
(IJC) if G([t, +∞)) is connected for all t ≥ 0.
Remark 3.1 [t, +∞)-joint connectedness for all t ≥ 0 is equivalent to that there exists an
unbounded time sequence 0 ≤ t1 < ··· < tk < tk+1 < . . . such that G([tk, tk+1)) is connected for
all k = 1, 2, . . . . Note that it does not require an upper bound for tk+1 − tk in the definition.
The objective for this group of autonomous agents is to reach a consensus, and meanwhile to
cooperatively solve the following optimization problem
N(cid:88)
i=1
min
z∈Rm
fi(z)
(8)
where fi : Rm → R represents the cost function of agent i, observed by agent i only, and z is
a decision vector. We suppose the optimal solution set of each component fi exists, denoted
Xi
= {v fi(v) = min
.
z∈Rm
We impose the following assumptions.
fi(z)}.
A2 (Convexity) X1, . . . , XN , are closed convex sets.
A3 (Nonempty Intersection) X0
.
=
Xi is nonempty and bounded.
N(cid:84)
i=1
Remark 3.2 The assumption that each Xi is a convex set is quite general, and it is not hard
to see that this assumption will be satisfied as long as each fi is a convex function. Moreover,
since the intersection of convex sets is a convex set itself, X0 is a convex set with the convexity
of each Xi. Additionally, with A3, it is obvious to see that X0 is compact, and it is the optimal
solution set of (8).
3.2 Distributed Control
1 , . . . , xT
Denote x = (xT
arc (j, i), for i, j ∈ V. Then we present the following distributed control law:
N )T ∈ RmN and let the continuous function aij(x, t) > 0 be the weight of
(cid:88)
aij(x, t)(xj − xi) + PXi(xi) − xi, i = 1, . . . , N
ui =
(9)
j∈Ni(σ(t))
Remark 3.3 We write the arc weight aij(x, t) in a quite general form showing that this weight
function can be time-varying and may depend nonlinearly on the state. Note that this doesn't
mean global information is required for the control design.
7
Remark 3.4 When Xi can be observed by node i, PXi(xi(t)) − xi(t) can be easily obtained.
For instance, node i may first establish a local coordinate system, and then construct a function
h(z) = z2
to compute ∇h(z) within this coordinate system. Then by (3), we have PXi(xi(t))−
xi(t) = −1/2∇h(z)z=xi(t).
Xi
Another assumption is made on each aij(x, t), i, j = 1, 2, ..., N .
A4 (Weights Rule) There are a∗ > 0 and a∗ > 0 such that
a∗ ≤ aij(x, t) ≤ a∗,
t ∈ R+, x ∈ RmN .
In this paper, we assume that Assumptions A1-A4 always hold. With (7) and (9), the closed
loop system is expressed by
(cid:88)
xi =
j∈Ni(σ(t))
aij(x, t)(xj − xi) + PXi(xi) − xi, i = 1, . . . , N.
(10)
Remark 3.5 By the non-expansiveness property (2), the convex projection PK(z) is continuous
for all z ∈ Rm for any closed convex set K ⊆ Rm. Therefore, a solution of (10) exists at least
over a finite interval for any initial condition x(t0). Note that the solution is not necessarily
unique. As will be shown in Remark 4.1, it also exists in [t0, +∞).
Remark 3.6 Since the projection term can be viewed as a subgradient for the special case
fi(z) = z2
which has the form of the sum of a consensus term and a subgradient term. On the other hand,
/2, (10) is actually a continuous-time version of the algorithm proposed in [27],
Xi
in [28], a "projected consensus algorithm" was presented to solve the same intersection com-
putation problem in which each agent takes consensus and projection steps alternatively. Note
that there is some essential difference between (10) and the "projected consensus algorithm" in
[28], because (10) takes advantage of the consensus and projection information at the same time
instant. It is not hard to construct examples in which each node i would never enter its own set
Xi along the trajectories of (10).
Let x(t) be the trajectory of (10) with initial condition x0 = x(t0) = (xT
1 (t0), . . . , xT
N (t0))T ∈
RmN . Then the considered optimal consensus is defined as following (see Fig. 1).
Definition 3.2 (i) A global optimal set convergence of (10) is achieved if for all x0 ∈ RmN , we
have
t→+∞xi(t)X0 = 0,
lim
i = 1, . . . , N.
(11)
8
Figure 1: The goal of the agents is to achieve a consensus in X0.
(ii) A global consensus of (10) is achieved if for all x0 ∈ RmN , we have
t→+∞xi(t) − xj(t) = 0,
lim
i, j = 1, . . . , N.
(12)
(iii) A global optimal consensus is achieved of (10) if both (i) and (ii) hold.
Remark 3.7 It is easy to find that, based on the analysis methods we provide, all the results
obtained in this paper will still hold if the control law (10) is replaced by
aij(x, t)(xj − xi) + bi(xi, t)(PXi(xi) − xi), i = 1, . . . , N
(cid:88)
xi =
j∈Ni(σ(t))
for some scalar functions 0 < b∗ ≤ bi(xi, t), i = 1,··· , N with b∗ > 0 being a constant. Here we
just choose the form of (10) to make the statements and proofs simplified.
3.3 Main Results
In this subsection, we present the main results on optimal consensus.
First the following conclusion is our main result for directed graphs.
Theorem 3.1 System (10) achieves a global optimal consensus if Gσ(t) is UJSC.
We say the communications over the considered multi-agent network are bidirectional if
Gσ(t) is a bidirectional graph for all t ≥ t0. Note that, this does not imply that the arc weights,
aij(x, t), i, j = 1, . . . , N , are symmetric. Then we have the following main result on optimal
consensus for the bidirectional case.
9
Theorem 3.2 System (10) with bidirectional communications achieves a global optimal consen-
sus if (and in general only if ) Gσ(t) is IJC.
Theorem 3.2 shows that the connectedness conditions to reach an optimal consensus can be
relaxed for bidirectional communications without requiring a uniform bound of the length of
intervals in the definition of connectivities.
Remark 3.8 Let us explain what "in general only if" means in Theorems 3.2. Clearly, the
connectivity condition proposed in Theorem 3.2 is not a necessary condition to ensure a global
optimal consensus for a particular optimization problem (8). However, in regard to a global
In fact, as long as (cid:84)N
optimal consensus for all possibilities of X1, . . . , XN , simple examples could show that this IJC
assumption is also necessary using the same idea studying state agreement problem in [25, 22].
i=1 Xi is not a singleton, it can be easily shown that consensus cannot be
guaranteed for all initial conditions. Therefore, from this perspective, Theorem 3.2 gives "sharp"
connectivity conditions for a global optimal consensus of system (10).
Remark 3.9 If A3, the nonempty intersection assumption, is removed, control law (10) becomes
i=1 Xi).
In this case, under proper connectivity assumptions (even each node cannot always obtain the
a special case of the target aggregation controller studied in [22] with respect to co((cid:83)N
information of Xi), it can be shown that (10) will lead the network to converge into co((cid:83)N
[22]. The dynamics within co((cid:83)N
i=1 Xi)
i=1 Xi) can be complicated, and the optimal consensus will
fail since there is no longer a simple expression of X∗, the real optimal solution set of (8).
However, we guess that in this case the control law (10) still implies a suboptimal convergence
such that there will be a constant B, which does not depend on the initial condition, satisfying
lim supt→∞ xi(t)X∗ ≤ B under UJSC connectivity conditions.
In order to prove Theorems 3.1 and 3.2, on one hand, we have to prove all the agents
converge to the global optimal solution set, i.e., X0; and, on the other hand, we have to verify
that a consensus is also achieved. In fact, the convergence analysis is quite challenging, due to
the nonlinearity nature of each weight function aij(x, t) and the convex projection part in the
control law. In the following two sections, we will focus on the optimal solution set convergence
and the consensus analysis, respectively, by which complete the proofs for Theorems 3.1 and
3.2.
10
4 Optimal Set Convergence
In this section, we prove the optimal solution set convergence for system (10). We first establish
a method to analyze the distance between the agents and the global optimal set with the help
of convex analysis, and then the convergence to X0 for all the agents is proposed under directed
and bidirectional communications, respectively.
4.1 Distance Function
Define di(t) = xi(t)2
and let
X0
d(t) = max
i∈V di(t)
be the maximum among all the agents. Although d(t) may not be continuously differentiable,
it is still continuous. Thus, we can analyze the Dini derivative of d(t) to study its convergence
property. Moreover, it is easy to see that d(t) is locally Lipschitz. Then the Dini derivative of
d(t) is finite for any t.
We prove several elementary lemmas for the following analysis. At first, the following lemma
indicates that d(t) is nonincreasing.
Lemma 4.1 D+d(t) ≤ 0 for all t ≥ 0.
Proof. According to (3), one has
d
dt
(cid:88)
di(t) = 2(cid:104)xi − PX0(xi), xi(cid:105)
= 2(cid:104)xi − PX0(xi),
j∈Ni(σ(t))
aij(x, t)(xj − xi) + PXi(xi) − xi(cid:105).
(13)
Then, based on Lemma 2.3 and denoting I(t) as the set containing all the agents that reach the
maximum in the definition of d(t) at time t, we obtain
D+d(t) = max
i∈I(t)
= 2 max
i∈I(t)
d
di(t)
dt
[(cid:104)xi − PX0(xi),
(cid:88)
j∈Ni(σ(t))
aij(xj − xi) + PXi(xi) − xi(cid:105)].
(14)
Furthermore, for any i ∈ I(t), according to (5) of Lemma 2.2, one has
(cid:104)xi − PX0(xi), xj − xi(cid:105) ≤ 0
for any j ∈ Li(σ(t)) since it always holds that xjX0 ≤ xiX0.
(15)
11
Moreover, in light of (1), we obtain
(cid:104)PXi(xi) − PX0(xi), PXi(xi) − xi(cid:105) ≤ 0
(16)
since we always have PX0(xi) ∈ Xi for all i = 1, . . . , N . Therefore, it is easy to see that for any
i ∈ V,
(cid:104)xi − PX0(xi), PXi(xi) − xi(cid:105) ≤ (cid:104)xi − PXi(xi), PXi(xi) − xi(cid:105) = −xi2
Xi.
Thus, with (14), (15) and (17), one has
D+d(t) ≤ 2 max
i∈I(t)
[−xi2
Xi] ≤ 0.
Then the proof is completed.
(17)
(18)
(cid:3)
Remark 4.1 According to Lemma 4.1, {y y2
≤ d(t0)} is a positively invariant set for system
(10). Since X0 is compact, {y y2
≤ d(t0)} is also compact. This leads to that each solution
of (10) exists in [t0, +∞). Moreover, if the weight functions aij, i, j = 1, . . . , N , are only state-
dependent, the continuity implies that there will be a∗ ≥ a∗ > 0 such that
X0
X0
a∗ ≤ aij(x(t)) ≤ a∗,
∀t > 0, i, j = 1, 2, ...N
(19)
along trajectory x(t) of system (10). In this case, A4 follows automatically, and then needs not
to be assumed.
With Lemma 4.1, for any initial condition, there exists a constant d∗ ≥ 0 such that limt→∞ d(t) =
d∗. Clearly, the optimal solution set convergence will be achieved for system (10) if and
only if d∗ = 0. Furthermore, since it always holds that di(t) ≤ d(t), there exist constants
0 ≤ θi ≤ ηi ≤ d∗, i = 1, . . . , N such that
lim inf
t→∞ di(t) = θi,
lim sup
t→∞ di(t) = ηi.
To establish the optimal set convergence, we also need the following lemmas, whose proofs
can be found in the Appendix.
Lemma 4.2 Assume that θi = ηi = d∗, i = 1, . . . , N . Then we have limt→+∞ xi(t)Xi = 0 for
all i = 1, . . . , N .
Lemma 4.3 Assume that either Gσ(t) being UJSC or Gσ(t) being IJC with bidirectional commu-
nications. Then θi = ηi = d∗ for all i = 1, 2, . . . , N .
12
Remark 4.2 If the network communication graph is undirected, i.e., i ∈ Nj(σ(t)) if and only if
j ∈ Ni(σ(t)) with aij(x, t) ≡ aji(x, t), i, j = 1, . . . , N , then according to (13) and (17), we have
di(t) ≤ 2
aij(x, t)(cid:104)xi − PX0(xi), xj − xi(cid:105) − 2
xi2
Xi
N(cid:88)
d
dt
i=1
=
+
=
i=1
j∈Ni(σ(t))
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
j=1
i=1
(cid:88)
(cid:88)
(cid:88)
(cid:88)
i=1
j∈Ni(σ(t))
j∈Ni(σ(t))
i∈Nj (σ(t))
N(cid:88)
i=1
N(cid:88)
i=1
xi2
Xi
aij(x, t)(cid:104)xi − PX0(xi), xj − xi(cid:105)
aji(x, t)(cid:104)xj − PX0(xj), xi − xj(cid:105) − 2
aij(x, t)(cid:104)xi − xj + PX0(xj) − PX0(xi), xj − xi(cid:105) − 2
N(cid:88)
i=1
xi2
Xi.
Furthermore, based on (1) and (2), we obtain
(cid:104)xi − xj + PX0(xj) − PX0(xi), xj − xi(cid:105) ≤ −xi − xj2 + xi − xj · PX0(xj) − PX0(xi) ≤ 0
for all i, j = 1, . . . , N . Therefore, we have
di(t) ≤ −2
N(cid:88)
i=1
xi(t)2
Xi,
d
dt
N(cid:88)
(cid:90) ∞
N(cid:88)
i=1
which implies
immediately based on Lemma 4.1.
i=1
0
xi(t)2
Xi <
N
2
d(t0)
(20)
As a result, with (20), we can apply Barbalat's lemma on xi(t)2
, and then it follows
Xi
immediately that limt→+∞ xi(t)Xi = 0, i = 1, . . . , N without the assumptions of Lemma 4.2.
Remark 4.3 Note that, Lemmas 4.1 and 4.2 hold without requiring any connectivity of the
system communication graph.
4.2 Directed Graphs
The following conclusion is for optimal set convergence with directed communications.
Proposition 4.1 System (10) achieves the global optimal solution set convergence if Gσ(t) is
UJSC.
13
Proof. According to Lemmas 4.2 and 4.3, we have limt→∞ xi(t)Xi = 0, i = 1, . . . , N . As a
result, for any ε > 0, there exists T1(ε) > 0 such that when t ≥ T1,
xi(t)Xi ≤ ε,
Take t1 = T1 and k0 ∈ V. Defining hk0(t)
i = 1, . . . , N.
.
= maxi∈V xi(t)Xk0
(21)
, similarly to the analysis of
(14), we have that for all t,
d
dt
k0(t) ≤ 2hk0(t) · max
h2
i=1,...,N
xi(t)Xi,
which implies D+hk0(t) ≤ ε, t ≥ t1. Thus, hk0(t) ≤ hk0(t1) + (N − 1)T0ε, t ∈ [t1, t1 + (N − 1)T0].
Since Gσ(t) is UJSC, we can find a node k1 such that (k0, k1) ∈ Eσ(t) for t ∈ [t1, t1 + τD) ⊆
[t1, t1 + T0), where T0 = T + 2τD. In light of Lemma 2.2 and (21), we have
(xk1), xk0 − xk1(cid:105) + 2(cid:104)xk1 − PXk0
= 2ak1k0(x, t)(cid:104)xk1 − PXk0
xk1(t)2
Xk0
d
dt
(xk1),
ak1j(xj − xk1) + PXk1
(xk1) − xk1(cid:105)
(cid:88)
j∈Nk1 (σ(t))\k0
≤ −2a∗xk1(t)Xk0
(xk1(t)Xk0
− ε) + 2(N − 2)a∗xk1(t)Xk0
) + 2xk1(t)Xk0
· ε,
· (hk0(t1)
t ∈ [t1, t1 + τD),
(22)
(23)
(24)
1 + [2(N − 2)(N − 1)T0 + 1]a∗
w0 =
; M0 =
.
Proceeding the estimation in time interval [t1 + τD, t1 + (N − 1)T0] will lead to
D+xk1(t)Xk0
for all t ∈ [t1 + τD, t1 + (N − 1)T0]. This implies
≤ −(N−1)a∗xk1(t)Xk0
a∗ + (N − 2)a∗
+(N−1)a∗[hk0(t1)+(N−1)T0ε]+ε, t ∈ [t1+τD, t1+(N−1)T0],
xk1(t)Xk0
≤ ς0(w0hk0(t1) + M0ε) + (1 − ς0)(w0hk0(t1) + (N − 1)T0ε +
≤ w0hk0(t1) + M0ε,
t ∈ [t1 + τD, t1 + (N − 1)T0]
ε
(N − 1)a∗ )
(25)
14
+ (N − 1)T0ε − xk1(t)Xk0
from which we obtain that for any t ∈ [t1, t1 + τD),
D+xk1(t)Xk0
Therefore, noticing that xk1(t1)Xk0
≤ −(a∗ + (N − 2)a∗)xk1(t)Xk0
one has
xk1(t1 + τD)Xk0
where
≤ ν0xk1(t1)Xk0
≤ w0hk0(t1) + M0ε,
(N − 2)a∗ + a∗ν0
(N − 2)a∗ + a∗
+ (N − 2)a∗[hk0(t1) + (N − 1)T0ε] + (1 + a∗)ε.
≤ hk0(t1) + (N − 1)T0ε and denoting ν0 = e−(a∗+(N−2)a∗)τD ,
+ (1 − ν0) · (N − 2)a∗[hk0(t1) + (N − 1)T0ε] + (1 + a∗)ε
a∗ + (N − 2)a∗
where
ς0 = e−(N−1)2a∗T0;
M0 =
2 + [3(N − 1)2T0 + 1]a∗
a∗ + (N − 2)a∗
.
(26)
Further, continuing the analysis on time interval [t1 + T0, t1 + 2T0], k2 can be found with a
neighbor in {k0, k1} during [t2, t2 + τD) ⊆ [t1 + T0, t1 + 2T0]. An upper bound for xk2(t)Xk0
can
be similarly obtained as
xk2(t)Xk0
≤ w1hk0(t1) + 2 M0ε, t ∈ [t2 + τD, t1 + (N − 1)T0]
(27)
where w1 = (N−2)a∗+a∗ν2
(N−2)a∗+a∗ .
0
Next, respectively, we repeat the analysis on time intervals [t1 + 2T0, t1 + 3T0], . . . , [t1 + (N −
2)T0, t1 + (N − 1)T0] for k3, . . . , kN−1 ∈ V, and we finally reach
xi(t1 + (N − 1)T0)Xk0
≤ wN−1hk0(t1) + (N − 1) M0ε, i = 1, . . . , N,
which implies
hk0(t2) ≤ wN−1hk0(t1) + (N − 1) M0ε,
(28)
(29)
where t2 = t1 + (N − 1)T0 and 0 < wN−1 = (N−2)a∗+a∗νN
(N−2)a∗+a∗ < 1.
0
Denoting w∗ = wN−1 and tn+1 = tn + (N − 1)T0 for n = 2, . . . , and by the same analysis on
time intervals [tn, tn+1], n = 2, . . . , one has
hk0(tn) ≤ wn−1∗ hk0(t1) +
≤ wn−1∗ hk0(t1) +
n−1(cid:88)
wj−1∗
(N − 1) M0ε
j=1
(N − 1) M0
1 − w∗
· ε
(30)
Since ε in (30) can be arbitrarily small, we see that limt→∞ xi(t)Xk0
1, . . . , N , which immediately implies limt→∞ xi(t)X0 = 0. The proof is completed.
= 0 for all i, k0 =
(cid:3)
4.3 Bidirectional Graphs
The following conclusion is for optimal set convergence under bidirectional graphs.
Proposition 4.2 System (10) achieves the optimal solution set convergence with bidirectional
communications if Gσ(t) is IJC.
Proof. Suppose d∗ > 0. According to Lemmas 4.2 and 4.3, we have that for all i = 1, . . . , N ,
t→∞xi(t)X0 =
lim
√
d∗,
15
t→∞xi(t)Xi = 0.
lim
(31)
.
= {y√
d∗ + ε ≤ yX0 ≤ √
This implies, for any ε > 0, we have that xi(t) ∈ B0(ε) ∩ Bi(ε) for sufficiently large t, where
B0(ε)
= {yyXi ≤ ε}, i = 1, . . . , N . Then we see
.
from (13) that the derivative of xi(t)2
is globally Lipschitz. Therefore, based on Barbalat's
lemma, we know
d∗ + ε} and Bi(ε)
X0
(32)
= {(i, j)(i, j) ∈ Eσ(t) for infinitely long time}. Then G∞ = (V,E∞) is connected
X0 = 0.
lim
t→∞
Define E∞ .
xi(t)2
d
dt
since G([t, +∞)) is connected for all t ≥ 0.
Let N∞
i be the neighbor set of node i in graph G∞. With Lemma 2.2, (31) and (32) yield
that for any i = 1, . . . , N and j ∈ N∞
,
i
t→∞(cid:104)xi(t) − PX0(xi(t)), xj(t) − xi(t)(cid:105) = 0.
lim
(33)
Taking i0 ∈ V, we define two hyperplanes:
H1(t)
= {v(cid:104)xi0(t) − PX0(xi0(t)), v − xi0(t)(cid:105) = 0};
.
H2(t)
= {v(cid:104)xi0(t) − PX0(xi0(t)), v − PX0(xi0(t))(cid:105) = 0}.
.
Then ∀j ∈ N∞
i0
, (33) implies that
t→∞xj(t)H1(t) = 0;
lim
t→∞xj(t)H2(t) =(cid:112)g∗,
lim
which leads to
t→∞PX0(xj(t)) − PH2(t)(xj(t)) = 0.
(34)
Because G∞ is connected, we can repeat the analysis over the network, then arrive that (34)
holds for all j = 1, . . . , N .
lim
Let Ci0(t) = co{PXi0
Therefore, with (31) and (34) and according to the structure of H1(t) and H2(t), there will
(xi0(t)), PX0(x1(t)), . . . , PX0(xN (t))}. Then Ci0(t) ⊆ Xi0,∀t ≥ 0.
be a point z∗ ∈(cid:84)N
i0=1 Ci0(t) ⊆ X0 for sufficiently large t such that
(cid:104)xi0(t) − PX0(xi0(t)), z∗ − PX0(xi0(t))(cid:105) > 0,
which contradicts (1). Therefore, d∗ > 0 does not hold, and then the optimal set convergence
(cid:3)
follows.
16
5 Global Consensus
In this section, we present the consensus analysis. In order to show the consensus, we have to
present a clear estimation of the influence on state agreement by terms xi−PXi(xi), i = 1, . . . , N .
We first introduce a class of positively invariant set for system (10) which characterizes the
agreement property in Subsection 5.1. Then the consensus analysis is investigated for directed
and bidirectional communication cases, respectively in Subsection 5.2.
5.1
Invariant Set
We define a multi-projection function: Pikik−1...i1 : Rm →(cid:83)N
i=1 Xi with i1, . . . , ik ∈ {1, . . . , N}, k =
1, 2, . . . , by
Pikik−1...i1(x) = PXik
PXik−1
. . . PXi1
(x).
Particularly, P∅ is denoted by P∅(x) = x as the case for k = 0. Let
= {Pikik−1...i1i1, . . . , ik ∈ {1, . . . , N}, k = 0, 1, 2, . . .}
.
Γ
be the set which contains all the multi-projection functions we define.
= co{P (y)y ∈ K, P ∈ Γ}.
.
, based on a similar analysis as the proof of Lemma 4.1, it is
Furthermore, let K be a convex set in Rm, and define ∆K as ∆K
Denoting g(t) = max
xi(t)2
i=1,...,N
∆K
not hard to find that
D+g(t) ≤ 0,
t ≥ 0.
This implies, g(t) ≡ 0 for all t ≥ t0 once we have g(t0) = 0, which leads to the following
conclusion immediately (see Fig. 2).
Lemma 5.1 Let K be a convex set in Rm. Then ∆N
K
= ∆K × ··· × ∆K is positively invariant
.
for system (10).
We next establish an important property of the constructed invariant set ∆N
K.
Lemma 5.2 yK ≤ 2 maxz∈K zX0, ∀y ∈ ∆K.
m+1(cid:88)
Proof. With Lemma 2.1, any y ∈ ∆K has the following form
y =
λiP (cid:104)i(cid:105)(zi),
i=1
17
Figure 2: Constructing an invariant set from K = co{y1, y2}.
where (cid:80)m+1
i=1 λi = 1 with λi ≥ 0, P (cid:104)i(cid:105) ∈ Γ and zi ∈ K, i = 1, . . . , m + 1. Then, by the
non-expansiveness property (2), we have that for any z ∈ Rm and P∗ ∈ Γ,
PX0(z) − P∗(z) = P∗(PX0(z)) − P∗(z) ≤ PX0(z) − z = zX0.
i=1
This leads to
m+1(cid:88)
λiP (cid:104)i(cid:105)(zi) − m+1(cid:88)
λizi ≤ m+1(cid:88)
≤ m+1(cid:88)
which implies the conclusion because(cid:80)m+1
i=1
i=1
i=1
zX0,
≤ 2 max
z∈K
i=1 λizi ∈ K.
λizi − P (cid:104)i(cid:105)(zi)
λizi − PX0(zi) +
m+1(cid:88)
i=1
λiPX0(zi) − P (cid:104)i(cid:105)(zi)
(cid:3)
Now we are ready to reach the global consensus for system (10). Let us focus on each
coordinate, and denote x(cid:96)
i(t) as the (cid:96)-th coordinate of xi(t). Moreover, let
φ(t) = min
i∈V {x(cid:96)
i(t)}, ϕ(t) = max
i∈V {x(cid:96)
i(t)}
be the minimum and the maximum within all the agents. Denote H(t) (cid:44) ϕ(t) − φ(t). Then a
consensus is achieved for system (10) if and only if limt→∞ H(t) = 0.
In the next subsection, we will prove the global consensus for system (10) with directed and
bidirectional communications, respectively by showing that limt→∞ H(t) = 0.
5.2 Consensus Analysis
In this subsection, we propose the consensus analysis. First we study the directed case.
18
Proposition 5.1 System (10) achieves a global consensus if Gσ(t) is UJSC.
Proof. Based on Proposition 4.1, we have that limt→∞ xi(t)X0 = 0 for all i = 1, . . . , N .
Therefore, for any ε > 0, there exists T1(ε) > 0 such that, when t ≥ T1,
As a result, according to Lemma 5.2, for any y ∈ ∆co{x1(t),...,xN (t)} with t > T1(ε), we have
xi(t)X0 ≤ 1
2
ε, i = 1, . . . , N
(35)
dist(y, co{x1(t), . . . , xN (t)}) ≤ ε.
Moreover, by Lemma 5.1, we see that xi(t) ∈ ∆co{x1(t),...,xN (t)}, i = 1, . . . , N for all t ≤ t ≤ ∞,
which implies that for all t ≥ t ≥ T1, we have
dist(xi(t), ∆co{x1(t),...,xN (t)}) ≤ ε,
i = 1, . . . , N.
(36)
We divide the following proof into three steps.
Step 1: Take t1 = T1 with x(cid:96)
i0
to x(cid:96)
i0
(t) during t ∈ [t1, t1 + (N − 1)T0].
Based on (36), we see that for all T1 ≤ t < t ≤ ∞
(t1) = φ(t1) and denote T0 = T + 2τD. In this step, we give bound
φ(t) ≥ φ(t) − ε; ϕ(t) ≤ ϕ(t) + ε.
Noting the fact that
i0(t) ≤ −(N − 1)a∗x(cid:96)
x(cid:96)
i0(t) + (N − 1)a∗(ϕ(t1) + ε) + ε, t ≥ t1,
d
dt
we obtain
(37)
(38)
i0(t) ≤ µ1 (cid:44) ς0φ(t1) + (1 − ς0)ϕ(t1) +
x(cid:96)
(39)
Step 2: Since Gσ(t) is UJSC, we can find i1 ∈ V and t1 ≥ t1 such that (i0, i1) ∈ Gσ(t) for t ∈
[t1, t1+τD) ⊆ [t1, t1+T0). In this step, we give bound to x(cid:96)
(t1) during t ∈ [t1+τD, t1+(N−1)T0].
· ε, t ∈ [t1, t1 + (N − 1)T0].
i1
(N − 1)a∗ + 1
(N − 1)a∗
Similarly to the analysis of (22), when t ∈ [t1, t1 + τD), one has
d
dt
which yields
i1(t) ≤ a∗(µ1 − x(cid:96)
x(cid:96)
i1(t)) + (N − 2)a∗(ϕ(t1) + ε − x(cid:96)
i1(t)) + ε,
i1(t1 + τD) ≤ ν0(ϕ(t1) + ε) + (1 − ν0) × a∗µ1 + (N − 2)a∗(ϕ(t1) + ε) + ε
x(cid:96)
a∗ + (N − 2)a∗
= (1 − w0)ς0φ(t1) + [1 − (1 − w0)ς0]ϕ(t1) + L0ε
(cid:44) θ1
(40)
19
after some simple manipulations by combining (39) and (40), where L0 = 1 +
Then, applying (38) on node i1 during t ∈ [t1 + τD, t1 + (N − 1)T0] will lead to
[a∗+(N−2)a∗](N−1) .
N
i1(t) ≤ ς0θ1 + (1 − ς0)ϕ(t1) +
x(cid:96)
(N − 1)a∗ + 1
(N − 1)a∗
· ε
= m1φ(t1) + [1 − m1]ϕ(t1) + L0ε,
(41)
for all t ∈ [t1 + τD, t1 + N T0], where m1 = (1 − w0)ς 2
0 and L0 = ς0 L0 + 1 +
Step 3: We proceed the analysis for i2, . . . , iN−1 with mk = ((1 − w0)ς 2
Denoting t2 (cid:44) t1 + (N − 1)T0, we obtain
1
(N−1)a∗ .
0 )k, k = 2, . . . , N − 1.
i(t2) ≤ mN−1φ(t1) + (1 − mN−1)ϕ(t1) + (N − 1)L0ε, = 0, . . . , N − 1,
x(cid:96)
which implies
ϕ(t2) ≤ mN−1φ(t1) + (1 − mN−1)ϕ(t1) + (N − 1)L0ε.
(37) and (43) lead to
H(t2) ≤ mN−1φ(t1) + (1 − mN−1)ϕ(t1) + (N − 1)L0ε − (φ(t1) − ε)
= (1 − mN−1)H(t1) + [(N − 1)L0 + 1]ε
(42)
(43)
(44)
Define a time sequence T1 = t1 < t2 < . . . with tk = tk−1 + (N − 1)T0. Applying the same
analysis on each interval [tk−1, tk) will lead to
H(tk) ≤ (1 − mN−1)H(tk−1) + [(N − 1)L0 + 1]ε, k = 1, 2, . . . .
(45)
As a result, we obtain
H(tk+1) ≤ (1 − mN−1)kH(t1) +
k−1(cid:88)
(1 − mN−1)j[(N − 1)L0 + 1]ε
≤ (1 − mN−1)kH(t1) +
j=0
(N − 1)L0 + 1
mN−1
· ε,
k = 1, 2, . . .
(46)
Therefore, noting the fact that 0 < mN−1 < 1, (37) and (46) yield
lim sup
t→∞ H(t) ≤ (2 +
(N − 1)L0 + 1
mN−1
) · ε.
Then limt→∞ H(t) = 0 since ε can be arbitrarily small. This completes the proof.
(cid:3)
Then the global consensus for bidirectional case is proved by the following conclusion.
Proposition 5.2 System (10) achieves a global consensus with bidirectional communications if
Gσ(t) is IJC.
20
Proof. Take t1 = T1 with x(cid:96)
i0
(t1) = φ(t1) as the proof of Proposition 5.1. Then (36) and (37)
still hold.
Denote the first time when i0 has at least one neighbor during t ≥ t1 as t1, and denote the
neighbor set of i0 for t ∈ [t1, t1 + τD) as V1. Next, we show the bound for i0 and j ∈ V1 during
t ∈ [t1, t1 + τD) .
Note that when i0 has no neighbor during t ∈ (t1, s) for t1 ≤ s ≤ ∞, one has that for any
t ∈ [t1, s),
Then, we see that
x(cid:96)
i0(t) − x(cid:96)
i0(s) ≤ ε.
(47)
i0(t) ≤ µ1 (cid:44) ς0φ(t1) + (1 − ς0)ϕ(t1) +
x(cid:96)
(N − 1)a∗ + 1
(N − 1)a∗
· ε
for all t ∈ [t1, t1 + τD], where ς0 = e−(N−1)a∗τD
By similar analysis with (40), we have that for any j ∈ V1,
j(t1 + τD) ≤ θ1 (cid:44) m1φ(t1) + (1 − m1)ϕ(t1) + L0ε
x(cid:96)
(48)
with m1 = (1 − w0)ς0.
When there is no link between V \ ({i0} ∪ V1) and {i0} ∪ V1 for t ∈ [t1 + τD, t), applying
Lemma 5.1 on the subsystem formed by nodes in {i0} ∪ V1, (36) leads to
j(t) ≤ θ1 + ε, t ∈ [t1 + τD, t), j ∈ {i0} ∪ V1.
x(cid:96)
(49)
Therefore, defining t2 as the first moment during t ∈ [t1 + τD,∞) when there is an edge
between j ∈ {i0} ∪ V1 and V \ ({i0} ∪ V1), we have
j(t) ≤ ς0(θ1 + ε) + (1 − ς0)ϕ(t1) +
x(cid:96)
(N − 1)a∗ + 1
(N − 1)a∗
· ε
(50)
for t ∈ [t2, t2 + τD].
Denoting V2 = {k ∈ Vthere is a link between k and {i0} ∪ V1 at t2}, bounds for x(cid:96)
k(t2 +
τD), k ∈ V2 can be similarly given by
k(t2 + τD) ≤ m2φ(t1) + (1 − m2)ϕ(t1) + L0ε,
x(cid:96)
(51)
where m2 = ((1 − w0)ς 2
0 )2.
with mj0 = ((1 − w0)ς 2
0 )j0, we have
Next, V3, . . . ,Vj0 can be defined until V = {i0} ∪ V1 ∪ ··· ∪ Vj0 since Gσ(t) is JC. Moreover,
i(tj0 + τD) ≤ mj0φ(t1) + (1 − mj0)ϕ(t1) + L0(N − 1)ε, = 1, . . . , N.
x(cid:96)
(52)
21
Therefore, denoting t2 (cid:44) tj0 + τD, we obtain
ϕ(t2) ≤ mj0φ(t1) + (1 − mj0)ϕ(t1) + L0(N − 1)ε,
which implies
H(t2) ≤ (1 − mj0)H(t1) + (N − 1)L0ε.
(53)
(54)
Then limt→∞ H(t) = 0 holds by similar analysis as the proof of Proposition 5.1. This
(cid:3)
completes the proof.
With Propositions 4.1, 4.2, 5.1 and 5.2, it is straightforward to see that the main results of
the paper, Theorems 3.1 and 3.2 hold.
6 Conclusions
This paper addressed an optimal consensus problem for multi-agent systems. With jointly
connected graphs, the considered multi-agent system achieved not only consensus, but also op-
timum by agreeing within the global solution set of a sum of objective functions. Assuming that
each agent can observe the projection information onto the solution set of its own optimization
component and the intersection of all solution sets is nonempty, the original unconstrained opti-
mization problem was converted to an intersection computation problem. Control laws applied
to the agents were simple and distributed. The results showed that a global optimization prob-
lem can be solved over a multi-agent network under time-varying communications and limited
interactions. Future work includes randomization in the nodes' decision-making and event-based
methods in the optimization algorithm design.
Appendix
A.1 Proof of Lemma 4.2
Based on the definitions of θi and ηi, when θi = ηi = d∗ holds for all i = 1, . . . , N , one has
t→+∞ di(t) = d∗,
lim
i = 1, . . . , N
Thus, for any ε > 0, there exists T1(ε) > 0 such that, when t ≥ T1(ε),
di(t) ∈ [d∗ − ε, d∗ + ε],
i = 1, . . . , N.
(55)
22
When d∗ = 0, then it is easy to see that the conclusion holds because xi(t)Xi ≤ xi(t)X0
for all t ≥ 0. Therefore, we just assume d∗ > 0 in the following.
According to (13) and (17), it is not hard to find that
di(t) ≤ −2xi2
Xi + 2(cid:104)xi − PX0(xi),
d
dt
aij(x, t)(xj − xi)(cid:105).
(cid:88)
j∈Ni(σ(t))
(56)
(57)
(58)
(59)
(60)
(61)
(62)
(cid:3)
Furthermore, based on (55) and Lemmas 2.2 and 4.1, one has that when t > T1(ε),
(cid:104)xi − PX0(xi), xj − xi(cid:105) ≤ xiX0 · xiX0 − xjX0 ≤ 2(cid:112)d(t0)ε
for all i = 1, . . . , N and j ∈ Ni(σ(t)).
If the conclusion does not hold, there exist a node i0 and a constant M0 > 0 such that
xi0(tk)Xi0
= M0
for a time serial 0 < t1 < ··· < tk < tk+1 < . . . with limk→∞ tk+1 = ∞. Noting the fact that
there is a constant L > 0 such that a − b ≤ L for all a, b ∈ {y y2
≤ d(t0)} since X0 is
compact, we have that for all for all i = 1, . . . , N ,
X0
(cid:12)(cid:12)(cid:12)(cid:12) d
dt
xi(t)2
Xi
(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2
(cid:12)(cid:12)(cid:12)(cid:12) =
≤ 2d(t0) + 2(N − 1)a∗(cid:112)d(t0)L.
≤ 2xi(t)2
j∈Ni(σ(t))
aij(cid:104)xi − PXi(xi), xj − xi(cid:105) − 2xi(t)2
Xi
X0 + 2(N − 1)a∗xi(t)X0 · xj(t) − xi(t)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Denoting τ0 (cid:44)
and according to (58) and (59), we obtain
2
M0
(cid:113)
d(t0)+(N−1)a∗√
xi0(tk)2
Xi0
d(t0)L
≥ 1
2
M 2
0 ,
t ∈ [tk, tk + τ0],
which leads to
d
dt
di0(t) ≤ − 1
2
0 + 2(cid:112)d(t0)ε ≤ − 1
M 2
4
M 2
0 ,
t ∈ [tk, tk + τ0].
√
for all tk > T1 and ε ≤ M 2
0
d(t0)
8
. As a result, we have
di0(tk + τ0) ≤ d(t0) − M 2
0 τ0
4
+ ε
Therefore, (62) contradicts (55) when ε < M 2
0 τ0
8
, which completes the proof.
A.2 Proof of Lemma 4.3
23
We prove the conclusion by contradiction. Suppose there exists a node i0 ∈ V such that
0 ≤ θi0 < ηi0 ≤ d∗. Then for any ε > 0, there exists T1(ε) > 0 such that, when t ≥ T1(ε),
di(t) ∈ [0, d∗ + ε],
i = 1, . . . , N.
(63)
(cid:113) 1
Take ζ0 =
2 (θi0 + ηi0). Then there exists a time serial
0 < t1 < ··· < tk < . . .
with limt→∞ tk = ∞ such that xi0(tk)X0 = ζ0 for all k = 1, 2, . . . .
According to (63) and Lemma 2.2, we have that for all t > tk0,
(cid:88)
di0(t) ≤ 2
d
dt
j∈Ni0 (σ(t))
ai0j(x, t)(cid:104)xi0 − PX0(xi0), xj − xi0(cid:105)
≤ 2(N − 1)a∗xi0(t)X0(
√
d∗ + ε − xi0(t)X0),
which will lead to
D+xi0(t)X0 ≤ −(N − 1)a∗xi0(t)X0 + (N − 1)a∗√
d∗ + ε.
As a result, for t ∈ [s,∞) with s ≥ tk0, we have
√
xi0(t)X0 ≤ e−(N−1)a∗(t−s)xi0(s)X0 + (1 − e(N−1)a∗(t−s))
d∗ + ε.
(64)
(65)
We divide the following proof into two cases: directed communications and bidirectional com-
munications.
Directed Case: Denote T0 = T + 2τD. Since Gσ(t) is UJSC, it is not hard to find that there exist
i1 ∈ V and t1 such that (i0, i1) ∈ Gσ(t) for t ∈ [t1, t1 + τD) ⊆ [tk0, tk0 + T0). Then based on (65),
we obtain
xi0(t)X0 ≤ ξ1 (cid:44) ς0ζ0 + (1 − ς0)
√
d∗ + ε, t ∈ [tk0, tk0 + (N − 1)T0],
(66)
where ς0 = e−(N−1)2a∗T0. Thus, for t ∈ [t1, t1 + τD), one has
ai1j(cid:104)xi1 − PX0(xi1), xj − xi1(cid:105) + ai1i0(cid:104)xi1 − PX0(xi1), xi0 − xi1(cid:105)]
√
d∗ + ε − xi1(t)X0) − a∗xi1(t)X0(xi1(t)X0 − ξ1),
(67)
(cid:88)
di1(t) ≤ 2[
d
dt
j∈Ni1 (σ(t))\i0
≤ 2(N − 2)a∗xi1(t)X0(
which leads to
D+xi1(t)X0 ≤ −((N − 2)a∗ + a∗)xi1(t)X0 + (N − 2)a∗√
d∗ + ε + a∗ξ1.
(68)
24
Therefore, we obtain
xi1(t)X0 ≤ e−((N−2)a∗+a∗)(t−t1)xi1(t1)X0 + (1− e−((N−2)a∗+a∗)(t−t1))· (N − 2)a∗√
for t ∈ [t1, t1 + τD), which implies
d∗ + ε + a∗ξ1
(N − 2)a∗ + a∗
xi1(t1 + τD)X0 ≤ ζ1 (cid:44) w0
d∗ + ε + (1 − w0)ξ1,
(69)
√
where w0 is defined in (24). Furthermore, applying the same analysis of (65) on node i1, one
has that when t ∈ [t1 + τD,∞),
√
xi1(t)X0 ≤ e−(N−1)a∗(t−(t1+τD))ζ1 + (1 − e−(N−1)a∗(t−(t1+τD)))
d∗ + ε,
(70)
Combing (66), (69) and (70), we obtain
√
xi1(t)X0 ≤ m1ζ0 + (1 − m1)
d∗ + ε,
for all t ∈ [t1 + τD, tk0 + (N − 1)T0], where m1 = (1 − w0)ς 2
0 .
0 < ς0 < m1 < 1.
(71)
(71) also holds for i0 since
We can proceed to find a node i2 ∈ V such that there is an arc leaving from {i0, i1} entering
i2 in G([tk0 + T0, tk0 + 2T0)) because Gσ(t) is uniformly jointly strongly connected. Meanwhile,
similar analysis will result in estimations for agent i2 with the form (71) by m2 = ((1− w0)ς 2
0 )2.
Repeating similar analysis on time intervals [tk0 + 2T0, tk0 + 3T0], . . . , [tk0 + (N − 2)T0, tk0 +
(N − 1)T0] respectively, and finally, by mN−1 = ((1 − w0)ς 2
0 )N−1, we obtain
√
xi(tk0 + N T0)X0 ≤ mN−1ζ0 + (1 − mN−1)
d∗ + ε, i = 1, . . . , N,
(72)
which yields
√
d(tk0 + N T0) ≤ mN−1ζ0 + (1 − mN−1)
d∗ + ε.
Note that, (73) contradicts the definition of d∗ since mN−1ζ0 + (1 − mN−1)
for sufficiently small ε. The conclusion holds.
Bidirectional Case: When i0 has no neighbor for t ∈ [tk0, s], by (17) we see that
√
d∗ + ε <
(73)
√
d∗
xi0(t)X0 ≤ xi0(tk0)X0 = ζ0, t ∈ [tk0, s].
(74)
Denote the first moment when i0 has at least one neighbor during t ∈ [tk0,∞) as t1, and
denote the neighbor set of i0 for t ∈ [t1, t1 + τD) as V1. Then, by a similar analysis as (66), one
has
√
xi0(t)X0 ≤ ξ1 (cid:44) ς0ζ0 + (1 − ς0)
d∗ + ε, t ∈ [t1, t1 + τD]
(75)
25
with ς0 = e−(N−1)a∗τD . Thus, according to the same process by which we obtain (69), one also
obtains
√
d∗ + ε,
(76)
xi1(t1 + τD)X0 ≤ w0
where m1 = ς0(1 − w0).
√
d∗ + ε + (1 − w0) ξ1 = m1ζ0 + (1 − m1)
Similarly, we can define t2 as the first moment when there is another node connected to
{i0} ∪ V1 during t ≥ t1 + τD. Let V2 be the node set which connect to {i0} ∪ V1 at t2. Since we
have the dwell time for σ(t), without loss of generality, we can always assume that all the links
between {i0}∪V1 and V2 last for at least τD time starting from t2. Moreover, similar estimations
will lead to
√
xi2(t2 + τD)X0 ≤ m2ζ0 + (1 − m2)
d∗ + ε
for all i2 ∈ {i0} ∪ V1 ∪ V2, where m2 = (ς0(1 − w0))2.
Furthermore, since Gσ(t) is JC, we can always proceed the upper process until V = {i0} ∪
V1 ∪ ··· ∪ Vj0, and then we obtain
√
xi(tj0 + τD)X0 ≤ mj0ζ0 + (1 − mj0)
d∗ + ε,
with mj0 = (ς0(1 − w0))j0, which contradicts the definition of d∗. Then the conclusion holds for
bidirectional case.
The proof is completed.
References
(cid:3)
[1] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY: Cambridge University
Press, 2004.
[2] J. Aubin and A. Cellina, Differential Inclusions. Berlin: Speringer-Verlag, 1984.
[3] R. T. Rockafellar, Convex Analysis. New Jersey: Princeton University Press, 1972.
[4] C. Godsil and G. Royle, Algebraic Graph Theory. New York: Springer-Verlag, 2001.
[5] N. Rouche, P. Habets, and M. Laloy, Stability Theory by Liapunov's Direct Method. New
York: Springer-Verlag, 1977.
[6] J. Danskin, "The theory of max-min, with applications," SIAM J. Appl. Math., vol. 14,
641-664, 1966.
26
[7] D. Cheng, J. Wang, and X. Hu, "An extension of LaSalle's invariance principle and its
applciation to multi-agents consensus," IEEE Trans. Autom. Control, vol. 53, no. 7, pp.
1765-1770, 2008.
[8] F. Clarke, Yu.S. Ledyaev, R. Stern, and P. Wolenski, Nonsmooth Analysis and Control
Theory. Berlin: Speringer-Verlag, 1998.
[9] S. Martinez, J. Cortes, and F. Bullo, " Motion coordination with distributed information,"
IEEE Control Syst. Mag., vol. 27, no. 4, pp. 75-88, 2007.
[10] R. Olfati-Saber, "Flocking for multi-agent dynamic systems: algorithms and theory," IEEE
Trans. Autom. Control, vol. 51, no. 3, pp. 401-420, 2006.
[11] M. Cao, D. A. Spielman and A. S. Morse, "A lower bound on convergence of a distributed
network consensus algorithm," in Proc. IEEE Conf. on Decision and Control, pp. 2356-
2361, 2005.
[12] R. Olfati-Saber and R. Murray, "Consensus problems in the networks of agents with switch-
ing topology and time delays," IEEE Trans. Autom. Control, vol. 49, no. 9, pp. 1520-1533,
2004.
[13] J. Fax and R. Murray, "Information flow and cooperative control of vehicle formations,"
IEEE Trans. Automatic Control, vol. 49, no. 9, 1465-1476, 2004.
[14] H. G. Tanner, A. Jadbabaie, G. J. Pappas, "Flocking in fixed and switching networks,"
IEEE Trans. Autom. Control, vol. 52, no.5, pp. 863-868, 2007.
[15] A. Jadbabaie, J. Lin, and A. S. Morse, "Coordination of groups of mobile autonomous
agents using nearest neighbor rules," IEEE Trans. Autom. Control, vol. 48, no. 6, pp.
988-1001, 2003.
[16] J. Tsitsiklis, D. Bertsekas, and M. Athans, "Distributed asynchronous deterministic and
stochastic gradient optimization algorithms," IEEE Trans. Autom. Control, vol. 31, no. 9,
pp. 803-812, 1986.
[17] F. Xiao and L. Wang, "State consensus for multi-agent systems with swtiching topologies
and time-varying delays," Int. J. Control, vol. 79, no. 10, pp. 1277-1284, 2006.
27
[18] F. Xiao and L. Wang, "Asynchronous consensus in continuous-time multi-agent systems
with switching topology and time-varying delays," IEEE Trans. Autom. Control, vol. 53,
no. 8, pp. 1804-1816, 2008.
[19] Y. Hong, L. Gao, D. Cheng, and J. Hu, "Lyapuov-based approach to multiagent systems
with switching jointly connected interconnection," IEEE Trans. Autom. Control, vol. 52,
no. 5, pp. 943-948, 2007.
[20] W. Ren and R. Beard. Distributed Consensus in Multi-vehicle Cooperative Control,
Springer-Verlag, London, 2008.
[21] W. Ren and R. Beard, "Consensus seeking in multi-agent systems under dynamically chang-
ing interaction topologies," IEEE Trans. Autom. Control, vol. 50, no. 5, pp. 655-661, 2005.
[22] G. Shi and Y. Hong, "Global target aggregation and state agreement of nonlinear multi-
agent systems with switching topologies," Automatica, vol. 45, no. 5, pp. 1165-1175, 2009.
[23] G. Shi, Y. Hong and K. H. Johansson, "Connectivity and set tracking of multi-agent systems
guided by multiple moving leaders," IEEE Trans. Autom. Control, vol. 57, no. 3, pp. 663-
676, 2012.
[24] Z. Lin, B. Francis, and M. Maggiore, "State agreement for continuous-time coupled nonlin-
ear systems," SIAM J. Control Optim., vol. 46, no. 1, pp. 288-307, 2007.
[25] L. Moreau, "Stability of multiagent systems with time-dependent communication links,"
IEEE Trans. Autom. Control, vol. 50, no. 2, pp. 169-182, 2005.
[26] A. Nedi´c, A. Olshevsky, A. Ozdaglar, and J. N. Tsitsiklis, "Distributed subgradient methods
and quantization effects," in Proc. IEEE Conf. on Decision and Control, Cancun, Mexico,
pp. 4177-4184, 2008.
[27] A. Nedi´c and A. Ozdaglar, "Distributed subgradient methods for multi-agent optimization,"
IEEE Trans. Autom. Control, vol. 54, no. 1, pp. 48-61, 2009.
[28] A. Nedi´c, A. Ozdaglar and P. A. Parrilo, "Constrained consensus and optimization in multi-
agent networks," IEEE Trans. Autom. Control, vol. 55, no. 4, pp. 922-938, 2010.
[29] M. Rabbat and R. Nowak, "Distributed optimization in sensor networks," in IPSN'04, pp.
20-27, 2004.
28
[30] B. Johansson, M. Rabi, and M. Johansson, "A simple peer-to-peer algorithm for distributed
optimization in sensor networks, in Proc. IEEE Conf. on Decision and Control, New Or-
leans, LA, pp. 4705-4710, 2007.
[31] B. Johansson, T. Keviczky, M. Johansson, and K. H. Johansson, "Subgradient methods
and consensus algorithms for solving convex optimization problems," in Proc. IEEE Conf.
on Decision and Control, Cancun, Mexico, pp. 4185-4190, 2008.
[32] S. S. Ram, A. Nedi´c, and V. V. Veeravalli, "Stochastic incremental gradient descent for
estimation in sensor networks," in Proc. Asilomar Conference on Signals, Systems, and
Computers, Pacific Grove, pp. 582-586, 2007.
[33] J. Lu, C. Y. Tang, P. R. Regier, and T. D. Bow, "Gossip algorithms for convex consensus
optimization over networks," IEEE Trans. Autom. Control, vol. 56, pp. 2917-2923, 2011.
[34] S. S. Ram, A. Nedi´c, and V. V. Veeravalli, "Incremental stochastic subgradient algorithms
for convex optimization," SIAM Journal on Optimization, vol. 20, no. 2, 691-717, 2009.
[35] B. Johansson, A. Speranzon, M. Johansson, K. H. Johansson, "On decentralized negotiation
of optimal consensus," Automatica, vol. 44, pp. 1175-1179, 2008.
[36] B. Johansson, M. Rabi and M. Johansson. A randomized incremental subgradient method
for distributed optimization in networked systems. SIAM Journal on Optimization, vol. 20,
no. 3, pp. 1157-1170, 2009.
[37] D. Jakoveti´c, J. Xavier and J. M. F. Moura, "Cooperative convex optimization in net-
worked systems: augmented lagrangian algorithms with directed gossip communication,"
http://arxiv.org/abs/1007.3706, 2011.
[38] N. Aronszajn, "Theory of reproducing kernels," Trans. Amer. Math. Soc., vol. 68, no. 3,
pp. 337-404, 1950.
[39] L. G. Gubin, B. T. Polyak, and E. V. Raik, "The method of projections for finding the
common point of convex sets," U.S.S.R Comput. Math. Math. Phys., vol. 7, no. 6, pp.
1211-1228, 1967.
[40] F. Deutsch, "Rate of convergence of the method of alternating projections," in Parametric
Optimization and Approximation, B. Brosowski and F. Deutsch, Eds. Basel, Switzerland:
Birkhauser, vol. 76, pp. 96-107, 1983.
29
[41] K. Arrow, L. Hurwicz and H. Uzawa, Studies in linear and non-linear programming. Stan-
ford University Press, 1958.
[42] R. W. Brockett, "Dynamical systems that sort lists, diagonalize matrices, and solve linear
programming problems," vol. 146, pp. 79-91, 1991.
[43] L. Ljung, "Analysis of recursive stochastic algorithms," IEEE Trans. Autom. Control, vol.
22, no. 4, pp. 551-575, 1977.
[44] M. Shamos and D. Hoey, "Geometric intersection problems," in 17th IEEE Conf. Founda-
tions of Computer Science (FOCS 1976), pp. 208-215, 1976.
[45] J. L. Bentley and T. Ottmann, "Algorithms for reporting and counting geometric intersec-
tions", IEEE Trans. Computers, vol. C-28, no. 9, pp. 643647, 1979.
30
|
1802.08534 | 2 | 1802 | 2018-04-14T16:34:29 | Weighted Double Deep Multiagent Reinforcement Learning in Stochastic Cooperative Environments | [
"cs.MA",
"cs.AI",
"cs.LG"
] | Recently, multiagent deep reinforcement learning (DRL) has received increasingly wide attention. Existing multiagent DRL algorithms are inefficient when facing with the non-stationarity due to agents update their policies simultaneously in stochastic cooperative environments. This paper extends the recently proposed weighted double estimator to the multiagent domain and propose a multiagent DRL framework, named weighted double deep Q-network (WDDQN). By utilizing the weighted double estimator and the deep neural network, WDDQN can not only reduce the bias effectively but also be extended to scenarios with raw visual inputs. To achieve efficient cooperation in the multiagent domain, we introduce the lenient reward network and the scheduled replay strategy. Experiments show that the WDDQN outperforms the existing DRL and multiaent DRL algorithms, i.e., double DQN and lenient Q-learning, in terms of the average reward and the convergence rate in stochastic cooperative environments. | cs.MA | cs | Weighted Double Deep Multiagent Reinforcement Learning in Stochastic
Cooperative Environments
Yan Zheng1, Jianye Hao1, Zongzhang Zhang2,
1 Tianjin University, Tianjin, China
2 Soochow University, Suzhou, China
[email protected], [email protected], [email protected]
8
1
0
2
r
p
A
4
1
]
A
M
.
s
c
[
2
v
4
3
5
8
0
.
2
0
8
1
:
v
i
X
r
a
Abstract
Recently, multiagent deep reinforcement learning
(DRL) has received increasingly wide attention.
Existing multiagent DRL algorithms are inefficient
when facing with the non-stationarity due to agents
update their policies simultaneously in stochastic
cooperative environments. This paper extends the
recently proposed weighted double estimator to
the multiagent domain and propose a multiagent
DRL framework, named weighted double deep Q-
network (WDDQN). By utilizing the weighted dou-
ble estimator and the deep neural network, WD-
DQN can not only reduce the bias effectively but
also be extended to scenarios with raw visual in-
puts. To achieve efficient cooperation in the mul-
tiagent domain, we introduce the lenient reward
network and the scheduled replay strategy. Ex-
periments show that the WDDQN outperforms the
existing DRL and multiaent DRL algorithms, i.e.,
double DQN and lenient Q-learning, in terms of the
average reward and the convergence rate in stochas-
tic cooperative environments.
1 Introduction
The goal of reinforcement learning (RL) is to learn an optimal
behavior within an unknown dynamic environment, usually
modeled as a Markov decision process (MDP), through trial
and error [Sutton and Barto, 1998]. Over the past years, deep
RL (DRL) has achieved great successes.
It has been prac-
tically shown to successfully master various complex prob-
lems [Mnih et al., 2013; Mnih et al., 2015]. To a large extent,
these successes can be credited to the incorporation of the
experience replay and target network that stabilizes the net-
work training [Mnih et al., 2016; Mnih et al., 2013; Mnih
et al., 2015; Schaul et al., 2016; Van Hasselt et al., 2016;
Wang et al., 2016].
Approaches like [Bloembergen et al., 2011; Matignon
et al., 2007; Matignon et al., 2012; Panait et al., 2006;
Wei and Luke, 2016] have been proposed by extending Q-
learning to address the coordination problems in cooperative
multiagent systems. They are able to achieve coordination
in relatively simple cooperative multiagent system. However,
none of them has been combined with deep learning tech-
niques.
Recently, increasing wide attention has been drawn in em-
ploying DRL in multiagent environments. Unfortunately,
these multiagent DRL algorithms still suffer from two intrin-
sic difficulties in the interactive environments [Gupta et al.,
2017; Lanctot et al., 2017; Matignon et al., 2012]: stochas-
ticity due to the noisy reward signals; and non-stationarity
due to the dynamicity of coexisting agents. The stochas-
ticity introduces additional biases in estimation, while the
non-stationarity harms the effectiveness of experience replay,
which is crucial for stabilizing deep Q-networks. These two
characteristics result in the lack of theoretical convergence
guarantees of most multiagent DRL algorithms and amplify
the difficulty of finding the optimal Nash equilibriums, espe-
cially in cooperative multiagent problems.
This work focuses on learning algorithms of independent
learners (ILs) in cooperative multiagent systems. Here, we
assume that agents are unable to observe other agents' ac-
tions and rewards [Claus and Boutilier, 1998]; they share a
common reward function and learn to maximize the com-
mon expected discounted reward (a.k.a. return). To handle
the stochastic and non-stationary challenges in the multiagent
systems, we propose the weighted double deep Q-network
(WDDQN) with two auxiliary mechanisms, the lenient re-
ward network and the scheduled replay strategy, to help ILs
in finding the optimal policy that maximizes the common re-
turn.
Our contributions are three-fold. First, we extend weighted
double Q-learning (WDQ) [Zhang et al., 2017], a state-of-
the-art traditional RL method, to the multiagent DRL set-
tings. Second, we introduce a lenient reward network in-
spired by the lenient Q-learning [Palmer et al., 2018; Panait
et al., 2006]. Third, we modify the exisitin prioritized expe-
rience replay strategy to stabilize and speed up the learning
process in complex multiagent problems with raw visual in-
puts. Empirical results demonstrate that on a fully coopera-
tive multiagent problem WDDQN with the new mechanisms
indeed contribute to increasing the algorithm's convergence,
decreasing the instability and helping ILs to find an optimal
policy simultaneously.
[0, 1] such that ∃s ∈ S,∃a ∈ A,(cid:80)
2 Preliminaries
This section briefly introduces the definition of cooperative
Markov games, Q-learning and its variants.
2.1 Cooperative Markov Game
Markov (stochastic) games, as an extension of repeated
games and MDPs, provide a commonly used framework for
modeling interactions among agents. They can be formalized
as a tuple < N, S, A, T r, R1, ...RN , γ >. Here, N is the
number of players (or agents), S is the set of states, A =
A1 × ... × AN is the joint action set, where Ai is the action
space of player i, T r is the transition function S × A × S →
s(cid:48)∈S T r(s, a, s(cid:48)) = 1, Ri
is the reward function S × A → R for player i, and γ ∈ [0, 1]
is a discount factor. The state s is assumed to be observable
for all players. A fully cooperative Markov game is a spe-
cific type of Markov games where all agents receive the same
reward under the same outcome, and thus share the same best-
interest action.
2.2 Q-learning and Its Variants
Q-learning
is based on the core idea of temporal difference (TD) learn-
ing [Sutton, 1988] and is well suited for solving sequen-
tial decision making problems [Claus and Boutilier, 1998;
Watkins, 1989]. Q-learning tries to find an accurate estimator
of the Q-Values, i.e. Qt(s, a), for state-action pairs [Claus
and Boutilier, 1998]. Each Q-value is an estimate of the dis-
counted sum of future rewards that can be obtained at time
t through selecting action a in state s. The iterative update
formula is outlined in Equation 1:
a(cid:48) Q(s(cid:48), a(cid:48))− Q(s, a)], (1)
Q(s, a) ← Q(s, a) + α[r + γ max
where r is the immediate reward and α ∈ [0, 1) is the learn-
ing rate. The updating process always chooses the action a(cid:48)
with the maximum Q value and updates Q with the saved Q
value. Once the process terminates, an optimal policy can
be obtained by selecting the action with the maximum Q-
value in each state [Bellman, 1957]. However, Q-learning
uses a single estimator to estimate E{maxa(cid:48) Q(s(cid:48), a(cid:48))},
to
which has been proved to be greater than or equal
maxa(cid:48) E{Q(s(cid:48), a(cid:48))} [Smith and Winkler, 2006]. Thus, a pos-
itive bias always exists in the single estimator.
Deep Q-Network (DQN)
extends Q-learning with neural network to solve complex
problems with extensive state spaces.
It uses an online
neural network parametrized by θ to approximate the vec-
tor of action values Q(s,·; θ) for each state s, and a target
network parameterized by θ(cid:48) which is periodically copied
from θ to reduce oscillation during training. The neu-
ral network is optimized by minimizing the difference be-
tween the predicted value Q(st, at; θt) and the target value
t = rt+1 + γ maxa Q(st+1, a; θ(cid:48)
Y Q
t), using experienced sam-
ples (st, at, rt+1, st+1) drawn from a replay memory. To
minimize the difference, the parameters of the network are
updated along with the direction of the target value Y Q
es-
t
timated by experienced samples (st, at, rt+1, st+1) drawn
from a replay memory using the following formula:
t − Q(st, at; θt))∇θtQ(st, at; θt)], (2)
θt+1 = θt + αE[(Y Q
where ∇θtQ(st, at; θt) is the gradient. Both the replay mem-
ory and the target network help DQN to stabilize learning and
can dramatically improve its performance. However, like tab-
ular Q-learning, using the single maximum estimator is prone
to cause overestimating, leading to poor performance in many
situations.
Double Q-learning
uses the double estimator to ease the overestimation. The
double estimator selects the action with the maximum Q
value and evaluates the Q values of different actions sepa-
rately in turn [Hasselt, 2010]. The double Q-learning algo-
rithm stores two Q-values, denoted QU and QV , and replaces
the estimated value maxa(cid:48) Q(s(cid:48), a(cid:48)) in Equation 1 with the
combination QU (s(cid:48), arg maxa(cid:48) QV (s(cid:48), a(cid:48))). Unfortunately,
Hasselt [Hasselt, 2010] proved that though the double esti-
mator can overcome the overestimation issue, a negative bias
is introduced in the same time which may harm the resulting
algorithm's performance and effectiveness.
Double DQN
incorporates the idea of double Q-learning into DQN to avoid
the overestimation [Van Hasselt et al., 2016]. It uses two sets
of Q-networks Q(s, a; θ) and Q(s, a, θ(cid:48)): one for selecting
action and the other for estimating the target Q-value. At each
time the Q-network Q(s, a; θ) is updated using the following
target value:
t ≡ Rt+1 + γQ(st+1, arg max
Y Q
a
Q(st+1, a, θt); θ(cid:48)
t).
(3)
By leveraging the above two Q-networks to select and eval-
uate the Q-values symmetrically in turn, this algorithm takes
advantage of the double estimator to reduce the overestima-
tion of Q values and lead to better performance in a variety of
complex RL scenarios.
Weighted Double Q-learning (WDQ)
uses a dynamic heuristic value β, which depends on a con-
stant c, to balance between the overestimation of the single
estimator and the underestimation of the double estimator
during the iterative Q-value update process:
Q(s, a)U,W DQ = βQU (s, a∗) + (1 − β)QV (s, a∗),
(4)
where a linear combination of QU and QV is used for
updating Q-value. When a∗ is chosen by QU , i.e., a∗ ∈
arg maxa QU (s, a), QU (s, a∗) will be positively biased and
QV (s, a∗) will be negatively biased, and vice versa. β ∈
[0, 1] balances between the positive and negative biases. Ex-
periments on tabular MDP problems show that more accu-
rate value estimation can indeed boost Q-learning's perfor-
mance. However, it is still not clear whether this idea can be
extended to the end-to-end DRL framework to handle high-
dimensional problems.
Lenient Q-learning
[Potter and De Jong, 1994] updates the policies of multi-
ple agents towards an optimal joint policy simultaneously by
letting each agent adopt an optimistic dispose at the initial
exploration phase. This has been empirically proved to be
efficient at increasing the likelihood of discovering the op-
timal joint policy in stochastic environments and avoiding
agents gravitating towards a sub-optimal joint policy [Bloem-
bergen et al., 2015; Palmer et al., 2018; Panait et al., 2006;
Wei and Luke, 2016].
During training, lenient agents keep track of the tempera-
ture Tt(s, a) for each state-action pair (s, a) at time t, which
is initially set to a defined maximum temperature value and
used for measuring the leniency l(s, a) as follows:
l(st; at) = 1 − e−K∗Tt(st,at),
(5)
where K is a constant determining how the temperature af-
fects the decay in leniency. As suggested by [Wei and Luke,
2016], Tt(st, at) is decayed using a discount factor κ ∈ [0, 1]
and Tt+1(st, at) = κTt(st, at). Given the TD error δ =
t − Qt(st, at; θt), the iterative update formula of lenient
Y Q
Q-learning is defined as follows:
(cid:26) Q(st, at) + αδ
Q(st, at)
Q(st, at) =
prioritized experience replay [Schaul et al., 2016] in multi-
agent DRL leads to poor performance, as stored transitions
can become outdated because agents update their policies si-
multaneously. To address this, we propose a scheduled replay
strategy to enhance the benefit of prioritization by adjusting
the priority for transition sample dynamically. In the remain-
der of this section, we will describe these facets in details.
3.1 Network Architecture
WDDQN outlined in Algorithm 1 is adapted from WDQ by
leveraging neural network as the Q-value approximator to
handle problems with high-dimensional state spaces. The
overall network architecture of the algorithm is depicted in
Fig. 1. To reduce the estimation bias, WDDQN uses the com-
bination of two estimators, represented as Deep Q-networks
QU and QV with the same architecture, to select action
a = maxa(cid:48) QU (s,a(cid:48))+QV (s,a(cid:48))
(line 5). Besides, the target
QTarget(s, a) (lines 12 and 17) used for Q-value updating in
back-propagation is replaced with a weighted combination as
well (lines 11 and 16). Intuitively, the combination balances
between the overestimation and the underestimation. In addi-
tion, we also propose to use a reward approximator and an ef-
ficient scheduled replay strategy in WDDQN to achieve bias
reduction and efficient coordination in multiagent stochastic
environments.
2
if δ > 0 or x > l(st, at),
otherwise.
(6)
The random variable x ∼ U (0, 1) is used to ensure that a neg-
ative update δ is performed with a probability 1 − l(st, at).
Due to the initial state-action pairs being visited more of-
ten than the later ones, the temperature values for states
(cid:80)
close to the initial state can decay rapidly. One solution
to address this is to fold the average temperature ¯T (s(cid:48)) =
ai∈A T (s(cid:48), ai) for next state s(cid:48) into the temperature that
1A
is being decayed for (st, at) [Wei and Luke, 2016], as below:
Tt+1(st, at) = κ∗
(cid:26) Tt(st, at)
(1 − η) ∗ Tt(st, at) + η ¯Tt(s(cid:48))
(7)
where η is a constant controlling the extent that ¯T (s(cid:48)) is
folded in. We absorb this interesting notion of forgiveness
into our lenient reward network to boost the convergence in
cooperative Markov games which will be explained later.
if s(cid:48) is terminal,
otherwise.
3 Weighted Double Deep Q-Networks
In the section, we introduce a new multiagent DRL algorithm,
weighted double deep Q-networks (WDDQN), with two aux-
iliary mechanisms, i.e., the lenient reward approximation and
the scheduled replay strategy, to achieve efficient coordina-
tion in stochastic multiagent environments. In these environ-
ments, reward could be extremely stochastic due to the envi-
ronments' inherent characteristics and the continuous change
of the coexisting agents' behaviors.
For the stochastic rewards caused by the environments,
WDDQN uses the combination of the weighted double es-
timator and the reward approximator to reduce the estima-
tion error. As for the non-stationary coexisting agents, we
incorporate the leniency from lenient Q-learning [Palmer et
al., 2018; Panait et al., 2006] into the reward approximator to
provide an optimistic estimation of the expected reward under
each state-action pair r(s, a). In addition, directly applying
Figure 1: Network Architecture of WDDQN
3.2 Lenient Reward Network
To reduce noise in stochastic rewards, we use a reward net-
work, which is a neural network estimator, to approximate the
reward function R(s, a) explicitly. The reward network can
reduce bias in immediate reward r yielded from stochastic
environments by averaging all rewards for distinct (s, a) pair
and be trained using the transitions stored in the experience
replay during the online interaction. When updating the net-
work, instead of using the reward r in transition (s, a, r, s(cid:48))
from experience memory, WDDQN uses the estimated re-
ward by the reward network (lines 12 and 17).
In addition to stochasticity, in a cooperative multiagent en-
vironment, the coexisting agents introduce additional bias to
r as well. The mis-coordination of coexisting teammates may
Algorithm 1 WDDQN
1: The maximum number of episodes: M axE, the max-
imum number of steps: M axS, global memory: DG,
episodic memory: DE, reward network: RN , deep Q-
networks: QU and QV
Initialize DE
for step = 1 to M axS do
2: for episode = 1 to M axE do
3:
4:
5:
6:
7:
a ← maxa(cid:48) QU (s,a(cid:48))+QV (s,a(cid:48))
Execute a and store transitions into DE
Sample mini-batch (s, a, r, s(cid:48)) of transitions from
(with ε-greedy)
2
DG
8:
9:
10:
11:
12:
13:
14:
15:
16:
Update QU or QV randomly
if update QU then
a∗ ← arg maxa QU (s(cid:48), a)
U (s(cid:48), a∗) ← βQU (s(cid:48), a∗) + (1 −
Qw
β)QV (s(cid:48), a∗)
QTarget(s, a) ← RN (s, a) + Qw
Update network QU towards QTarget
a∗ ← arg maxa QV (s(cid:48), a)
V (s(cid:48), a∗) ← βQV (s(cid:48), a∗) + (1 −
Qw
β)QU (s(cid:48), a∗)
QTarget(s, a) ← RN (s, a) + Qw
Update network QV towards QTarget
V (s(cid:48), a∗)
U (s(cid:48), a∗)
else
end if
Update RN according to transitions in DG
17:
18:
19:
20:
21:
22:
23: end for
end for
Store DE into DG
lower the reward r for (s, a∗) despite the agent has adopted
the optimal action. To address this, we use a lenient reward
network (LRN) enhanced with the lenient concept in [Pot-
ter and De Jong, 1994] to allow the reward network to be
optimistic during the initial exploration phase. The LRN is
updated periodically (line 20) as follows:
(cid:26) Rt(st, at) + αδ
Rt(st, at)
Rt+1(st, at) =
if δ > 0 or x < l(st, at),
otherwise.
i
i
t
t
i=1...n r(s,a)
1/n(cid:80)
(8)
where Rt(st, at) is the reward approximation of state s and
− R(st, at) is the TD
action a at time t, and δ = ¯r(s,a)
error between the Rt(st, at) and the target reward ¯r(s,a)
=
obtained by averaging all immediate re-
ward r(s,a)
of (s, a) pairs in experience memory. Note that
l(st, at) inherits from Equation 6 and has the same meaning,
which is gradually decayed each time a state-action (s, a) pair
is visited. Consequently, the LRN contributes to reduce bias
by reward approximation and can help agents to find optimal
joint policies in cooperative Markov games.
3.3 Scheduled Replay Strategy
Prioritized experience replay (PER) can improve the DQN
algorithm's training efficiency by allocating samples with
Figure 2: Comparison between the prioritized experience re-
play and the scheduled replay strategy: each dot represents a
sample (s, a, r, s), and a trajectory consists of an ordered se-
quence of samples. The x-axis represents the order that each
sample comes into the relay memory and the y-axis is the pri-
ority of each sample.
different priorities according to their TD error. Samples
with higher priorities are more likely to be chosen for net-
work training. However, in stochastic multiagent environ-
ments, due to the noisy reward and the continuous behavior
changes of coexisting agents, PER may deteriorate the algo-
rithm's convergence and perform poorly. Given a transition
(s, a, r, s, d) with an extremely biased reward r, PER will
treat it as an important sample for its large TD error and will
frequently select it to update the network, though it is incor-
rect due to the big noise in r at the beginning. To address
this, we replace r with an estimation RN (s, a) using LRN
to correct TD error, by which the PER can distinguish truly
important samples.
Another potential problem is that PER gives all samples
in the new trajectory the same priority, thus resulting in the
indistinguishability of importance for all new samples. To
be specific, in Fig. 2, the sample with the maximum prior-
ity is colored by red dot. PER gives all samples (blue dots)
in the latest trajectory with an identical priority 1. However,
in cooperative multiagent environments, the trajectories that
agents succeed in cooperation are relatively rare, and in these
trajectories the samples closer to the terminal state is even
more valuable than the ones far from the terminal state. Be-
sides, the Q(s, a) = r + Q(s(cid:48), a∗) far from the terminal state
can further deteriorate if bootstrap of action value Q(s(cid:48), a∗)
is already highly inaccurate, since inaccurate estimation will
propagate throughout the whole contiguous samples. These
two traits explain why samples that are close to the terminal
state should be frequently used for network training. To this
end, we develop a scheduled replay strategy (SRS) using a
precomputed rising schedule [w0, w1, ..., wn] with size n to
assign different priorities according to the sample's position i
in the trajectory with n samples.
The values for wi = eρc∗ui are computed using an expo-
nent ρc which grows with a rising rate u > 1 for each i,
0 (cid:54) i < n. The priority pi assigned to sample with index i is
obtained by multiplying the current maximum priority pmax
in experience memory (priority of the red point in Fig. 2) by
1See
source
https://github.com/openai/baselines.
OpenAI
code
for
details:
Figure 3: Comparisons of DDQN, WDDQN w.o. LRN+SRS and WDDQN on pacman with 4 different sizes. The X-axis is the
number of training episodes and the Y-axis is a ratio of the number of minimum steps to the goal to the number of steps that the
agent actually used during training.
wi:
pi = pmax × wi
The SRS assigns higher priority to samples near the ter-
minal state (the green dot in Fig. 2) to ensure they are more
likely to be sampled for network training. In this way, the
estimation bias of the Q(s, a) near the terminal state is ex-
pected to decrease rapidly. This can significantly speed up
the convergence and improve the training performance, as to
be experimentally verified in the following section.
4 Experiments
Empirical evaluation is conducted to verify the effectiveness
of WDDQN in terms of reducing bias and achieving coordi-
nation in stochastic multiagent domains.
First, we present comparisons of double DQN (DDQN)
and WDDQN with /without LRN and SRS, denoted by WD-
DQN and WDDQN w.o. LRN+SRS, in terms of the bias re-
duction, learning speed and performance on a gridworld game
with raw visual input. Then, we use a cooperative Markov
game to investigate WDDQN's effectiveness of finding an
optimal cooperative policy. A discussion about benefits of
WDDQN, LNR and SRS is given in the end.
Figure 4: Gridworld game.
Figure 5: Predator game.
Table 1: Network architectures in WDDQN
# Network Visual input
84 * 84 * 3
84 * 84 * 3
DQN
LRN
Filters in Conv. 1/2/3 Unit in F.C
32/64/64
16/16/16
512
128
We set the constant c in β to 0.1 in WDDQN, parameters
K, κ, η in lenient Q-learning to 2, 0.95 and 0.6 respectively.
Besides, the learning rate α for network training of DDQN,
lenient Q-learning is set to 0.0001. Table 1 depicts the ar-
chitecture of deep Q-networks and LRN in WDDQN. We use
three hidden convolution layers (using rectifier non-linearities
between each two consecutive layers), and a fully-connected
hidden layer. The output layer of DQN and LRN is a fully-
connect linear layer with a single output layer for each valid
action Q(s, a) and reward R(s, a). For exploration purpose,
DQN(-greedy) is adopted with the annealed linearly from
1 to 0.01 over the first 10000 steps. We used the Adam algo-
rithm with 0.0001 learning rate and the minibatches of size
32. We trained for a total of 2500 episodes and used a replay
memory of 8192 most recent frames. Last, to be fair, K, κ, η
in LRN is the same as the lenient Q-learning while ρc and µ
in SRS is set to 0.2 and 1.1.
4.1 Pacman-like Grid World
The first experiment is an n×n pacman-like grid-world prob-
lem (Fig. 4), where the agent starts at the s0 (top left cell),
and moves towards the goal cell (pink dot at right bottom cell)
using only four actions: {north, south, east, west}. Every
movement leads the agent to move one cell in the correspond-
ing direction, except that a collision on the edge of the grid
results in no movement. The agent tries to search the goal
cell which may appear randomly in any position in the grid
world. The agent receives a stochastic reward of -30 or 40
with equal probability for any action entering into the goal
and ending an episode. Choosing north or west will get a re-
ward of -10 or +6, and south or east get a reward of -8 or +6
at a non-goal state. The environment is extremely noisy due
to the uncertainty in the reward function.
Empirical results in Figure 3 demonstrate that, under ex-
tremely stochastic environments, DDQN takes a long time
to optimize the policy, while WDDQN w.o. LRN+SRS and
WDDQN need much less episodes to get a better policy due
to the weighted double estimator. DDQN and WDDQN w.o.
LRN+SRS oscillate too frequently to converge to an optimal
policy, while WDDQN performs steadily and smoothly be-
cause of the use of LRN. Another finding is that the train-
ing speed of WDDQN is faster than the others, which is at-
tributed to the SRS. In general, WDQ works not as well as
in relatively simple RL problems and both DDQN and WD-
(a) deterministic rewards.
(b) stochastic rewards.
Figure 6: (Left) Comparisons of WDDQN and its variants using the predator game with deterministic rewards; and (right)
comparisons of WDDQN and other algorithms using the predator game with stochastic rewards. Note that, each point in the
x-axis consists of 50 episodes, and the y-axis is the corresponding averaged reward. The shadow area ranges from the lowest
reward to the highest reward within the 50 episodes.
DQN w.o. LRN+SRS may not converge even after a very
long training time. By contrast, as shown in Fig. 1, WDDQN
learns efficiently and steadily due to the use of both LRN and
SRS.
4.2 Cooperative Markov Game
In this section, we consider the two predators pursuit prob-
lem.
It is a more complex cooperative problem and firstly
defined in [Benda et al., 1986]. Here we redefine it in a sim-
ple way. The robots in Figure 5 represent two agents trying
to enter into the goal state at the same time. The cell with
letter S is a suboptimal goal with a reward of +10 while G
is a global optimal with a reward of +80. There is a thick
wall (in gray) in the middle that separates the area into two
zones.
In each episode, two agents start at the left bottom
cell and right bottom cell separately and try to go to the green
goal cell together. Each agent has four actions: {north, south,
east, west}. Every movement leads the agent to move one
grid in the corresponding direction, except that a collision on
the edge of the grid or thick wall results in no movement.
A reward of 0 is received whenever entering into a non-goal
state. The agent receives a positive reward for any action en-
tering into the goal together and ending an episode, otherwise
a negative reward of -1 is received with miscoordination.
There are two types of cooperative policies moving to-
wards the suboptimal goal cell S or the global optimal cell
G, as shown in the Fig. 5.
In the remaining part, we in-
vestigate whether WDDQN and related algorithms can find
cooperative policies, especially the optimal policy.
Evaluation on WDDQN
Our goal is to train two agents simultaneously to coordinate
in order to get higher rewards. The performance of WDDQN
w.o. LRN+SRS, WDDQN(LRN)2, and WDDQN in terms of
the average reward is depicted in Figure 6(a). As WDDQN
w.o. LRN+SRS's convergence is no longer guaranteed in the
neural network representation, it is not surprising that it fails
in finding the cooperative policy by directly combining WDQ
with neural network. By contrast, WDDQN(LRN), due to
the LRN, achieves coordination more quickly and finds the
optimal policy after a period of exploration. By leveraging
the SRS, WDDQN shows a more promising result that the
optimal policy is learned much faster than the two others.
Evaluation Against Other Algorithms
Here, we compare WDDQN against DDQN, a DRL algo-
rithm, and lenient Q-learning, a multiagent RL algorithm on
the same game except that the agent receives a reward of +10
or +100 with the possibility of 60% or 40% at goal S and a
deterministic reward of +80 at goal G. Goal S is still subop-
timal as its average reward is 46. This slight adjustment may
affect the algorithm's performance by misleading the agent to
converge to the suboptimal goal where a higher reward may
appear accidentally.
Results in terms of the average reward are depicted in Fig.
6(b), where two dashed lines indicate optimal and suboptimal
policy with the expected rewards of 80 and 46, respectively.
Both WDDQN and lenient Q-learning outperform DDQN in
terms of the convergence speed and the average reward in all
experiments, which confirms the infeasibility of directly ap-
plying DRL algorithms in multiagent problems. Note that,
WDDQN, due to the use of both LRN and SRS, is more sta-
ble, performs better and is more likely to find the optimal so-
lution with the average reward of 80 than lenient Q-learning
2WDDQN(LRN) uses only LRN and is identical to WDDQN
w.o. SRS
with the average reward of 46 in such a stochastic multiagent
environment.
5 Conclusion
This paper proposes WDDQN with the lenient reward net-
work and the scheduled replay strategy to boost the training
efficiency, stability and convergence under stochastic multia-
gent environments with raw image inputs, stochastic rewards,
and large state spaces. Empirically, WDDQN performs bet-
ter than WDDQN w.o. LRN+SRS, DDQN and lenient Q-
learning in terms of the average reward and convergence rate
on the pacman and two predators pursuit domains.
One downside to our approach is that it only uses one agent
to explore the large-scale RL problems and train network at
the same time. These can significantly slow down the ex-
ploration procedure and affect WDDQN's performance and
efficiency. This could be remedied in practice by accelerating
the training procedure of WDDQN using asynchronization,
as being used in the A3C algorithm [Mnih et al., 2016]. We
leave this investigation to future work.
References
[Bellman, 1957] Richard Bellman. Dynamic programming.
Princeton University Press., 1957.
[Benda et al., 1986] M. Benda, V. Jagannathan, and R. Dod-
hiawala. On optimal cooperation of knowledge sources
- an empirical investigation.
Technical Report BCS–
G2010–28, Boeing Advanced Technology Center, Boeing
Computing Services, 1986.
[Bloembergen et al., 2011] Daan Bloembergen, Michael
Kaisers, and Karl Tuyls.
Empirical and theoretical
support for lenient learning. In International Conference
on Autonomous Agents and Multiagent Systems, pages
1105–1106, 2011.
[Bloembergen et al., 2015] Daan Bloembergen, Karl Tuyls,
Daniel Hennes, and Michael Kaisers. Evolutionary dy-
namics of multi-agent learning: A survey. Journal of Arti-
ficial Intelligence Research, 53:659–697, 2015.
[Claus and Boutilier, 1998] Caroline Claus
and Craig
Boutilier. The dynamics of reinforcement learning in
cooperative multiagent systems. In AAAI Conference on
Artificial Intelligence, pages 746–752, 1998.
[Gupta et al., 2017] Jayesh K Gupta, Maxim Egorov, and
Mykel Kochenderfer. Cooperative multi-agent control us-
In International Con-
ing deep reinforcement learning.
ference on Autonomous Agents and Multiagent Systems,
pages 66–83, 2017.
[Hasselt, 2010] Hado V. Hasselt. Double Q-learning. In Ad-
vances in Neural Information Processing Systems, pages
2613–2621, 2010.
[Lanctot et al., 2017] Marc Lanctot, Vinicius Zambaldi, Au-
drunas Gruslys, Angeliki Lazaridou, Julien Perolat, David
Silver, Thore Graepel, et al. A unified game-theoretic ap-
proach to multiagent reinforcement learning. In Advances
in Neural Information Processing Systems, pages 4193–
4206, 2017.
[Matignon et al., 2007] Laetitia Matignon, Guillaume J Lau-
rent, and Nadine Le Fort-Piat. Hysteretic q-learning: an
algorithm for decentralized reinforcement learning in co-
operative multi-agent teams. In International Conference
on Intelligent Robots and Systems, pages 64–69, 2007.
[Matignon et al., 2012] Laetitia Matignon, Guillaume J Lau-
rent, and Nadine Le Fort-Piat. Independent reinforcement
learners in cooperative markov games: a survey regarding
coordination problems. The Knowledge Engineering Re-
view, 27(1):1–31, 2012.
[Mnih et al., 2013] Volodymyr Mnih, Koray Kavukcuoglu,
David Silver, Alex Graves, Ioannis Antonoglou, Daan
Wierstra, and Martin Riedmiller. Playing atari with deep
reinforcement learning. arXiv preprint arXiv:1312.5602,
2013.
[Mnih et al., 2015] Volodymyr Mnih, Koray Kavukcuoglu,
David Silver, Andrei A. Rusu, Joel Veness, Marc G. Belle-
mare, Alex Graves, Martin Riedmiller, Andreas K. Fid-
jeland, Georg Ostrovski, Stig Petersen, Charles Beattie,
Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan
Kumaran, Daan Wierstra, Shane Legg, and Demis Has-
sabis. Human-level control through deep reinforcement
learning. Nature, 518(7540):529–533, 2015.
[Mnih et al., 2016] Volodymyr Mnih, Adria Puigdomenech
Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim
Harley, David Silver, and Koray Kavukcuoglu. Asyn-
chronous methods for deep reinforcement learning.
In
International Conference on Machine Learning, pages
1928–1937, 2016.
[Palmer et al., 2018] Gregory Palmer, Karl Tuyls, Daan
Bloembergen, and Rahul Savani. Lenient multi-agent deep
In International Conference on
reinforcement learning.
Autonomous Agents and Multi-Agent Systems, page to ap-
pear, 2018.
[Panait et al., 2006] Liviu Panait, Keith Sullivan, and Sean
Luke. Lenient learners in cooperative multiagent systems.
In International Conference on Autonomous Agents and
Multiagent Systems, 2006.
[Potter and De Jong, 1994] Mitchell A Potter and Kenneth A
De Jong. A cooperative coevolutionary approach to func-
tion optimization. In International Conference on Parallel
Problem Solving from Nature, pages 249–257, 1994.
[Schaul et al., 2016] Tom Schaul,
Ioannis
Antonoglou, and David Silver. Prioritized experience re-
play. In International Conference on Learning Represen-
tations, 2016.
John Quan,
[Smith and Winkler, 2006] James E Smith and Robert L
Winkler. The optimizer's curse: Skepticism and postde-
cision surprise in decision analysis. Management Science,
52(3):311–322, 2006.
[Sutton and Barto, 1998] Richard S Sutton and Andrew G
Barto. Reinforcement learning: An introduction. MIT
press Cambridge, 1998.
[Sutton, 1988] Richard S Sutton. Learning to predict by
the methods of temporal differences. Machine learning,
3(1):9–44, 1988.
[Van Hasselt et al., 2016] Hado Van Hasselt, Arthur Guez,
and David Silver. Deep reinforcement learning with dou-
In AAAI Conference on Artificial Intelli-
ble q-learning.
gence, pages 2094–2100, 2016.
[Wang et al., 2016] Ziyu Wang, Tom Schaul, Matteo Hes-
sel, Hado Van Hasselt, Marc Lanctot, and Nando De Fre-
itas. Dueling network architectures for deep reinforcement
learning. In International Conference on Learning Repre-
sentations, 2016.
[Watkins, 1989] Christopher John Cornish Hellaby Watkins.
Learning from delayed rewards. PhD thesis, King's Col-
lege, University of Cambridge, 1989.
[Wei and Luke, 2016] Ermo Wei and Sean Luke. Lenient
learning in independent-learner stochastic cooperative
games. Journal of Machine Learning Research, 17(84):1–
42, 2016.
[Zhang et al., 2017] Zongzhang Zhang, Zhiyuan Pan, and
Mykel J Kochenderfer. Weighted double Q-learning. In
International Joint Conference on Artificial Intelligence,
pages 3455–3461, 2017.
|
1811.12557 | 1 | 1811 | 2018-11-30T00:38:18 | Deep Multi-Agent Reinforcement Learning with Relevance Graphs | [
"cs.MA",
"cs.LG"
] | Over recent years, deep reinforcement learning has shown strong successes in complex single-agent tasks, and more recently this approach has also been applied to multi-agent domains. In this paper, we propose a novel approach, called MAGnet, to multi-agent reinforcement learning (MARL) that utilizes a relevance graph representation of the environment obtained by a self-attention mechanism, and a message-generation technique inspired by the NerveNet architecture. We applied our MAGnet approach to the Pommerman game and the results show that it significantly outperforms state-of-the-art MARL solutions, including DQN, MADDPG, and MCTS. | cs.MA | cs |
Deep Multi-Agent Reinforcement Learning with
Relevance Graphs
Aleksandra Malysheva∗†
JetBrains Research
National Research University
Higher School of Economics
St. Petersburg, Russia
[email protected]
Tegg Taekyong Sung∗†
Department of Electronics and
Communications Engineering
Kwangwoon University
Seoul, Republic of Korea
[email protected]
Chae-Bong Sohn
Department of Electronics and
Communications Engineering
Kwangwoon University
Seoul, Republic of Korea
[email protected]
Daniel Kudenko
University of York
York, United Kingdom
[email protected]
Aleksei Shpilman
JetBrains Research
National Research University
Higher School of Economics
St. Petersburg, Russia
[email protected]
Abstract
Over recent years, deep reinforcement learning has shown strong successes in
complex single-agent tasks, and more recently this approach has also been applied
to multi-agent domains. In this paper, we propose a novel approach, called MAGnet,
to multi-agent reinforcement learning (MARL) that utilizes a relevance graph
representation of the environment obtained by a self-attention mechanism [17], and
a message-generation technique inspired by the NerveNet architecture [18]. We
applied our MAGnet approach to the Pommerman game [11] and the results show
that it significantly outperforms state-of-the-art MARL solutions, including DQN,
MADDPG, and MCTS.
1
Introduction
A common difficulty of reinforcement learning in a multi-agent environment is that in order to
achieve successful coordination, agents require information about the relevance of environment
objects to themselves and other agents. For example, in the game of Pommerman it is important
to know how relevant bombs placed in the environment are for teammates, e.g. whether or not the
bombs can threaten them. While such information can be hand-crafted into the state representation
for well-understood environments, in lesser-known environments it is preferable to derive it as part of
the learning process.
∗Equal contribution.
†Work performed while at Deep Learning Camp Jeju.
32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada.
In this paper, we propose a novel method, named MAGNet, to learn such relevance information in
form of a relevance graph and incorporate this into the reinforcement learning process. Furthermore,
we propose the use of message generation techniques from this graph, inspired by the NerveNet
architecture [18]. NerveNet has been introduced in the context of robot locomotion, where it has
been applied to a graph of connected robot limbs. MAGNet uses a similar approach, but basing the
message generation on the learned relevance graph.
We applied MAGNet to the popular Pommerman [11] multi-agent environment, and achieved sig-
nificantly better performance than a baseline heuristic method and state-of-the-art RL techniques
including DQN [12], MADDPG [9] and MCTS [5]. Additionally, we empirically demonstrate the
effectiveness of self-attention, graph sharing and message generating modules with an ablation study.
2 Deep Multi-Agent Reinforcement Learning
In this section we describe the state-of-the-art (deep) reinforcement learning techniques that were
applied to multi-agent domains. The algorithms introduced below (DQN, MCTSNet, and MADDPG)
were also used as evaluation baselines in our experiments.
The majority of work in the area of reinforcement learning applies a Markov Decision Process (MDP)
as a mathematical model [13]. An MDP is a tuple (S, A, T, R), where S is the state space, A is the
action space, T (s, a, s(cid:48)) = P r(s(cid:48)s, a) is the probability that action a in state s will lead to state s(cid:48),
and R(s, a, s(cid:48)) is the immediate reward r received when action a taken in state s results in a transition
to state s(cid:48). The problem of solving an MDP is to find a policy (i.e., mapping from states to actions)
which maximises the accumulated reward. When the environment dynamics (transition probabilities
and reward function) are available, this task can be solved using policy iteration [2].
The problem of solving an multi-agent MDP is to find policies π = π1, . . . , πN that maximize the
expected reward Ji = Es∼pπ,a∼πi[R] for every agent i, where pπ is the distribution of states visited
with policy π.
2.1 Deep Q-Networks
Q-learning is a value iteration method that tries to predict future rewards from current state and an
action. This algorithm apply so called temporal-difference updates to propagate information about
values of state-action pairs, Q(s, a). After each transition, (s, a) → (s(cid:48), r), in the environment, it
updates state-action values by the formula:
Q(s, a) → Q(s, a) + α[r + γ max Q(s(cid:48), a(cid:48)) − Q(s, a)],
(1)
where α is the rate of learning and γ is the discount factor. It modifies the value of taking action a in
state s, when after executing this action the environment returned reward r, and moved to a new state
s(cid:48).
Deep Q-learning utilizes a neural network to predict Q-values of state-action pairs [12]. This so-called
deep Q-network is trained to minimize the difference between predicted and actual Q-values as
follows:
y = r + γ max
a
Qpast(s(cid:48), a(cid:48))
L(θ) = Es∼ρπ,a∼p(s)[yi − Q(s, aθ)]2,
(2)
(3)
where y is best action according to the previous deep Q-network, θ is the parameter vector of the
current Q-function and a ∼ p(s) denotes all actions that are permitted in state s.
The simplest way to apply this approach in multi-agent systems is to use an independent network for
every agent [15]. However this approach has been shown to not perform well with more complex
environments [10]. One of the shortcomings of DQN learning in multi-agent settings is that past
experience replay is less informative, because unlike a single-agent setting, the same action in the
same state may produce a different result based on the actions of other agents. A way to alleviate this
problem is passing parameters of other agents as additional environmental information [16].
2
2.2 Monte-Carlo Tree Search nets (MCTSNet)
An alternative approach to reinforcement learning is to directly find an optimal policy without the
intermediate step of computing a value function. Policy gradient methods (e.g. [14]) have been
developed to do just this.
Policy gradient methods have been shown to be successful in combination with Monte-Carlo tree
search (MCTS) [3], which is a general, powerful, and widely used decision making algorithm, most
commonly applied to games. In MCTS a sample tree of simulated future states is created, and
evaluations of those states are backed-up to the root of this so-called search tree to compute the best
action.
A recent study [5] incorporates a neural network inside the tree-search by expanding, evaluating and
backing-up a vector embedding of the states. The key idea is to assign a feature or a "memory" vector
h ∈ Rn to an internal state (search tree node) that is then propagated up the tree and used to calculate
the value or action in the root node. This MCTSNet approach has been shown to outperform other
MCTS methods.
2.3 Multi-agent Deep Deterministic Policy Gradient
When dealing with continuous action spaces, the methods described above can not be applied. To
overcome this limitation, the actor-critic approach to reinforcement learning was proposed [14]. In
this approach an actor algorithm tries to output the best action vector and a critic tries to predict the
value function for this action.
Specifically, in the Deep Deterministic Policy Gradient (DDPG [8]) algorithm two neural networks
are used: µ(s) is the actor network that returns the action vector. Qw(s, a) is the critic network, that
returns the Q value, i.e. the value estimate of the action of a in state s.
The gradient for the critic network can be calculated in the same way as the gradient for Deep
Q-Networks described above (Equation 3). Knowing the critic gradient ∇aQw we can then compute
the gradient for the actor as follows:
∇θµJ = Es∼ρπ [∇aQw(s, aθw)s = st, a = µ(sθµ)],
(4)
where θw and θµ are parameters of critic and actor neural networks respectively, and ρπ(s) is the
probability of reaching state s with policy π.
The authors of [9] proposed an extension of this method by creating multiple actors, each with it's
own critic with each critic taking in the respective agent's observations and actions of all agents.
3 MAGnet approach and architecture
The overall network architecture of our MAGNet approach is shown in Figure 1. The whole process
can be divided into a relevance graph generation stage (shown in the left part) and a decision making
stages (shown in the right part). We see them as a regression and classification problem respectively.
In this architecture, the concatenation of the current state and previous action forms the input of the
models, and the output is the next action. The details of the two processes are described below.
3.1 Relevance graph generation stage
In the first part of our MAGNet approach, a neural network is trained that produces a relevance graph.
The relevance graph represents the relationship between agents and between agents and environment
objects. The higher the weight of an edge between an agent a and another agent b or object o is, the
more important b or o are for the achievement of agent a's task. The graph is generated by MAGNet
from the current and previous state together with the respective actions.
Figure 4B shows an example of such a graph for two agents. The displayed graph only shows those
edges which have a non-zero weight (thus there are objects to which agent 1 is not connected in the
graph).
In MAGNet, a neural network is trained via back-propagation to output a relevance graph represented
as an A × (A + O) matrix, where A is the number of agents and O is the maximum number
3
Figure 1: The overall network architecture of MAGNet. Left section shows the graph generation
stage. Right part shows the decision making stage.
of environment objects. The input to the network are the current and the two previous states (denoted
by X(t), X(t − 1), and X(t − 2) in Figure 1), the two previous actions (denoted by a(t − 1) and
a(t − 2)), and the relevance graph produced at the previous time step (denoted by graph(t − 1)). For
the first learning step (i.e. t = 0), the input consists out of three copies of the initial state, no actions,
and a random relevance graph. The inputs are passed into a convolution and pooling layer, followed by
a padding layer, and then concatenated and passed into fully connected layer and finally into the graph
generation network (GGN). The GGN can be either a multilayer perceptron (MLP) or a self-attention
network, which uses an attention mechanism to catch long and short term time-dependencies, and is
an analogue to a recurrent network such as LSTM, but takes much less time to compute [17]. The
result of the GGN is fed into a two-layer fully connected network with dropout, which produces the
relevance graph matrix, as described above.
The loss function for the back-propagation training is composed of two parts:
(5)
L = (cid:107)Wt − Wt−1(cid:107)2
2 +
(wt(ξ) − s(ξ))2
(cid:88)
ξ∼Ξ
The first component is based on the difference between the current graph Wt and the one generated
in the previous state Wt1. It is important to note that graph on each step is the same only weights are
changing. The second component comes into play when a special pre-defined event ξ ∈ Ξ occurs,
and is based on the difference between a selected edge weight updated according to heuristic rules
s(ξ) and the weight of the same edge in the current graph wt(ξ). For example, a heuristic rule would
specify that if a bomb explodes and kills the agent, the edge weight between the agent and the bomb
is set to a high value (i.e. the bomb is clearly of high relevance of the agent).
The training of the neural network can be performed in two stages: first with a default rule-based AI
agent, and then with a learning agent.
3.2 Decision making stage
The agent AI responsible for decision making is also represented as a neural network whose inputs
are accumulated messages (generated by a method inspired by NerveNet [18] and described below)
and the current state. The output of the network is an action to be executed.
4
The graph G generated at the last step is G = (V, E) where edges represent relevance between agents
and objects. Every vertex v has a type: b(v) ∈ {0, 1, 2, 3, 4, 5, 6} that in our case corresponds to:
"ally", "enemy", "placed bomb" (about to explode), "increase kick ability", "increase blast power",
"extra bomb" (can be picked up). Every edge has a type as well: c(e) ∈ {0, 1}, that corresponds to
"edge between the agents" and "edge between the agent and the object in the environment".
The final (action) vector is computed in 4 stages through message passing system, similar to a system
used for distributed computing and described in [1]. Stages 2 and 3 are repeated for a specified
number of message propagation steps.
1. Initialization of information vector. Each vertex v has an initialization network M LP b(v)
init
associated with it according to it's type b(v) that takes as input the current individual
observation Ov and outputs initial information vector µ0
v for each vertex.
v = M LP b(v)
µ0
init (Ov)
(6)
2. Message generation. At message propagation step t + 1 message networks M LP c(v,u)
mess
compute output messages for every edge (v, u) ∈ E based on type of the edge c(v, u).
mt
(v,u) = M LP c(v,u)
m
(µt
v)
(7)
3. Message processing. Information vector mt+1
at message propagation step t is updated by
update network LST M b(v)
associated with it according to it's type b(v), that takes as input
up
a sum of all message vectors from connected to v edges multiplied by the edge relevance
v.
w(v,∗) and information at previous step mt
µt+1
v = LST M b(v)
(cid:88)
up (µt
v,
v
mt
(v,∗)w(v,∗))
(8)
4. Choice of action. All vertices that are associated with agents have a decision network
v and compute the mean
choice which takes as an input its final information vector mt
M LP b(v)
of the action of the Gaussian policy.
All networks are trained using back-propagation following the DDPG actor-critic approach [8].
av = M LP b(v)
choice(µt
v)
(9)
4 Experiments
4.1 Environment
In this paper, we use popular Pommerman game environment which can be played by up to 4
players [11]. This game has been used in many empirical evaluations of multi-agent algorithms,
and therefore is especially suitable for a comparison to state-of-the-art techniques. In Pommerman,
the environment is a grid-world where each agent can move in one of four directions, lay a bomb,
or do nothing. A grid square is either clear (which means that an agent can enter it), wooden, or
rigid. Wooden grid squares can not be entered, but can be destroyed by a bomb (i.e. turned into clear
squares). Rigid squares are indestructible and impassable. When a wooden square is destroyed, there
is a probability of items appearing, e.g., an extra bomb, a bomb range increase, or a kick ability. Once
a bomb has been placed in a grid square it explodes after 10 time steps. The explosion destroys any
wooden square within range 1 and kills any agent within range 4. The last surviving agent wins the
gameast
The map of the environment is randomly generated for every episode. The game has two different
modes: free for all and team match. Our experiments were carried out in the team match mode in
order to evaluate the ability of MAGnet to exploit the discovered relationships between agents (e.g.
being on the same team).
4.2 Network training
We first trained the graph generating network on 50,000 episodes with the default Pommerman AI as
the decision making agent. After this initial training, the default AI was replaced with the learning
5
decision making AI described in section 3. All learning graphs show the training episodes starting
with this replacement (except the ones which explicitly show the relevance graph learning).
Table 1 shows results for different MAGNet variants in terms of achieved win percentage against
a default agent after 50,000 episodes. The MAGNet variants are differing in the complexity of the
approach, starting from the simplest version which takes the relevance graph as a direct input, to the
version incorporating message generation, graph sharing, and self-attention. The table clearly shows
the benefit of each extension.
Table 1: Influence of different modules on the performance of the MAGnet model.
MAGnet modules
Self-
attention
+
+
+
+
-
-
-
-
Graph
Sharing
+
+
-
-
+
+
-
-
Message
Generation
+
-
+
-
+
-
+
-
Win %
71.3±0.7
56.7 ± 1.8
62.4 ± 1.7
54.5 ± 2.6
67.1 ± 1.9
52.0 ± 1.7
45.2 ± 3.6
32.7 ± 5.9
Each of the three extensions with their hyper-parameters are described below:
Graph Generating Network (GGN): we used a MLP (number of layers and neurons was varied, and
a network with 3 layers 512-128-128 neurons achieved the best result) and a self-attention (SA)
layer [17] with default parameters.
Graph Sharing (GS): relevance graphs were trained individually for both agents, or in form of a
shared graph for both agents.
Message Generation (MG): the message generation module was implemented as either a MLP or a
message generation (MG) architecture, as described in Section 3. We tested the MLP and message
generation network with a range of hyper-parameters. For the MLP with 3 fully connected layers
1024-256-64 neurons achieved the best result, while for the message generation network 2 layers
with 128-32 neurons and 5 message passing iterations showed the best result.
Dropout layers were individually optimized by grid search in [0, 0.2, 0.4] space.
We tested two convolution sized: [3x3] and [5x5]. [5x5] convolutions showed the best result.
Rectified Linear Unit (ReLU) transformation was used for all connections.
4.3 Evaluation Baselines
In our experiments, we compare the proposed method with state-of-the-art reinforcement learning
algorithms simulated in team match mode. Figure 2a shows a comparison with DQN [12], MCT-
SNets [5], MADDPG[9], and a default heuristic AI. The latter algorithm is provided as part of the
Pommerman environment [11]. Each of the reinforcment learning algoritms played a number of
games (i.e. episodes) against the heuristic AI, and the respective win rates are shown.
All graphs display a 95% confidence interval to illustrate the statistical significance of our results.
The parameters chosen for the baselines were set as follows.
For DQN we implemented multi-agent deep Q-learning approach which has been shown to be
successful in past work [4]. In this method training is performed in two repeated steps: first, one agent
is training at a time, while policies of other agents are kept fixed; second, the agent that was trained
in the previous step distributes its policy to all of its allies as an additional environmental variable.
The network consists of five convolutional layers with 64 3x3 filters in each layer followed by three
fully connected layers with 128 neurons each with residual connections [6] and batch normalization [7]
that takes an input an 11x11x4 environment tensor and one-hot encoded action vector (a padded 1x6
6
(a)
(b)
Figure 2: (a) The best performing MAGnet variant (MAGnet-Att-NerveNet-GS) compared to state-
of-the-art MARL techniques. (b) The effectiveness MAGNet with various module combinations:
Message Generation (MG), shared relevance graph (GS), and Self-Attention (SA). MADDPG is
currently the best performing state-of-the-art algorithm.
vector) that are provided by the Pommerman environment and outputs a Q-function for that state.
This network showed the best result at the parameter exploration stage.
Parameter exploration on MCTSNet led to the following settings: The backup network β is a
multilayer perceptron (MLP) with 5 fully connected layer with 64 neurons in each layer that takes in
current "memory" vectors of the node and updated "memory" vector of the child and updated the
node's "memory" vector. The embedding network , is consists of 7 convolutional layers with 64
3x3 filters followed by 3 fully connected layers with 128 neurons each with residual connections [6]
and batch normalization [7] that takes an input an 11x11x4 environment tensor and one-hot encoded
action vector (a padded 1x6 vector) that are provided by Pommerman and outputs a "memory" vector.
The policy network has the same architecture, but with 5 convolutional layers with 32 3x3 filters
each and it outputs an action for simulation. The readout network, is a multilayer perceptron with 2
fully connected layer with 128 neurons in each layer that inputs root "memory" vector and outputs an
action.
For our implementation of MADDPG we used a multilayer perceptron (MLP) with 5 fully connected
layer with 128 neurons in each layer and for the critic we used a 3 layer network with 128 neurons in
each layer.
4.4 Self-attention and graph sharing in training a relevance graph
Figure 3 shows the shared graph loss value (Equation 5) with and without self-attention module
and with or without graph sharing. As we can see from this figure, both self-attention and graph
sharing significantly improve graph generation in terms of speed of convergence and final loss value.
Furthermore, their actions are somewhat independent which is seen in that using them together gives
additional improvement.
To provide further evidence for the usefulness of the shared graph approach, we let a MAGNet-
AttNerveNet team play against a MAGNet-Att-NerveNet-GS team. As the graph in 3a shows, even
though both have the same base architectures, the graph sharing method yields a higher win-rate after
10,000 episodes.
4.5 Relevance graph visualization
Figure 4 shows examples of relevance graphs with the corresponding environment states. Red nodes
denote friendly team agents, the purple nodes denote the agents on the opposing team, and the other
nodes denote environment objects such as walls (green) and bombs (black). The lengths of edges
represent their weights (shorter edge equals higher weight, i.e. higher relevance). The graphs in
Figure 4B are shared, while the graphs in Figure 4C are agent-individual.
As can be seen when comparing the individual and shared graphs, in the shared case agent 1 and agent
2 have different strategies related to the opponent agents (agents 3 and 4). Agent 4 is of relevance to
7
(a)
(b)
Figure 3: (a) Win rate of a MAGNet team with a shared relevance graph vs a MAGNet team with
agent-individual relevance graphs. (b) Loss value in training the graph generator with and without a
self-attention module (SA+/-) and with or without graph sharing (GS+/-).
agent 1 but not to agent 2. Similarly, agent 3 is of relevance to agent 2, but not to agent 1. In contrast,
when considering the individual graphs, both agents 3 and 4 have the same relevance to agents 1
and 2. Furthermore, it can be seen from all graphs that different environment objects are relevant to
different agents.
5 Conclusion
In this paper we presented a novel method, MAGNet, for deep multi-agent reinforcement learning
incorporating information on the relevance of other agents and environment objects to the RL
agent. We also extended this basic approach with various optimizations, namely self-attention, shared
relevance graphs, and message generation inspired by Nervenet. The MAGNet variants were evaluated
on the popular Pommerman game environment, and compared to state-of-the-art MARL techniques.
Our results show that MAGNet significantly outperforms all competitors.
6 Acknowledgments
This was supported by Deep Learning Camp Jeju 2018 which was organized by TensorFlow Korea
User Group. This work was partially supported by the Ministry of Trade, Industry & Energy (MOTIE,
Korea) under Industrial Technology Innovation Program No.10077659, 'Development of artificial
intelligence based mobile manipulator for automation of logistics in manufacturing line and logistics
center'.
8
References
[1] H. Attiya and J. Welch. Distributed computing: fundamentals, simulations, and advanced topics,
volume 19. John Wiley & Sons, 2004.
[2] D. P. Bertsekas, D. P. Bertsekas, D. P. Bertsekas, and D. P. Bertsekas. Dynamic programming
and optimal control, volume 1. Athena scientific Belmont, MA, 2005.
[3] G. Chaslot, S. Bakkes, I. Szita, and P. Spronck. Monte-carlo tree search: A new framework for
game ai. In AIIDE, 2008.
[4] M. Egorov. Multi-agent deep reinforcement learning, 2016.
[5] A. Guez, T. Weber, I. Antonoglou, K. Simonyan, O. Vinyals, D. Wierstra, R. Munos, and
D. Silver. Learning to search with mctsnets. arXiv preprint arXiv:1802.04697, 2018.
[6] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition.
In
Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770 --
778, 2016.
[7] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing
internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
[8] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra.
Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015.
[9] R. Lowe, Y. Wu, A. Tamar, J. Harb, O. P. Abbeel, and I. Mordatch. Multi-agent actor-critic for
mixed cooperative-competitive environments. In Advances in Neural Information Processing
Systems, pages 6379 -- 6390, 2017.
[10] L. Matignon, G. J. Laurent, and N. Le Fort-Piat. Independent reinforcement learners in coopera-
tive markov games: a survey regarding coordination problems. The Knowledge Engineering
Review, 27(1):1 -- 31, 2012.
[11] T. Matiisen.
pommerman-baselines, 2018.
Pommerman
baselines.
https://github.com/tambetm/
[12] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller.
Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013.
[13] M. L. Puterman. Markov decision processes: discrete stochastic dynamic programming. John
Wiley & Sons, 2014.
[14] R. S. Sutton, D. A. McAllester, S. P. Singh, and Y. Mansour. Policy gradient methods for rein-
forcement learning with function approximation. In Advances in neural information processing
systems, pages 1057 -- 1063, 2000.
[15] M. Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings
of the tenth international conference on machine learning, pages 330 -- 337, 1993.
[16] G. Tesauro. Extending q-learning to general adaptive multi-agent systems. In Advances in
neural information processing systems, pages 871 -- 878, 2004.
[17] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and
I. Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems,
pages 5998 -- 6008, 2017.
[18] T. Wang, R. Liao, J. Ba, and S. Fidler. Nervenet: Learning structured policy with graph neural
networks. Proceedings of the International Conference on Learning Representations, 2018.
9
A Visualization of Relevance Graph
Figure 4: Visualization of relevance graph. (A) Corresponding game states. (B) Shared graph. (C)
Agent-individual graphs.
10
|
1302.1937 | 1 | 1302 | 2013-02-08T04:03:50 | Embedding agents in business applications using enterprise integration patterns | [
"cs.MA"
] | This paper addresses the issue of integrating agents with a variety of external resources and services, as found in enterprise computing environments. We propose an approach for interfacing agents and existing message routing and mediation engines based on the endpoint concept from the enterprise integration patterns of Hohpe and Woolf. A design for agent endpoints is presented, and an architecture for connecting the Jason agent platform to the Apache Camel enterprise integration framework using this type of endpoint is described. The approach is illustrated by means of a business process use case, and a number of Camel routes are presented. These demonstrate the benefits of interfacing agents to external services via a specialised message routing tool that supports enterprise integration patterns. | cs.MA | cs | Embedding agents in business applications using
enterprise integration patterns
Stephen Cranefield and Surangika Ranathunga
Department of Information Science, University of Otago, Dunedin, New Zealand
{scranefield,surangika}@infoscience.otago.ac.nz
Abstract. This paper addresses the issue of integrating agents with a variety of
external resources and services, as found in enterprise computing environments.
We propose an approach for interfacing agents and existing message routing and
mediation engines based on the endpoint concept from the enterprise integration
patterns of Hohpe and Woolf. A design for agent endpoints is presented, and
an architecture for connecting the Jason agent platform to the Apache Camel
enterprise integration framework using this type of endpoint is described. The
approach is illustrated by means of a business process use case, and a number of
Camel routes are presented. These demonstrate the benefits of interfacing agents
to external services via a specialised message routing tool that supports enterprise
integration patterns.
1
Introduction
Much of the research in multi-agent systems (MAS) is based on a conceptual model
in which the only entities are agents and an abstracted external environment. This is in
contrast to modern enterprise computing environments, which comprise a diverse range
of middleware and server technologies.
The current solutions for integrating agents with external computing infrastructure
are: (a) to access these resources and services directly from agent code (if using a
conventional programming language), (b) to implement user-defined agent actions or an
environment model to encapsulate these interactions, (c) to provide custom support in an
agent platform for specific types of external service, or (d) to provide a generic interface
for calling external resources and services, either using a platform-specific API [9] or by
encapsulating them as agents [4], artifacts [8] or active components [7]. However, none
of these approaches are a good solution when agents need to be integrated with a range
of technologies. They either require agent developers to learn a variety of APIs, or they
assume that agent platform developers or their users will provide wrapper templates for
a significant number of commonly used technologies.
This paper proposes an alternative approach: the use of a direct bridge between
agents and the mainstream industry technology for enterprise application integration:
message routing and mediation engines, and in particular, those that support the enterprise
integration patterns (EIP) of Hohpe and Woolf [5]. Our integration approach is illustrated
in Figure 1. In this figure, each "pipes" graphic represents a messaging-based service
coordination tool, such as an enterprise service bus [3]. The larger one represents an
3
1
0
2
b
e
F
8
]
A
M
.
s
c
[
1
v
7
3
9
1
.
2
0
3
1
:
v
i
X
r
a
Fig. 1. The proposed MAS integration model1
organisation's existing message-based infrastructure for managing business processes
by coordinating information passing between applications and services. We propose
that agents can be embedded into this infrastructure by integrating them with their
own local message-routing and mediation engines, such as the lightweight Java-based
Apache Camel enterprise integration framework [6]. This integration is based on the EIP
notion of an endpoint, and we present the design of endpoints that can translate agent
requests (encoded as agent communication language messages or action executions) to
EIP messages, and from EIP messages to agent messages and percepts.
We describe an implemented architecture for connecting the Jason agent platform [2]
to Camel using these "agent endpoints". The approach is illustrated by means of a
business process use case requiring the integration of Jason agents with a database
management system, a mail server, a message broker and the Apache ZooKeeper coordi-
nation server. A number of Camel routes handling aspects of this use case are presented
to demonstrate the benefits of interfacing agents to external services via a specialised
message routing tool that supports enterprise integration patterns.
2 Enterprise Integration Patterns
Enterprise computing environments typically comprise hundreds and possibly thousands
of applications [5]. These may use a variety of communication protocols and interface
technologies due to departmental autonomy (e.g. to acquire "best of breed" applications
for specific business problems), incremental and opportunistic growth, mergers, etc. To
preserve loose coupling between the diverse applications involved in the automation of
business processes, and thus facilitate maintenance and extensibility, the use of middle-
ware products based on asynchronous message-passing has emerged as the mainstream
approach for enterprise application integration. In this approach, applications interact
1 Pipes photo by Herv´e Cozanet, source: http://commons.wikimedia.org/wiki/File:Piping system
on a chemical tanker.jpg (CC BY-SA 3.0)
...Centralised service integration and orchestrationLocal agent/ external system integration……………Service(s) used only by agentsEnterprise services(e.g. message brokers, mail servers, DBMSs)Enterprise applications by sending and receiving structured messages to and from named queues or publish-
subscribe 'topics' managed by (possibly federated) message brokers. Message routing
and transformation rules can be executed by the message broker or by specialised mes-
sage routing and mediation engines, thus providing a single location for the specification
of business processes. The concept of the enterprise service bus extends this idea further
by integrating message brokers with middleware for deploying and interacting with
various type of service, such as web services [3].
Hohpe and Woolf [5] have identified 65 "enterprise integration patterns" (EIPs) for
solving basic problems that commonly arise in messaging-based enterprise application
integration, such as the scatter-gather pattern: "How do you maintain the overall message
flow when a message needs to be sent to multiple recipients, each of which may send a
reply?" A number of middleware tools have direct support for these patterns, including
Apache Camel.
3 Apache Camel
Camel is an open source Java framework for executing message routing and mediation
rules that are defined using domain-specific languages (DSLs) based on Java and Scala,
or by using XML configuration files. In the work reported in this paper we have used the
Java DSL.
Camel is based on the EIP concepts of routes and endpoints. A Camel application
comprises a set of route definitions. Each route receives messages from a consumer
endpoint, and performs a sequence of processing steps on each message, such as filtering
and transforming messages, before sending the processed messages to one or more
producer endpoints. Endpoints can be "direct" links to other routes in the application
(i.e. messages leaving one route may flow directly into another route) or they may
represent connections to external resources and services. For example, a mail endpoint
may be used as a consumer to receive messages representing unread mail in a specified
account on a mail server, or as a producer that sends mail to a server. Camel has more
than 130 different components defined to provide a variety of endpoint types. These
enable sending and/or receiving messages to and from external resources such as files,
databases, message brokers, generic web services, specific Amazon and Google services,
RSS and Atom feeds, and Twitter. To enable this diversity of endpoint types, Camel's
concept of a message is very general: a message has headers (a map from names to Java
objects), a body (which can be any Java object) and optional attachments.
The code below defines two simple Camel routes. These use the agent component
described in this paper to enable "local" agents (those running within the same process
as the Camel routes) to communicate with remote agents via a message broker.
from("agent:message")
.setHeader("CamelJmsDestinationName",
simple("$headers.receiver.split(\" \")[0]"))
.to("jms:dummy")
from("jms:"+containerId).to("agent:message");
These routes are defined using Camel's Java DSL. This is a Java API for constructing
routes via a sequence of method calls. The from method creates a consumer endpoint
and the to method creates a producer endpoint. Endpoints are specified using uniform
resource identifiers (URIs), with the first part of the URI (the scheme) identifying the
type of the endpoint. Other parts of the URI provide additional details, and the various
endpoint types provided by Camel make use of URI parameters to provide configuration
details for the instantiation of the endpoint. The routes shown above use two types of
endpoint: the agent endpoint described in this paper, and Camel's JMS endpoint for
sending and receiving messages from a message broker using the Java Message Service.
The first route definition above creates an endpoint that receives all messages sent by
local Jason agents. For each Jason message received, this endpoint copies the message
content into the body of a new Camel message, and records the other message details
using sender, receiver and illoc_force headers (these correspond directly to
Jason message properties).
The routes are run within a Camel context object. Our architecture allows
multiple distributed Camel contexts, each with their own set of local agents run-
ning within an agent container, so all agents are created with names of the form
containerId__localName. The second and third lines of the first route above
use Camel's "Simple" expression language to extract the first part of the name, which
identifies the agent container that the message recipient is attached to, and stores this
as the value of a specific header predefined by the JMS component. When the message
is processed by the JMS producer endpoint, this header is used to override the queue
or topic name that appears as a mandatory component of a JMS endpoint URI (hence
the "dummy" message queue name at the end of the first route above). This illustrates
two aspects of the use of message headers in Camel: they are commonly used within
routes to store information needed later in the route, and they can affect the handling of
messages by endpoints.
The second route definition above creates a JMS endpoint that receives messages
from a message broker (the address of the broker is provided to Camel's JMS component
on initialisation). The endpoint listens to a specific queue, which is named after the
unique identifier for the local agent container (note that there may be agent containers
associated with other Camel contexts running elsewhere on the network or in other
processes). The JMS consumer endpoint copies the body and the message headers from
the received JMS messages to create Camel message objects. The route specifies that
these messages flow from the JMS consumer endpoint directly to a Jason producer
endpoint. This endpoint generates Jason messages corresponding to the Camel messages
and delivers them to the appropriate agents. The Jason producer endpoint does the reverse
of the Camel to Jason message mapping described above.
Camel supports a number of message exchange patterns (MEPs), with the most
commonly used being InOnly and InOut. The pattern to use for handling a message
arriving at a consumer endpoint is set by that endpoint, possibly based on information in
the message (such as a JMSReplyTo header on incoming JMS messages). The MEP
can also be manually set by a route using methods of the Java DSL. If a message reaches
the end of a route with the InOut MEP, it is returned to the consumer endpoint. If that
endpoint supports it, that message will be treated as the reply to the initial request. Thus
Fig. 2. The architecture of our Jason/Camel bridge
Camel can be used to implement both synchronous and asynchronous processing of
messages.
Note that Camel routes can be significantly more complicated than those shown
above, as later examples in this paper will demonstrate. In particular, the Java DSL
includes methods for conditional branching, exception-handling and for starting, stop-
ping, suspending and resuming routes. In addition, an important feature of Camel is the
provision of methods that can be used singly or in combination to implement enterprise
integration patterns such as splitting and aggregating messages, or to "enrich" messages
with content obtained by making synchronous calls to other endpoints.
4 A Jason/Camel bridge
In this section we briefly describe the architecture of our Jason/Camel bridge and discuss
how we map between the conceptual models of Jason and Camel. In particular, we
describe the design and interpretation of agent endpoints.
4.1 Application architecture
Our Jason/Camel bridge2 consists of an "agent component" for Camel and an application
template that integrates the Jason BDI interpreter with a Camel context. The agent
component for Camel is a factory for creating agent consumer and producer endpoints.
Its implementation consists of the component class and classes that are instantiated to
create producer and consumer endpoints for Jason.
The architecture of the bridge is shown in Figure 2 using UML3. A Camel application
initialises any required components, creates a CamelContext object, passes it a Route-
Builder object with a method that defines the routes, and then starts the context. Our
2 http://github.com/scranefield/camel-agent
3 Classes and interfaces developed or adapted by us are shaded in the figure.
ApplicationDefaultCamelContextAgentComponent«interface»Component«interface»Endpoint«instantiate»«instantiate»«instantiate»«interface»Producer«interface»Consumer«thread»AgentProducer«thread»AgentConsumerAgentContainer«thread»SimpleJasonAgentConcurrentLinkedQueue«interface»AgentActionListenerAgentEndpoint3*«interface»Route***** integration architecture extends this by adding to the application an agent container.
On initialisation, this container locates all Jason agent source (.asl) files in a given
directory4 and, for each agent, instantiates our extension of the SimpleJasonAgent
class5. This class allows the Jason BDI interpreter to be used without any of the existing
Jason "infrastructures" for agent communication. It is responsible for providing the BDI
interpreter with methods to call to get percepts, to perform actions, and to send and
check for messages. We chose this as the most lightweight approach for embedding
Jason agents into business processes via Camel
Our SimpleJasonAgent class maintains concurrently accessible queues for per-
cepts of two types (transient and persistent) and for incoming messages. Messages on
these queues are read (and consumed in the case of transient percepts) when the BDI in-
terpreter calls the class's methods for getting percepts and messages. Note that each agent
and endpoint runs in a separate thread. The agent container writes messages and percepts
to the queues for the relevant agents after receiving them from agent:message and
agent:percept endpoints that appear in Camel routes. An endpoint for producing
percepts chooses whether percepts are transient or persistent based on the endpoint URI
parameters and/or the headers of the Camel message being processed. Transient percepts
are cleared after an agent has perceived them, whereas persistent ones will repeatedly
perceived (but may be overwritten by other percepts with the same functor -- see the
discussion of the updateMode URI parameter and message header in Section 4.2).
On construction, each agent is passed a list of agent consumer endpoints, and these
are used to deliver messages and actions -- the endpoints are responsible for selecting
which of these match their configuration parameters. Camel messages generated by the
consumer endpoints are processed using the InOnly message exchange pattern, unless
specified otherwise by a route or an endpoint URI.
Inter-agent messaging via a message broker, as implemented by the routes shown
above in Section 3, requires the existence of a separate message queue for each agent
container. To enable this functionality, our application class has a optional configuration
parameter specifying that an Apache ZooKeeper6 server should be used to dynamically
obtain a unique identifier for the container.
A ZooKeeper server maintains a set of named nodes, arranged in a tree structure, to
which system configuration information can be read and written by clients. The nodes
are kept in memory to enable high performance, but transaction logs and persistent
snapshots are also used to provide reliability. The data can be replicated across a cluster
of ZooKeeper servers. Nodes can be persistent or ephemeral -- a node of the latter type
is automatically deleted if the client session that created it is no longer maintaining a
"heartbeat". Nodes can also be sequential. These have a unique number appended to the
specified node name, based on a counter associated with the parent node. A client can
place a watch for changes to the data recorded in a node, the existence of a node, or
the set of children of a node. Together, these features can be used to implement a range
4 This simple approach will be replaced in the future by the use of OSGi "bundles" to package
and deploy Camel contexts together with their associated agents.
5 http://jason.sourceforge.net/faq/faq.html#SECTION00057000000000000000
6 http://zookeeper.apache.org/
Consumer endpoints
Endpoint type
agent:message illoc force,
Optional parameters Camel headers set
sender, receiver,
annotations,
match, replace
agent:action actor,
annotations,
match, replace,
resultHeaderMap
illoc force,
sender, receiver,
annotations,
msg id
actor,
annotations,
actionName,
params
Endpoint type
Optional parameters Camel headers used
Producer endpoints
agent:message illoc force,
sender, receiver,
annotations
agent:percept receiver,
annotations,
persistent,
updateMode
illoc force,
sender, receiver,
annotations
receiver,
annotations,
persistent,
updateMode
Table 1. Agent endpoint types
Camel body contains
The message content
(as a string)
The action term
(as a string)
Camel body expected
to be
The message content
(as a string)
The percept (as a string)
of distributed coordination mechanisms, such as distributed queues, barriers and locks,
maintaining lists of active group members, and electing group leaders.
Our application class obtains the agent container identifier by requesting the creation
of an ephemeral sequence node with the path containers/container and receives
in response the name of the created node with a sequence number appended.
ZooKeeper servers can also be accessed from within Camel routes, via ZooKeeper
endpoints. A use case for this functionality is illustrated in our MAS application scenario
in Section 5.
Another option provided by our bridge is to directly deliver messages between agents
that are in the same context, if preferred, rather than sending these to the Camel context
for routing via JMS or any other means specified by the provided routes.
4.2 Agent endpoint design
We support two types of Jason consumer endpoints to handle local agent messages
and actions delivered to them from our Jason/Camel bridge. Endpoints of these types
generate Camel messages that correspond (respectively) to messages sent by the local
agents and to actions executed by them. The details of the Jason messages and actions
are encoded in the headers and body of the Camel message, as shown in Table 1. For
example, the content of a Jason message is placed in the body of the Camel message,
and the illoc_force (illocutionary force), sender, receiver, msg_id and
annotations properties of the Jason message are stored on the Camel message using
headers with these names.
A route definition creates these types of endpoints by calling the from method
with an argument that is string of the form "agent:message?options" or
"agent:action?options". The options are specified using the standard URI query
parameter syntax ?opt1=v1&opt2=v2.... Camel messages are only generated by
these endpoints if the selection criteria specified by the optional parameters are satisfied.
The parameters recognised by these endpoint types are shown in Table 1 and explained
below.
We also support two types of Jason producer endpoints, which generate messages and
percepts, respectively, for the local agents. These messages and percepts are created from
Camel messages that reach the endpoints via Camel routes, and their content is taken
from the body and headers of those Camel messages and the endpoint URI parameters.
As shown in Table 1, the URI parameters supported for the producer endpoints are
mirrored by the headers that the endpoints check. This is because these message headers
can be used to override the URI parameters when converting a Camel message to a Jason
message or percept. This allows Camel routes to dynamically control the delivery and
construction of Jason messages and percepts.
The URI endpoint parameters and Camel message headers are used as agent message
and action selectors (for consumer endpoints) or to specify generated percepts or agent
messages (for producer endpoints). Below, we provide some additional details for some
of the parameter and header options.
receiver: We interpret the value "all" for this URI parameter and message header
as meaning that only broadcast messages should be selected by a message consumer
endpoint or that the message should be sent to all local agents from a message or percept
producer endpoint. This is the default value for a producer. No agent can have this name
because the agent container identifier is prepended to the names of all agents on creation.
The receiver value can also be a comma-separated list of recipients when provided
to a message or percept producer endpoint.
annotations: Jason supports the attachment of a list of annotation terms to a
literal. An annotations URI parameter or header can be specified for controlling the
selection of messages or actions by a consumer endpoint or to trigger the generation of
annotations by a producer endpoint. The values are specified as a comma-separated list
of literal strings (for the parameter) or as a Java list of strings (when using a header).
match and replace: These are used on consumer endpoints. A match parameter
specifies a regular expression, and a Camel message is only generated if this matches
the incoming message or action (in string format). The Java regular expression syntax
is used, and pairs of parentheses may be used to specify 'groups' in the pattern. The
values corresponding to these groups in the matched string are recorded and used when
processing a replace parameter (if present). A replace parameter specifies a string
to be used as the body of the generated Camel message. This can contain group variables
(in the form $n), and these are replaced with the values that were recorded during
matching.
resultHeaderMap: An action consumer endpoint supports both synchronous and
asynchronous actions. An asynchronous action corresponds to a Jason external action
Fig. 3. Architecture of our use case
(which cannot contain variables), and the endpoint always returns the result true to
the agent that performed the action. In order to handle actions that map to routes with
an InOut message exchange pattern, we implemented a Java class that provides a new
Jason internal action jasoncamel.syncInOut-Exchange. This is used to send
terms that represent actions with unbound variables to Jason action consumer endpoints.
Once the route (which is run with an InOut message exchange pattern) is completed, the
endpoint unifies the variables in the action term with the resulting Camel message body.
An endpoint processing this type of action must have a resultHeaderMap endpoint
parameter. Its value should be a comma-separated list of header-name:argument-index
pairs. When a Camel message completes the route, for each of these pairs the value of
the header with name header-name header is unified with the action term's argument at
index argument-index.
persistent and updateMode: As described in Section 4.1, percepts delivered to
agents by a percepts producer endpoint can be transient or persistent. The choice is
controlled by the updateMode URI parameter or a message header with that name.
Persistent percepts with the same functor and arity but different argument values can
either accumulate in an agent's persistent percepts list, or each new percept of that form
can replace previous ones. The latter case is useful for percepts that represent the state of
an external resource. A value of "-+" for an updateMode URI parameter or a Camel
message header with that name can be used to specify the percept replacement behaviour.
This can also apply to transient beliefs to prevent multiple percepts with the same functor
and arity being queued up between consecutive perceptions by an agent.
5 A business process use case
In this section we illustrate the use of the Jason/Camel bridge by describing a hypothetical
business process in which agents could play a valuable role. This use case addresses
the problem of achieving more targeted information flow within an organisation and
Mail serverZooKeeperserver clusterPrefs.Prefs.Prefs.Prefs.Prefs.Prefs.Application serverDBMSAgent clustersMessage broker Listing 1. Camel route for implementing an action as a database query
from("agent:action?exchangePattern=InOut" +
"&actionName=get_email_accounts" +
"&resultHeaderMap=result:1"
)
.setBody(constant("select email from users"))
.to("jdbc:dataSource")
.setHeader("result").groovy(
"exchange.in.body.collect{'\"'+it['email']+'\"'}")
reducing the overuse of the CC header in email messages. Our solution, shown in
Figure 3, assumes the existence of a specific "to.share" email account. Users with
information they think may be of interest to others can mail it to this account. Agents
monitor this account and evaluate the relevance of each new message to other users, with
each agent responsible for considering the interests and needs of a specific subset of
users. The sets of users assigned to the agents form a partition of the complete user base.
The agents base their decision on knowledge of the roles of users and the organisational
structure (stored in a database), as well as specific plans that may optionally be provided
by users to encode their goals for receiving information. We assume that these plans are
created using a graphical web interface that provides a user-friendly abstraction layer on
top of Jason's plan syntax. When agents determine which users might be interested in an
email message, they deliver the message to those users' mail accounts via SMTP.
Our system design for implementing this business process involves coordinated use
of Jason agents, a mail server, a database management system, a message broker and
ZooKeeper, with the coordination performed by Camel routes. The key routes are as
follows7.
1. On start-up, each agent performs an action get_email_ accounts(Accounts)
that is mapped to a database query by a Camel route. The route sets a message header
to hold the list of accounts, and the agent consumer endpoint instantiates the argu-
ment Accounts with this value. After this action succeeds, the agent records this
account list in a belief. This route is shown in Listing 1.
2. On start-up, each agent also performs a register action. A route maps this to the
creation of an ephemeral sequential node in ZooKeeper (under the node /agents).
3. A route is watching the children of the ZooKeeper node /agents. Whenever there
is a change (due to Camel contexts and their associated agent containers starting
and stopping), the route sends an updated list of active agents to its local agents as a
persistent percept in -+ update mode. This, and the route described in the previous
paragraph, are shown in Listing 2.
4. Whenever a new email account is created or deleted by the administrator, the
database is updated, and in addition a notification of the change is sent to a specific
publish-subscribe topic on a message broker (a topic is needed rather than a queue
7 Note that the two example routes presented earlier in Section 3 are not used because agents in
this application do not send messages to each other, but rather interact with external services
via actions and percepts.
to allow all running Camel contexts to receive the message). A route monitors this
topic and sends any received messages as transient percepts to all local agents.
5. Similar routes are provided for agents to obtain from the database information
about users' roles and their positions in the organisation structure, as well as a set
of default email-forwarding plans. This is done by database queries from routes
that are triggered by agent actions. Notifications of changes to this information are
sent by an administrative tool (or a database trigger) to a message broker topic. A
route monitoring that topic generates updates to the corresponding persistent agent
percepts, which then cause agents to call the actions to load this information again.
6. Similar routes are also provided for agents to retrieve from a database, for a specified
set of users, their information relevance assessment plans, and to receive notifications
of changes to these plans via a message broker topic.
7. The agents have plans that react to changes in their beliefs about the currently active
agents and the list of email accounts. When a change occurs, they each run an
algorithm (common to all agents) to divide the list of email accounts amongst them,
based on their own position in the list of agents. They maintain a belief recording
the accounts they are responsible for.
8. A set of routes polls the "to.share" email account for new mail using a mail consumer
endpoint, sends a message to all local agents asking them to evaluate which of their
allocated email accounts the message is relevant for, aggregates the reply with the
email message, and forwards it to the nominated users, via a mail producer endpoint.
These routes are shown in Listing 3.
9. When the list of accounts that an agent is responsible for changes, or it is notified
of changes to the plans for any of the accounts it handles, it must re-fetch plans for
the relevant agents. The routes handling these notifications suspend the mail-polling
route and start another route that uses a timer endpoint to resume the mail-polling
route after a fixed amount of time. This gives agents time to fetch any required new
plans from the database.
Listings 1, 2 and 3 show the routes we have implemented and tested for three aspects
of the system's functionality. We underline the beginnings of the agent endpoint URIs to
highlight where the integration with agents occurs. Listing 1 shows how the execution
of an agent action literal with a free variable can be implemented by a Camel route
with the InOut message exchange pattern (note the use of the standard Camel URI
parameter exchangePattern). The route sends an SQL query to a pre-configured
database connection, the returned result is converted to an AgentSpeak list of strings
using a Groovy expression, and then the result header is used to store the result.
The consumer endpoint URI has a resultHeaderMap parameter specifying that the
endpoint should unify the value of the result header with the first argument of the
action literal.
Listing 2 illustrates how Camel provides a convenient way to use ZooKeeper to
monitor the active members of a distributed group of agents, and to map this information
to agent percepts. The first route (lines 3 -- 11) implements the agent register action
by creating an ephemeral sequential node (see Section 4.1) in a ZooKeeper server to
represent the agent, and storing its name in that node. Camel's support for the idempotent
receiver enterprise integration pattern provides a simple way to filter out duplicate
Listing 2. Camel routes for tracking active agents via ZooKeeper
1 // Implement registration by creating a new ZooKeeper sequence
2 // node with the agent name as its content
3 from("agent:action?actionName=register")
4
// Process only one register action from each agent
.idempotentConsumer(
5
6
7
8
9
10
11
header("actor"),
MemoryIdempotentRepository.memoryIdempotentRepository(100)
).eager(true)
.setBody(header("actor")) // Put actor name in message body
.to("zookeeper://" + zkserver + "/agents/agent" +
"?create=true&createMode=EPHEMERAL_SEQUENTIAL");
12
13 // Watch agents node in ZooKeeper for changes to list of children
14 from("zookeeper://" + zkserver + "/agents" +
15
"?listChildren=true&repeat=true")
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
.setHeader("numChildren", simple("${body.size}"))
.split(body()) // Split agent node list into separate messages
.process(new Processor() {
public void process(Exchange exchange) throws Exception {
// Map the ZooKeeper node name for an agent to the agent
// name by getting the content of the ZooKeeper node
ConsumerTemplate consumer = camel.createConsumerTemplate();
String agentName =
consumer.receiveBody("zookeeper://"+zkserver+"/agents/"
+ exchange.getIn().getBody(),
String.class);
exchange.getIn().setBody(agentName);
}})
// Aggregate mapped names into a single message containing a
// list of names. All messages will have the same headers - any
// will do as the message correlation id
.aggregate(header("numChildren"),
new ArrayListAggregationStrategy()
).completionSize(header("numChildren"))
.setBody(simple("agents(${bodyAs(String)})"))
.to("agent:percept?persistent=true&updateMode=-+");
Listing 3. Camel routes for forwarding email based on agent recommendations
1 // Poll for email messages
2 from("imaps://mail.bigcorp.com?username=to.share"
3
+ "&password="+mailPassword+"&delete=true©To=processed")
4
5
.setHeader("id", simple("\"${id}\""))
.to("seda:forward-message", "direct:ask-agents");
6
7 // Request agents to evaluate message on behalf of their
8 // allocated users
9 from("direct:ask-agents")
10
.setBody(
11
12
13
14
15
16
simple("check_relevance(" +
"${header.id}, \"${header.from}\", " +
"\"${header.subject}\", \"${bodyAs(String)}\")"))
.setHeader("receiver", constant("all"))
.setHeader("sender", constant("router"))
.to("agent:message?illoc_force=achieve");
17
18 // Receive responses from agents and aggregate them to get a
19 // single lists of relevant users
20 from("agent:message?illoc_force=tell" +
21
"&receiver=router" +
"&match=relevant\\((.*),(.*)\\)" +
"&replace=$1:$2")
22
23
24
25
26
27
28
29
30
.setHeader("id", simple("${body.split(\":\")[0]}"))
.setBody(simple("${body.split(\":\")[2]}"))
.aggregate(header("id"),
new SetUnionAggregationStrategy()
).completionTimeout(2000)
.setHeader("to", simple("${bodyAs(String)}"))
.to("seda:forward-message");
31
32 // Aggregate original mail message with message summarising
33 // interested users in "to" header, and send it
34 from("seda:forward-message")
35
.aggregate(header("id"),
36
37
38
39
new CombineBodyAndHeaderAggregationStrategy("to")
).completionSize(2)
.setHeader("from", constant("[email protected]"))
.to("smtp://[email protected]?password="+mailPassword);
registration requests from agents. The second route (lines 14 -- 36) is triggered by changes
to the set of ZooKeeper sequence nodes representing agents. On each change, it receives a
message listing the current sequence nodes. The splitter pattern is used (line 17) to obtain
a separate message for each node, and each of these triggers a query to ZooKeeper to get
the agent name stored at that node (lines 22 -- 27). Finally (lines 32 -- 34), the aggregator
pattern is used to combine the names into a list stored in the body of a single message,
and that is sent to the local agents as the argument of a percept (lines 35 -- 36).
In the first route in Listing 3, the to.share mail account is polled for new mail (lines
2 -- 3). A Camel message representing each new mail message is generated and the Camel
message exchange identifier is written to a message header for latter use in correlating
the agent responses with this Camel message (line 4). The message is then forwarded to
two other routes (line 5). One is started asynchronously (via a "seda" endpoint, which
queues incoming messages) and the other synchronously (via a "direct" endpoint). The
second route (lines 9 -- 16) sends an achieve request to the local agents, asking them to
consider whether the mail is relevant to any of their allocated users. The third route (lines
20 -- 28) handles messages sent by agents in response to this goal, which contain lists of
potentially interested users. The aggregator pattern (lines 24 -- 26) is used to produce, for
each email message, a single message containing a combined list of users to forward it
to. This is sent to the final route (lines 32 -- 37), which also has (in a queue) the Camel
message containing the email message that is waiting to be forwarded. This route uses
the aggregator pattern again to combine the email message and the list of users to send
the message to (stored in the to header). Finally, an SMTP endpoint is used to send the
mail to these users.
The routes discussed in this paper have been tested using Jason stubs and the
necessary external services, but the full Jason code for this scenario has not yet been
developed and is not the focus of this paper. However, because the coordination logic
is factored out and encoded in the Camel routes, the agent code required will be much
simpler than would otherwise be needed without the use of our Jason/Camel bridge.
Most of the agent behaviour is to react to percepts sent from Camel by performing
actions (e.g. to fetch an updated list of email accounts), and to use Jason's .add_plan
and .remove_plan internal actions to update the plans used to evaluate the relevance
of email messages to users. In response to the goal to evaluate a message, the agent must
call the user plans, collect the users for whom these plans succeed, and send these in
a message to Camel. The agents must also recompute the allocation of users to agents
whenever the set of agents changes or new users are added, which they detect via 'new
belief' events.
6 Related Work
One of the oldest approaches to integrating agents with other technologies is the use of
wrappers or transducers that make the functionality of all the tools to be interconnected
available through agent communication [4]. The overall system coordination can then be
treated as a pure multi-agent system coordination problem. However, this approach has
not gained traction in industry and we do not see it as a viable approach for integrating
agents into enterprise computing environments.
A pragmatic but low-level approach for integrating agents with external systems is
to call them directly from the agent program. If an agent platform is a framework for
using a mainstream programming language for agent development (e.g. JADE8), then
it is possible for agents to use whatever protocols and client libraries are supported in
that language to invoke external services directly from within agents or to monitor for
external events. An interpreter for a specialised agent programming language may allow
user-defined code in the underlying implementation language to implement functionality
called by the agent program. For example, new "internal actions" for Jason can be
developed in Java, and these can use any Java communication libraries for external
interaction. An agent's environment abstraction is another potential location for user
customisation. For example, a Jason developer can implement an environment class that
acts as a facade for external interaction.
The integration of agents with web services has been an important topic over the last
decade, and some agent platforms provide specific support for this. For example, the
online documentation for the JADE platform includes tutorials on calling web services
from JADE and exposing agent services as web services, and the Jack WebBots [1]
framework allows web applications to be built using agents.
More generally, it would be possible for the developers of an agent platform (or its
community) to provide support for connecting agents to a range of external resource
and service types. For example, the IMPACT agent platform [9] includes a module that
provides a uniform interface for connecting agents to external services, with support for
a small number of service types already implemented.
The A&A (Agents and Artifacts) meta-model extends the concept of an agent
environment to include artifacts. These represent resources and tools with observable
properties and specific operations that agents can invoke. These can be used to provide
services internal to an MAS, or as an interface to external services, such as web services
[8]. However, it is unlikely that the developer and user community for any agent-specific
technology, whether a specific platform like IMPACT or a more general approach such
as A&A, could rival the scale and diversity provided by a more mainstream integration
technology such as Camel, which supports more than 130 endpoint types. Also, for the
case of A&A, an agent developer would need to learn multiple APIs (for each artifact
type) when integrating agents with different types of external service. This is not the
case in our approach (see Section 7).
The active components paradigm is a combination of a component model with agent
concepts [7]. Active components can communicate via method calls or asynchronous
messages and may be hierarchically composed of subcomponents. They run within a
management infrastructure that controls non-functional properties such as persistence
and replication. They may have internal architectures of different types, and this hetero-
geneity, combined with a uniform external interface model, facilitates the interoperation
of different types of system that are encapsulated as active components. As with artifacts,
the success of this approach for large-scale integration rests on the availability of active
components encapsulating a wide range of service types. However, a Camel context
could be encapsulated within an active component (or, alternatively, an artifact).
8 http://jade.tilab.com/
7 Conclusion
In this paper we have proposed a novel approach for integrating agents with external
resources and services by leveraging the capabilities of existing enterprise integration
technology. By using a mainstream technology we can benefit from the competitive
market for robust integration tools (or the larger user base for open source software),
and can have access to a much larger range of pre-built components for connecting
to different resource and service types. This is evidenced by Camel's large number of
available endpoint types.
We presented the design of an interface between agents and the Camel integra-
tion framework in terms of the EIP endpoint concept. This can serve as a pattern for
interconnecting agents with any type of message-based middleware.
We described an implemented architecture for this approach and illustrated its
practical use in a hypothetical (but, we think, plausible) business process use case. The
Camel routes we presented demonstrate the benefits of using a specialist coordination
tool such as Camel for handling the coordination of distributed agents and services,
and leaving the agent code to provide the required core functionality. This division
of responsibilities also enables a division of implementation effort: the coordination
logic can be developed by business process architects using a programming paradigm
that directly supports common enterprise integration patterns, and less development
time is needed from (currently scarce) agent programmers. An agent programmer using
our framework does not need to learn any APIs for client libraries or protocols -- the
agent code can be based entirely on the traditional agent concepts of messages, actions
and plans. The developer of the message-routing logic does not need to know much
about agents except the basic concepts encoded in the agent endpoint design (message,
illocutionary force, action, percept, etc.) and the syntax of the agent messages to be sent
from and received by the message routes.
References
1. Agent Oriented Software: JACK Intelligent Agents WebBot manual. http://www.aosgrp.com/
documentation/jack/WebBot Manual WEB/ (2011)
2. Bordini, R.H., Hubner, J.F., Wooldridge, M.: Programming Multi-Agent Systems in Agent-
Speak using Jason. Wiley (2007)
3. Chappell, D.: Enterprise Service Bus: Theory in Practice. O'Reilly (2004)
4. Genesereth, M.R., Ketchpel, S.P.: Software agents. Communications of the ACM 37(7), 48 -- 53
(1994)
5. Hohpe, G., Woolf, B.: Enterprise Integration Patterns: Designing, Building, and Deploying
Messaging Solutions. Addison-Wesley (2004)
6. Ibsen, C., Anstey, J.: Camel in Action. Manning (2010)
7. Pokahr, A., Braubach, L., Jander, K.: Unifying agent and component concepts. In: Multiagent
System Technologies, LNAI, vol. 6251, pp. 100 -- 112. Springer (2010)
8. Ricci, A., Piunti, M., Viroli, M.: Environment programming in multi-agent systems: an artifact-
based perspective. Autonomous Agents and Multi-Agent Systems 23(2), 158 -- 192 (2011)
9. Rogers, T.J., Ross, R., Subrahmanian, V.: IMPACT: A system for building agent applications.
Journal of Intelligent Information Systems 14, 95 -- 113 (2000)
|
1906.09874 | 1 | 1906 | 2019-06-05T17:44:36 | Escaping the State of Nature: A Hobbesian Approach to Cooperation in Multi-agent Reinforcement Learning | [
"cs.MA",
"cs.AI",
"cs.GT"
] | Cooperation is a phenomenon that has been widely studied across many different disciplines. In the field of computer science, the modularity and robustness of multi-agent systems offer significant practical advantages over individual machines. At the same time, agents using standard reinforcement learning algorithms often fail to achieve long-term, cooperative strategies in unstable environments when there are short-term incentives to defect. Political philosophy, on the other hand, studies the evolution of cooperation in humans who face similar incentives to act individualistically, but nevertheless succeed in forming societies. Thomas Hobbes in Leviathan provides the classic analysis of the transition from a pre-social State of Nature, where consistent defection results in a constant state of war, to stable political community through the institution of an absolute Sovereign. This thesis argues that Hobbes's natural and moral philosophy are strikingly applicable to artificially intelligent agents and aims to show that his political solutions are experimentally successful in producing cooperation among modified Q-Learning agents. Cooperative play is achieved in a novel Sequential Social Dilemma called the Civilization Game, which models the State of Nature by introducing the Hobbesian mechanisms of opponent learning awareness and majoritarian voting, leading to the establishment of a Sovereign. | cs.MA | cs | ESCAPING THE STATE OF NATURE: A HOBBESIAN APPROACH TO
COOPERATION IN MULTI-AGENT REINFORCEMENT LEARNING
A Thesis Presented
By
William F. Long
To
The Departments of Computer Science and Government
in partial fulfillment of the requirements
for a degree with honors of Bachelor of Arts
Harvard College
March 2019
Acknowledgments
This thesis is the product of an incredible network of supporters and advisors. I owe a huge debt
of gratitude to my advisors Prof. Michael Rosen, for his invaluable feedback and tireless editing,
and Prof. David Parkes and Max Kleiman-Weiner, for their patient direction and generous
availability. The work here is dedicated to my parents, without whose love and countless
sacrifices it would not have been possible. And as with all things, ad majorem Dei gloriam.
Abstract
Cooperation is a phenomenon that has been widely studied across many different
disciplines. In the field of computer science, the modularity and robustness of multi-agent
systems offer significant practical advantages over individual machines. At the same time, agents
using standard independent reinforcement learning algorithms often fail to achieve long-term,
cooperative strategies in unstable environments when there are short-term incentives to defect.
Political philosophy, on the other hand, studies the evolution of cooperation in humans who face
similar incentives to act individualistically, but nevertheless succeed in forming societies.
Thomas Hobbes in Leviathan provides the classic analysis of the transition from a pre-social
State of Nature, where consistent defection results in a constant state of war, to stable political
community through the institution of an absolute Sovereign. This thesis argues that Hobbes's
natural and moral philosophy are strikingly applicable to artificially intelligent agents and aims
to show that his political solutions are experimentally successful in producing cooperation
among modified Q-Learning agents. Cooperative play is achieved in a novel Sequential Social
Dilemma called the Civilization Game, which models the State of Nature by introducing the
Hobbesian mechanisms of (1) opponent learning awareness and (2) majoritarian voting, leading
to (3) the establishment of a Sovereign.1
1 The code implementation for this research is openly available at www.github.com/wlong0827/state_of_nature
TABLE OF CONTENTS
1
5
10
12
16
18
23
24
26
29
31
33
36
39
44
46
49
51
54
55
58
59
62
65
68
INTRODUCTION
OVERVIEW
COOPERATION AND GAME THEORY
MULTI-AGENT SYSTEMS
SEQUENTIAL SOCIAL DILEMMAS
REINFORCEMENT LEARNING
HOBBES AND LEVIATHAN
METHODOLOGY
AI AS HOBBESIAN MAN
MECHANISTIC MATERIALISM
STRICT DETERMINISM
MORAL SUBJECTIVISM
RATIONAL SELF-INTEREST
EQUALITY AND INDIVIDUALISM
THE STATE OF WAR
THE CIVILIZATION GAME
HOBBESIAN Q-LEARNING
OPPONENT LEARNING AWARENESS
MAJORITARIAN VOTING
SOVEREIGN REWARD SHAPING
RESULTS
Q-LEARNING VS HOBBESIAN Q-LEARNING
LEARNING CURVES
POLICIES
MATRIX GAME ANALYSIS
CONCLUSION
DISCUSSION
FUTURE WORK
WORKS CITED
I.
Introduction
Overview
The central goal of the field of Artificial Intelligence is to produce fully autonomous
agents that interact with their environments to learn optimal behaviors, improving over time
through trial and error. A principled mathematical framework for generating that experience-
driven learning is reinforcement learning (RL) which teaches software agents to take actions in
an environment so as to maximize some notion of cumulative reward. Incredible strides have
been taken with recent state-of-the-art RL algorithms like deep Q-networks and asynchronous
advantage actor-critic that enable intelligent agents to master previously intractable problems2
including, famously, playing graphical video games3 and beating humans at Go4. But as we
apply AI to larger and more complex problems like coordinating entire transportation or
manufacturing infrastructures, it becomes increasingly untenable for a single agent to master an
entire problem by itself.
Consequently, multi-agent systems (MAS) is one of the most important frontiers of
research within the field of AI. Research in this subfield is concerned with how to produce
desired global properties in "heterogeneous groups of autonomous agents that pursue partially
conflicting goals in an autonomous fashion."5 Even when each of a group of agents has identical
interests -- for example, when acting on behalf of a single user or organization -- physical or
computational considerations can make it desirable to have these agents make decisions
independently.6 With the advent of autonomous vehicles and the explosive growth of software
2 Arulkumaran, K et al. A Brief Survey of Deep Reinforcement Learning. 2017. 1.
3 Mnih, V et al. Playing Atari with Deep Reinforcement Learning. 2013.
4 Silver, D et al. Mastering the Game of Go without Human Knowledge. 2017.
5 Ossowski, S and Serrano, A. Social Coordination among Autonomous Problem-solving Agents. 1998. 134-148.
6 Groves, T. 1973. Incentives in Teams. 617 -- 631.
1
systems, MAS offer solutions to enable intelligent agents, while pursuing their own objectives
(e.g. driving a car from point A to B), to interact with other independent agents and humans in a
shared environment in a way that leads to optimal global behavior (e.g. minimizing total travel
time) without requiring external coordination.
When we place RL agents in a MAS setting, however, we get the problem of cooperation.
Traditional learning algorithms are self-regarding and have only the objective of optimizing their
behavior so as to maximize their own cumulative reward. So, when placing multiple self-
interested agents in a shared environment, often there is divergence between what is best for the
individual and what is best for the group. Such scenarios are called social dilemmas and they are
a matter of interdisciplinary interest from game theory and economics7 to biology8 and
sociology9. Multi-agent reinforcement learning faces an additional challenge because of the
unique perception-action-learning feedback loop within which each individual agent operates.
When the reward received by an agent in a state is a function of not only its action, but the
actions of other agents beyond its control, it can have difficulty determining the optimal policy to
follow. As a result, a research area of critical importance focuses on how to overcome these
obstacles and achieve cooperation in order to make MAS feasible. In the last several decades,
this field of research has developed "a vast array of techniques for cooperation and collaboration
as well as for agents to handle adversarial or strategic situations," but even so, "current
generation agents are unlikely to meet this new challenge except in very simple situations."10
7 Curtis Eaton, B. The elementary economics of social dilemmas. 2004.
8 Ale, S. B. Evolution of Cooperation: Combining Kin Selection and Reciprocal Altruism into Matrix Games with
Social Dilemmas. 2013.
9 Kollock, P. Social dilemmas: The anatomy of cooperation. 1998.
10 Grosz, B. A Multi-Agent Systems 'Turing Challenge'. 2013.
2
The field of political philosophy, likewise, is concerned with the cooperation of
autonomous, self-interested individuals in pursuit of common goods. Human society, or political
community, is the vehicle by which we achieve these collective goods and, consequently, it is
said that "society is the complete political good."11 While we often take the fact of human
cooperation within the nation-state for granted, political philosophers have long argued about
whether humans naturally form political communities and which preexists the other. Aristotle
summarizes one view when he claims that "men journey together with a view to particular
advantage, and by way of providing some particular thing needed for the purposes of life, and
similarly the political association seems to have come together originally, and to continue in
existence, for the sake of the general advantages it brings."12
Thomas Hobbes, one of the central modern political philosophers and the father of
modern liberalism, on the other hand, takes the opposite view. Gauthier summarizes Hobbes's
view as claiming that "individual human beings not only can, but must, be understood apart from
society. The fundamental characteristics of men are not products of their social existence." 13
Much of modern political philosophy and social contract theory exists in response to Hobbes'
ideas about human nature and interaction. His famous contention is that life in the State of
Nature is so terrible that it is overwhelmingly in the best interest of individuals "to confer all
their power and strength upon one Man,"14 the Sovereign, whose exclusive and absolute powers
of retributive punishment finally establish stable human society. Once pre-social man comes to
realize this necessity, he will desire to subject himself to laws and limitations as long as others
are also similarly desirous, and when a majority of them come to this agreement, they ratify the
11 Strauss, L. What is Political Philosophy? And other Studies. 1959. 10.
12 Aristotle, Ethics viii. 9.1160a
13 Gauthier, David, The Social Contract as Ideology. 138.
14 Id. at 120.
3
Social Contract and institute the Sovereign. This individualistic train of thinking shaped a strand
of later political thought as well, which Macpherson labels "possessive individualism,"15 in
which an individual is conceived as the sole proprietor of his or her skills and owes nothing to
society for them.
While it may be unclear at first sight, the problem of cooperation in the fields of MAS
and political philosophy overlap in significant ways. Both see a tension between individual and
collective interests and both are interested in developing practical mechanisms for moving
individuals away from acting on their short-sighted interests, which we might call defection,
towards a joint policy of cooperation that's strictly better for everyone involved in the long-run.
It stands to reason, then, that the political solutions that seem to have been successful in
promoting civil society for humans might also be conducive to producing cooperation in MAS.
Hobbes's theory of the state is well suited to AI because of the applicability of his theory of
human nature and the mathematical approach he takes to deducing the formation of the state
from those foundational assumptions.
Given these similarities, I suggest that, to the degree that the mechanisms in Hobbes's
political philosophy are actually effective in prompting humans to cooperate, those same
solutions should be beneficial for teaching MAS to find cooperative policies. The overarching
objectives of the thesis, therefore, can be listed as follows:
1) Present an original case for the applicability of Hobbes' philosophy of human nature and
moral psychology to AI.
2) Introduce the Civilization Game, a new Markov game modelled on the Hobbesian State
of Nature to measure the performance of a MAS composed of RL agents.
15 Macpherson, C. B. The political theory of possessive individualism: Hobbes to Locke. 1962.
4
3) Show that an augmentation of the Civilization Game and standard Q-Learning algorithms
with Hobbesian political mechanisms with Opponent Learning Awareness, Majoritarian
Voting, and Sovereign reward-shaping produces a significant improvement in the ability
of agents to find the long-term cooperative strategy.
This project applies methodologies from these two disparate fields, AI and political
philosophy, to synthesize solutions for producing cooperation. This project, beyond its
immediate research goals, also seeks to motivate further interdisciplinary exploration that draws
on long-standing philosophical ideas about human nature to inform contemporary work on
reproducing intelligence in machines. My hope is that the research and solution produced in this
thesis might offer a compelling instance of such a synthesis.
Cooperation and Game Theory
Cooperation is a wide-ranging concept that has been studied across many different
disciplines from economics and social science to evolutionary biology and philosophy. Olson
famously summarized the central problem by saying that "unless the number of individuals in a
group is quite small, or unless there is coercion or some other special device to make individuals
act in their common interest, rational, self-interested individuals will not act to achieve their
common or group interests."16 Hobbes likewise agreed that men are not like bees and ants for
whom "the Common good differs not from the Private."17 Even in evolutionary biology, it seems
that "Darwinian selection should preclude cooperation from evolving,"18 because of the constant
drive for individual survival.
16 Olson, Mancur. The Logic of Collective Action. 2009. 2.
17 Hobbes, Thomas. Leviathan. Tuck, R, Cambridge: Cambridge University Press. 2016. XVII. 119.
18 Supra, Ale at 1.
5
Popular consensus had generally been that groups of individuals with common interests
usually attempt to further those common interests. Groups of individuals with common interests
are expected to act on behalf of their common interests much as single individuals are expected
to act on behalf of their personal interests. This opinion about group behavior is frequently found
in both popular discussions and scholarly work. Such a view has, for example, been important in
many theories of labor unions, in Marxian theories of class action19, and in concepts of "counter-
vailing power."20 It also underpins the Madisonian concept of faction built into the American
political system wherein citizens are supposed to be "united and actuated by some common
impulse of passion, or of interest, adverse to the rights of other citizens"21 in a way that translates
into policy outcomes.
Olson's thesis motivated both the further study of the dynamics of individual behavior in
collective settings and the concept of the organization which "performs a function when there are
common or group interests," although organizations often also serve purely personal, individual
interests, "their characteristic and primary function is to advance the common interests of groups
of individuals."22 The formation and operation of organizations is often far from optimal,
however, as is shown by "the fact that profit maximizing firms in a perfectly competitive
industry can act contrary to their interests as a group."23 In the same way that market pressures
cause competing companies to produce and sell goods at lower and lower profit margins though
it is against the interest of each company individually, individuals with aligned common interests
may fail to realize them in an organization because of competing private interests.
19 Przeworski, A. Capitalism and social democracy. 1985.
20 Galbraith, J. American capitalism: the concept of countervailing power. 1956.
21 Hamilton, A. and Madison, J. The federalist papers. X.
22 Supra, Olson.
23 Chamberlin, E. Monopolistic Competition. 4.
6
So, then, "if the members of a large group rationally seek to maximize their personal
welfare, they will not act to advance their common or group objectives unless there is coercion to
force them to do so"24 Coercion is ultimately what enables communities to prevent the Tragedy
of the Commons (e.g. coordinating communal resource use, according to Ostrom25),
governments to pursue public goods (e.g. taxing citizens to build highways), and, crucially for
Hobbes's political system, enables individuals in the State of Nature to enforce contracts with
one another, for "Covenants, without the Sword, are but words, and of no strength to secure a
man at all."26 Indeed, we will see that for Hobbes, once a central coercive power has been
established, the situation individuals face in the State of Nature changes dramatically, and their
actions become calculated based not on what they think conduces to their self-interest, but what
the Sovereign knows to be best for the society.
Game theory offers a powerful, formal setting within which to study the individual vs
collective tensions that exist within organizations like human society or MAS. Hobbes is often
said to be the father of game theory since his broader project was to resolve the complex whole
of political society into simple social interactions. In De Cive, he writes:
For as in a watch, or some such small engine, the matter, figure, and motion of the wheels
cannot well be known, except it be taken asunder and viewed in parts; so to make a more
curious search into the rights of states and duties of subjects, it is necessary, I say, not to take
them asunder, but yet that they be so considered as if they were dissolved; that is, that we
rightly understand what the quality of human nature is, in what matter it is, in what not, fit to
24 Supra, Olson at 2.
25 Ostrom, E. Governing the commons: The evolution of institutions for collective action. 1990.
26 Supra, Leviathan at XVII. 117
7
make up a civil government, and how men must be agreed amongst themselves that intend to
grow up into a well-grounded state.27
Hobbes attempts to reduce the whole of human existence in the State of Nature into individual,
game-theoretic encounters between one individual and another, and by examining these
encounters, culminate with the Social Contract, which represents the establishment of civil
government. Depending on what kind of dilemma is faced by pre-social humans, we should see a
proper objective, function, and scope of the state.
Skyrms argues that "if one simple game is to be chosen as an exemplar of the central
problem of the social contract… the most appropriate choice is not the prisoner's dilemma, but
rather the stag hunt."28 The story is first told by Rousseau in 1750: "If it was a matter of hunting
a deer, everyone well realized that he must remain faithful to his post; but if a hare happened to
pass within reach of one of them, we cannot doubt that he would have gone off in pursuit of it
without scruple."29
The game is formalized into a game-theoretic environment when (1) the hunters each
have just the choice of hunting the hare or the deer; (2) the deer is much more valuable than the
hare; (3) the chance of getting the hare is independent of the actions of other hunters and so is
called the risk-dominant strategy; (4) the chances of getting the deer is directly dependent on the
number of cooperating hunters but is the payoff-dominant strategy.
27 Hobbes, Thomas. De Cive. II, xiv.
28 Skyrms, Brian. The Stag Hunt and the Evolution of Social Structure. 2004. xii
29 Rousseau, Jean-Jacques, A Discourse on Inequality, III.
8
Figure 1: Matrix game social dilemmas. A cell of X, Y represents a payoff of X for the row player
and Y for the column player. Each player chooses either to cooperate (C) or defect (D). In the Stag
Hunt, agents defect out of fear of a non-cooperative partner whereas in the Prisoner's Dilemma,
agents motivated by both fear and greed.
Represented in this way, the stag hunt interaction can be seen in many other places.
David Hume postulated a similar dilemma in his Treatise on Human Nature: "Two neighbors
may agree to drain a meadow, which they possess in common; because 'tis easy for them to
know each other's mind, and each may perceive that the immediate consequence of failing in his
part is the abandoning of the whole project. But 'tis difficult, and indeed impossible, that a
thousand persons should agree in any such action."30 Hume observed that cooperation in the stag
hunt is consistent with rationality, but that the viability of cooperation depends on mutual beliefs
and trust. Consequently, the Stag Hunt is also referred to equivalently as the Assurance Game.
Rational agents are pulled in one direction by considerations of mutual benefit and in the other
by considerations of personal risk.31 For these reasons, the greater the number of players or
agents in the game, the more difficult it is to achieve cooperation.
In this way, the problem of cooperation in the State of Nature can be seen as a Stag Hunt
dilemma between individuals where working together will result in common benefit but
cooperating when others refuse may lead to personal risk. In the Hobbesian account, then, both
the State of Nature and the state of society can be called Nash equilibria since it's always better
30 Hume, David, A Treatise of Human Nature, Book III, Pt II, Section VII, 538.
31 Supra, Skyrms at 3.
9
to cooperate when others cooperate and to defect when others defect. Thus, formulated in game-
theoretic terms, the fundamental question of both Hobbesian political philosophy and MAS
cooperation becomes: "how do we get from the hunt-hare equilibrium to the hunt-stag
equilibrium"?
Multi-agent Systems
A multi-agent system is defined as the structure in which multiple agents share a common
environment, each interacting with every other and having its internal state changed by its own
actions as well as those of others. An agent can be a physical or virtual entity that can act,
perceive its environment and communicate with others, is autonomous and has skills to achieve
its goals and tendencies; even human can be considered components of a MAS32. A MAS
contains an environment, objects and agents, relations between all the entities, a set of operations
that can be performed by the entities and the changes of the universe in time due to these
actions.33 Because each agent has no control over the actions of others and "no global control is
applied to the participating agents, it has a subjective view of the evolution of the world."34 Since
this "evolution" is partly a function of actions beyond any individual agent's control, we can
understand why a learning algorithm might find it difficult to converge to long-term cooperative
behavior since an agent's action only imperfectly maps to a resulting outcome or reward,
meaning that it must learn to reason about other agents' actions and intentions.
32 Mao A et al. Human Computation and Multiagent Systems: An Algorithmic Perspective. 2011.
33 Ferber, J. Multi-agent systems: an introduction to distributed artificial intelligence. 1999.
34 Schumacher, M. Objective Coordination in Multi-agent Systems. 2.
10
Figure 2: A typology of cooperation in MAS. The type that is aimed at in this experimental
setting is emergent cooperation since each agent in the Civilization Game maintains its own
independent agenda, self-preservation. (Doran et al.)
Cooperation is "one of the key concepts which differentiates MAS from other related
disciplines such as distributed computing, object-oriented systems, and expert systems."35 A
MAS is independent if each agent pursues its own agenda36 independently of the others. It is
discrete if it is independent and the agendas of the agents bear no relation to one another; these
systems involve no cooperation. On the other hand, a system's behavior is considered emergent
if it can only be specified using descriptive categories which are not to be used to describe the
behavior of the constituent components.37 This behavior maintains an intrinsically "subjective
aspect by directly depending on the cognitive properties of an observer."38 For example, the puck
gathering robots of Beckers et al.39 represent an independent system, each following the agenda
of moving in a straight line till an obstacle is encountered and then changing direction. Puck
gathering is an emergent behavior of the system in that, from an observer's viewpoint, the agents
appear to be working together, but from the agent's viewpoint they are not. They are simply
carrying out their own individual behavior. This emergent cooperation is ultimately the objective
35 Doran J. E. et al. On cooperation in multi-agent systems. 1996. 2.
36 Franklin, S. Is it an Agent, or just a Program?: A Taxonomy for Autonomous Agents. 1996
37 Forrest, S. Emergent Computation: Self-organizing, Collective and Cooperative Phenomena in Natural and
Artificial Computing Networks. 1990. 42.
38 Chantemargue, F, Lerena, P and Courant, M. Autonomy-based multi-agent systems: statistical issues.
39 Beckers, R. From Local Actions to Global Tasks: Stigmergy in Collective Robotics. 1994.
11
being pursued for the MAS playing the Civilization Game, and the relevant benchmark for
success will be something measurable only at the system level.
When MAS find successful strategies for cooperation, "the interactions of individual
agents possess a value-add compared with each simple agent capability."40 Specifically, these
advantages include natural implementations of problems that ask for distributed data and control,
greater robustness, scalability, and reusability. Looking ahead to an increasingly automated
future, MAS will also offer a valuable paradigm for human-computer cooperation.
Sequential Social Dilemmas
Before delving into the agents themselves, we'll lay out the structure of the environment
within which they will operate. In order to meaningfully explore the concept of cooperation in
MAS, agents must face a social dilemma: an environment or scenario which exposes the tension
between collective and individual rationality. If individual objectives align with the common
good, then cooperation is achieved trivially. Historically, the theory of repeated general-sum
matrix games has provided a powerful framework for understanding cooperation through a
mathematical model of social dilemmas and has produced famous strategies like Tit-for-Tat in
Iterated Prisoner's Dilemma.41
40 Nwana, H.S. and Ndumu, D. T. A Perspective on Software Agents Research. 1999.
41 Rapoport, A. Prisoner's dilemma-recollections and observations. 1974. 17-34.
12
Figure 3: Outcome variables R, P, S, and T are mapped to the cells of a generic game matrix.
Incentives are calculated as Fear = P -- S and Greed = T -- R and games are classified based on
those incentives that exist in a game.
At each stage of the game, there are precisely four possible outcomes: the player will
receive a reward for mutual cooperation, 𝑅, a punishment arising from mutual defection, 𝑃, a
"sucker" outcome obtained by the player who cooperates with a defecting partner, 𝑆, and a
temptation reward achieved by defecting against a cooperating player, 𝑇. A matrix game is a
social dilemma when its four payoffs satisfy the following social dilemma inequalities42:
1) 𝑅 > 𝑃. Mutual cooperation is preferred to mutual defection
2) 𝑅 > 𝑆 . Mutual cooperation is preferred to being exploited by a defector
3) 2𝑅 > 𝑇 + 𝑆. Mutual cooperation is preferred to an equal probability of unilateral
cooperation and defection
4) Either of greed: 𝑇 > 𝑅 or fear: 𝑃 > 𝑆. Mutual defection is preferred over being
exploited
These Matrix Game Social Dilemmas (MGSD) have been fruitfully utilized to study a
wide variety of phenomena and scenarios in social science and biology43, but game-theoretic
frameworks like the two-player Iterated Prisoner's Dilemma, Stag Hunt or other multi-player
matrix games44 necessarily treat the choice to cooperate or defect as a discrete, atomic action.
Either a hunter will choose to hunt hare or to hunt stag. In real-world social dilemmas, however,
42 Macy, Michael and Flache, Andreas. Learning dynamics in social dilemmas. 2002. 7229-7236.
43 Supra, Ale et al.
44 Broom, M, Cannings, C, and Vickers G. T. Multi-player matrix games. 1997. 931-952.
13
these choices are temporally extended. Cooperativeness is a property that applies to policies not
elementary actions and consequently exists on a graded spectrum. Decisions to cooperate or
defect are made with only impartial information about other players' activities and the state of
the world.45 A real hunter scouts, tracks, runs, and shoots, but we don't label these actions as
intrinsically cooperative or defecting. Rather when we observe the broader policy of observing
other hunters pursuing the stag and then joining in the chase, we call the hunter's strategy
cooperative, and when he chooses the easier route to the hare, disregarding the actions and
intentions of his fellow hunters, we call it defecting.
Recent research in social dilemmas centers around Markov game-based Sequential Social
Dilemmas (SSD) that treat cooperativeness as "a property that applies to policies, not elementary
actions" and requires "agents to learn policies that implement their strategic intentions." 46
Because cooperation and defection are measured through observed behavior and not simply
through binary choice, they offer the most powerful framework today for measuring true
cooperation in realistic ways.47 SSDs are a recent innovation on Markov Games that share the
mixed incentive structure of MGSDs but also require agents to learn policies that implement
their strategic intentions.
A Markov game is formally defined, in the two-player case for simplicity, as 𝑀 =
(𝑆, 𝑂, 𝐴1, 𝐴2, 𝑇, 𝑟) where 𝑆 is the set of states and 𝑂 is the observation function 𝑂 ∶
𝑆 × {1, 2} → ℝ𝑑 specifying each player's partial 𝑑-dimensional view of the current state of the
world, 𝐴1and 𝐴2 denoting the two players' sets of legal actions, a transition function 𝑇 ∶
𝑆 × 𝐴1 × 𝐴2 → Δ(𝑆) where Δ(𝑆) is the set of discrete probability distributions over the state
45 Leibo, Joel et al. Multi-agent Reinforcement Learning in Sequential Social Dilemmas. 2017. 2.
46 Id at 1.
47 Kleiman-Weiner, M. et al. Coordinate to cooperate or compete: abstract goals and joint intentions in social
interaction. 2016.
14
space 𝑆, and the reward function 𝑟𝑖 ∶ 𝑆 × 𝐴1 × 𝐴2 → ℝ for player 𝑖. If we further let each
player's observation function be defined as 𝑂𝑖 = {𝑜𝑖 𝑠 ∈ 𝑆, 𝑜𝑖 = 𝑂(𝑠, 𝑖)}, then to choose
actions, each player uses a policy 𝜋𝑖 ∶ 𝑂𝑖 → Δ(𝐴𝑖). Formalized in this way, it becomes apparent
that matrix games are simply the special case where the state is perfectly observable (𝑂𝑖(𝑠) = 𝑠)
and the size of the state 𝑆 = 1. In general, we can split the universe of all possible legal policies
into two disjoint sets of cooperative and defecting policies Π𝐶 and Π𝐷 where there is no
guarantee that Π𝐶 ∪ Π𝐷 = Π since for sequential behavior, cooperativeness will usually be a
graded property. Thus, we define these two sets by thresholding a continuous social behavior
metric 𝛼 ∶ Π → ℝ with values 𝛼𝑐 and 𝛼𝑑 such that 𝛼(𝜋) < 𝛼𝑐 ↔ 𝜋 ∈ Π𝐶 and 𝛼(𝜋) >
𝛼𝑑 ↔ 𝜋 ∈ Π𝐷. This behavior metric will represent our key benchmark for achieving
cooperation.
Formally, we can define an SSD as the tuple (𝑀, Π𝐶, Π𝐷) when the Markov Game 𝑀
satisfies the condition that there exist states 𝑠 ∈ 𝑆 for which the induced empirical payoff matrix
satisfies the social dilemma inequalities above. SSD games studied include the Gathering game,
where players choose between shooting other players which temporarily disables them and
gathering apples located around a grid-like map, and the Wolfpack game, where players can
choose to capture prey alone or capture together with other players to obtain an extra bonus.48
Both these games expose tensions between immediate short-term rewards and optimal
cooperative behavior that leads to long-term bonuses but requires adjusting one's policies to
coordinate with others. I introduce here a new SSD, the Civilization game, that faithfully models
the dynamics that Hobbes posits in the State of Nature and enables the study of a more complex
range of incentive structures than previous SSDs.
48 Supra, Leibo et al.
15
Reinforcement Learning
Turning now to the mechanics of the agents within a MAS, we see that there are three
broad approaches to learning: (1) supervised learning where the correct output is provided by a
teacher during a training phase; (2) un-supervised learning where there is no knowledge of the
correct output but where it is possible to distinguish equivalence classes in the input itself to
guide the learning process; and (3) reward-based or reinforcement learning where only a
feedback is provided by the environment on the utility of the performance. Of these, the theory
of reinforcement learning provides the most natural framework, deeply rooted in psychological49
and neuroscientific50 perspectives on animal behavior and how agents can optimize their control
of an environment. The family of RL algorithms, as a subset of all machine learning algorithms,
is particularly concerned with how to act in a given environment so as to maximize cumulative
reward. From this family, the Q-Learning algorithm represents an extremely powerful RL
technique because it is model-free, meaning it doesn't require complete knowledge of the
environment, and consequently well suited to play partially-observable Markov games.
Q-Learning computes a value for every state and action combination with a function 𝑄 ∶
𝑆 × 𝐴 → ℝ where the magnitude of the Q-value is proportional to the perceived value of
taking action 𝐴 at state 𝑆. At any state, a QL agent simply looks at the set of all possible actions
that can be taken and, under the epsilon greedy strategy, chooses with some probability 𝜖 to take
the action with the highest Q-value and takes a random legal action with probability 1 − 𝜖. This
strategy enables the agent to both explore the state space by performing actions it wouldn't
otherwise take while also exploiting learned experience. It also embodies the classic explore vs
49 Thorndike, E. L. Animal Intelligence: Experimental studies. 1911
50 Schultz, W., Dayan, P. & Montague, P. R. A neural substrate of prediction and reward. 1997. 1593 -- 1599.
16
exploit trade-off where taking actions that are known from experience to yield rewards may
prevent an agent from finding global maximums.
Procedurally, all Q-values are initially assigned a random value before learning begins.
Then, at every time 𝑡, the agent selects an action 𝑎𝑡 from state 𝑠𝑡 and receives reward 𝑟𝑡 and
updates the Q-value according to the Bellman update equation:
𝑄(𝑠𝑡, 𝑎𝑡) ← (1 − 𝛼)𝑄(𝑠𝑡, 𝑎𝑡) + 𝛼(𝑟𝑡 + 𝛾 ∙ max
𝑎
𝑄(𝑠𝑡+1, 𝑎))
The parameter 𝛼 ∈ [0,1] is the learning rate where a value of 𝛼 = 0 means that the agent
essentially learns nothing, exploiting only previous experience, and a value of 𝛼 = 1 means that
the agent considers only the most recent experience, ignoring prior knowledge. The discount
factor 𝛾 ∈ [0,1] determines the importance of future rewards where 𝛾 = 0 makes the agent
myopic, or short-sighted, by considering only immediate reward, and 𝛾 = 1 will make it strive
towards future reward. Taken altogether, the update equation describes a simple value iteration
update ran for a fixed length of time that utilizes a weighted average of old values and new
information. We will see later how RL and the learning parameters in particular find a parallel in
Hobbes' theory of the nature of human perception and decision-making.
Figure 4: The perception-action-learning loop where the action 𝑎𝑡 taken translates directly into
the reward 𝑟𝑡+1 and new state 𝑠𝑡+1. In a MAS, learning becomes unstable because the reward and
state are functions of other agents' actions over which a player has no control. (Sutton and Barto)
17
For Q-Learning in the multi-agent system setting, the i-th agent stores its own Q-
function 𝑄𝑖 ∶ 𝑆 × 𝐴𝑖 → ℝ in a Q-value table. The policy for that agent can then be written as:
𝜋𝑖(𝑠) = {
argmax𝑎 ∈ 𝐴𝑖 𝑄𝑖(𝑠, 𝑎),
𝑈(𝐴𝑖),
𝑝 = 𝜖
𝑝 = 1 − 𝜖
Where 𝑈(𝐴𝑖) denotes a sample of legal actions from the uniform distribution. In independent
multi-agent reinforcement learning, each agent regards the actions of others simply as part of the
environment. This independence assumption can be considered an assumption of bounded
rationality in that agents do no reasoning about other agents' learning and it represents a major
obstacle for achieving cooperation in MAS. Some promising work has been done recently to
make agents aware of opponents' learning51, and that paradigm will also play an important role
in implementing a Hobbesian mechanism for achieving cooperation.
Hobbes and Leviathan
From what we've considered so far, it might not yet be obvious why the political
philosophy of Hobbes would play an intuitive or valuable role in offering new solutions to the
problem of cooperation in MAS. Providing some historical and philosophical background on this
thinker and his unique ideas will hopefully help make the case for their applicability to the
problem at hand.
Thomas Hobbes (1588-1679) was an English political philosopher writing in a
tumultuous period of European history. From the French Wars of Religion and the bloody Thirty
Years' War to the English Civil War culminating in Cromwell's Protectorate, the 17th century
was one characterized by extraordinary instability and savagery. Given this kind of political
51 Foerster, Jakob et al, Learning with Opponent-Learning Awareness. 2018
18
turmoil, it is not surprising that Hobbes came to hold a view of human beings as creatures who
will, if unchecked, inevitably behave violently toward one another. The chief political struggle in
England during this period was between the Parliamentarians, who advocated the rights of
Parliament, and the Royalists, who supported King Charles I.
In particular, Charles's claim to the right to imprison nobles per special mandatum
domini reges (by special command of the King) raised the question of whether the King was
above the Law and could therefore subvert it in order to secure the Commonwealth or whether he
himself was subject to the Law. Hobbes, seeking to settle this disagreement which was causing
so much conflict and bloodshed, sided strongly with the King, presenting his famous argument in
Leviathan that the Sovereign embodied the body politic and consequently exercised complete
and absolute authority. No entity could behave with a single will and reject its Sovereign; he
could not be tried and executed by the authority of 'the people' as Charles I was. Hobbes's
royalist friends in 1640-2 were well satisfied with this argument, for it secured to the king all the
powers vis-à-vis Parliament which he claimed in those years,52 although they eventually realized
that those same arguments would go on to support obedience to Cromwell's republican
government.53 Hobbes himself was likely a staunch royalist whose radical arguments ended up
undermining his own monarchist position.54
Leviathan is both Hobbes's systematic construction of the social state and consequently,
it is the work this thesis will build its cooperative mechanism upon. It is also the culmination of
his metaphysical, moral, and natural philosophy begun in De Cive and Elements of Law, and so
these two works will primarily form the basis of the argument motivating the application of
52 Supra, Leviathan at xix
53 Skinner, Q. Conquest and consent: Hobbes and the engagement controversy. Visions of Politics. 2002. 287-307
54 Hamilton, J. Hobbes the royalist, Hobbes the republican. History of Political Thought. 2009. 411-454.
19
Hobbesian political philosophy to Artificial Intelligence. To the degree that his natural and moral
philosophy is applicable to artificially intelligent agents, the mechanism or process he proposes
in his political philosophy for the formation of society should in principle be applicable to
producing cooperation in MAS. Consequently, a substantial portion of this thesis will offer an
interpretation of his metaphysics and natural and moral philosophy as uniquely fitting for
artificially intelligent agents, specifically reinforcement learning-based AI. Finally, on the
strength of that comparison, it will implement Hobbes's political solutions as modifications to
standard RL to test in the Civilization Game.
Before that, understanding Hobbes's philosophical methodology is important. For
Hobbes, philosophy is divided into three branches: natural, moral, and political, each of which
study a different kind of "body". His natural philosophy is guided by a fundamental belief in
mechanistic materialism, claiming that the universe is machine-like, and strict determinism,
acting according to strict scientific laws of cause and effect and being composed exclusively of
material bodies. Similarly for humans, there is no such thing as soul or spirit; rather, everything
from our actions, emotions, and states of mind, are series of internal "motions" caused by
external sensation. This materialism enables Hobbes to attempt a "science of politics", deriving
political truths from fundamental, self-evident principles through a logical, almost mathematical,
methodology. He makes a case for political argumentation based on reason, or "ratiocination",
alone, since "the principles of politics consist in the knowledge of the motions of the mind."55
As Hampton puts it, "Hobbes characterizes his political argument as a 'demonstration' in
the style of geometry, in which his political conclusions are derived from his psychological
55 Supra, De Corpore at 6.74.
20
theories, which in turn can be derived from his physiological theories of human behavior." 56 In
his biography of Hobbes, John Aubrey explains the encounter that inspired this radical approach:
He was... 40 years old before he looked upon geometry; which happened accidentally. Being in a
gentleman's library..., Euclid's Elements lay open, and 'twas the 47 El. libri I. He read the
proposition. 'By G -- ,' said he (he would now and then swear, by way of emphasis), 'this is
impossible!' So he reads the demonstration of it, which referred him back to such a proposition,
which proposition he read. That referred him back to another, which he also read. Et sic deinceps,
that at last was demonstrably convinced of the truth. This made him in love with geometry. 57
Impressed with the incontrovertible conclusions which geometry enabled mathematicians to
show, Hobbes set his mind to deriving a system of government in a similar fashion. In his mind,
the serious political problems of his time resulted from disagreements over who had the ultimate
authority in political and religious affairs.
Consequently, a purely rational methodology could avert the constant quarrels amongst
philosophers about the proper structure and scope of government. This system of careful,
deductive, one might say, mathematical reasoning, seeks to establish definitively the proper
organization and scope of political society. Beginning with Hobbes's fundamental premises
about human nature and the universe, he forms a chain of consequences that passes through
morality and society before reaching politics. If we can show that Hobbes's premises in natural
and moral philosophy hold, then his political prescription for the problem of Cooperation should
also prove effective.
It's worth noting that this rationalist epistemology and materialist metaphysics radically
contrasted with the Christian Scholastic philosophy of Hobbes's immediate predecessors and to
other traditional thinkers like Augustine. Augustine, for example, also defined civitas as "nothing
56 Hampton, Jean. Hobbes and the Social Contract Tradition. 30.
57 Aubrey, John. A Brief Life of Thomas Hobbes, 1588-1679. 1898.
21
else than a multitude of men bound together by some associating tie,"58 but he emphasized that
this was insufficient to distinguish civitas from other forms of communal life. It is only when we
give due consideration to the religious and moral implications of the adjective "earthly," that the
distinctive traits of the civitas terrena become plain.59
And, for a modern example, Tocqueville insisted that he would "never admit that men
form a society by the sole fact that they acknowledge the same leader and obey the same laws;
there is a society only when men consider a great number of objects in the same way; when they
have the same opinions on a great number of subjects; when, finally, the same facts give rise
among them to the same impressions and the same thoughts."60 Disregarding mores, religion,
and common culture as superfluous to the formation of political society, Hobbes was freed to
pursue his geometric proof for the foundation of the state, but arguably lost the characteristics
that made his subjects uniquely human. But, regardless of whether it represents a compelling
model for human nature, I will show that his materialist and rationalist framework can apply to a
class of standard AI methodologies.
58 Augustine, City of God. XV. 73.
59 Loetscher, Frederick. St. Augustine's Conception of the State. 1935. 16-42.
60 Tocqueville, Alexis, Democracy in America. 598
22
II.
Methodology
AI as Hobbesian Man
The field of Computer Science and specifically its subfield Artificial Intelligence offers a
unique opportunity for Hobbesian scholarship, because, arguably, the tenets of Hobbes's natural
and moral philosophy apply remarkably well to computational intelligent agents. Six core
propositions, two from his foundational natural philosophy and four from his intermediate
human or moral philosophy, will enable us to both examine this applicability and understand the
full geometric logic of Hobbes's political argument for the necessity of the Sovereign:
1) Mechanistic Materialism -- The universe consists solely of sensible material and motion
2) Strict Determinism -- Action in the universe is entirely determined by material causes
3) The Moral Vacuum -- Humans can/should do whatever is necessary for self-preservation
4) Rational Self-Interest -- Humans are rational and driven by subjective interest
5) Radical Individualism -- Humans should be considered as fundamentally individuals.
6) Equality of Condition -- Humans are essentially equal in mental and physical capacity
Hobbes's mechanistic materialism and determinism (the first two propositions) lead him to see
human nature as characterized by moral subjectivism, rational self-interest, radical
individualism, and equality of condition, which ultimately serves to prove "by experience known
to all men and denied by none, to wit, that the dispositions of men are naturally such, that except
they be restrained by some coercive power, every man will dread and distrust each other."61 I
shall examine each in turn.
61 Supra, De Cive at 103.
23
Mechanistic Materialism
For Hobbes, all that exists is physical: "for the Universe, being the aggregate of all
bodies, there is no real part thereof that is not also Body, nor anything properly a Body, that is
not also part of the Universe."62 The idea of metaphysical realities like soul or spirit are reducible
to physical, dispositional, or imaginative phenomena since substances are corporeal by
definition. "Substance and body signify the same thing and therefore substance incorporeal are
words, which when they are joined together, destroy one another, as if a man should say, an
incorporeal body"63 Even God Himself is not what is commonly called Spirit, and when we refer
to him as such, we do not "signify our opinion of his Nature, but our desire to honor him with
such names as we conceive most honorable amongst ourselves."64 In effect, he deflects man's
claims about the reality of the supernatural or metaphysical as simply a pious means of
expressing our intentions and opinions.
This materialist premise quickly translates into a mechanical interpretation of human
action: since life "is but a motion of limbs, the beginning whereof is in some principal part
within, why may we not say, that all Automata have an artificial life? For what is the Heart, but a
spring, and the Nerves, but so many strings, and the Joints, but so many wheels, giving motion to
the whole both, such as was intended by the Artificer?"65 Man is no more than a mechanical
automaton controlled and impelled by an internal force. And what is that internal force? Mere
Imagination, simply the impression of some object upon the senses, the "conception remaining,
and by little and little decaying from and after the act of sense,"66 which he calls "the first
62 Supra, Leviathan at 269.
63 Id at 270.
64 Id at 271.
65 Id at Introduction, 9.
66 Hobbes, Thomas. The Elements of Law, Natural and Politic. III. 27.
24
internal beginning of all Voluntary Motion." By the movement of the Imagination, "these small
beginnings of Motion, within the body of Man, before they appear in walking, speaking, striking,
and other visible actions, are commonly called Endeavour" and finally, "this endeavor, when it is
toward something which causes it, is called Appetite or Desire"67 or their opposite, Aversion. In
this way, all human desire, and consequent action, is reduced to exclusively material causes
through a mechanistic operation. It is in this sense that we can call Hobbes's natural philosophy a
mechanistic materialist one.
The implications of this view can be difficult to appreciate in our modern context.
Hobbes has overturned the traditional teleological view of the universe begun with Aristotle for
whom "every craft and every investigation, and likewise every action and decision, seems to aim
at some good" and consequently, the good can be meaningfully defined as "that at which
everything aims."68 Instead of seeing natural ends as the impetus of action, Hobbes's mechanical
model sees action merely as the result of blind matter and motion. The Platonic and Christian
conception of the incorporeal soul which balances reason with appetite and spirit is rejected.
Most pertinently for his political philosophy, Hobbes's reduction of the motivation of all human
action to the desires elevates the centrality of "power", the ability to satisfy or obtain the objects
of Appetite or avoid the objects of Aversion.69 And thus, his famous assertion in Leviathan that
"I put for a general inclination of all mankind, a perpetual and restless desire of Power after
power, that ceaseth only in Death."70
Hobbes's mechanistic materialist view of the universe is the aspect of his thought most
obviously and immediately applicable to AI. Strictly defined, "modern computers are simply
67 Supra, Leviathan at VI. 38.
68 Supra, Nicomachean Ethics at 1094a1.
69 Supra, Leviathan at X.
70 Id at XI. 70.
25
Turing machines that operate on an alphabet consisting of the two symbols '0' and '1'."71 A
Turing machine is simply a mathematical model of computation that defines an abstract machine
which manipulates symbols on a strip of tape according to a table of rules. As its creator, British
mathematician Alan Turing, described it, "the machine is supplied with a 'tape' (the analogue of
paper) running through it and divided into sections (called 'squares') each capable of bearing a
'symbol'."72 It shuttles the tape back and forth, reading data, making marks on its scratch tape,
and inspecting the tape for information on what to do next. Turing proved that such a seemingly
naïve and rudimentary contraption, in fact, could embody every imaginable computational
operation. Nothing significantly changes when we consider reinforcement learning AI. Every
intelligent operation for a Q-Learning agent is simply a computation that considers memories of
the quantified "pleasure" or "pain" of being in some state and a further operation to decide what
action is best fitted to attain maximal utility. So, then, we can easily see the applicability of
Hobbesian materialism to AI.
Strict Determinism
Mechanistic materialism is also closely related to strict determinism. If every extant body
and potential cause of motion is perfectly contained within the physical, sensible universe, then
every subsequent effect or observable motion is reducible to an enumerable set of causes, which
are themselves the result of observable causes, and so. Consequently, free will is also cast aside
"because every act of man's will and every desire and inclination proceedeth from some cause,
and that from another cause, in a continual chain, they proceed from Necessity." 73 Likewise, if
we consider that at the root of this chain lies "one first mover, that is, a first and eternal cause of
71 Weizenbaum, Joseph. Computer Power and Human Reason. 1976. III. 73.
72 Turing, A. M. On Computable Numbers with an Application to the Entscheidungsproblem. 1936. 2.
73 Supra, Leviathan at XXI. 146.
26
all things, which is that which men mean by the name of God,"74 then we can also understand
why Hobbes might consider himself a sort of compatibilist instead when he explains that "the
liberty of man in doing what he will, is accompanied with the necessity of doing that which God
will, and no more, nor less" 75 Now, when we apply this Hobbesian premise of strict determinism
to human beings, we form a portrait of human cognition as a sole function of a set of past
experiences: "conception of the future is but a supposition of the same, proceeding from
remembrance of what is past… And that anything hath power now to produce another thing
hereafter, we cannot conceive, but by remembrance that it hath produced the like heretofore"76
The epistemological implications of viewing human learning in such a deterministic way seem to
preclude the possibility of transcending our mechanical minds and grasping real truth, and indeed
Hobbes himself comes out and explicitly states that "experience concludes nothing
universally."77 Rather, truth, is simply a matter of linguistic convention and "consists in speech
and not in the things spoken of."78 Strict Determinism in Hobbesian natural philosophy, far from
being a marginal and inconsequential tenet, clearly has profound implications upon the nature of
truth and the capacity for humans to attain it.
Similarly, "machines, when they operate properly, are not merely law abiding; they are
embodiments of law."79 In the same way that materialism implies that every action by a physical
bodies is a result of a fixed and observable set of physical motion and cause, so for computers,
the output of any Turing Machine is simply the conclusion of a set of fixed computational
operations upon an input. "Every effective procedure can be reduced to a series of nothing but
74 Id at XII. 77.
75 Id at XXI. 146.
76 Supra, Elements of Law at VIII. 48.
77 Id at I. 10.
78 Supra, De Corpore at 35.
79 Supra, Weizenbaum at II. 40.
27
commands (i.e. statements of the form 'do this' and 'do that') interlaced with conditional-branch
instructions"80 which simply perform different commands based on whether some claim is true.
For Artificial Intelligence, the details of the computational decision process that determines
action has been laid out thoroughly above and it is unnecessary to recreate it here.
Thus, there is a striking similarity between Hobbes account of human perception and
cognition and the reinforcement learning model. In the same way that Hobbesian man consumes
sense information about the world and then translates it into "endeavor" by means of an internal
motion, so it is precisely with AI. And just as Hobbesian man is restricted in his cognition to the
operation of his imagination which is no more than shadows and memories of past experience, so
reinforcement learning, by definition, draws up its actions solely from the outcomes of previous
interactions with its environment in order to find the optimal policy of action that results in the
obtainment of reward and the avoidance of punishment.
There is, however, one potential deviation from the deterministic nature of that process:
the introduction of stochasticity, or randomness, that enables agents to occasionally explore new
states that their decision process would otherwise avoid. The conditional comparison of a
random number and the exploration parameter in an epsilon-greedy learning strategy
determines whether the agent will decide their action based on what has been historically
desirable (exploitation) or without factoring in any past experience (exploration). The question of
whether true randomness is deterministic is an exceedingly difficult one for which there is no
clear consensus amongst philosophers or physicists.
80 Id, at III. 97.
28
Moral Subjectivism
Having examined the nature and role of materialism and determinism, we now turn to its
effects on Hobbes's moral philosophy. How are humans living in a Hobbesian universe supposed
to behave? What injunctions exist in the State of Nature? What are the prerogatives and
limitations in such a state?
The first and most natural consequence is moral subjectivity. In precisely the same way
that truth is simply a matter of usage and the combination of words into definitions, so "these
words of Good, Evil, and Contemptible are ever used with relation to the person that useth them,
there being nothing simply and absolutely so, nor any common Rule of Good and Evil to be
taken from the nature of the objects themselves."81 Having dispensed with Aristotle's
teleological model of the universe by assuming a materialist framework, Hobbes destroys any
independent conception of the Good by precluding metaphysics and then removing the
possibility of deriving it from natural ends.
The Laws of Nature that he proposes in Leviathan are simply rules discovered by reason
for best maintaining one's self-preservation that are binding in foro interno, or within one's own
thinking and intentions, but not in foro externo, or the actual performance in practice. For
example, while all men must strive for peace internally, that doesn't necessarily obligate you to
work towards it. "For he that should be modest and tractable, and preform all he promises, in
such time and place where no man else should do, should be make himself prey to others and
procure his own certain ruin, contrary to the ground of all laws of nature, which tend to nature's
preservation."82
81 Supra, Leviathan at VI. 39.
82 Id at XV. 107.
29
This is quite different from what we might normally associate with conscience or natural
law, which is both objective, existing independently of one's own conceptions, and binding on
external activity. What results, then, is a moral vacuum in the State of Nature with no source of
authority or judgment to restrain human action. And so, in the "war of every man against every
man, this also is consequent: that nothing can be Unjust. The notions of Right and Wrong,
Justice and Injustice have there no place."83 With these premises securely in place, Hobbes
secures himself from the charge that his proposed state of war is prohibited by human conscience
and the natural law, since what actually exists are only reason, power, and appetite.
Computers, too, have no equivalent to conscience or natural law traditionally understood.
This is perhaps one of the strongest reasons for interpreting AI in Hobbesian terms. In many
other systems of natural and moral philosophy, from the Platonic and Aristotelian to the
Thomistic, writers assert the existence of both body and soul in humans and that there is an
imprint of the Law of God "written on their hearts, while their conscience also bears witness, and
their conflicting thoughts accuse or even excuse them."84 Obviously, such a picture of human
nature would be entirely unfit to be modelled by a machine (at least in the existing state of
Computer Science), but the mechanistic and deterministic Hobbesian system, by throwing away
the concept of natural morality, clears the way for the comparison.
The incapacity of AI to navigate questions of morality is a subject of significant scholarly
and policy interest.85 With the advent of autonomous agents driving cars, determining terms of
parole, and fighting wars, policymakers must grapple with "the capacity of algorithms or trained
systems to reflect human values such as fairness, accountability, and transparency." 86 The
83 Id at XIII. 90.
84 Rom 2:15
85 Kleiman-Weiner, M. Learning a commonsense moral theory. 2017.
86 Calo, Ryan. Artificial Intelligence Policy: A Primer and Roadmap. 2017. 9.
30
Hobbesian political system offers a framework for achieving cooperation in multi-agent social
dilemmas even without reference to something like moral reasoning.
Rational Self-Interest
From the moral vacuum of the State of Nature, follows the primacy of rational self-
interest in human action. This Hobbes calls "the Right of Nature, which writers commonly call
Jus Naturale, the Liberty each man hath, to use his own power, as he will himself, for the
preservation of his own Nature… and consequently, of doing anything, which in his own
Judgment and Reason he shall conceive to be the aptest means thereunto"87 Because Law and
Right are opposed, with the complete absence of law, natural or civil, in the State of Nature
comes the monopoly of right in determining human action.
While it may seem that certain ethical norms like prohibitions upon the murdering of the
innocent, for example, may be universally held, Hobbes rejects any intrinsic metaphysical
consideration separate from Reason and self-interest. In fact, for him, the variance in these norms
across culture and time is greater than the similarity; since a "common standard of virtues and
vices," Hobbes says, "does not appear except in civil life; this standard cannot, for this reason, be
other than the laws of each and every state."88 Instead, humans always can and do pursue their
self-interest, subjectively defined as that which satisfies their Appetites and removes their
Aversions, which in the vast majority of cases (with a few notable exceptions) includes self-
preservation.
In this sense, Hobbes falls under the category of what contemporary philosophers might
term behavioralism. In place of the prescriptive traditional natural law of Aristotle and Aquinas
he substitutes new Laws of Nature derived purely from reason upon the sole premise self-interest
87 Id at XIV. 91.
88 Hobbes, Thomas. De Homine. B. Gert, Indianapolis: Hackett Publishing, 1991. 69.
31
which counsel among other things the pursuit of peace, the performance of covenants, and
gratitude for benefits rendered, all rationally calculated to achieve one's self-interest. By freeing
man from all the constraints of traditional ethical norms, natural law, and conscience, Hobbes
removes any factors from the State of Nature that would have been outside the scope of his
geometric and deductive methodology. Most importantly, the absence of objective Justice to
arbitrate human interaction and check each man's self-interest in the State of Nature allows
Hobbes to paint it as a state of complete license to invade a neighbor's territory or renege on
one's vows, leading ultimately to continual fear and reprisal.
The decision process for AI agents takes as input the current state of the world (e.g. my
location relative to other agents, how much territory I occupy, whether I've been invaded by
another, etc.) and compares it to the "memory" of past experience taking action in that state. In
the standard case where the agent's objective function is to maximize its own reward, it
considers only that single factor when choosing which action to take from a given state. Just as
for Hobbesian man, then, rational self-interest generally represents the exclusive principle of
action for AI. My decision to invade my neighbor's land and thereby inflict injury upon him is a
calculation made solely with a view towards what I consider is best for me and not informed by
anything like Sympathy or Justice. If I choose not to invade, it must only be because I expect
either that he will invade me in retaliation or that I expect for some other reason to receive
greater long-term benefit for taking a peaceful policy. In the primacy of rational self-interest,
then, we also see the close applicability of the Hobbesian moral framework to computer
intelligence.
32
Equality and Individualism
Finally, this leads us to consider the radical individualism and equality of physical and
mental condition that characterizes humans in the State of Nature. The two premises are closely
related and represent a dramatic reversal of the traditional understanding of human nature.
Hobbes sees man as fundamentally isolated from family and society as well as conflict-
prone, not least because of the differing appellations of words which causes misunderstanding.
Contrast this with Aristotle who held that "man is by nature a social animal" and so "society is
something that precedes the individual."89 Hobbes's Ninth Law of Nature prohibits Pride and
obliges "every man to acknowledge the other for his Equal by Nature"90 since, unless men were
equal, they would never enter into mutual Peace, but would find that the self-interest of the
strong would be to dominate their inferiors. At the same time, because "the difference between
man, and man, is not so considerable, as that one man can thereupon claim to himself any
benefit, to which another may not pretend, as well as he,"91 there is a consequent mistrust or
"diffidence" about the intentions of others. If I declare a certain plot of land to be my property, I
will immediately suspect my neighbors, who are my equals in strength and ambition, of making
designs upon it.
The connection with Hobbes's individualism is most strikingly observed in his view of
the family within which "there is not always that difference of strength or prudence between the
man and the woman, as that the right can be determined without War."92 By positing the
equivalence of the father and mother, he destroys the natural paternal authority associated with
the father in the family. And so also with children, that authority "is not so derived from the
89 Aristotle, Politics. 1253a.
90 Supra, Leviathan at XV. 107.
91 Id at XIV. 87.
92 Id at XX. 139.
33
Generation, as if therefore the parent had Dominion over his child because he begat him, but
from the child's Consent."93 Rather than admitting the existence of some natural authority
separate from the political, Hobbes claims that the dominion of a father over his children is of the
same sort of that of the Sovereign over his citizens: a covenant entered into by consent in order
to achieve self-preservation. This has the further effect of breaking the natural familial ties that
would otherwise enable each member of the family to rely on every other member for support
and protection. Equality thus begets individualism and both, in turn, make the State of Nature
one of isolation, distrust, and paranoia.
That equality and individualism apply equally to AI, is very plausible. In some MAS, we
can imagine agents that are completely equal in cognitive and active capacities. In a given
experimental setup, every "instance" of a learning agent can be the result of the same code and
thus share precisely the same decision processes, set of potential actions, and motivating
principle. The only thing separating each agent would be the set of lived experiences unique to it
and the accidents of chance and position that may cause individual deviances in observed
behavior. And while individualism does not follow from equality of condition for AIs as it does
for the Hobbesian man, it applies just the same, but for different reasons.
Intelligent agents have no concept of the family or natural authority that Hobbes goes to
such lengths to deny for humans. Whereas one might reasonably consider that a father has a
natural tendency to protect and provide for his child in the State of Nature, no such relation exists
amongst agents in a MAS except by convention. For agents that behave in a rational, self-
interested way, any tendency towards cooperation and mutual support that arises is calculated to
achieve that ideal of individual self-preservation which Hobbes places at the foundation of every
93 Id at XX. 139.
34
relation of dominion between father and child, father and mother, and citizen and Sovereign. Just
as the premises of equality and independence produce distrust and violence in humans, so they
may also potentially turn out to be a cause of defecting behavior in multi-agent systems. If this is
the case, we should expect to see improvements in cooperation with the introduction of an agent
with dramatically superior power to dominate or reward and thus shape the behavior of other
agents (roughly a picture of natural authority).
In summary, I have shown how, starting from Hobbes's materialist and determinist
natural philosophy, he incrementally deduces the moral subjectivism, rational self-interest,
radical individualism, and equality of condition that constitutes the human condition.
Furthermore, I've correlated each premise with his infamous conclusion that the State of Nature
is a state of "continual fear, and danger of violent death; and the life of man, solitary, poor, nasty,
brutish, and short."94 By drawing out the causal links in his deductive methodology, we clearly
see that for subsequent philosophers to avoid his political conclusions, they must either reject his
foundational assumption about a material and deterministic universe or else uncover a flaw in
any of the four succeeding moral philosophical premises. Since these properties have been
shown to be applicable to the standard form of AI methodology (with agents adopting a rational,
self-interested viewpoint), we should reasonably expect to find Hobbes's description of a brutal
State of Nature and his political prescription, the Sovereign, to be important components of the
problem and solution for cooperation in MAS.
94 Supra, Leviathan at XIV. 89.
35
The State of War
From the preceding model of human nature, Hobbes goes on to show that the state of pre-
social man is a state of war where all involved lose out. In this, it is notable that he stands apart
from other social contractarians like Locke and Rousseau who paint the State of Nature in almost
idyllic tones preferable to their 18th century political societies. It is arguably the case that, for AI
in social dilemmas like the Prisoner's Dilemma, the consequences of the cycle of defection make
it far closer to the Hobbesian prediction. Consequently, there should be great interest in
understanding his particular political solution, and before that, his diagnosis for why the State of
Nature is a state of war and how it follows from human nature.
Hobbes begins by distancing himself from a long-standing strand of thinking about
cooperation and human nature: "It is true that certain living creatures such as bees and ants, live
sociably with one another (… by Aristotle numbered amongst political creatures), … and
therefore some man may perhaps desire to know why mankind may not do the same."95 For
animals, the "social contract" is written into the very chromosomes of their cells; rationality is
clearly not necessary for the emergence of cooperation. Hobbes believed, however, that not only
was it unnecessary, but rational self-interest was the very source of the problem since "the
agreement of these creatures is natural; but that of men is by covenant only, which is artificial." 96
Humans do not naturally work together and furthermore, "the difference between man,
and man, is not so considerable, as that one man can thereupon claim to himself any benefit, to
which another may not pretend, as well as he." The result of equality of condition is diffidence,
or distrust, amongst men, since "if any two men desire the same thing, which nevertheless they
cannot both enjoy, they become enemies; and in the way to their End, endeavor to destroy, or
95 Supra. Leviathan at Pt. II, Ch. 17.
96 Id
36
subdue one another."97 Faced with this position of vulnerability with respect to his fellow men,
each man inevitably pursues the rational policy of pre-emptive invasion. Owing to the high
potential for conflict in the State of Nature, rational individuals recognize that striking first in
their struggle for life provides them with the necessary advantage to survive. Consequently, the
State of Nature can be called a state of war and "Force and Fraud are in war, the two Cardinal
virtues."98 Propositions 3 and 5 make these seeming vices morally justifiable since each man has
the Right of Nature "to use his own power, as he will himself, for the preservation of his own
Nature; that is to say, of his own Life."99
Of course, when everyone has a Right of Nature to seize whatever it is in one's power to
obtain, human wellbeing is fleeting and building civilization is impossible. "In such condition,
there is no place for Industry; because the fruit thereof is uncertain"100 So then, if others are
willing to lay down this right to achieve security, then the rational policy is always to seek peace
as well. "But if other men will not lay down their Right, as well as he; then there is no Reason
for anyone, to divest himself of his: for that were to expose himself to Prey, rather than to
dispose himself to Peace."101 But after enough experience of the state of war, there may come a
point where a majority of individuals decide that they would rather submit to any authority or
limitation on their power than go on any longer in that state. When such a plurality of agreement
has been reached, each man, by "the mutual transferring of Right… which men call Contract,"102
enacts what we call the Social Contract.
97 Id at 87.
98 Id at 90.
99 Id at 91.
100 Id at 89.
101 Id at 92.
102 Supra, Leviathan at 94.
37
Now, once enacted, what is necessary is the performance of the obligations of that
contract "without which Covenants are in vain and but empty words; and the right of all men to
all things remaining, we are still in the condition of War."103 Hobbes finally introduces the role
of the Sovereign, explaining that "there must be some coercive Power, to compel men equally to
the performance of their Covenants, by the terror of some punishment, greater than the benefit
they expect by the breach of their Covenant."104 Only a figure with absolute and unquestionable
authority could possibly coerce unscrupulous, opportunistic men to perform their contracts with
one another. By "conferring all their power and strength upon one man" in an act which "is more
than Consent, or Concord; it is a real Unity of them all," political society is finally established.
The Commonwealth culminates in "one Person, of whose Acts a great Multitude, by mutual
Covenants one with another, have made themselves every one the Author, to the end he may use
the strength and means of them all, as he shall think expedient, for their Peace and Common
Defense."105
From his model of human nature, Hobbes has produced a political argument for the state
of war based on several key underlying factors: the inherent distrust that each agent has for the
other, the consensus of a majority to limit their own rights in order to escape this plight, and
subsequently, the necessity of an authority vested with the power to reward and punish agents in
a way that makes them amenable to living in society. These are the three core components that
comprise the force of Hobbes's solution.
103 Id at 100.
104 Id at 101.
105 Id at 121.
38
The Civilization Game
The Civilization Game is a game that mimics the agrarian conditions of pre-social
humanity and represents an experimental model of the experience of primitive man in the
Hobbesian State of Nature. Players move to acquire territory, farm rewards from existing
territory, or invade other player's territory to loot their rewards. Long-term cooperative behavior
(i.e. farming rather than invading) leads to overall higher cumulative rewards for all players,
while defecting behavior leads to higher short-run rewards from invading, but lower long-run
rewards since being invaded in turn incurs a much larger penalty. The game is turn-based and
agents navigate a 2-dimensional grid-world, moving in the four cardinal directions. These actions
translate into two broad categories of action, farming one's own land or invading others', that
represent cooperation and defection, respectively. As they traverse the grid world, they acquire
"territory" by labelling each square they occupy with their unique player ID. In order to mimic
the incentives that exist in the State of Nature, an "invasion bonus," a positive reward given to
the agent who has invaded another agent's territory, and an "invasion penalty," a negative fine
levied on the agent who has been invaded, are instituted. Choosing to farm, meanwhile, produces
a reward for the agent proportional to the amount of territory it owns on its turn (i.e. one point
for each square held).
This setup captures the fundamental dynamics that Hobbes posits in the State of Nature.
Man is isolated and self-sufficient, and, left to his own devices, can secure a reasonable
subsistence for himself by farming his property. Once he finds himself in the company of
neighbors and subject to a scarcity of resources, however, a social dilemma rapidly develops: "it
comes to pass, that where an Invader hath no more to fear, than another man's single power; if
one plant, sow, build, or possess a convenient Seat, others may probably be expected to come
39
prepared with forces united, to dispossess and deprive him, not only of the fruit of his labor, but
also of his life or liberty. And the Invader again is in the like danger of another."106
Greed emerges as the incentive to invade others' territory, both because of the immediate
ability to steal their goods, but also because of the expansion of one's own territory leads to more
farming rewards each turn. Fear, also, is the incentive not to be invaded by another, for the same
two reasons; it is also the incentive that is completely essential for Hobbes since the prospect of
invasion or death is far more painful than the profits of invasion are pleasant. To reflect this, the
invasion penalty is greater in magnitude than the invasion bonus.
Figure 5: The Civilization Game with b = 4 and p = 2. Player i = 0 has name 𝑃0 = 𝑃0 and Player
𝑖 = 1 has name 𝑃1 = 𝑃1. On his move 𝑚 = 1, if 𝑃1 takes action 𝑎 = 𝑙𝑒𝑓𝑡, then he has invaded
𝑃0's territory. 𝑃1 will then gain a 25-point invasion bonus in addition to an 8-point farming
reward and set 𝑛𝑡+1
1 = 𝑇𝑟𝑢𝑒 to indicate that 𝑃0 should incur a -10-invasion penalty on 𝑚 = 0.
In practice, the grid-world is square sized (e.g. a 4x4 grid with 16 squares) and holds up
to four agents, initially placed at the corners of the grid. Formally, for every moment in game
time 𝑡 ∈ [0, 1, 2, … 𝑇], where we end the game after 𝑇 steps, the state is given by 𝑠⃑𝑡 ∈ 𝑆, where 𝑆
is the set of all possible states, the action taken is 𝑎𝑡 ∈ 𝐴(𝑠⃑𝑡), where 𝐴(𝑠⃑𝑡) gives the set of all
106 Id at 87.
40
legal actions available in state 𝑠⃑𝑡, and the subsequent reward is some real number 𝑟𝑡 ∈ ℝ. For an
instance of the Civilization Game with board size, 𝑏, and number of players, 𝑝, any state 𝑠⃑𝑡 can
be represented as a single vector consisting of three components:
𝑠⃑𝑡 = 〈𝑏⃑⃑
𝑡, 𝑛⃑⃑𝑡, 𝑚𝑡〉
The vector 𝑏⃑⃑
𝑡 represents the 𝑏 × 𝑏 grid-world board, with each location containing either a
player or a player's territory. 𝑛⃑⃑𝑡 stores information about whether each player has been invaded
in an array of booleans of size 𝑝 where the 𝑖th entry represents whether player 𝑖 has been invaded
by another player since its last move. 𝑚𝑡 ∈ [0, 1, … , 𝑝 − 1] is an integer that tracks which
player's move it is at time 𝑡 such that on move 𝑚𝑡 = 𝑚, player 𝑖 = 𝑚 takes an action. Player
names that are placed in the locations of players in the board are stored in an array 𝑃⃑⃑ of size 𝑝
(e.g. [𝑃0, 𝑃1, … 𝑃3] for a game with four players), and if a location on a board is the territory of
player 𝑖 ∈ [0, 1, … 𝑝], then we simply place the integer 𝑖 at that location. So, in total, with a
representation consisting of (1) a 𝑏 × 𝑏 board with each player appearing once, (2) a 𝑝 sized
array tracking invasions, and (3) a single byte storing the current move, there are
𝑏2!
(𝑏2 − 𝑝)!
× 𝑏𝑝(𝑏 − 1) × 2𝑝 × 𝑝
possible legal states. For a game with a 3x3 board with 2 players, this translates into 𝑆 =
6,912 possible states, but with a 4x4 board and 4 players, the state size balloons to 𝑆 =
67,092,480 states.
The possible actions are any of the four cardinal directions, or if none are legally
available, then the agent can choose to stay in its current location. Formally, 𝑎 ∈
[𝑢𝑝, 𝑑𝑜𝑤𝑛, 𝑙𝑒𝑓𝑡, 𝑟𝑖𝑔ℎ𝑡, 𝑠𝑡𝑎𝑦], and in a given time 𝑡 and state 𝑠⃑, we have 𝑎𝑡 ∈ 𝐴(𝑠⃑𝑡) where 𝐴(𝑠⃑𝑡)
returns the legal actions for player 𝑖 = 𝑠⃑𝑡,𝑚 that meet two criteria: (1) taking action 𝑎 will not
41
cause the agent to move out of the boundaries of the board, and (2) taking action 𝑎 will not cause
the agent to move onto the location of another player. Because there are 𝑝 players in the
Civilization Game, at each time 𝑡, the collective actions for each player are stored in an action
array 𝑎⃑𝑡 = 〈𝑎𝑡
0, 𝑎𝑡
1, … , 𝑎𝑡
𝑝−1〉. Importantly, the game is also asynchronous, or turn-based. So,
while 𝑎⃑𝑡 contains an action for each player at every time 𝑡 for ease of notation, only the action
for player 𝑖 = 𝑚𝑡 actually executes at time 𝑡.
Traditionally, most standard MAS methodology utilizes a synchronous model where
"agents repeatedly and simultaneously take action, which leads them from their previous state to
a new one," but a problem in defining the transition functions of agents arises "due to the fact
that the state in which the agent ends up after taking a particular action at a particular state
depends also on actions and states of other agents." There are two standard ways to address this
problem: we can "think of the transition function of the entire system, a mapping from the states
and actions of all agents to the new states of all agents" or else we have to "define the transition
function of each agent independently, and account for the effects of other agents by making the
function nondeterministic."107 Instead of taking either of these routes, agents play the
Civilization Game in an asynchronous, turn-based way that allows the transition function to
remain deterministic.
We define that transition function 𝑇(𝑠⃑𝑡, 𝑎⃑𝑡) as follows: (1) take the single action 𝑎𝑡
𝑖 for
player 𝑖 = 𝑠𝑡,𝑚 from location 𝑙 = 𝐼𝑁𝐷𝐸𝑋(𝑏⃑⃑
𝑡, 𝑃𝑖), (2) move the player from position 𝑏𝑡
𝑙 to a new
location 𝑙′ = 𝑀𝑂𝑉𝐸(𝑏𝑡
𝑙, 𝑎𝑡
𝑖 , 𝑖) defined by:
107 Shoham, Y. and Tennenholtz, M. On Social Laws for Artificial agent societies. 1995. 243.
42
𝑀𝑂𝑉𝐸(𝑏⃑⃑
𝑡, 𝑎𝑡
𝑖 , 𝑖) =
𝑙 − 𝑏 𝑎𝑡
𝑙 + 𝑏 𝑎𝑡
𝑙 − 1 𝑎𝑡
𝑙 + 1 𝑎𝑡
𝑎𝑡
𝑙
𝑖 = 𝑢𝑝
𝑖 = 𝑑𝑜𝑤𝑛
𝑖 = 𝑙𝑒𝑓𝑡
𝑖 = 𝑟𝑖𝑔ℎ𝑡
𝑖 = 𝑠𝑡𝑎𝑦
{
And setting 𝑏𝑡+1
𝑙′
= 𝑃𝑖 (3) if another player's territory was in the location moved to, or 𝑏𝑡+1
𝑙′ ∈
[0, 1, … 𝑝] − [𝑖], then set that player's invaded state to be True by setting 𝑛𝑡+1
𝑖′
= 𝑇𝑟𝑢𝑒 for
player 𝑖′ = 𝑃𝑏𝑡+1
𝑙
, (4) set the previous location to be the player's territory 𝑏𝑡+1
𝑙′
= 𝑖 unless 𝑎𝑡
𝑖 =
𝑠𝑡𝑎𝑦, (5) if player 𝑖 was invaded, or 𝑛𝑡
𝑖 = 𝑇𝑟𝑢𝑒, then reset it to 𝑛𝑡
𝑖 = 𝐹𝑎𝑙𝑠𝑒, and finally, (6)
increment the move to the next player's turn 𝑚𝑡+1 = (𝑚𝑡 + 1) % (𝑝 − 1). Notice that we
assume that 𝑎𝑡
𝑖 ∈ 𝐴(𝑠⃑𝑡) is taken from the set of legal actions available to agent 𝑖 at time 𝑡 so that
we preclude the possibility of moving to a location outside the board or onto the same location as
another player.
The reward a player receives on its turn in the Civilization Game is determined by three
factors: the amount of territory the agent controls and can farm, the invasion bonus accrued if it
invades its neighbor's territory (i.e. gains +10 points), and the invasion penalty if the agent was
invaded itself by another player (i.e. loses -25 points if 𝑛𝑡
𝑖 = 𝑇𝑟𝑢𝑒). Formally, for player 𝑖 we
can write the reward function as:
𝑅𝑖(𝑠⃑𝑡, 𝑎⃑𝑡) = {
𝑇𝐸𝑅𝑅𝑖(𝑠𝑡,𝑏) + 10 𝐼𝑁𝑉𝐴𝐷𝐸𝑖(𝑠𝑡,𝑏, 𝑎𝑡
𝑠𝑡,𝑚 = 𝑖
0 𝑠𝑡,𝑚 ≠ 𝑖
𝑖 , 𝑖) − 25 𝐼𝑁𝑉𝐴𝐷𝐸𝐷𝑖(𝑠𝑡,𝑛)
where the amount of territory is calculated by 𝑇𝐸𝑅𝑅𝑖(𝑏⃑⃑
𝑡) = 𝐶𝑂𝑈𝑁𝑇(𝑏⃑⃑
𝑡, 𝑖), whether the player
has invaded another player's territory is calculated by:
𝐼𝑁𝑉𝐴𝐷𝐸𝑖(𝑏⃑⃑
𝑡, 𝑎, 𝑖) = {
𝑙′
∈ [0, 1, … 𝑝] − [𝑖]
1 𝑏𝑡
0 𝑜/𝑤
And whether the player has been invaded is determined by
43
𝐼𝑁𝑉𝐴𝐷𝐸𝐷𝑖 (𝑛⃑⃑𝑡) = {
1 𝑛𝑡
𝑛𝑡
0
𝑖 = 𝑇𝑟𝑢𝑒
𝑖 = 𝐹𝑎𝑙𝑠𝑒
Note that, defined in this way, the Civilization Game satisfies the Markov property since the
deterministic transition from any state-action pair 𝑇: 𝑆 × 𝐴 → 𝑆′ does not depend upon prior
states. Furthermore, it satisfies the definition of a Markov game 𝑀 = (𝑆, 𝑂, 𝐴⃑, 𝑇, 𝑟) with a
perfectly transparent observation function 𝑂𝑖(𝑠⃑𝑡) = 𝑠⃑𝑡 and a reward function 𝑟: 𝑆 × 𝐴⃑ → ℝ.
Because of the incentives for both greed and fear, the Game also contains an embedded Matrix
Game Social Dilemma, but this must be proved experimentally by observing the long-term
payoffs achieved by actual agent gameplay.
Hobbesian Q-Learning
As I have already described, Q-Learning is a model-free reinforcement learning
algorithm whose goal is to learn a policy that determines which actions to take under what
circumstances. Normally, Q-Learning converges to an optimal policy when playing a finite
Markov Decision Process,108 but in non-stationary environments like the Civilization Game
where other actors can interfere with state transitions and rewards, it generally can't distinguish
between the effects caused by its actions and those caused by external forces, and consequently,
fails to converge. In these types of unstable environments, there is no general theory for the
performance of Q-Learning. In general-sum games there can be multiple equilibria and thus
initial conditions can alter the long-run payoffs that an agent and others involved can earn.
Consequently, in some classes of games, many different kinds of dynamics can converge on
equilibria with low payoffs even though high payoff equilibria exist.109
108 Melo, Francisco. Convergence of Q-Learning: A Simple Proof. February 2017.
109 Kandori, M et al. Learning, mutation, and long run equilibria in games. 1993. 29.
44
The Civilization game represents an example of a game where a policy of invading offers
high immediate incentives but results in negative long-run reward while a policy of farming may
produce low rewards but leads to a positive long-run score. In a 4x4 grid world and 4 Q-Learning
agents with the following value iteration update parameters: the discount factor 𝛾 = 0.99, the
soft update parameter 𝑎 = 0.5, and the annealing discovery parameter 𝜀 = 0.1, we observe that
agents don't converge to a high payoff cooperative strategy (i.e. avoiding invasion in favor of
farming, which leads to positive long-term rewards). Rather the learning agents hardly out-
perform agents that behave purely randomly, thus invading one another regularly, and any
improvements in score are sporadic and temporary.
Figure 6: Collective scores of 4 Q-Learning agents playing on a 4x4 board versus collective
scores of 4 Random agents. Collective score sums all players' rewards earned in a 𝑏𝑖𝑛 = 2.5𝑘
game step period and divides by 𝑏𝑖𝑛 to get a time average. Error bars are the max and min values
from 𝑡𝑟𝑖𝑎𝑙𝑠 = 3 runs while the line represents the median scores.
From the Hobbesian game-theoretic perspective, Moehler succinctly summarizes this problem in
the following way:
45
If a situation of social interaction among rational individuals, who behave as if they were to
maximize their expected utility, has the form of a one-shot PD game, cooperation does not take
place, assuming common knowledge of rationality and complete information. Rational individuals
alone are not able to realize the possible gains of cooperation in a one-shot PD game. Instead, they
end up with a suboptimal outcome, both individually and collectively. To realize the (collectively)
optimal outcome, an external authority, such as the state, is needed to transform the one-shot PD
game into another game that makes mutual cooperation rational.110
In order to remedy this fundamental issue, we need to introduce additional social mechanisms
that equip Q-Learning players with means of learning and interacting with other players in the
game. Making learning agents social by implementing Hobbesian mechanisms for producing
cooperation in MAS avoids explicitly requiring either a single programmer to coordinate and
control all agents or, in the other extreme, allowing agents to behave in an unconstrained manner
and appealing to a supervisor for resolving conflicts.111 We call a learning agent with the basic
Q-learning framework, but equipped with the Hobbesian mechanisms of 1) Opponent Learning
Awareness, 2) Majoritarian Voting, and 3) Sovereign Reward-shaping, a Hobbesian Q-learner or
HQ-learner.
Opponent Learning Awareness
The three following social mechanisms introduced closely follow the Hobbesian account
of the formation of political society from the state of consistent defection that learning agents
find themselves stuck in. The low payoff equilibrium of defection is, as Hobbes would call it, a
state of war that "consists not in actual fighting, but in the known disposition thereto,"112
transitions in his view to a situation in which each player is "willing, when others are so too, as
110 Moehler, M. Why Hobbes' State of Nature is Best Modeled by an Assurance Game. 2009. 297.
111 Supra, Shoham at 232.
112 Id at XIII. 88.
46
far-forth, as for Peace, and defense of himself he shall think it necessary, to lay down this right to
all things, and be contented with so much liberty against other men, as he would allow other men
against himself."113 The problem facing an agent who would like to stop invading and being
invaded, leading to negative rewards, is that as soon as he "lays down his arms," another agent
has an incentive to invade him for short-term gain. "Rational individuals recognize, according to
Hobbes, that striking first in their struggle for life provides them with the necessary advantage to
survive in the State of Nature."114 The Hobbesian term for this phenomenon is Diffidence, or the
distrust of other player's intentions, and plays a central role in the unattainability of cooperation
for him. This is also why Moehler calls the Hobbesian State of Nature an Assurance game and
considers the primary role of the Sovereign to be one of eliminating Diffidence and
institutionalizing social trust. "In order for rational individuals to leave the State of Nature, and
in this sense to cooperate with each other, the Sovereign does not have to change the structure of
the game that rational individuals play in Hobbes' State of Nature… Instead, the Sovereign only
must assure the individuals that if they lay down their rights to everything and agree to keep their
contracts, all other individuals will do so, too. The Sovereign must only assure the individuals'
trust in one another to establish society."115
The first thing the Sovereign does is create a channel of communication that makes each
agent's intentions accessible to every other player in the game. This Opponent Learning
Awareness mechanism mitigates the mistrust and uncertainty surrounding intentions by shaping
the anticipated learning of other agents in the environment. It modifies the learning rule to
include an additional term ∆𝑡
𝑖 that accounts for the impact of agent 𝑖's learning update on the
113 Id at XIV. 92.
114 Supra, Moehler at 300.
115 Id at 311.
47
anticipated parameter update of the other agents. That term is the update from the Bellman
equation:
𝑖
𝑄𝑖(𝑠𝑡, 𝑎𝑡) ← (1 − 𝛼)𝑄𝑖(𝑠𝑡, 𝑎𝑡) + ∆𝑡
𝑖 = 𝛼(𝑟𝑡 + 𝛾 ∙ max
∆𝑡
𝑎
𝑄𝑖(𝑠𝑡+1, 𝑎𝑡))
Every time 𝑄𝑖(𝑠𝑡, 𝑎𝑡) gets updated on move 𝑚 = 𝑖, we also update every other player's Q-table
with ∆𝑡
𝑖 . If, for example, a player 𝑖 chooses to take the defer action (i.e. cooperate and enter
political society) in state 𝑠𝑡 and earns some reward 𝑟𝑡, which might be positive or negative, then
the Sovereign enables every other agent to learn by applying the same update term ∆𝑡
𝑖 in 𝑠𝑡 but
putting them "in the shoes" of player 𝑖 by adjusting the current move 𝑠𝑡,𝑚 and the invasion store
𝑠𝑡,𝑖 to reflect an alternate reality in which the other agent is the one deciding to take an action in
the state.
In other words, OLA induces a sort of "empathy" in which, on every player's turn, all
other players can "imagine" what it might be like to be in that situation and take that action, and
their proclivity towards or away from the action is adjusted in the same direction as the player
who actually does take the action. This mechanism is an internal feature of learning agents
themselves, not an adjustment in the framework of the game or social dilemma. Formally, it
modifies each player's Q-value update to adjust at every time step 𝑡 and every move 𝑚, rather
than just on its turn 𝑠𝑡,𝑚 = 𝑚 = 𝑖 for player 𝑖, and gives us the new update equation:
𝑄𝑖(𝑠𝑡
𝑖, 𝑎𝑡) ← (1 − 𝛼)𝑄𝑖(𝑠𝑡
𝑖, 𝑎𝑡) + ∆𝑡
𝑚 for 𝑖 ∈ [0, 1, … 𝑝 − 1]
Where ∆𝑡
𝑚 is the learning update for the player whose move it currently is and
𝑖 = 〈𝑆𝑊𝐴𝑃(𝑏⃑⃑
𝑠𝑡
𝑡, 𝐼𝑁𝐷𝐸𝑋(𝑏⃑⃑
𝑡, 𝑃𝑖), 𝐼𝑁𝐷𝐸𝑋(𝑏⃑⃑
𝑡, 𝑃𝑚)), 𝑆𝑊𝐴𝑃(𝑛⃑⃑𝑡, 𝑖, 𝑚𝑡), 𝑚𝑡 = 𝑖〉
𝑆𝑊𝐴𝑃(𝑣⃑, 𝑖, 𝑖′) = {𝑣⃑ 𝑡𝑒𝑚𝑝 = 𝑣𝑖, 𝑣𝑖 = 𝑣𝑖′, 𝑣𝑖′ = 𝑡𝑒𝑚𝑝}
48
is the "alternate reality" state where player 𝑖 is put in the place of player 𝑚 by swapping board
locations, invasion states, and current move.
As another consequence of this new OLA update, every player performs a Q-update
every turn, rather than as before, each player performing its update only every 𝑝 turns. With this
Hobbesian learning awareness mechanism built in, we should expect to see faster and steadier
convergence to equilibrium policies across all players by increasing the frequency of learning
and sharing those updates. It can still be the case that the equilibrium converged to may be a low
payoff one, (i.e. mutual defection) so in order to reach the high payoff equilibrium, we need a
further mechanism to ensure that the policy convergence for agents goes toward cooperation.
Majoritarian Voting
The only way to escape this State of Nature for Hobbes, is to form a political union under
a Sovereign with sufficient instruments and power to shape the policy of individual agents. To
erect this Commonwealth is "to confer all their power and strength upon one Man;" an act which
"is more than Consent, or Concord; it is a real Unity of them all." This action is "the Generation
of that great Leviathan, or rather of that Mortal God, to which we owe under the Immortal God,
our peace and defense."116 Practically, how does this political society come about? In the
Hobbesian account, after a while, agents come to recognize the frustrating predicament in which
they are stuck and, as we see in the Civilization game, rewards are consistently negative. What is
required is to offer players a vote in order for them to signal to others their desire, whether to
remain in or to escape the State of Nature. After the final tally, "if the representative consist of
many men, the voice of the greater number, must be considered as the voice of them all."117 Each
116 Supra, Leviathan at XVII. 120.
117 Id at XVI. 114.
49
agent votes to "authorize and give up my Right of Governing myself to this [Sovereign], on this
condition, that thou give up thy Right to him and authorize all his Actions in like manner."118 If a
majority of agents vote to "defer" their right to choose their own action in favor of the
Sovereign's choice, then for Hobbes, that vote is binding on all players since the threat of
preemptive invasion disappears. "In a civil state, where there is a Power set up to constrain those
that would otherwise violate their faith, that fear is no more reasonable, and for that cause, he
which by the Covenant is to perform first, is obliged so to do."119 With the Sovereign in power
through that vote, invasion becomes sine jure because the Right of Nature no longer exists, and
consequently the Sovereign enforces the social law that no player may enter another player's
territory on its turn, but rather should stay on its own property and farm it, which leads to
positive long-term payoff equilibrium. As Hobbes puts it, "the Sovereign assigns to every man a
portion according as he, and not according as any Subject or number of them, shall judge
agreeable to Equity and the Common Good."120
The Majoritarian Vote mechanism lowers the threshold of the number of agents disposed
to cooperative action to achieve system-wide cooperation and once reached, universalizes that
policy by making it binding on non-cooperative agents. Instead of allowing a single invading
player to upset the cooperative equilibrium, a simple majority vote can commit all other players
to take the "defer" action, and once players become habituated in this way to choosing to
cooperate, they learn from the reward feedback that the policy is good for them in the long run.
Formally, we adjust the Civilization Game dynamics to account for a new potential action
"defer" that, with a simple majority vote every 𝑝 turns, commits all players to taking the "defer"
118 Id at XVII. 120.
119 Id at XIV. 96.
120 Id at XXIV. 171.
50
action. The new action set is 𝑎𝑡 ∈ [𝑢𝑝, 𝑑𝑜𝑤𝑛, 𝑙𝑒𝑓𝑡, 𝑟𝑖𝑔ℎ𝑡, 𝑠𝑡𝑎𝑦, 𝑑𝑒𝑓𝑒𝑟] and we increment the
number of possible moves to 𝑚 ∈ [0, 1, … 𝑝], making move 𝑚 = 𝑝 the period where all 𝑝 agents
take the Majoritarian Vote for whether to institute the Sovereign. The Hobbesian transition
function becomes piecewise based upon the current move:
𝑇 𝐻(𝑠⃑𝑡, 𝑎⃑𝑡) = {
𝑇 𝑉𝑂𝑇𝐸(𝑠⃑𝑡, 𝑎⃑𝑡)
𝑇(𝑠⃑𝑡, 𝑎⃑𝑡)
𝑠𝑡,𝑚 = 𝑝
𝑠𝑡,𝑚 ≠ 𝑝
Where 𝑇(𝑠⃑𝑡, 𝑎⃑𝑡) is the standard transition function from the standard Q-Learning model and the
voting procedure is described by:
𝑇 𝑉𝑂𝑇𝐸(𝑠⃑𝑡, 𝑎⃑𝑡) = {
𝑠⃑𝑡
〈𝑠𝑡,𝑏, 𝑠𝑡,𝑛, 𝑠𝑡,𝑚 = 0〉
𝑉𝑂𝑇𝐸(𝑎⃑𝑡) > 𝑝/2
𝑉𝑂𝑇𝐸(𝑎⃑𝑡) ≤ 𝑝/2
The function to count votes is given by 𝑉𝑂𝑇𝐸(𝑎⃑𝑡) = ∑ [𝑎𝑡
𝑝−1
𝑖=0
𝑖 = 𝑑𝑒𝑓𝑒𝑟]
. A successful vote
simply forces each player to take the "defer" action by setting 𝐴(𝑠𝑡) = [𝑑𝑒𝑓𝑒𝑟] until the next
vote at 𝑡′ = 𝑡 + 𝑝, not moving from their current board location since 𝑇(𝑠𝑡, 𝑑𝑒𝑓𝑒𝑟) = 𝑠𝑡, and
thus changing nothing about the state except taking 𝑝 turns. A failed vote simply advances the
move to player 𝑖 = 0 but with 𝑑𝑒𝑓𝑒𝑟 ∉ 𝐴(𝑠𝑡) for all players until the next vote at 𝑡′. The reward
each player accrues by playing "defer" after a successful vote is simply the farming bonus
proportional to the amount of territory held 𝑅𝑖(𝑠⃑𝑡, 𝑑𝑒𝑓𝑒𝑟) = 𝑇𝐸𝑅𝑅𝑖(𝑠𝑡,𝑏). As long as a majority
of agents choose to defer to the Sovereign, we should see an improved convergence to
cooperation in the MAS playing the Civilization Game.
Sovereign Reward-shaping
With OLA, we have an intra-agent mechanism for modifying the way that agents learn
and share their intentions leading to faster and more stable convergence while with Majoritarian
Voting, we've introduced a modification in the game environment that pushes players towards
the high payoff equilibrium. In the Hobbesian narrative, we're halfway to the state of society:
51
individuals who are empowered to overcome their mutual Diffidence subsequently elect, by
common assent, a single authority to whom they defer their actions. What remains is to
institutionalize the behavior of this Sovereign and determine, from the Hobbesian perspective,
how it conduces to the security and benefit of political society. Famously, Hobbes believed that
the Sovereign exercised absolute power over its subjects and no policy it executed could be
considered unjust. Even so, he still held that a good Sovereign knew that its own majesty and
dominion went hand-in-hand with the flourishing of the state, for "the Nutrition of a
Commonwealth consists in the Plenty and Distribution of Materials conducing to Life."121 To
this end, this primary prerogative is committed to the Sovereign: "the Power of Rewarding with
riches, or honor; and of Punishing with corporal, or pecuniary punishment, or with ignominy
every Subject according… as he shall judge most to conduce to the encouraging of men to serve
the Commonwealth, or deterring of them from doing disservice to the same."122 The institution
of reward for players when they choose to defer to the Sovereign and penalty when they defect is
the last component in the Hobbesian political prescription and requires modifying the
Civilization Game.
In particular, after every 𝑝 turns, if the Majoritarian Vote succeeds, or 𝑉𝑂𝑇𝐸(𝑎⃑𝑡) > 𝑝/2,
then every player receives a 10 point bonus and performs a Q-update for taking the 𝑑𝑒𝑓𝑒𝑟 action
in that state. Even players who didn't vote 𝑑𝑒𝑓𝑒𝑟 still receive the bonus and perform the update
as if they had voted to cooperate. If the Vote fails, then only those players who did vote to 𝑑𝑒𝑓𝑒𝑟
receive a -10-point penalty in line with the Hobbesian and Assurance game view that those who
attempt to lay down their arms while others choose to invade, suffer terrible consequences.
Stated formally, we add a new integer component to the game state, 𝑟 ∈ [−1, 1], that represents
121 Id at XXIV. 170.
122 Id at XVIII. 126.
52
whether we should apply the Sovereign reward or penalty to players after a vote. The new game
state is then:
𝑠⃑𝑡 = 〈𝑏⃑⃑
𝑡, 𝑛⃑⃑𝑡, 𝑚𝑡, 𝑟𝑡〉
And the transition vote function 𝑇 𝑉𝑂𝑇𝐸 only is modified to reflect the new state, since it is the
only function that touches the new state component, not the overall Hobbesian transition 𝑇 𝐻 or
the standard transition 𝑇:
𝑇 𝑉𝑂𝑇𝐸(𝑠⃑𝑡, 𝑎⃑𝑡) = {
〈𝑠𝑡,𝑏, 𝑠𝑡,𝑛, 𝑠𝑡,𝑚, 𝑠𝑡,𝑟 = 1〉
〈𝑠𝑡,𝑏, 𝑠𝑡,𝑛, 𝑠𝑡,𝑚 = 0, 𝑠𝑡,𝑚 = −1〉
𝑉𝑂𝑇𝐸(𝑎⃑𝑡) > 𝑝/2
𝑉𝑂𝑇𝐸(𝑎⃑𝑡) ≤ 𝑝/2
The reward function is also adjusted to account for the Sovereign's reward/punishment
𝑅𝑖
𝑆(𝑠⃑𝑡, 𝑎⃑𝑡) function:
𝑅𝑖(𝑠⃑𝑡, 𝑎⃑𝑡) = {
𝑇𝐸𝑅𝑅𝑖(𝑠𝑡,𝑏) + 10 𝐼𝑁𝑉𝐴𝐷𝐸𝑖(𝑠𝑡,𝑏, 𝑎𝑡
𝑅𝑖
𝑆(𝑠⃑𝑡, 𝑎⃑𝑡)
𝑖 , 𝑖) − 25 𝐼𝑁𝑉𝐴𝐷𝐸𝐷𝑖(𝑠𝑡,𝑛)
𝑠𝑚 = 𝑖
𝑠𝑚 = 𝑝
0 𝑜/𝑤
Where the Sovereign reward function is given by:
𝑅𝑖
𝑆(𝑠⃑𝑡, 𝑎⃑𝑡) = {
10𝑠𝑡,𝑟
𝑎𝑡
max (10𝑠𝑡,𝑟, 0) 𝑎𝑡
𝑖 = 𝑑𝑒𝑓𝑒𝑟
𝑖 ≠ 𝑑𝑒𝑓𝑒𝑟
Finally, we utilize the same Q-update equation from OLA, but add the additional voting move
𝑚 = 𝑝 to 𝑖 ∈ [0, 1, … 𝑝] with the further Sovereign reward from 𝑅𝑝
𝑆(𝑠⃑𝑡, 𝑎⃑𝑡).
These three mechanisms, then, Opponent Learning Awareness, Majoritarian Voting, and
Sovereign Reward-shaping form the backbone for a Hobbesian strategy for overcoming
defection and entering and preserving the state of society. For Hobbes, each one is closely
related to the others in this narrative and without all of them, we theoretically lose either the
stability, speed, or direction of joint policy convergence for the MAS. Consequently, the HQ-
Learning agent proposed will be deployed experimentally with a single, comprehensive toolkit.
53
III.
Results
Q-Learning versus Hobbesian Q-Learning
Having seen the problems that independent multi-agent reinforcement learning faces in
overcoming the low payoff equilibrium to achieve the high payoff long-term cooperative policy
and shown their similarity to the ones faced in the Hobbesian State of Nature, I have proposed
that the three aforementioned political mechanisms offered by Hobbes should also be applicable
solutions in MAS. The problem of mistrust and the ever-present fear of being invaded is
mitigated by making each agent aware of its opponents' learning and using that to align all
agents' policies. The next step is to lower the barrier for system-wide joint cooperation by means
of making a majority vote for cooperative action binding on all agents in the game, thus also
teaching the disruptive players about the benefits of cooperation. Finally, reward-shaping by the
Sovereign ensures that when votes succeed, a central figure has authority to reward cooperation
and punish defection.
In order to compare the performance of simple Q-Learning and Hobbesian Q-Learning,
which is Q-Learning with the aforementioned Hobbesian mechanisms for shaping game and
learning dynamics, the Civilization Game is played with 𝑏 = 4 and 𝑝 = 4 with an invasion
bonus of 10 points and invasion penalty of -25 points as the experimental standard. The value
iteration update parameters used are the discount factor 𝛾 = 0.99, the soft update parameter 𝑎 =
0.5, and the annealing discovery parameter 𝜀𝑡 = 0.9 × 0.9999𝑡. Majoritarian Voting activates
the 𝑑𝑒𝑓𝑒𝑟 action with a strict majority 𝑉𝑂𝑇𝐸(𝑎⃑𝑡) > 𝑝/2 and the Sovereign rewards 𝑑𝑒𝑓𝑒𝑟 with
a 15 point incentive while players who get duped (i.e. vote for 𝑑𝑒𝑓𝑒𝑟 while other players decide
to defect) incur a -10 point hit.
54
Learning Curves
Figure 7: Collective rewards for all 𝑝 = 4 players every 𝑏𝑖𝑛 = 2500 steps for 𝑇 = 250𝑘 steps.
Error bars indicate the results of running the game 𝑡𝑟𝑖𝑎𝑙𝑠 = 3 times with the median line
displayed. Score is a function of 𝑇𝐸𝑅𝑅𝑖(𝑠𝑡,𝑏), 𝐼𝑁𝑉𝐴𝐷𝐸𝑖(𝑠𝑡,𝑏, 𝑎𝑡
𝑖 , 𝑖), and 𝐼𝑁𝑉𝐴𝐷𝐸𝐷𝑖(𝑠𝑡,𝑛).
The first point of comparison is the collective score earned over time. With a sample rate
of 0.01 and 𝑇 = 250𝑘 game steps, we get a bin size of 𝑏𝑖𝑛 = 2500. Collective Score earned
starting at game step 𝑡 is then defined as
𝑡+𝑏𝑖𝑛
𝑝
𝐶𝑆𝑡,𝑏𝑖𝑛
𝑎
𝐵 = ∑ ∑ 𝑟𝑖
𝑖=𝑡
𝑎=0
where 𝑟𝑖
𝑎 is the reward earned by agent 𝑎 at time 𝑖. Random play consistently incurs negative
reward and acts as the base comparison. The rewards available to Q-Learning and Hobbesian Q-
Learning are made equal even though the learning remains distinct in order to make a fair apples-
to-apples comparison. Specifically, the rewards that the Sovereign provides apply to Q-Learners
if a majority decide to take the 𝑑𝑒𝑓𝑒𝑟 action, but unlike HQ-Learning, that reward does not
cause a Q-update. As we can see from Figure 7, the scores earned by Q-Learning players every
𝑏𝑖𝑛 steps do improve over time, even jumping into positive territory a handful of times, but
55
learning is slow and sporadic as expected. For HQ-Learning, on the other hand, we see consistent
and fast convergence to the high payoff equilibrium. The reward growth demonstrated in the
learning curve is a dramatic improvement over independent reinforcement learning, but we still
want to examine the policy underlying this performance improvement.
Figure 8: Collective invasions for all 𝑝 = 4 players every 𝑏𝑖𝑛 = 2500 steps for 𝑇 = 250𝑘 steps.
Error bars indicate the results of running the game 𝑡𝑟𝑖𝑎𝑙𝑠 = 3 times with the median line
displayed. Invasions measured by tracking changes to the game state 𝑛⃑⃑𝑡.
For any compelling case of cooperation in the Civilization Game, we should see
that agents learn to stop invading one another. With the same game parameters and bin
size of 𝑏𝑖𝑛 = 2500, we see that collective invasions drop off dramatically for HQ-
Learning, converging to 0 invasions per 𝑏𝑖𝑛 steps. Collective Invasions starting at game
step 𝑡 is measured every 𝑝 turns at times 𝐸𝑡 ∈ [𝑡, 𝑡 + 𝑝, … 𝑡 + 𝑏𝑖𝑛] as
𝑝
𝐶𝐼𝑡,𝑏𝑖𝑛
𝐵 = ∑ ∑[𝑛𝑖
𝑎 = 𝑇𝑟𝑢𝑒]
𝑖∈𝐸𝑡
𝑎=0
where 𝑛𝑖
𝑎 = 𝑇𝑟𝑢𝑒 if agent 𝑎 has been invaded at time 𝑖. As expected, invasions roughly
inversely follow game score because of the invasion penalty that detracts from the score.
56
Even so, if agents are invading one another less often in HQ-Learning than standard Q-
Learning, they must be taking some other action instead.
Figure 9: Collective successful defer votes for all 𝑝 = 4 players every 𝑏𝑖𝑛 = 2500 steps for 𝑇 =
250𝑘 steps. Error bars indicate the results of running the game 𝑡𝑟𝑖𝑎𝑙𝑠 = 3 times with the median
line displayed. Successful votes measured by tracking the game state when 𝑟𝑡 = 1.
The other metric that shows that agents have learned to cooperate in the
Civilization Game is the number of successful votes where more than half of agents
choose to defer their actions to the Sovereign. Formally, collective successful defers is
measured every 𝑝 turns at time 𝐸𝑡 ∈ [𝑡, 𝑡 + 𝑝, … 𝑡 + 𝑏𝑖𝑛] by:
𝑆𝐷𝑡,𝑏𝑖𝑛
𝐵 = ∑[𝑠𝑡,𝑟 = 1]
𝑖∈𝐸𝑡
Just as in the collective score and invasions over time, Q-Learning learns to vote 𝑑𝑒𝑓𝑒𝑟
only inconsistently while HQ-Learning learns the cooperative policy stably, because of
the alignment of agent learning with OLA, and quickly, because of the Sovereign reward-
shaping. OLA is especially important here because only cooperative action by 𝑝/2 agents
at any time 𝑡 ∈ 𝐸𝑡 is needed to induce a learning step for defecting agents that increases
the Q-value of deferring at state 𝑠⃑𝑡. With more and more successful votes, defectors will
57
eventually discover that cooperating is more beneficial, even without necessarily having
taken the 𝑑𝑒𝑓𝑒𝑟 action.
Policies
Figure 10: Breakdown of actions played by Q-Learning player 𝑖 = 0 every 𝑏𝑖𝑛 = 2500 game
steps. The policy choice apparently fails to converge, with significant variance in the number of
𝑑𝑒𝑓𝑒𝑟 actions taken. Note that the 𝑠𝑡𝑎𝑦 action is only legal when the other four cardinal
directions are not.
Figure 11: Breakdown of actions played by HQ-Learning player 𝑖 = 0 every 𝑏𝑖𝑛 = 2500 game
steps. This player gradually learns the value of rejecting short-term incentives to navigate and
invade in favor of deferring to the Sovereign.
58
Lastly, to verify that the policy agents are learning is in fact to choose the
Sovereign and political society rather than pursuing their own independent actions, we
drop below the collective metrics and analyze the actions of a single agent playing the
Civilization Game, arbitrarily chosen from the 4 players. Q-Learning vacillates between
the high and low payoff equilibria as we anneal 𝜀 over time while HQ-Learning learns to
achieve the long-term cooperation policy by reliably deferring to the Sovereign.
Matrix Game Analysis
We can also perform an analysis to determine what kind of matrix game the Civilization
Game SSD would translate into by playing agents with cooperative policies against agents who
have learned to defect. As is the case for matrix games, Markov games like the Civilization
Game become social dilemmas when the payoff outcomes 𝑅, 𝑃, 𝑆 and 𝑇 defined below satisfy the
social dilemma inequalities. These outcomes are derived from:
𝑅(𝑠) ∶= 𝑉1
𝜋𝐶,𝜋𝐶
(𝑠) = 𝑉2
𝜋𝐶,𝜋𝐶
(𝑠)
𝑃(𝑠) ∶= 𝑉1
𝜋𝐷,𝜋𝐷
(𝑠) = 𝑉2
𝜋𝐷,𝜋𝐷
(𝑠)
𝑆(𝑠) ∶= 𝑉1
𝜋𝐶,𝜋𝐷
(𝑠) = 𝑉2
𝜋𝐷,𝜋𝐶
(𝑠)
𝑇(𝑠) ∶= 𝑉1
𝜋𝐷,𝜋𝐶
(𝑠) = 𝑉2
𝜋𝐶,𝜋𝐷
(𝑠)
Where 𝜋𝐶 and 𝜋𝐷are cooperative and defecting policies respectively. The long-term payoff 𝑉 for
player 𝑖 when the joint policy 𝜋⃑ = (𝜋1, 𝜋2) is followed from an initial state 𝑠0 is:
𝜋⃑⃑ (𝑠0) = Ε𝑎⃑ 𝑡∽𝜋⃑⃑ (𝑂(𝑠𝑡)), 𝑠𝑡+1∽𝑇(𝑠𝑡, 𝑎⃑ 𝑡) [∑ 𝛾𝑡 𝑟𝑖(𝑠𝑡, 𝑎 𝑡)
𝑉𝑖
]
𝑡=0
∞
with the temporal discount factor 𝛾 ∈ [0,1]. The social dilemma inequalities as previously
defined are 𝑅 > 𝑃 (mutual cooperation is preferred to mutual defection), 𝑅 > 𝑆 (mutual
59
cooperation is preferred to being exploited by a defector), 2𝑅 > 𝑇 + 𝑆 (mutual cooperation is
preferred to an equal probability of unilateral cooperation and defection), and either greed 𝑇 >
𝑅 or fear 𝑃 > 𝑆 (mutual defection is preferred over being exploited).
As discussed in the SSD section, we can define what a cooperation versus defection
policy is by thresholding a continuous social behavior metric 𝛼 ∶ Π → ℝ with values 𝛼𝑐 and 𝛼𝑑
such that 𝛼(𝜋) < 𝛼𝑐 ↔ 𝜋 ∈ Π𝐶 and 𝛼(𝜋) > 𝛼𝑑 ↔ 𝜋 ∈ Π𝐷. An obvious candidate for this
metric would be invasions per hundred moves; if we set the threshold to 𝛼𝑐 = 5 and 𝛼𝑑 = 15,
then plugging in the Q-Learning and HQ-Learning policies learned after 𝑇 = 250𝑘 steps, 𝜋𝑄 and
𝜋𝐻𝑄, we have 𝜋𝑄 ∈ Π𝐷 and 𝜋𝐻𝑄 ∈ Π𝐶. Having shown that the standard Q-Learning policy is a
defecting policy and the HQ-Learning policy is cooperative, we can now play pairs of these
policies against one another and calculate the long-term payoff 𝑉 and thus the matrix payoff
values. For example, we can find the mutual cooperation payoff 𝑅 = 𝑉0
𝜋𝐶,𝜋𝐶
(𝑠0) by playing two
HQ-Learning agents against one another. With 𝑇 = 100𝑘 game steps and the same game
dynamics and payoffs as before, we get the payoff matrix:
Figure 12: The Civilization Game SSD matrix game with long-term payoff 𝑉𝑖
𝜋⃑⃑ (𝑠0) divided by
𝑇 = 100𝑘. These values characterize the embedded MGSD within the Game.
In this run, by calculating 𝐹𝑒𝑎𝑟 = 𝑃 − 𝑆 = 0.029 and 𝐺𝑟𝑒𝑒𝑑 = 𝑇 − 𝑅 = −0.013, we see that
the Civilization Game, and, to the degree that it conceptually resembles it, the Hobbesian State of
Nature, is a Stag Hunt MGSD. However, the matrix payoffs change each trial run because of the
existence of stochasticity in the discovery parameter, 𝜀. The parameter is retained in order to
prevent play from being "stuck" in a single state-action cycle, somewhat in the spirit of the
60
trembling hand perfect equilibrium123 that also accounts for some randomness. Consequently, 15
trials are run to calculate different payoff matrices.
Figure 13: Civilization matrix game analysis with 𝑡𝑟𝑖𝑎𝑙𝑠 = 15 runs where each run consists of
three match-ups: HQL vs HQL, HQL vs QL, and QL vs QL with 𝑇 = 100𝑘 steps. The resulting
𝜋⃑⃑ (𝑠) long-term payoff values are divided by 𝑇 for scale, then used to calculate 𝐹𝑒𝑎𝑟 and 𝐺𝑟𝑒𝑒𝑑.
𝑉𝑖
The results for the greed and fear calculations differ by around 5% and two-thirds of runs end up
as Stag Hunts. The fact that almost all runs have at least some fear factor while only one-third
have a positive greed factor can be seen in the game reward structure: players are only rewarded
10 points for an invasion while being invaded incurs a -25-point penalty that is over twice as
large. In addition, the existence of alternative channels for reward (i.e. farming and deferral
bonuses) means that an invasion penalty only consists of about a third of the maximum possible
reward on a given turn.
123 Bielefeld, R. S. A Reexamination of the Perfectness Concept for Equilibrium Points in Extensive Games. 1975.
25 -- 55.
61
IV.
Conclusion
Discussion
The central goal of this thesis has been to offer a new, interdisciplinary solution to the
problems involved in producing cooperation in multi-agent reinforcement learning by looking to
the way in which humans learn to cooperate in political society in the presence of incentives to
defect. After implementing a new SSD and learning mechanisms inspired by that comparison,
we see a surprisingly large improvement in game performance and convergence to a joint
cooperative policy when compared to standard learning algorithms. Individually, there has been
theoretical work in the field of MAS proposing frameworks similar to opponent learning
awareness124, learning sharing125, reward shaping126, and voting127, but approaching the problem
from a Hobbesian perspective merges and motivates them in a way that produces real
experimental results. Specifically, the addition of an opponent learning awareness term ∆𝑡
𝑖 to
each player's Q-update helps share information about how the behavioral intentions of other
players changes and enables stable policy convergence. Secondly, the institution of a binding
majoritarian voting procedure reduces the threshold of cooperative players needed to achieve a
joint cooperative policy and thus increases the speed of convergence. Finally, when agents
successfully vote to defer their actions, the Sovereign is empowered to disburse reward and
punishment in a way that shapes the learning of defecting players.
The interdisciplinary application of solutions from Hobbesian political philosophy to the
problem of cooperation in MAS is also unprecedented. While some philosophers interested in a
computational theory of mind have seen connections to Hobbes's comment that "when a man
124 Supra, Foerster.
125 Hwang, K. S. Model Learning and Knowledge Sharing for a Multiagent System with Dyna-Q Learning. 2015.
126 Grzes, Marek. Reward Shaping in Episodic Reinforcement Learning. 2017.
127 Pitt, Jeremy. Voting in Multi-Agent Systems. 2005.
62
reasoneth, he does nothing else but conceive a sum total, from addition of parcels; or conceive a
remainder, from subtraction of one sum from another,"128 no attempt has been made to
instantiate his insights in an experimental setting. In order to motivate that application, I've also
offered a new argument that applies Hobbes's natural and moral philosophy, specifically his
materialist, determinist, morally subjectivist, rationally self-interested, individualist, and
egalitarian understanding of human nature, to artificially intelligent agents. This falls in line
with, although certainly doesn't justify, Haugeland's dramatic assertion that Hobbes was
"prophetically launching Artificial Intelligence"129 when he published Leviathan. The argument
raises questions about the degree to which AI agents and humans share the same nature and
where differences might lie. Importantly, Hobbes also emphasizes the significance of the
passions like vainglory, contempt, courage, and benevolence that don't fit neatly into existing
computational frameworks.130 This is certainly an exception to the otherwise striking
comparison, but I will note that, for Hobbes, the passions are merely components of a
psychology that is still materialist and completely determined by sense perception.
Consequently, there is theoretically no reason why they couldn't be incorporated into a learning
agent's computational decision process.
The Civilization Game SSD introduced in this thesis also offers a new blend of incentives
and punishment meant to reflect the social dynamics that exist in the State of Nature and a
unique opportunity to test Hobbes's philosophical framework. This experimental analysis
contributes to contemporary Hobbes scholarship that understands it as a Stag Hunt, otherwise
known as the Assurance Game. Moehler famously claimed that: "the situation of prudent
128 Supra, Leviathan at V.
129 Haugeland, J. Artificial intelligence: the very idea. 1985. 23.
130 Supra, Leviathan at VI.
63
individuals in Hobbes' State of Nature is best modeled by an assurance game, [because] the
highest payoff for her is not achieved by unilateral defection, as suggested by the one-shot
Prisoner's Dilemma game, but by mutual cooperation."131 Skyrms also argued that if the
shortsighted Foole were to consider the 'shadow of the future' in a repeated game, the alleged
Prisoner's Dilemma game would be transformed into an assurance game.132 Gauthier, also,
believes that the adequate long-term representation of the game that rational individuals play in
Hobbes' State of Nature is an assurance game, although he does not call the game by its name.133
Fascinatingly, however, we can see from this thesis that it isn't always or necessarily an
Assurance Game, but rather, depending on the way one adjusts the ratio between how tempting
the invasion bonus is and how painful the invasion penalty is, the game might transform into a
Prisoner's Dilemma instead.
Importantly, the matrix game analysis presented here is the first attempt, to my
knowledge, to go beyond a close exegesis of Hobbes to an SSD analysis of the game-theoretic
dynamics of the State of Nature. While scholars have studied Hobbes quantitatively via an
iterated matrix game analysis134, those investigations do not account for the fact that 1) the
complexity of implementing cooperative versus defecting policies may be unequal, 2)
successfully implementing a policy often requires solving coordination sub-problems, and 3)
cooperativeness is a gradated quantity, not an atomic choice of action. Utilizing the Civilization
Game SSD enables us to understand how agents in a realistic setting learn to choose and
implement policies, and with the help of Hobbesian mechanisms, achieve a cooperative
equilibrium.
131 Supra, Moehler at 315.
132 Supra, Skyrms at 1-13.
133 Gauthier, D. The Logic of Leviathan. 85 -- 86.
134 Palumbo, A. Playing Hobbes: The Theory of Games and Hobbesian Political Theory. 1996.
64
Future Work
This thesis illustrates the promising practical value of combining knowledge from the
fields of computer science and political philosophy in understanding the evolution of cooperation
and lays out some of the experimental groundwork to make that possible. Even so, a great deal
remains for future work. Within just the bounds of the research question considered here, there
would be great value in taking the time to implement a more state-of-the-art RL algorithm like
deep Q-networks that utilize the same theoretical framework as Q-Learning but offers scalability
to larger state spaces. This would open the door to testing more complicated games with more
players and a bigger grid-world to explore. The dynamics of an 8 player Civilization Game, for
example, may look very different from the 4 player and 2 player scenarios tested here.
There are also additional ideas in Hobbes that could be translated into promising
mechanisms, like the detailed process for deriving and enforcing binding Covenants between
different players and between players and the Sovereign. Implementing the Sovereign as a
separate agent itself instead of a component of the game environment might also allow a
researcher to analyze its interactions with other players and ask, for example, how good the
Sovereign must be at punishing defection before the MAS devolves back to the low payoff
equilibrium. Finally, varying the reward structures within the Civilization Game including the
farming bonus, invasion bonus, and invasion penalty as well as the Sovereign reward-shaping
incentives may likely result in different embedded MGSDs and consequently, different learning
behavior.
Although here the three Hobbesian political mechanisms are considered together as a
unified strategy, further MAS research may analyze the individual merits of each considered
separately. Whether individually or taken as a whole, it would also be of great interest to observe
65
the performance of modified Hobbesian Q-Learning in other popular SSDs as well like the Wolf
Pack and Gathering games from Leibo135, or especially the Stag Hunt-inspired Markov games
like Harvest and Escalation from Peysakhovich and Lerer136. We would expect to see similar
improvements in performance, especially for those games that preserve high level properties of
the Stag Hunt. The unique asynchronous turn-based dynamic of the Civilization Game might
also offer an opportunity to contrast experimentally considerations of player memory length and
complexity cost required to compute equilibria versus in the standard synchronous model. 137
Generally, my hope is to encourage research in the field of AI or MAS to look beyond traditional
methodologies and search out interdisciplinary solutions, even in fields as unconventional as
political philosophy, to interdisciplinary problems.
Likewise, scholars of Hobbes, in line with his thoroughly empirical method, have at their
disposal a powerful new experimental tool that models the State of Nature in many important
respects, although admittedly simplified in others, and deploys Hobbesian, intelligent agents to
play that game. By varying the incentive structure in the Civilization Game, we can see how
varying certain underlying assumptions that Hobbes makes might affect his political conclusions.
I have already shown that increasing the invasion bonus / invasion penalty ratio transforms the
embedded MGSD in the Civilization Game from an Assurance game to a Prisoner's Dilemma.
Further research might also examine the consequences of reducing the effectiveness of the
Sovereign by either reducing the size of the reward or punishment or by making it apply only
once every 𝑛 times and observing whether agents still learn to cooperate. The results of such an
experiment might inform our perspective on Hobbes's infamous position that life under a
135 Supra, Leibo et al.
136 Supra, Peysakhovich et al.
137 Bhaskar, V. Asynchronous Choice and Markov Equilibria. 1998. 1.
66
Sovereign, no matter how harsh, is always better than life in the State of Nature. This thesis also
offers the first argument for the applicability of Hobbes's natural and moral philosophy to
modern computational artificial intelligence. The obvious exception that has been highlighted is
the role of the passions in directing human action that finds no direct corollary in mechanical RL.
That being said, in Hobbes's mechanistic materialist psychology, there should almost certainly
exist a way to mechanize the passions in his account in a way that makes them feasible to model
or implement in computers.
In conclusion, it's clear that the evolution of cooperation is a fascinating and difficult
problem of great practical importance for both humans and AI. Hobbes's own answer profoundly
influenced the way in which humans organize our modern, liberal societies, and the solutions we
develop to enable cooperation amongst artificially intelligent agents will similarly affect what
technology in the 21st century will look like and how it will interface with us. There is good
reason to be optimistic about a near future in which humans and computers work hand-in-hand to
solve the biggest challenges we face, and my hope is that the work presented here might
represent a small contribution to that larger project.
67
Works Cited
Ale, S. B., Brown, J. S., & Sullivan, A. T. (2013). Evolution of cooperation: combining kin
selection and reciprocal altruism into matrix games with social dilemmas. PloS one, 8(5),
e63761.
Aristotle. Aristotle's Politics. Oxford :Clarendon Press.
Arulkumaran, K., Deisenroth, M. P., Brundage, M., & Bharath, A. A. (2017). A brief survey of
deep reinforcement learning. arXiv preprint arXiv:1708.05866.
Aubrey, J., & Dick, O. L. (1949). Brief lives (p. 157). London: Secker and Warburg.
Augustine, S. (2008). The city of God. Hendrickson Publishers.
Beckers, R., Holland, O. E. and Deneubourg J_L (1994). "From Local Actions to Global Tasks:
Stigmergy in Collective Robotics," in R. Brooks and P. Maes eds. Artificial Life IV,
Cambridge, Mass.: MIT Press.
Bhaskar, V., & Vega-Redondo, F. (2002). Asynchronous choice and Markov equilibria. Journal
of Economic Theory, 103(2), 334-350.
Bielefeld, R. S. (1988). Reexamination of the perfectness concept for equilibrium points in
extensive games. In Models of Strategic Rationality (pp. 1-31). Springer, Dordrecht.
Broom, M., Cannings, C., & Vickers, G. T. (1997). Multi-player matrix games. Bulletin of
mathematical biology, 59(5), 931-952.
Calo, R. (2017). Artificial Intelligence policy: a primer and roadmap. UCDL Rev., 51, 399.
Chamberlin, E. H. (1949). Theory of monopolistic competition: A re-orientation of the theory of
value. Oxford University Press, London.
Chantemargue, F., Lerena, P., & Courant, M. (1999, July). Autonomy-based multi-agent
systems: statistical issues. In Proceedings of Third World Multiconference of Systemics,
68
Cybernetics and Informatics (SCI'99) Fifth International Conference of Information
Systems Analysis and Synthesis (ISAS'99).
Curtis Eaton, B. (2004). The elementary economics of social dilemmas. Canadian Journal of
Economics/Revue canadienne d'économique, 37(4), 805-829.
De Tocqueville, A. (2003). Democracy in America (Vol. 10). Regnery Publishing.
Doran, J. E., Franklin, S. R. J. N., Jennings, N. R., & Norman, T. J. (1997). On cooperation in
multi-agent systems. The Knowledge Engineering Review, 12(3), 309-314.
Ferber, J., & Weiss, G. (1999). Multi-agent systems: an introduction to distributed artificial
intelligence (Vol. 1). Reading: Addison-Wesley.
Foerster, J., Chen, R. Y., Al-Shedivat, M., Whiteson, S., Abbeel, P., & Mordatch, I. (2018, July).
Learning with opponent-learning awareness. In Proceedings of the 17th International
Conference on Autonomous Agents and MultiAgent Systems (pp. 122-130). International
Foundation for Autonomous Agents and Multiagent Systems.
Forrest, S. (1990). Emergent computation: self-organizing, collective, and cooperative
phenomena in natural and artificial computing networks: introduction to the proceedings
of the ninth annual CNLS conference. Physica D: Nonlinear Phenomena, 42(1-3), 1-11.
Franklin, S., & Graesser, A. (1996, August). Is it an Agent, or just a Program?: A Taxonomy for
Autonomous Agents. In International Workshop on Agent Theories, Architectures, and
Languages (pp. 21-35). Springer, Berlin, Heidelberg.
Fudenberg, D., Drew, F., Levine, D. K., & Levine, D. K. (1998). The theory of learning in games
(Vol. 2). MIT press.
Galbraith, J. (1956). American capitalism : The concept of countervailing power (Sentry edition ;
18). Boston: Houghton Mifflin Company.
69
Gauthier, D. (1977). The social contract as ideology. Philosophy & Public Affairs, 130-164.
Gauthier, D. P. (1969). The logic of Leviathan: the moral and political theory of Thomas
Hobbes. Oxford University Press.
Grosz, B. J. (2013, May). A multi-agent systems Turing challenge. In Proceedings of the 2013
international conference on Autonomous agents and multi-agent systems(pp. 3-4).
International Foundation for Autonomous Agents and Multiagent Systems.
Groves, T. (1973). Incentives in teams. Econometrica, 41(4), 617-631.
Grześ, M. (2017, May). Reward shaping in episodic reinforcement learning. In Proceedings of
the 16th Conference on Autonomous Agents and MultiAgent Systems (pp. 565-573).
International Foundation for Autonomous Agents and Multiagent Systems.
Hamilton, A., Madison, J., & Jay, J. (2008). The federalist papers. Oxford University Press.
Hamilton, J. (2009). Hobbes the royalist, Hobbes the republican. History of Political Thought,
30(3), 411-454.
Hampton, J. (1988). Hobbes and the social contract tradition. Cambridge University Press.
Haugeland, J. (1989). Artificial intelligence: The very idea. MIT press.
Hobbes, T. (1991). Man and Citizen: De homine and De cive. Hackett Publishing.
Hobbes, T. (1996). Leviathan. Richard Tuck, ed. Cambridge, Cambridge UP, 87.
Hobbes, T. (2013). Elements of law, natural and political. Routledge.
Hont, I., Kapossy, B., & Sonenscher, M. (2015). Politics in commercial society : Jean-Jacques
Rousseau and Adam Smith. Cambridge, Massachusetts: Harvard University Press.
Hume, D. (2003). A treatise of human nature. Courier Corporation.
70
Hwang, K. S., Jiang, W. C., & Chen, Y. J. (2015). Model learning and knowledge sharing for a
multiagent system with Dyna-Q learning. IEEE transactions on cybernetics, 45(5), 978-
990.
Kandori, M., Mailath, G. J., & Rob, R. (1993). Learning, mutation, and long run equilibria in
games. Econometrica: Journal of the Econometric Society, 29-56.
Kleiman-Weiner, M., Ho, M. K., Austerweil, J. L., Littman, M. L., & Tenenbaum, J. B. (2016,
January). Coordinate to cooperate or compete: abstract goals and joint intentions in social
interaction. In CogSci.
Kleiman-Weiner, M., Saxe, R., & Tenenbaum, J. B. (2017). Learning a commonsense moral
theory. cognition, 167, 107-123.
Kollock, P. (1998). Social dilemmas: The anatomy of cooperation. Annual review of sociology,
24(1), 183-214.
Leibo, J. Z., Zambaldi, V., Lanctot, M., Marecki, J., & Graepel, T. (2017, May). Multi-agent
reinforcement learning in sequential social dilemmas. In Proceedings of the 16th
Conference on Autonomous Agents and MultiAgent Systems (pp. 464-473). International
Foundation for Autonomous Agents and Multiagent Systems.
Loetscher, F. W. (1935). St. Augustine's conception of the state. Church History, 4(1), 16-42.
Macpherson, C. B., & Cunningham, F. (1962). The political theory of possessive individualism:
Hobbes to Locke (p. 331145263). Oxford: Clarendon Press.
Macy, M. W., & Flache, A. (2002). Learning dynamics in social dilemmas. Proceedings of the
National Academy of Sciences, 99(suppl 3), 7229-7236.
71
Mao, A., Parkes, D. C., Procaccia, A. D., & Zhang, H. (2011, August). Human computation and
multiagent systems: an algorithmic perspective. In Proceedings of the twenty-fifth AAAI
conference on artificial intelligence.
Melo, F. S. (2001). Convergence of Q-learning: A simple proof. Institute Of Systems and
Robotics, Tech. Rep, 1-4.
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., & Riedmiller,
M. (2013). Playing atari with deep reinforcement learning. arXiv preprint
arXiv:1312.5602.
Moehler, M. (2009). Why Hobbes' state of nature is best modeled by an assurance game. Utilitas,
21(3), 297-326.
Nwana, H. S., & Ndumu, D. T. (1999). A perspective on software agents research. The
Knowledge Engineering Review, 14(2), 125-142.
Olson, M. (2009). The logic of collective action (Vol. 124). Harvard University Press.
Ossowski, S., & García-Serrano, A. (1997, December). Social co-ordination among autonomous
problem-solving agents. In Australian Workshop on Distributed Artificial
Intelligence (pp. 134-148). Springer, Berlin, Heidelberg.
Ostrom, E. (1990). Governing the commons : The evolution of institutions for collective action
(Political economy of institutions and decisions). Cambridge ; New York: Cambridge
University Press.
Pakaluk, Michael. (2005). Aristotle's Nicomachean Ethics (Cambridge introductions to key
philosophical texts). Cambridge University Press - M.U.A.
Palumbo, A. (1996). Playing Hobbes. The Theory of Games and Hobbesian Political Theory.
72
Pitt, J., Kamara, L., Sergot, M., & Artikis, A. (2006). Voting in multi-agent systems. The
Computer Journal, 49(2), 156-170.
Przeworski, A. (1985). Capitalism and social democracy (Studies in Marxism and social theory).
Cambridge ; New York : Paris: Cambridge University Press ; Editions de la Maison des
sciences de l'homme.
Rapoport, A. (1974). Prisoner's Dilemma -- Recollections and observations. In Game Theory as a
Theory of a Conflict Resolution (pp. 17-34). Springer, Dordrecht.
Rousseau, J. J. (1984). A discourse on inequality. Penguin.
Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward.
Science, 275(5306), 1593-1599.
Schumacher, M. (2001). Objective coordination in multi-agent system engineering: design and
implementation. Springer-Verlag.
Shoham, Y., & Tennenholtz, M. (1995). On social laws for artificial agent societies: off-line
design. Artificial intelligence, 73(1-2), 231-252.
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., ... & Chen, Y.
(2017). Mastering the game of go without human knowledge. Nature, 550(7676), 354.
Skyrms, B. (2004). The stag hunt and the evolution of social structure. Cambridge University
Press.
Strauss, L. (2017). What is political philosophy?. In Plato and Modern Law (pp. 71-96).
Routledge.
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction. MIT press.
Thorndike, E. L. (1911). Animal intelligence: Experimental studies. Macmillan.
73
Turing, A. M. (1937). On computable numbers, with an application to the
Entscheidungsproblem. Proceedings of the London mathematical society, 2(1), 230-265.
Weizenbaum, J. (1976). Computer power and human reason: From judgment to calculation.
74
|
1109.0601 | 1 | 1109 | 2011-09-03T07:43:09 | Application of distributed constraint satisfaction problem to the agent-based planning in manufacturing systems | [
"cs.MA"
] | Nowadays, a globalization of national markets requires developing flexible and demand-driven production systems. Agent-based technology, being distributed, flexible and autonomous is expected to provide a short-time reaction to disturbances and sudden changes of environment and allows satisfying the mentioned requirements. The distributed constraint satisfaction approach underlying the suggested method is described by a modified Petri network providing both the conceptual notions and main details of implementation. | cs.MA | cs | Application of distributed constraint satisfaction problem
to the agent-based planning in manufacturing systems
S. Kornienko, O. Kornienko, P. Levi
Institute of Parallel and Distributed High-Performance Systems,
University of Stuttgart, Universitatsstr. 38, 70569 Stuttgart, Germany. ∗
1
1
0
2
p
e
S
3
]
A
M
.
s
c
[
1
v
1
0
6
0
.
9
0
1
1
:
v
i
X
r
a
Abstract
Nowadays, a globalization of national markets re-
quires developing flexible and demand-driven pro-
duction systems. Agent-based technology, being
distributed, flexible and autonomous is expected to
provide a short-time reaction to disturbances and
sudden changes of environment and allows satis-
fying the mentioned requirements. The distributed
constraint satisfaction approach underlying the sug-
gested method is described by a modified Petri net-
work providing both the conceptual notions and
main details of implementation.
1 Introduction
An internationalization of market outlets, an individualiza-
tion of consumer needs, a rapid implementation of technical
innovation become a key factor of success in modern eco-
nomic market. In order to survive in international competi-
tion, enterprises are forced to react dynamically to new re-
quirements, to make permanent modifications and adaptation
of own structures.
In particular this concerns the planning
processes.
A planning in manufacturing systems is traditionally orga-
nized top-down. The strategical level of planning transmits
the results to the tactical level, which in turn triggers the op-
erational level of planning process. The final result of the
planning process is a detailed schedule of manufacturing to
be implemented on the shop floor. Every encountered dis-
turbance, unforeseen in the schedule, triggers a new planning
cycle. Efforts, costs and time required for replanning can be
essentially reduced by making the planning process adaptive
to disturbances. In this way a modern flexible manufactur-
ing requires a new approach that introduces elements of self-
organization into operational control. Taking into account a
spatial distribution of manufacturing elements and require-
ment of flexibility to the whole system, the concept of au-
tonomous agents has found some applications in this field [1].
This approach seems to be very promising to guarantee the re-
quired robustness, fault tolerance, adaptability and agility in
the field of transformable manufacturing systems.
Application of agents to manufacturing requires also a
development of new and adaptation of known approaches
∗Post-Proceeding Version 2004. Appeared in proceedings of
the International Scientific Congress "Intelligent Systems (IEEE
AIS'03)" and "Intelligent CAD's (CAD-2003)", p.124-140, Di-
vnomorsk, Russia, 2003
towards typical problems of multi-agents technology, like
distributed problem solving, planning or collective decision
making [2]. This paper deals with the distributed constraint-
based short-term planning (assignment) process supported by
an multi-agent system (MAS).
2 Assignment problem
The assignment problem is often encountered in manufactur-
ing, it is a part of Operations Research / Management Sci-
ence (OR/MS). It can be classified into scheduling, resources
allocation and planning of operations order (e.g. [3]). This
is a classical N P -hard problem, there are known solutions
by self-organization [4; 5], combinatorial optimization [6],
evolutionary approaches [7], constraint satisfaction and opti-
mization [8], discrete dynamic programming [9]. However
these methods are developed as central planning approaches,
the distributed or multi-agents planning for the assignment
problem is in fact not researched (overview e.g. in [10]).
Generally, this problem consists in assigning a lot of low-
level jobs (like "to produce one piece with defined specifica-
tion") to available machines so that all restrictions will be sat-
isfied. The solution consists of four steps. The first step is to
prove whether the machine is technological able to manufac-
ture, whereas the second step is to prove whether the machine
is organizational available, e.g. it is without prearrangements.
The first two steps formalize the constraint problem, where
the following criteria should be taken into account [11]:
• Technical criteria determine necessary features of a
piece and on this basis it can be decided whether a ma-
chine is able to manufacture this kind of feature e.g. the
drilling feature.
• Technological criteria determine whether the machine
can operate with a necessary quality or determine a tech-
nologically necessary order, e.g.
the statutory toler-
ances.
• Geometrical criteria result from a geometrical descrip-
tion of the workpiece, e.g. whether a necessary chucking
is possible.
• Organizational criteria are a set of specifications of pro-
duction orders and machines. Firstly, it defines whether
a machine has an available time slot and, secondly,
whether this time slot is suitable to manufacture the
given production order.
• Optimization criteria like a cost or delivery time.
All these criteria determine the agent-based process plan-
ning in the form of corresponding constraints, which reduce
the decisions space in the assignment problem. The third
and fourth steps of solution are the distributed constraint-
satisfaction problem (CSP) and composition/optimization
(COP) correspondingly. In the considered example it needs
to manufacture a multitude of parts and variants of produc-
tion orders. Totalling there are different types of workpieces
(5-20 pieces of each type, see Fig. 1) that have to be manu-
Figure 1: Example of a workpiece to be manufactured.
factured on available machines. Table 1 shows an exemplary
sequence of working steps, where all mentioned technologi-
cal constraints are already considered.
WS L/M 1 L/M 2 L/M 3 Order
1
2
3
4
5
6
7
8
9
10
11
1
2
1
3
0
2
1
3
2
0
1
1
2
0
3
1
2
0
3
2
1
1
1
2
2
2
2
2
3
4
4
4
4
0
0
1
0
1
2
1
0
0
1
1
Table 1: Technological table for a workpiece, WS - working step,
L/M - length/machine. Zero in a length (of a working step) at cor-
responding machine means that this machine cannot perform the re-
quested operation. Order of working steps means, that e.g. the steps
2,3,4,5,6 should be produced after the step 1 and before the step 7.
It is natural to assume these steps cannot be performed at the same
time on different machines.
Processing of each workpiece consists of several working
steps (defined by a technological process), all these working
steps cannot be processed on one machine. Each from the
working steps has different length and also cost. Moreover
each type of the pieces has own technology, i.e. a processing
consists of different working steps. For simplification it is as-
sumed that available machines are of the same type, therefore
the cost and length of the same working step do not differ on
these machines (in general case they are different). The aim
is to generate a plan of how to manufacture these workpieces
Let us denote a working step as W Si
with minimal cost, minimal time (or other optimization crite-
ria), taking into account restrictions summarized in Table 1.
j, where i is a type of
workpieces and j a number of the working step, an available
machine is denoted as Mk, where k is a number of machine.
We need also to introduce a piece P m
n , where m is a priority
of production and n is a number of this piece. In this way
j)) denote a start and a final po-
st(P m
sitions of the corresponding working step that belongs to the
corresponding piece (st(W Si
j) for all pieces). We
start with the definition of these values
n∈[1−20] (W Si∈{A,B,C,D,E}
P m∈[1−20]
j), f n(W Si
) = o ∈ operation,
j)), f n(P m
j∈[1,...,11]
n (W Si
n (W Si
Mk∈{1,2,3,4} = {o ∈ operation},
st(P m
f n(P m
j)) = {t ≥ 0, t ∈ R},
j)) = {st(P m
n (W Si
n (W Si
n (W Si
j)) +
n (W Si
n (W Si
+ length(P m
j), o2 ∈ Mk, o1 = o2}.
j))}.
The first constraint determines a correspondence between op-
erations of working step and of the k-machine
C1 = {(o1, o2)o1 ∈ P m
The technological restrictions given by the Table 1 can be
rewritten in the following form
C2 = {(f n(W SA
j × W SA
j },
j },
j × W SA
C3 = {(f n(W SA
C4 = {(f n(W SA
j × W SA
j },
[2−6] (for all pieces) cannot be performed at the
where W SA
same time
C5 = {(j ∈ [st(W SA
[2−6])) ⊂ W SA
[7])) ⊂ W SA
[8−11])) ⊂ W SA
[1]) < st(W SA
[2−6]) < st(W SA
[7]) < st(W SA
[w]), ..., f n(W SA
(cid:54)= j ∈ [st(W SA
⊂ W SA
[8−11]
j × W SA
and also W SA
C6 = {(j ∈ [st(W SA
(cid:54)= j ∈ [st(W SA
⊂ W SA
j × W SA
[w])] (cid:54)=
[w(cid:48)])]) ⊂
[w(cid:48)]), ..., f n(W SA
j ; w, w(cid:48) = 2, 3, 4, 5, 6; w (cid:54)= w(cid:48)}
[w])] (cid:54)=
[w(cid:48)])])
[w]), ..., f n(W SA
[w(cid:48)]), ..., f n(W SA
j ; w, w(cid:48) = 8, 9, 10, 11; w (cid:54)= w(cid:48)}.
C7 = {(m ∈ P m
Priority of production can be expressed by
n(cid:48) st(P m
n × P m
n > m ∈ P m
j ))) ⊂ P m
> st(P m
n(cid:48) (W SA
n (W SA
j )) >
n , n (cid:54)= n(cid:48)}.
As soon as the variable, the domains of values and constraints
are defined, a propagation approach can be started. The aim
is to restrict the values of variables (or to find such values of
variables) that will satisfy all constraints. This propagation
can be represented in the way shown in Fig. 2.
All working steps that belong to the same workpiece build
a sequence. Every node in this sequence gets a "finish"-
position of a working step from the previous node. Using this
value, a current node looks for "start"-positions of the next
working step that satisfy local constraints, calculates "finish"-
positions and propagates them to the next node. If no position
satisfying local constraint can be found, the node requests an-
other "finish"-position from the previous node. In this way
the network can determine locally consistent positions of all
working steps. After that the obtained values should be tested
for a global consistence.
key problem of multi-agent technology. On the one hand
if the system is hard restricted in the behaviour, the advan-
tage of MAS is lost. On the other hand, if the system has
too many degrees of freedom it can communicate an in-
finitely long time. In other words only several combinations
of agents states have a sense and the point is how to achieve
and to manage these states. This is a hard problem arising in
many branches of science and engineering and correspond-
ingly there are several ways to solve it. The suggested here
solution is based on a hierarchic software architecture that
supports agent's autonomy.
Before starting to describe an approach, it needs to men-
tion one methodological point concerning decentralization of
multi-agent system, shown in Fig. 3. The MAS solves a prob-
Figure 3: Methodological approach towards agent-based applica-
tions.
lem by using some methodological basis. For example the
CSP and COP approaches basically underlie the solution of
constraint problems. The point is that a methodological ba-
sis, in almost all cases, is formulated in a centralized way.
It looks like a "battle plan", where all agents and their inter-
actions are shown. Therefore this global description is often
denoted an interaction pattern.
However the agents do not possess such a global point of
view and the interaction pattern has to be distributed among
agents. This decentralization concerns global information,
messages transfer, synchronization, decision making and so
on. The decentralized description of the chosen method
should determine an individual activity of an agent as well
as its interaction with other agents. It is also important that
all agents behave in ordered way, i.e. to include cooperation
mechanisms (protocols) into this distributed description [13].
In order to enable a transition from the interaction pattern to
the cooperation protocol (see Fig. 3) a notion of a role is in-
troduced [14]. A role is associated with a specific activity
needed to be performed (according to a methodological ba-
sis). Agent can "play" one role or a sequence of roles. In this
way interactions are primarily determined between roles, an
agent (with corresponding abilities) handles according to the
role playing at the moment. An advantage of this approach is
that the centralized description (familiar for human thinking)
is preserved whereas the roles in the interaction pattern are "in
fact" already distributed, i.e. a mapping "agent-on-role" may
be performed in a formalized way by a program. Thus, an in-
Figure 2: Constraint network for the assignment problem.
3 Application of the autonomous agents
The CSP approach described in the previous section is nec-
essary and sufficient in solving the discussed kind of assign-
ment problem. Being implemented by one of programming
techniques, it will generate the required plan. However work-
ing in a presence of disturbances (like machine failure, tech-
nological change and so on) requires additional efforts to
adapt the planning approach to these changes. The princi-
ples of such an adaptation are not contained in the plan it-
self, an additional mechanism is needed. As mentioned in the
introduction, the multi-agent concept can be used as such a
mechanism that lends to manufacturing system more flexibil-
ity to adapt to disturbances. But what is the cause and cost
of this additional feature ? There is a long discussion of this
point based e.g. on the decentralization (e.g. [12]) or several
dynamics properties of MAS (e.g in [1]). We would like to
add the following argument into this discussion.
The multi-agent system can be considered from a view-
point of theory of finite-state automata. Transition of m-states
automaton (with or without memory, it does not change the
matter) from one state to another is determined by some rules
(by a program), therefore the automaton behaves in com-
pletely deterministic way. If a control cycle is closed (see e.g.
[1]) the automaton is autonomous, i.e. behaves independently
of environment (other automata). Now consider a few such
automata coupled into a system in the way that keeps their
autonomy. Forasmuch as each automaton behaves according
to own rules, there is no a central program that determines
a states transition of the whole system. In the "worst case"
coupling n automatons with m states, the coupled system can
demonstrate nm states.
Evidently this "worst case" has never to arise in the system,
but how to control a behaviour of the distributed system with-
out a central program (without a centralized mediator) ? The
point is that all automata are continuously communicating in
order to synchronize own states in regard to environment, to
the solving task, etc. (in this case the notion of an automaton
is replaced by the notion of an agent). The agents during com-
munication "consider" all possible states and then "choose"
such a state that is the most suitable to a currently solving
problem. This is a main difference to the "centralized pro-
gramming" approach. The central program can react only in
such a way that was preprogrammed. For example 10 agents
with 10 states can demonstrate 1010 different combinations.
However no programmer is able to predict all situations to
use all these states. Thus, the "centralized programming" ap-
proach restricts the multi-agent system although there are es-
sentially more abilities to react.
The sufficient number of degrees of freedom represents a
oped. Using this approach, the roles executed by agents are
described in a formal way by the Perti network. Details of the
RoPE system as well as description of cooperations protocols
in [11], [17]. As al-
via Perti networks can be found e.g.
ready mentioned, the primary algorithm consists of two parts,
parameterization and propagation, that represent a linear se-
quence of activities. The parameterization part, shown in
Fig. 4 has three phases p0, p1, p2 whose main result consists
Figure 4: Primary algorithm: Parameterization part.
in determining a structure, neighbourhood relations and pa-
rameterization of nodes of the constraint network. The roles
γ0, γ1 are "Initializers" of WS-order and WS-nodes corre-
spondingly. The role γ0 is activated by the first production
order, this role reads resource-objects and determines how
much nodes (WS-roles) are required. The transition t0 proves
whether the result of j.returnEN D() is true (action is suc-
cessful) and activates γ1 with parameter n as a number of
required nodes. The γ1 initializes each node according to
all restrictions (technology, propagation rules, number of ma-
chines and so on). If this activity is also successful (transition
t1), the third role γ2 is activated. It connects the created nodes
(return a pointer to previous node), composing in this way a
network. This interaction plan is finished (transition t3) if
there exists no next node needed to be connected.
The propagation part, shown in Fig. 5, consists of three
teraction pattern is a "mosaic image" that from afar looks like
a common picture (method), but at a short distance is a set of
separate fragments (roles). Moreover a concept of roles al-
lows decoupling the structure of cooperation processes from
agents organization, so that any modification of agents does
not affect the cooperation process and vice versa [14].
The interaction pattern determines a primary activity (pri-
mary algorithm) of multi-agent system [15]. The primary al-
gorithm includes also some parameters whose modifications
can be commonly associated with disturbances. Variation of
these parameters does not disturb an activity of agents.
In
this case these are the expected disturbances, a reaction of the
system on them is incorporated into the primary algorithm.
However due to specific disturbances every agent can reach
such a state that is not described by a primary algorithm and
where a performing of the next step is not possible. In this
case the agent is in emergency state and tries to resolve the
arisen situation all alone or with an assistance of neighbour
agents (secondary activity). If the abilities of an agent are not
sufficient or it requires additional resources it calls a rescue
agent. The rescue agent is an agent that possesses specific
(usually hardware) abilities. Anyway, the aim of agents in
emergency state is to change a part of the primary algorithm
so that to adapt it to disturbances. The disturbances causing
local emergency are expected (predicted) but not introduced
into the primary algorithm.
The primary algorithm as well as its parameterization
is optimal only for specific conditions (e.g.
combinato-
rial/heuristic methods for solutions of combinatorial prob-
lems, CSP/COP for constraints, etc.). If disturbances change
these conditions the primary algorithm became non optimal
and it has no sense to repair it. All agents have collectively
to recognize such a global change and to make a collective
decision towards replacement of the primary algorithm. This
change corresponds to a global emergency. The disturbances
causing the global emergency are not expected (predicted),
however they influence the conditions of primary algorithms
and in this way can be recognized. Finally, there are such
disturbances that cannot be absorbed by any changing of an
algorithm, they remain irresolvable. Forasmuch as all these
points cannot be considered in details in the framework of the
paper, we restrict ourselves only by treating the mentioned
primary algorithm in the language of cooperative processes
by the modified Petri nets. The secondary activity as well as
emergency states and rescue agents are not considered here.
3.1 The primary algorithms
In the case of an assignment planning, the primary algorithm
is determined by the CSP approach described in Sec. 2 (gen-
erally usage of constraint-based approaches for MAS is not
new, see e.g. [16]). Each working step in the approach is rep-
resented by a node in the constraint network shown in Fig. 2.
These nodes are separated from one another, moreover their
behaviour is determined by propagations. Therefore it is nat-
ural to give a separate role (R) to each node. However, before
starting a propagation, this network has to be created and pa-
rameterized by technology, machines, number of workpieces
and so on. These two steps (parameterization and propaga-
tion) will be described by interaction patterns using corre-
sponding roles.
For description of agent's activities the RoPE (Role ori-
ented Programming Environment) methodology is devel-
Figure 5: Primary algorithm: Propagation part.
blocks:
local (the phases p3, p4, p5), global (the phase p6)
propagations and an activity (the phase p7) in the case of
empty sets. The roles γ3, γ5 determine the propagation in
the first and last nodes whereas γ4 does the same for all other
nodes. The transition t7 proves whether the local propaga-
tion was successful for all nodes and activates then the global
propagation in γ6. We emphasize the local propagation re-
quires a sequential executing of roles whereas in global prop-
agation all roles can be in parallel executed. Finally, the tran-
sition t9 proves whether the values set (WS-positions) of each
node is empty. In the case of empty sets the role γ7 tries to in-
crease initial areas of values, first locally in neighbour nodes,
then globally by restart of the local propagation.
Thus, agents, executing the roles described in the param-
eterization and propagation parts, "know how to solve" the
CSP problem. Therefore all disturbances associated with a
change of resources and constraints can be absorbed by a
change of parameters in the interaction patterns. In this way
the MAS has enough degrees of freedom to be adaptive to
disturbances in a framework of the primary algorithm.
4 Agent-based optimization
The steps described in the previous sections allow generat-
ing the sequence of working steps that satisfies all local con-
straints. However such global characteristics of a plan as cost,
manufacturing time and so on are not considered. Therefore,
as pointed out by some authors, the next step consists in op-
timizing the obtained sequences. Generally speaking, con-
straint satisfaction and optimization cannot be separated into
two different steps, rather, it represents a sui generis combina-
tion. Before to start a discussion of agent-based optimization
it needed to mention two features of such an approach.
The first feature of agent-based optimization consists in a
local character of used data. Combinatorial or heuristic ap-
proaches assume the data, required to be optimized, are glob-
ally available. Optimization in this case looks like a "chess
play" where all pieces are visible and it needs to find some
combination of pieces positions. In the multi-agent variation
there is no this central viewpoint, each agent makes only a lo-
cal decision about to occupy a positions or not (see Fig. 6). In
this way the agent performs optimization of local decisions
instead of global positions of the working steps. Moreover,
from agents viewpoint, any of their decisions has not foresee-
able perspectives for a global optimum.
guarantee a global optimum. Therefore the agent has to com-
pute what will happen if the next processing step will begin
not immediately after the previous step. It can be achieved
by shifting a manufacturing of a workpiece on some steps
that increase a local cost of a plan (e.g. intermediate storage)
but reduce global costs (see Figs. 7, 8). This approach (fore-
casting) is similar to a decision tree in distributed form. As
known, an increasing of the depth of tree rapidly increases the
search space.
After discussing the features of agents-based optimization,
one can focus on the problem of assignment plan. There are
two important steps, that the optimization needs to be per-
formed on. Firstly, an order of the working steps in the group
2 and 4 (see Fig. 2). Forasmuch as there are only 2881 com-
bination between W S1-W S11, this optimization step can be
performed by exhaustive search. The second point of opti-
mization concerns the local decisions (concerning machine
and position) made by agents. However search space (taking
into account the forecasting effect) grows in this case expo-
nentially and e.g. even for 22 agent (2 production workpieces,
forecast for next 5 positions) comes to ∼ 1010. Therefore
exhaustive search methods like constraints optimization are
inefficient even on very fast computer. The search space can
be essentially reduced if to take into account the following
observation.
(a)
(c)
(b)
(d)
Figure 6: Example of the plan making for the minimal time and
minimal transportation cost. There are three pieces to be manufac-
tured (P1 grey, P2 white, P3 black) with three working steps (with
length 1, 2, 1). The W S2 cannot be processed on M2. (a) Deci-
sions W S2(P3) → M1 in position 5, W S3(P2) → M3 in position
4; (b) Decisions W S2(P3) → M3 in position 4, W S3(P2) → M2
in position 4; (c) Final plan of (a); (d) Final plan of (b).
The second feature of agent-based optimization is caused
by the local nature of optimization problem. Each agent dur-
ing the CSP phase tries to occupy a position immediately after
the previous working step. This strategy is motivated from the
manufacturing side in trying to avoid a waiting time at pro-
cessing elements (machines). Evidently, this strategy cannot
Figure 7: The "forecasting" effect in assignment planning.
(a)
Each working step begins immediately after the previous step. The
length of a whole plan is equal to 47; (b) The start of the working
step 8 (3th piece) is delayed on one step, that allows reducing the
common length to 43 steps.
The assignment planning for different workpieces repre-
sents an iterative process where all iterations are very sim-
ilar to one another. In this way the whole assignment plan
represents a periodic pattern, that can be observed in Fig. 7.
Here there are two main patterns shown by black and white
colors (order of the working steps as well as their positions
on machines) that however differ in the last workpieces. It
means that in case the optimal (or near optimal) scheme for
the first iteration is found, next iteration can use the same
scheme. The distributed approach being able to treat this kind
of pattern-like problem is known as the ant colony optimiza-
tion algorithm (ACO) [18]. This method originated from ob-
servation of ants in the colony. Going from the nest to the
food source, every ant deposits a chemical substance, called
pheromone, on the ground. In the decision point (intersection
between branches of a path) ants make a local decision based
on the amount of pheromone, where a higher concentration
means a shorter path. The result of this strategy represents
a pattern of routes where thick line points to a shorter path.
Similar strategy can be applied to local decisions of agents,
participating in the plan making.
5 Conclusion
The presented approach enables to react reasonably to dis-
turbances in manufacturing by using the constraint-based ap-
proach in a multi-agent way. It does not require any central-
ized elements, that essentially increases a reliability of com-
mon system.
References
[1] G. Weiss. Multiagent systems. MIT Press, 1999.
[2] T. Sandholm. Negotiation among self-interested computation-
ally limited agents. PhD thesis, University of Massachusetts,
Amherst, 1996.
[3] M. Pinedo.
Scheduling: Theory, Algorithms and Systems.
Prentice Hall, 1995.
[4] S. Kornienko O. Kornienko and P. Levi. Collective decision
making using natural self-organization in distributed systems.
In Proc. of CIMCA'2001, pages 460 -- 471, Las Vegas, USA,
2001.
[5] P. Levi, M. Schanz, S. Kornienko, and O. Kornienko. Ap-
plication of order parameter equation for the analysis and the
control of nonlinear time discrete dynamical systems. Int. J.
Bifurcation and Chaos, 9(8):1619 -- 1634, 1999.
[6] S. Graves, A.H.G. Rinnooy Kan, and P. Zipkin. Logistics of
Production and Inventory. Volume 4 of Handbooks in Op-
erations Research and Management Science, North Holland,
1993.
J. Blazewicz, W. Domschke, and E. Pesch. The job shop
scheduling problem. European Journal of Operational Re-
search, 93:1 -- 33, 1996.
[7]
[8] K. Alicke. Modellierung und Optimierung von mehrstufigen
Umschlagsystemen. PhD thesis, University of Karlsruhe, Karl-
sruhe, 1999.
[9] R. Bellman and S.E. Dreyfus. Applied dynamic programming.
Princeton, New Jersey, 1962.
[10] E. H. Durfee. Distributed problem solving and planing.
In
G. Weiss, editor, Multiagent Systems, pages 121 -- 164. MIT
Press, 1999.
[11] S. Kornienko, O. Kornienko, and J. Priese. Application of
multi-agent planning to the assignment problem. Computers
in Industry, 54(3):273 -- 290, 2004.
[12] H.K. Tonshoff, P.-O. Woelk, I.J. Timm, and O. Herzog. Flexi-
ble process planning and production control using co-operative
agent systems. In Proc. of COMA' 01, pages 442 -- 449, South
Africa, 2001. Stellenbosch.
[13] C. Constantinescu, S. Kornienko, O. Kornienko,
and
U. Heinkel. An agent-based approach to support the scalabil-
ity of change propagation. In Proc. of ISCA04, pages 157 -- 164,
San-Francisco, USA, 2004.
[14] M. Muscholl. Interaction und Kooperation in Multiagentsys-
temen. PhD thesis, University of Stuttgart, Stuttgart, 2001.
[15] S. Kornienko, O. Kornienko, and P. Levi. Multi-agent repairer
of damaged process plans in manufacturing environment. In
Proc. of the 8th Conf. on Intelligent Autonomous Systems (IAS-
8), Amsterdam, NL, pages 485 -- 494, 2004.
[16] A. Nareyek. Constraint-Based Agents, volume 2062 of Lecture
Notes in Computer Science. Springer-Verlag, 2001.
[17] S. Kornienko, O. Kornienko, and P. Levi. Flexible manufactur-
ing process planning based on the multi-agent technology. In
Proc. of Int. Conf. AIA'03, Innsbruck, Austria, pages 156 -- 161,
2003.
[18] D. Corne, M. Dorigo, and F. Glover (eds.). New ideas in opti-
mization. McGraw-Hill, 1999.
Figure 8: Non optimal assignment plans with different number of
jumps that result in different length and transportation costs.
Agents after the CSP approach choose several assignment
plans from the generated set of them and form an optimiza-
tion pool. These assignment plans can represent also only
segments of plans (these connected working steps represent
independent parts of assignment plan) that satisfy all formu-
lated constraints. These segments/plans can be combined into
a common plan so that to satisfy the postulated optimization
criterion. Thus, the more optimal segments are included into
this pool, the more optimal common plan will be obtained.
The ACO algorithm marks (like a pheromone rate) the opti-
mal segments obtained on the previous step. The fragments
with the highest pheromone rate are included into the top of
pool. In this way agents consider first the ACO-obtained se-
quence and try to modify it (e.g. using forecasting effect).
Thus, an optimization pool has always solutions with a high
pheromone rate, from them the most optimal one will be then
chosen.
The optimality of a plan is also influenced by a number of
transportations of a workpiece from one machine to another.
These transportations are represented by so-called "jumps"
in the plan making, as shown in Fig. 8. The minimal num-
ber of jumps for a workpiece is defined by technological re-
quirements and e.g. for a plan shown in Fig. 7 is equal to
2. However the number of jumps can be increased that wors-
ens a cost but improves other characteristics of an assignment
plan. This mechanism is utilized in combined optimization
criteria, e.g. the minimal cost at defined length (constant de-
livery date). Dependence between the number of jumps and,
for example, the length of generated plans is shown in Fig. 9.
Figure 9: Dependence between the number of jumps and the length
of generated plans. The curves 1, 2, 3 represent the cases without
forecasting, with one-position forecasting for only the first work-
piece, and with two-positions forecasting for all workpieces.
|
1002.0170 | 2 | 1002 | 2010-02-02T18:24:34 | Spectral Analysis of Virus Spreading in Random Geometric Networks | [
"cs.MA",
"cs.CE",
"cs.DM",
"nlin.AO"
] | In this paper, we study the dynamics of a viral spreading process in random geometric graphs (RGG). The spreading of the viral process we consider in this paper is closely related with the eigenvalues of the adjacency matrix of the graph. We deduce new explicit expressions for all the moments of the eigenvalue distribution of the adjacency matrix as a function of the spatial density of nodes and the radius of connection. We apply these expressions to study the behavior of the viral infection in an RGG. Based on our results, we deduce an analytical condition that can be used to design RGG's in order to tame an initial viral infection. Numerical simulations are in accordance with our analytical predictions. | cs.MA | cs | Spectral Analysis of Virus Spreading in Random Geometric Networks
Victor M. Preciado and Ali Jadbabaie
0
1
0
2
b
e
F
2
]
A
M
.
s
c
[
2
v
0
7
1
0
.
2
0
0
1
:
v
i
X
r
a
Abstract -- In this paper, we study the dynamics of a viral
spreading process in random geometric graphs (RGG). The
spreading of the viral process we consider in this paper is
closely related with the eigenvalues of the adjacency matrix
of the graph. We deduce new explicit expressions for all the
moments of the eigenvalue distribution of the adjacency matrix
as a function of the spatial density of nodes and the radius of
connection. We apply these expressions to study the behavior of
the viral infection in an RGG. Based on our results, we deduce
an analytical condition that can be used to design RGG's in
order to tame an initial viral infection. Numerical simulations
are in accordance with our analytical predictions.
I. INTRODUCTION
The analysis of spreading processes in large-scale complex
networks is a fundamental dynamical problem in network
science. The relationship between the dynamics of epi-
demic/information spreading and the structure of the under-
lying network is crucial in many practical cases, such as
the spreading of worms in a computer network, viruses in a
human population, or rumors in a social network. Several
papers approached different facets of the virus spreading
problem. A rigorous analysis of epidemic spreading in a
finite one-dimensional
linear network was developed by
Durrett and Liu in [3]. In [10], Wang et al. derived a sufficient
condition to tame an epidemic outbreak in terms of the
spectral radius of the adjacency matrix of the underlying
graph. Similar results were derived by Ganesh et al. in [4],
establishing a connection between the behavior of a viral
infection and the eigenvalues of the adjacency matrix of the
network.
In this paper, we study the dynamics of a viral spreading
in an important type of proximity networks called Random
Geometric Graphs (RGG). RGG's consist of a set of vertices
randomly distributed in a given spatial region with edges
connecting pairs of nodes that are within a given distance
r from each other (also called connectivity radius). In this
paper, we derive new explicit expressions for the expected
spectral moments of the random adjacency matrix associated
to an RGG. Our results allow us to derive analytical condi-
tions under which an RGG is well-suited to tame an infection
in the network.
The paper is structured as follows. In Section II, we
describe random geometric graphs and introduce several
useful results concerning their structural properties. We also
present the spreading model in [10] and review an important
This work was supported by ONR MURI N000140810747, and AFOR's
The
authors
complex networks program.
and Sys-
tems Engineering, University of Pennsylvania, 3451 Walnut Street,
{preciado,jadbabai}@seas.upenn.edu
are with the Department of Electrical
result that relates the behavior of an initial infection with the
spectral radius of the adjacency matrix. In Section III, we
study the eigenvalue spectrum of random geometric graphs.
We derive explicit expressions for the expected spectral
moments in the case of one- and two-dimensional RGG's.
In Section IV, we use these expressions to study the spectral
radius of RGG's. Our results allow us to design RGG's
with the objective of taming epidemic outbreaks. Numerical
simulations in Section IV validate our results.
II. VIRUS SPREADING IN RANDOM GEOMETRIC GRAPHS
In this section, we briefly describe random geometric
graphs and introduce several useful results concerning their
structural properties (see [7] for a thorough treatment). We
then describe the spreading model introduced in [10] and
show how to study the behavior of an infection in the network
from the point of view of the adjacency eigenvalues.
A. Random Geometric Graphs
Consider a set of n nodes, Vn = {v1, ...,vn}, respectively
c n = {x1, ...,xn}, where xi
located at random positions,
are i.i.d. random vectors uniformly distributed on the d-
dimensional unit torus, Td. We use the torus for convenience,
to avoid boundary effects. We then connect two nodes vi,v j ∈
xi − x j(cid:13)(cid:13) ≤ r, where r is the so-called
Vn if and only if (cid:13)(cid:13)
connectivity radius. In other words, a link exists between
vi and v j if and only if v j lies inside the sphere of radius
r (n) centered at vi. We denote this spherical region by Si (r),
and the resulting random geometric graph by G (c n;r). We
define a walk of length k from v0 to vk as an ordered set of
(possibly repeated) vertices (v0,v1, ...,vk) such that vi ∼ vi+1,
for i = 0,1, ...,k− 1; if n k = n 0 the walk is said to be closed.
is the number of edges
The degree di of a node vi
connected to it. In our case, the degrees are identical random
variables with expectation [7]:
E [di] = nV (d)rd,
(1)
where V (d) is the volume of a d-dimensional unit sphere,
V (d) = p d/2(cid:14)G (d/2 + 1), and G (·) is the Gamma function.
The clustering coefficient is a measure of the number of
triangles in a given graph, where a triangle is defined by
the set of edges {(i, j) , ( j,k) , (k,i)} such that i ∼ j ∼ k ∼ i.
For one- and two-dimensional RGG's we can derive an
explicit expression for the expected number of triangles,
E [ti], touching a particular node vi (details are provided in
Section III).
The adjacency matrix of an undirected graph G, denoted
by A(G) = [ai j], is defined entry-wise by ai j = 1 if nodes
i and j are connected, and ai j = 0 otherwise. (Note that
aii = 0 for simple graphs.) Denote the eigenvalues of a n× n
symmetric adjacency matrix A(G) by l 1 ≤ ... ≤ l n. The k-th
order moment of the eigenvalue spectrum of A(G) is defined
as:
1
N
l k
i
mk(G) =
n(cid:229)
i=1
(which is also called the k-th order spectral moment).
We are interested in studying asymptotic properties of the
[7], two particularly interesting regimes are introduced: the
sequence G(c n;r (n)) for some sequence {r (n) : n ∈ N}. In
thermodynamic limit with nr (n)d → a ∈ (0,¥ ), so that the
connectivity regime with r (n) → g (cid:16) log n
expected degree of a vertex tends to a constant, and the
with a constant
g , so that the expected degree of the nodes grows as clog n.
In this paper, we focus on studying the spectral moments
in the connectivity regime. In Section III, we derive explicit
expressions for the expected spectral moments of G (c n;rn)
for any network size n. We then use this information to bound
the spectral radius of the adjacency matrix of G(c n;r (n)).
B. Spectral Analysis of Virus Spreading
n (cid:17)1/d
In this section, we briefly review an automaton model
that describes the dynamics of a viral infection in a spe-
cific network of interactions. This model was proposed and
analyzed in [10], where a connection between the growth of
an initial infection in the network and the spectral radius of
the adjacency matrix was established. This model involves
several parameters. First, the infection rate b
represents the
probability of a virus at an infected node i spreading to
another neighboring node j during a time step. Also, we
denote by d
the probability of recovery of any infected node
at each time step. For simplicity, we consider b and d
to be
constants for all the nodes in G. We also denote by pi [k] the
probability that node i is infected at time k. The evolution
of the probability of infection is modeled by means of the
following system of non-linear difference equation:
pi [k + 1] = [1− (cid:213)
j∈Ni
(1− b p j [k])] + (1− d ) pi [k] ,
(2)
for i = 1, ...,n, where Ni denotes the set of nodes connected
to node i. We are interested in studying the dynamics of the
system for a low-density level of infection, i.e., b p j [k] ≪
1. In this regime, a sufficient condition for a small initial
infection to die out is [10]:
l max (A(G)) <
d
b
.
(3)
One can prove that (3) is a sufficient condition for local
stability around the disease-free state. Thus, we can use
condition (3) to design networks with the objective of taming
initial low-density infections.
III. SPECTRAL ANALYSIS OF RANDOM GEOMETRIC
GRAPHS
where
In this paper, we study the eigenvalue distribution of the
random adjacency matrix associated to G(c n;r (n)) for n →
. In this section, we characterize eigenvalue distribution
using its sequence of spectral moments. In our derivations,
we use an interesting graph-theoretical interpretation of the
spectral moments [1]: the k-th spectral moment of G is
proportional to the number of closed walks of length k in G.
This result allows us to transform the algebraic problem of
computing spectral moments of the adjacency matrix into the
combinatorial problem of counting closed walks in the graph.
In the following subsection, we compute the expected value
of the number of closed walks of length k in G(c n;r (n)).
A. Spectral Moments of One-Dimensional RGG's
As we mentioned above, we can compute the k-th spectral
moment of a graph by counting the number of closed walks
of length k. In the case of an RGG G(c n;r (n)), this number
is a random variable. In this subsection, we introduce a novel
technique to compute the expected number of closed walks
of length k. For clarity, we introduce our technique for the
first three expected spectral moments k = 1,2,3. We then use
these results to induce a general expression for higher-order
moments in one-dimensional RGG's.
The first-order spectral moment is equal to the number
of closed walks of length k = 1. Since G(c n;r) is a simple
graphs with no self-loops, we have that m1 (G(c n;r)) is a
deterministic quantity equal to 0.
the
We
expected
now study
second moment,
E [m2 (G(c n;r))], by counting the number of
closed
walks of length two. In simple graphs, the only possible
closed walks of length two are those that start at a given
node vi, visit a neighboring node v j ∈ Ni, and return back
to vi. Hence, the number of closed walks of length two
starting at vi is equal to di. Thus, from (1), we have
E [m2] =
1
n
n(cid:229)
i=1
E [di] = nV (d)rd,
where this result is valid for any dimension d ≥ 1.
The third spectral moment is proportional to the number
of closed walks of length three in the graph. We now derive
an expression for the expected number of triangular walks
starting at a given node vi in a one-dimensional RGG. Since
all nodes are statistically equivalent, our result is valid for
any other starting node. For simplicity in our calculations,
we consider that vi is located at the origin. A triangular walk
starting at node vi exists if and only if there exist two nodes
v j anv vk such that (cid:12)(cid:12)
x j(cid:12)(cid:12) ≤ r, xk ≤ r, and (cid:12)(cid:12)
since the random distribution of vertices on T1 is uniform
(with density n), the probability of nodes v j and vk being
respectively located in the differential lengths [x j + dx j) and
[xk + dxk) is equal to n2 dx jdxk. Hence, one can compute the
expected number of triangular walks starting at node vi as
xk − x j(cid:12)(cid:12) ≤ r. Also,
E [ti] =Z Z(x j ,xk)∈H2(r(n))
n2 dx jdxk,
H2 (r) =(cid:8)(x j,xk) ∈ T2 s.t. (cid:12)(cid:12)
x j(cid:12)(cid:12) ≤ r,
xk − x j(cid:12)(cid:12) ≤ r,xk ≤ r(cid:9) .
(cid:12)(cid:12)
(4)
¥
our derivation, due to space limitations, an explicit expres-
sion for the volume of Hk (1) is given by [9]:
Vol (Hk (1)) =
2
k!
k−1
j=1(cid:18) k
j− 1(cid:19) Ek, j,
(7)
where Ed,k denotes the Eulerian numbers. Substituting (7) in
(6) we obtain the statement of our lemma.
In [6], Lasserre proposed an algorithm to compute the
volume of a polyhedron defined by a set of linear inequali-
ties. We can use this algorithm to verify the validity of (5).
Applying this algorithm to the set of inequalities in (5), we
compute the following volumes for k = 1, ...,10:
H1 = 2, H2 = 3, H3 = 5.333..., H4 = 9.58333...,
H5 = 17.6000..., H6 = 32.70555..., H7 = 61.3587...,
H8 = 115.947..., H9 = 220.3238..., H10 = 420.825...
as
These numerical values match perfectly with our analytical
expression in Theorem 1.
If nr (n) = W
logn, or
(log n) (i.e.,
faster), one
the average degree grows
can prove
that E [mk] =
]. Hence, from (6) and (7), we
have the following closed-form expression for the asymptotic
expected spectral moments:
(cid:0)1 + O(cid:0)log−1 n(cid:1)(cid:1) E[W (k)
i
E [mk] ≍ (nr)k−1
1
2 (k− 1)!
k−2
j=1(cid:18)k− 1
j− 1(cid:19) Ek−1, j.
(8)
In the following table we compare the analytical result
in (8) with numerical realizations of the empirical spectral
moments. In our simulations, we distribute n = 1000 nodes
uniformly in T1 and choose a connectivity radius r = 0.01
(which results in an average degree E[di] = 20). The second,
third, and forth column in the following table represent
the analytical expectations of the spectral moments,
the
empirical average of the spectral moments from 10 random
realizations of the RGG, and the corresponding empirical
typical deviation, respectively.
(5)
k E [mk] Empirical Average Typical Deviation
1
2
3
4
1.38e-16
19.9326
297.284
5,956.30
1.3e-15
0.0976
4.3598
196.94
0
20
300
5,733
Thus, E [ti] can be computed as n2Vol[H2 (r (n))] (where
Vol(H) denotes the volume contained by the polyhedron H.)
Notice that H2 (r) can be defined by a set of linear inequalies;
hence, H2 (r) is a convex polyhedron that depends on r.
Furthermore, the set of linear inequalities in (4) presents a
homogeneous dependency with respect to the parameter r.
Therefore, we can write Vol(H2 (r)) as r2Vol(H2 (1)). Finally,
one can easily compute the volume of H2 (1) to be equal
to 3. Thus, the expected third spectral moment of a one-
dimensional RGG is given by
n(cid:229)
E [ti] = 3n2r2.
E [m3] =
1
n
i=1
In the following, we extend the above technique to com-
pute higher-order expected spectral moments. Denote by
W (k)
the number of closed walks of length k starting at node
i
vi in G(c n;r (n)). Regarding W (k)
, we derive the following
result.
i
Theorem 1: The expected number of closed walks of
, in a random geometric graph, G(c n;r), on
length k, W (k)
T1 is given by
i
k−2
1
Proof:
2 (k− 1)!
EhW (k)
i i = (nr)k−1
j− 1(cid:19) Ek−1, j,
j=1(cid:18)k− 1
where Ek−1, j are the Eulerian numbers 1.
Consider a particular closed walk, wk =
(v1,v2,v3, ...,vk,v1), of length k starting and ending at node
v1 (which we locate at zero for computational convenience).
A walk wk exists if and only if there exists a set of k− 1
x j+1 − x j(cid:12)(cid:12) ≤ r for
nodes, {v2,v3, ...,vk} , such that x1 ≤ r, (cid:12)(cid:12)
j = 2, ...,k− 1, and xk ≤ r. Since the distribution of vertices
on T1 is uniform (with density n) one can compute the
expectation of W (k)
as
i
where
nk−1 dx2...dxk,
i i =Z(x2,...,xk)∈Hk−1(r(n))
EhW (k)
Hk−1 (r) =n(v2,v3, ...,vk) ∈ Tk−1 s.t. v2 ≤ r,
x j+1 − x j(cid:12)(cid:12) ≤ r for j = 2, ...,k− 1,
(cid:12)(cid:12)
xk ≤ r} .
i
Thus, E[W (k)
] can be computed as nk−1Vol[Hk−1 (r)], where
Hk−1 (r) is a convex polyhedron defined by a set of linear
inequalities. Finally, note that the homogeneous structure of
the system of linear inequalities defining Hk−1 (r) allows us
to write Vol(Hk−1 (r)) = rk−1Vol(Hk−1 (1)). Therefore,
EhW (k)
i i = (nr)k−1 Vol (Hk−1 (1)) .
(6)
The volume of Hk−1 (1) is a particular number, independent
of the RGG parameters, i.e., n and r. Furthermore, we have
found an explicit analytical expression for the volume of
Hk (1) for any k ≥ 1. Although we do not provide details of
1The Eulerian number E (n,k) gives the number of permutations of
{1,2,...,n} having k permutation ascents [5].
Our numerical results present an excellent match with our
analytical predictions.
B. Spectral Moments of Two-Dimensional RGG's
In this subsection, we derive expressions for the first three
expected spectral moments of G(c n;r (n)) when the nodes
are uniformly distributed in T2. The expressions for the
first and second expected spectral moments are m1 = 0 and
E [m2] = p nr2. The third spectral moment is proportional to
the number of closed walks of length three in the graph. In
the two-dimensional case, we count the number of triangular
walks using a technique that we illustrate in Fig. 1. In this
figure, we plot two nodes vi and v j. The parameters r and
f
in Fig. 1 denote the distance and angle between these
(cid:229)
(cid:229)
(cid:229)
Fig. 1. This figure illustrates the technique proposed in Section III.B to
count the number of triangular walks in a two-dimensional RGG.
x j − xi(cid:13)(cid:13)
two nodes, i.e., r ,(cid:13)(cid:13)
and f = ∡ (x j − xi). An edge
between vi and v j exists if an only if v j is located inside the
circle Si (r). In this setting, the probability of existence of a
triangle touching both vi and v j is equal to the probability of
a third node vk being in the shaded area Al (see Fig. 1). This
area is the result of intersecting the circles Si (r) and S j (r),
and the resulting probability is equal to n Al. The intersecting
region Al is a symmetric lens which area can be computed
as a function of r and r as follows:
Al (r ;r) =(cid:26) 2r2 cos−1(cid:0)
2r(cid:1)−
0,
r
r
2p4r2 − r 2,
(9)
Therefore, we can compute the expected number of triangles
by integrating over the set of all possible positions of v j, i.e.,
h ∈ [0,r] and f ∈ [0,2p ), as follows
for r ≤ r,
for r > r.
n2Al (r ;r) r dr df .
(10)
E [ti] =Z r
r =0Z 2p
f =0
(11)
After substituting (9) in (10), we can explicitly solve the
resulting integral to be
3√3
2
2
E [ti] = p −
4 !p (cid:0)nr2(cid:1)
≈ 5.78(cid:0)nr2(cid:1)
.
(cid:229) n
Consequently, we have the following expression for the third
expected spectral moment E [m3] = 1
n
i=1 E [ti] = E [ti].
In the following, we extend the technique introduced above
to compute closed walks of arbitrary length. Denote by W (k)
the number of closed walks of length k starting at node v1
in G(c n;r (n)). The idea behind our technique is illustrated
in Fig. 2, where we represent a particular closed walk of
length 6. We denote this walk by wk = (v1,v2, ...,vk−1,vk,v1).
We define the following set of relative distances and angles
between every pair of connected vertices: r i , kxi+1 − xik
and f i = ∡ (xi+1 − xi) for i = 1, ...,k− 2. We also define the
i
following parameter
k−2
j=1
r =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
r
jeia
,
j(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(12)
Fig. 2. This figure illustrates the technique proposed in Section III.B to
count the number of closed walks of length k in a two-dimensional RGG.
(i = √−1) which is the resulting distance between nodes
vk−1 and v1 given a particular set of distances and angles
{(ri,f i)}i=1,...,k−2 (see Fig. 2). In this setting, the conditional
probability of existence of a walk wk = (v1,v2, ...,vk−1,vk,v1)
given the set of relative positions, {(ri,f i)}i=1,...,k−2, is equal
to the probability of vk being in the shaded area Al in Fig.
2. We have an expression for this area in (9), where r
is
defined in (12). Finally, we can compute the expectation of
W (k)
by performing an integration over the set of all possible
i
positions (i.e., r
as follows
j ∈ [0,r] and f j ∈ [0,2p ) for j = 2, ...,k− 1),
EhW (k)
i i = nk−1Z(h ,j )∈Ck−2
Al (r ;r)
k−1
h
j=2
j dh dj ,
i
] ≈ 14.2511(cid:0)nr2(cid:1)3.
where h = (r 2, ...,r k−1),
j = (f 2, ...,f k−1), and Ck−2 =
{(h ,j ) : h ∈ [0,r]k−2 and j ∈ [0,2p )k−2}. Although a
closed-form for the above expression can only be computed
for k ≤ 3, we can always find a good approximation via
numerical integration. For example, the integration for k = 4
gives us E[W (4)
In the following table, we compare our analytical results
with numerical realizations of the empirical spectral mo-
ments of a two-dimensional RGG. In our simulations, we
distribute n = 1000 nodes uniformly on T2 and choose a
connectivity radius r =p50/p n ≈ 0.1784 (which results in
an average degree E[di] = 50). The second, third, and forth
columns in the following table represent the analytical expec-
tation of the spectral moments, the empirical average from 10
random realizations, and the corresponding empirical typical
(cid:229)
(cid:213)
deviation, respectively.
0
50
k E [mk]
1
2
3
4
1,464.1
59,452
Empirical Average Typical Deviation
-9.2e-16
50.0820
1,475.8
60,127
1.1e-15
0.3908
37.3777
2,955.3
Our numerical results present an excellent match with our
analytical predictions.
In the following section, we use the results introduced in
this section to study the spreading of an infection in a random
geometric network.
IV. SPECTRAL ANALYSIS OF VIRUS SPREADING
In this section, we use the expressions for the expected
spectral moments to design random geometric networks to
tame an initial viral infection in the network. In our design
problem, we consider that the size of the network n and
the parameters in (2), i.e., b and d , are given. Hence, our
design problem is reduced to studying the range of values of
r for which the RGG is well-suited to tame an initial viral
infection.
A sufficient condition for
local stability around the
disease-free state was given in (3). Thus, we have to find
the range of values of r for which the associated spectral
radius l max is smaller than the ratio d /b . In the following
subsection, we show how to derive an analytical upper
bound for the spectral radius based on the expected spectral
moments.
A. Analytical Upper Bound for the Spectral Radius
In order to upper-bound the spectral radius, we use
Wigner's high-order moment method [11]. This method
provides a probabilistic upper bound based on the asymptotic
behavior of the k-th expected spectral moments for large k.
We present the details for a one-dimensional RGG, although
the same technique can be applied to RGG's in higher
dimensions. For a one-dimensional RGG in the connectivity
regime, we derived an explicit expression for the expected
spectral moments in (8). A logarithmic plot of Vol(Hk (1))
for k = 1,2, ...,9 unveils that Vol(Hk (1)) → b 1ck
1 for large-
order moments (a line in logarithmic scale), where, from a
numerical fitting, we find that b 1 = 0.35 and c1 = 1.9192.
Therefore, from (8) we have
E [mk] ≍ b 1 (c1nr)k ,
for large k.
For even-order expected spectral moments (i.e., k = 2s for
s ∈ N), the following holds
n(cid:229)
E [m2s] =
1
n
i=1
E[l 2s
i
1
n
] ≥
E[l 2s
max].
Define f (n) = n1−d log n; thus, for any e ,d > 0 (and c1 =
1.9192), we can apply Markov´s inequality as follows
P(cid:0)l 2s
max ≥ (c1nr + e r f (n))2s(cid:1) ≤
≤
E[l 2s
max]
(c1nr + e r f (n))2s
n E [m2s]
(c1nr + e r f (n))2s ,
Fig. 3.
Comparison between the empirical spectral radius of an RGG
(circles in the plot) and the values of our analytical upper bound (solid
line) for n = 1000 and r (n) = ¯d/2n, with expected degrees ¯d = [10:1:100].
For large s, one can prove that [8]
e
c1
sn−d logn(cid:19) .
P (l max ≥ c1nr + e r f (n)) ≤ nb 1 exp(cid:18)−
Assuming that s grows as b 2nd , for b 2,d > 0, we have
P (l max ≥ c1nr + e r f (n)) ≤ nb 1 exp(cid:18)−
for all sufficiently large e . Thus,
logn(cid:19) = o (1) ,
b 2e
c1
lim
n→¥ P(cid:16)l max < c1nr + e rn1−d logn(cid:17) = 1.
(13)
In other words, l max is upper-bounded by cnr +e rn1−d logn
with probability 1 for n → ¥
. In practice, for a large (but
finite) n, we can use 1.9192 nr as an upper bound of l max.
In Fig. 4, we plot the empirical spectral radius of an RGG
with n = 1000 and r (n) = ¯d/2n, with expected degrees ¯d =
[10:1:100] (circles in the figure). We also plot the values of
our analytical upper bound, 1.9192 nr, in solid line.
The technique introduced in this subsection is also valid
for RGG's in higher-dimensions. In general, one can prove
that for a d-dimensional RGG that the expected spectral
moment grows as E [mk] → b d(cid:0)cdnrd(cid:1)k. Applying Wigner's
high-order moment method to this sequence, one can derive
a probabilistic upper bound similar to (13). In particular, we
have that l max < cdnrd for large n with high probability. In
the following subsection, we use our results to design the
connectivity radius of an RGG in order to tame an initial
viral infection.
B. Spectral Radius Design
Once the spectral radius is upper-bounded, our design
problem becomes trivial. Since (3) represents a sufficient
condition for local stability around the disease-free state, we
have the following condition to tame an initial viral infection
for a d-dimensional RGG:
l max (G(c n;r)) < cdnrd <
d
b
,
Fig. 4. Color map representing the evolution of the probabilities of infection
pi [n] for i = 1,...,1000 in an RGG with n = 1000 nodes, connectivity
radius r = 0.005, rate of infection b = 0.020, and recovery rate d = 0.018.
Each horizontal line represents the value of pi [n] for a particular i. In this
color map, blue represents a zero value, green and yellow tones represent
intermediate values, and red represents values close to one. In this case, we
observe an epidemic outbreak.
which implies the following design condition for the con-
nectivity radius:
r <(cid:18) d
b cdn(cid:19)1/d
,
(14)
where cd is a positive constant that depends on the dimension
of Td. For example, in the one-dimensional case, we have
c1 = 1.9192; hence, (14) becomes r < d / (1.9192 b n). We
now validate this result with several numerical simulations
of a viral infection in a one-dimensional RGG.
Consider an RGG with n = 1000 nodes and a connectivity
radius of r = 0.005 (which implies an average degree of
10). The resulting spectral radius in this RGG is l max =
17.2629. In our numerical simulations, we choose the initial
probability of infection to be pi [0] ∼ 0.01Unif[0,1]; hence,
approximately 1% of the nodes in the network are initially
infected. In our first experiment, we choose a rate of infection
b = 0.020, and a recovery rate d = 0.018. Since the sufficient
condition for viral control in (14) is not satisfied, we cannot
guarantee an initial infection to be tamed. In Fig. 5 we show
an image of the evolution of the probability of infection for
this case. This figure is a color map for the simultaneos
evolution of pi [n] for i = 1, ...,1000. Each horizontal line
represents the value of pi [n] for a particular i. In this color
map, blue represents a zero value, green and yellow tones
represent intermediate values, and red represents values close
to one. On the other hand, if we increase the recovery rate to
d = 0.35 keeping the rest of parameters fixed, we have that
d /b = 17.50 > l max and we satisfy condition (3). Hence,
the probability of infection of every node is guaranteed to
converge towards zero. In Fig. 6, we observe the color map
for the evolution of the probability of infection in this case,
where we clearly observe how pi [n] → 0 for all i. Hence,
this latter RGG is well-suited to tame initial viral infections.
V. CONCLUSIONS
In this paper, we have studied the spreading of a viral
infection in a random geometric graph from a spectral point
Fig. 5. Color map representing the evolution of the probabilities of infection
pi [n] when we increase the recovery rate to d = 0.35 (the rest of parameters
are the same as we used for Fig. 5). We observe how the probability of
infection of every node converges towards zero in this case.
of view. We have focused our attention on studying the
eigenvalue distribution of the adjacency matrix. We have
derived, for the first time, explicit expressions for the spectral
moments of the adjacency matrix as a function of the density
of nodes and the connectivity radius. We have then applied
our results to the problem of viral spreading in a network
with a low-density infection. Using our expressions, we have
derived upper bounds for the spectral radius of the adjacency
matrix. Finally, we have applied this upper bound to design
random geometric graphs that are well-suited to tame an
initial low-density infection. Our numerical results match our
predictions with high accuracy.
REFERENCES
[1] N. Biggs, Algebraic Graph Theory, 2nd Edition. Cambridge University
Press, 1993.
[2] P. Blackwell, M. Edmondson-Jones, and J. Jordan, "Spectra of Adja-
cency Matrices of Random Geometric Graphs," Preprint.
[3] R. Durrett and X.-E. Liu, "The Contact Process on a Finite Set,"
Annals of Probability, vol. 16, pp. 1158-1173, 1988.
[4] A. Ganesh, L. Massoulie, and D. Towsley, "The Effect of Network
Topology on the Spread of Epidemics," Proc. IEEE INFOCOM '05,
pp. 1455-1466, 2005.
[5] R.L. Graham, D.E. Knuth, and O. Patashnik, Concrete Mathematics:
A Foundation for Computer Science, Second Edition, Addison-Wesley,
1994.
[6] J.B. Lasserre, "An Analytical Expression and an Algorithm for the
Volume of a Convex Polyhedron in Rn," J. Optim. Theor. Appl., vol.
39, pp. 363 -- 377, 1983.
[7] M. Penrose, Random Geometric Graphs, Oxford University Press,
2003.
[8] V.M. Preciado, Spectral Analysis for Stochastic Models of Large-Scale
Complex Dynamical Networks, Ph.D. dissertation, Dept. Elect. Eng.
Comput. Sci., MIT, Cambridge, MA, 2008.
[9] V.M. Preciado and A. Jadbabaie, "Moment-Based Spectral Analysis
of Random Geometric Graphs," in preparation.
[10] Y. Wang, D. Chakrabarti, C. Wang, and C. Faloutsos, "Epidemic
Spreading in Real Networks: An Eigenvalue Viewpoint," Proc. Int.
Symp. Reliable Distributed Systems, pp. 25-34, 2003.
[11] E.P. Wigner, "On the Distribution of the Roots of Certain Symmetric
Matrices," Ann. Math., vol. 67, pp. 325 -- 327, 1958.
|
1510.02225 | 1 | 1510 | 2015-10-08T08:16:59 | Comparison between purely statistical and multi-agent based ap-proaches for occupant behaviour modeling in buildings | [
"cs.MA"
] | This paper analyzes two modeling approaches for occupant behaviour in buildings. It compares a purely statistical approach with a multi-agent social simulation based approach. The study concerns the door openings in an office. | cs.MA | cs | Comparison between purely statistical and multi-agent based ap-
proaches for occupant behaviour modeling in buildings
Conférence IBPSA France-Arras-2014
Khadija Tijani∗1,2,Ayesha Kashif1,3,Quoc Dung Ngo1,Stéphane Ploix1,Benjamin
Haas2,Julie Dugdale3
1 G-SCOP Sciences pour la Conception l'Optimisation et la Production à Greno-
ble
46 avenue Félix Viallet 38031 Grenoble France
2 CSTB Centre Scientifique et Technique du Bâtiment
84 avenue Jean Jaurès 77420 France
3 LIG Laboratoire d'Informatique de Grenoble
110 avenue de la chimie Domaine Universitaire de Saint Martin d'hères 38041 Greno-
ble France
∗[email protected]
RESUME. Ce papier analyse deux approches de modélisation du comportement d'occupants dans
le bâtiment. Il compare une approche purement statistique avec une approche basée sur la simu-
lation sociale dans un environnement multi-agent. L'étude concerne les ouvertures de porte dans
un bureau.
MOTS-CLÉFS. bâtiment, simulation d'occupants, approche statistique, simulation sociale, sys-
tème multi-agent
ABSTRACT. This paper analyzes two modeling approaches for occupant behaviour in buildings.
It compares a purely statistical approach with a multi-agent social simulation based approach.
The study concerns the door openings in an office.
KEYWORDS. building, occupant simulation, statistical approach, social simulation, multi-agent
system
1 Introduction and context
Because of the reduction of energy consumptions, the relative impact of occupants is becoming
more important. Thus, the design and the operation of building systems have to take into account
the occupant behaviours to improve the ratio between services provided to occupants and the
required energy: this is the concept of usage efficiency proposed in (Chenailler, 2012). Regarding
solutions driven by numerical models, occupant behaviours have to be modelled properly as the
main function of a building is to provide services to its users. Occupant behaviours have to be
considered in the study of building energy efficiency by modelling needs in terms of comfort,
energy and health to deduce how these affect the indoor environment.
Furthermore, Building Energy Management Systems (BEMS) help to optimize energy con-
1
5
1
0
2
t
c
O
8
]
A
M
.
s
c
[
1
v
5
2
2
2
0
.
0
1
5
1
:
v
i
X
r
a
Conférence IBPSA France-Arras-2014
sumption and also allow occupants to take better decisions regarding energy use. The way it
operates involves interactions with the occupants. The BEMS receives signals from the smart
grid and gives information about the availability of energy, the price details and energy con-
sumption. Therefore, the occupants are reactive because they may adapt their behaviours and
interact with the BEMS to adapt the proposed energy strategy.
The purpose of this paper is to show that a multi-agent based model can do what a statistical
model does and even more by comparing these two approaches: purely statistical approach and
multi-agent based design of occupant models. To carry out this comparison, the study focuses on
air quality and in particular on occupants' interactions with door openings in an office setting.
Such human behaviour depends on several parameters and constraints arising from the external
and social environment. A statistical model based on Markov chains is first developed to predict
the average state of the door (open or closed) each hour of the day for two months. These results
are validated by sensor data. Then, a model of human behaviour focusing on how occupants
interact with doors is proposed. The model has been developed using the BRAHMS multi-agent
software platform.
It is pointed out that a global Markov chain model can be implemented
directly in an agent. Then, it is shown that using social simulation with multi-agent systems,
the statistic model can be enriched to model interactions between occupants and equipments
that can be captured by field study.
2 State of the art
The literature suggests that occupant behaviour strongly influences energy consumption pat-
terns and is an important factor for energy waste reduction in buildings (R.V. Andersen and
Olesen, 2009). Models of human behaviour in building simulation tools are usually based on
statistical algorithms that predict the probability of an action or event. (Dong and Andrews,
2009) developed an event based pattern detection algorithm for sensor-based modelling and pre-
diction of user behaviour. They have connected behavioural models based on a Markov model
to building energy and comfort management through the EnergyPlus simulation tool for energy
calculations. Building simulation tools are based on heat transfer, thermodynamic equations and
a human model. The latter is traditionally based on fixed schedules and predefined rule; this
does not reflect the actual human behaviour complexity nor reactive and deliberative behaviours.
In the scientific literature, two different kinds of approaches of occupant behaviour can be
found: stochastic approaches, usually based on Markov chains, and multi-agent system ap-
proaches.
A first attempt (C.A.Roulet and P.Cretton, 1991) to develop a statistical model to predict
the state of windows was based on a discrete-time Markov process model to predict transitions
between angles of openings. They used Markov chains with six states, each corresponding to
a class of opening angles.(J.Page, 2007) built a time series of presence/absence from the data
collected from single person offices and used a Markov chain to reproduce the presence profiles
through simulations. (D.Robinson and F.Haldi, 2009) have developed a hybrid stochastic model
of window opening, based on three modelling approaches:
logistic probability distributions,
Markov chains and continuous-time random processes.
The multi-agent approach allows more complex reactive behaviours to be modelled but more
parameters must be customized. (S.Abras and M.Jacomino, 2010) gave the control of appliances
and sources to the software agents that are used in a home automation system.
(A.Kashif,
2014) describes a detailed inhabitant model that represents cognitive, reactive and deliberative
behaviours.
2
Conférence IBPSA France-Arras-2014
3 Problem statement
3.1 Study case
3.1.1 The field survey
In this section, the experiment, which yields data for model design, is presented. Data have
been collected from the Grenoble Institute of Technology that contains an engineering school
and a research laboratory. The building has been equipped with 165 ENOCEAN sensors on an
installation area of 1500 m2, which is divided into three distinct patterns of use:
• Administration (offices, meeting room, hall, etc)
• Teaching (classroom, computer room, corridors, etc)
• Research (offices, meeting room, cafeteria, open-space, etc)
This paper focuses on a particular office occupied by researchers because it is equipped
with many sensors. An air quality sensor is used for measuring CO2, VOC (Volatile organic
compounds), humidity (relative and absolute) and temperature. A presence sensor detects the
movement of a person in the room. This type of sensor is useful to validate the presence of a
person in the office. This sensor does not return a signal of presence if someone is sitting behind
a computer, or if someone is not in the detection area. The contact sensors give the state of doors
or windows (open=1/closed=0), the data from these sensors translate the interactions between
the occupants and the environment. These sensors provide data using the ENOCEAN radio
protocol: information may be lost during communication.
Only the data related to October and November 2013 have been used. All data from the
sensors are transmitted to a hub. These data are transmitted at random times, which are different
for each sensor. Therefore, each set of data has a variable size, depending on the day and given
data. Each sensor provides two vectors, a « time » vector and a « value » vector. Time vector
provides the dates in hours that corresponds to the values in the same position in the « value »
vector.
3.2 Study objective
This paper proposes a common framework to simulate both Markov chains at the group occupant
level, and a fine reactive modelling at the individual occupant level with a multi-agent system.
The proposed approach focuses firstly on the occupant actions on doors. A group model is de-
signed and simulated and within a multi-agent system using Markov chains. Secondly, the model
is enriched by modelling each occupant independently taking into account their interactions with
the door.
4 Modelling of the group
4.1 Purely statistical model
The Markov chain model is used to predict the door states at each hourly time step. Data
processing is carried out to standardize the contact sensor data in setting the data time step to
one hour and computing the door states in terms of the ratio of opening time within an hour.
Three states of the door are created:
• door is considered open when the opening duration percentage ≥ 80%
• door is considered as moving, if the door state changes many times within the considered
hour, when 20% ≤ opening duration percentage ≤ 80%
3
4.2 Multi-agent system implementation
Conférence IBPSA France-Arras-2014
Figure 1. Comparison between door states for real data and two Markov simulations
• door is considered closed when the opening duration percentage ≤ 20%
In the data processing, it assumed that the door is closed on weekends and during week days
from 8pm to 8am. Two time slots per day have been taken:
• 1st time slot (working time): 8am → 12pm and 2pm → 8pm
• 2nd time slot (lunch time): 12pm → 2pm
The next step is to build the transition matrix for each time slot. The transitions from one
state to another have been counted. Since there are 3 distinct states, the matrix is a 3 * 3 square
matrix, which models 9 transitions:
Too
T M =
Tom Toc
Tmo Tmm Tmc
Tcm Tcc
Tco
(1)
To get the door state for each hour, two steps are required. The first multiplies the transition
matrix by the vector representing the previous state. The second one selects a random number at
each time step and compare it with the probabilities obtained from the previous step to determine
the current state. The vector of door state probabilities is written as: P (t) = (Po(t), Pm(t), Pc(t))
with Po, the probability to have the state open, Pm, the probability to have the state move and
Pc, the probability to have the state close.
Assume the initial state of the door is closed: S(0) = (0, 0, 1). The next sample time t = 1h
is computed by: S(1) = D (S(0) × T M ) where the function D(.) stands for selecting a unified
random number r and comparing it with each probability element of P (1) = S(0) × T M:
• if r (cid:54) Po(1) then state is Open
• if Po(1) (cid:54) r (cid:54) Po(1) + Pm(1) then state is Move
• if Po(1) + Pm(1) (cid:54) r (cid:54) Po(1) + Pm(1) + Pc(1) then state is Close
Because of random processes, it has been simulated several times to check whether these
simulations are close to the recorded data. Figure 1 shows a comparison between the recorded
data and two simulation results for the 2 considered months.
4.2 Multi-agent system implementation
In the above section the actions of occupants on the door are simulated using Markov processes.
It can also be implemented using a multi agent system approach. The design process consists of
agent's perception from the outside environment and internal physical comfort. This perception
4
4.2 Multi-agent system implementation
Conférence IBPSA France-Arras-2014
Figure 2. Design process to select the next state of the door
of the environment are modelled as agents' beliefs. These beliefs lead the agent to go through the
cognition phase where the agent makes some desires based on the beliefs. Finally, based on the
social and environmental constraints, the desires are transformed into an agent's intention. This
finally leads the agent to take some actions on the environment. A change in the environment
leads the agent to revise its previous beliefs about the environment and based on the new beliefs
repeat the above process.
In a first step, all of the occupants in the office are considered as one global group agent, that
means that the behaviour of the group agent is actually the behaviour of the Markov process.
This group agent perceives some limited information from the outside environment, e.g. the
current day and time, the location and the current state of the door which are transformed into
agent's beliefs (fig.2). The door could be OPEN, CLOSE or in a MOVE state. If the current
state of the door is open then it has a set of probabilities for the next possible states, that means
that the group agent has desires for the door to be in a particular state. Some constraints convert
these desires into an intention of the agent to select the next state of the door. This constraint
states that the probability for the door to be open at next time step is higher than 0.8. In case
of probabilities between 0.2 and 0.8, a random process decides the selection of the next state of
the door. Based on this process, the group agent performs the action on the door and changes
its state for the next hour.
Figure 3 shows the screen shot of simulation results from BRAHMS simulation environment.
Only 2 out of 60 days are shown in the figure with actions of the group agent at each hour.
The object Door's state when changed is displayed by different colours, e.g. the red colored
workframe shows that the door is closed.
From the design process and the simulation results, it can be seen that the next state of the
door depends on some simple decision making variables, such as the current state of the door,
the time, etc. However, for individual agents that represent actual humans, some additional
knowledge, to that of Markov process, is required to predict the future actions on the door. This
additional knowledge represents the decisions based on the comfort levels, the influence of other
agents around, the social and group behaviours, etc. In the next section a scenario that takes
into account more complex behaviour of inhabitants and their resulting actions to change the
door state is explained. This shows that not only Markov processes can be implemented using
5
Conférence IBPSA France-Arras-2014
Figure 3. The state of the door at each hour due to actions by the group agent
a multi agent approach but also more sophisticated cognitive elements of occupants' behaviours
can be introduced in the simulation.
5 Modelling of individuals: application of a multi-agent based
design approach
5.1 Field study
In this section, the office occupants are modelled more precisely. The office is occupied by three
persons, Stephane, Khadija and Audrey. Stephane often goes for lectures at the university.
Audrey comes to the office one week out of two. Khadija comes to the laboratory every day.
In the morning, Khadija usually arrives first, then Stephane and then Audrey. The agent that
arrives first opens the door and then either leaves it open or closes it. The three agents have
different door opening and closing behaviours during different times of the day. Khadija mostly
closes the door after opening it in the morning. However, sometimes, she leaves it open. Audrey
always closes the door as she sits very close to it, it disturbs her. Stephane's behaviour is mostly
dependent upon the presence of Audrey in the office and he mostly leaves the door open if Audrey
is not in the laboratory, otherwise he closes the door if Audrey is present. While Stephane is in
the office, visitors sometimes come to see him. Visitors mostly leave the door open while they
are in the office and when they leave. Audrey, if present, closes the door after their departure,
otherwise it remains open. The agents, however, close the door while going to lunch. In the
afternoon, Khadija and Audrey usually go to the cafeteria for a coffee break. Khadija uses to
ask Stephane if he wants to go for a coffee. Sometimes he accepts Khadija's proposal but when
he is busy, he prefers Khadija to bring him a coffee from cafeteria. If Khadija has to bring coffee
for Stephane, and Audrey is not in the university, Khadija leaves the door open as she leaves as
she believes that it would be difficult on the way back to open the door with a coffee in each
hand. However, if Audrey is in the office and if she is accompanying Khadija, she closes the door
before going to cafeteria to get coffee. Stephane sometimes has meetings in a nearby meeting
room and he usually leaves the door open when leaving, except if Audrey is in the office: in that
case, he closes the door when leaving. Finally, the door is closed at night.
5.2 Simulation and results
Each actor is represented by an agent in the environment. This environment monitors the
movements of the agent, the activities that the agent performs and its thought processes.
Figure 4 shows a screen shot from the simulation where agent Audrey is present in the office.
The state of the door in the presence of agent Audrey is kept mostly closed. When the visitors
arrive at around 11am, they leave the door open but as soon as they leave, agent Audrey closes
the door. Figure 4 also shows the communication between the agents Khadija and Stephane,
where agent Khadija asks agent Stephane to have a coffee and agent Stephane replies. As agent
6
Conférence IBPSA France-Arras-2014
Audrey is in the office, agent Stephane closes the door before leaving to the nearby meeting
room.
Figure 4. Impact of occupants' behaviour on the state of the door in the presence of Audrey agent
Figure 5 shows a comparison between the recorded data and the multi-agent approach model
results.
6 Analysis of the results
The previous sections show that a multi-agent model provides the same results as a purely
statistical model. Therefore, theoretically speaking, a multi-agent model could be implemented
as a purely statistical model. Multi-agent based social simulation makes it possible to easily
implement the behaviours captured during field studies, to model the reasons behind an agent's
actions i.e. cognition and reactive behaviours, and to model the interactions between individuals.
Figure 6 shows a comparison between the two approaches with the recorded data. This shows
that a multi-agent based approach can model the complex behaviours and multiple interactions
between occupants.
7 Conclusion and future works
Multi-agent approaches are able to capture the same level of complexity as Markov chain pro-
cesses. Combined with fields studies, multi-agent approach can propose models which goes
Figure 5. Comparison between the real data and the multi-agent approach model results
7
REFERENCES
Conférence IBPSA France-Arras-2014
Figure 6. Comparison between Multi-Agent Approach results and the recorded data for 2 months
beyond Markov chain statistical approaches. However, the complexity of current multi-agent
description make it difficult to apply to building energy and indoor air quality simulation. The
future work will focus on devising a simplified multi-agent approach for building simulation. A
first step will be a connection with C02 modelling for an integrated evaluation of indoor air
quality and actions on doors and windows.
This work is funded by Agence Nationale de la Recherche project MAEVIA under contract
ANR-12-VBDU-0005. Julie Dugdale would like to acknowledge the support of the University of
Adger, Norway, to which she is affiliated.
References
A.Kashif (2014). Modélisation du comportement humain réactif et délibératif avec une approche
multi-agent pour la gestion énergétique dans le bâtiment. PhD thesis, Université de Grenoble.
C.A.Roulet, R.Fritsch, J. and P.Cretton (1991). Stochastic model of inhabitant behavior with
regard to ventilation. Technical report.
Chenailler, H. (2012). L'efficacité d'usage énergétique : pour une meilleure gestion de l'énergie
électrique intégrant les occupants dans les bâtiments. PhD thesis, Université de Grenoble.
Dong, B. and Andrews, B. (2009). Sensor based occupancy behaviour pattern recognition for
energy and comfort management in intelligent buildings. 11th Int. Building Perfor- mance
Simulation Association (IBPSA) Conf. (Glasgow, Scotland; 2009),, pages 1444 -- 1451.
D.Robinson and F.Haldi (2009). Interactions with window openings by office occupants. Energy
and Buildings, 44:2378 -- 2395.
J.Page, D.Robinson, J. (2007). Stochastic simulation of occupant presence and behaviour in
buildings. Proc. Tenth Int. IBPSA Conf : Building Simulation, pages 757 -- 764.
R.V. Andersen, J. Toftum, K. A. and Olesen, B. (2009). Survey of occupant behaviour and
control of indoor environment in danish dwellings. Energy and Buildings, 41:11 -- 16.
S.Abras, S.Ploix, S. and M.Jacomino (2010). Advantages of mas for the resolution of a power
management problem in smart homes, in advances in intelligent and soft computing. Springer,
Berlin, Heidelberg, pages 269 -- 278.
8
|
1904.04007 | 1 | 1904 | 2019-04-08T12:38:05 | Coupling agent based simulation with dynamic networks analysis to study the emergence of mutual knowledge as a percolation phenomenon | [
"cs.MA"
] | The emergence of mutual knowledge is a major cognitive mechanism for the robustness of complex socio technical systems. It has been extensively studied from an ethnomethodological point of view and empirically reproduced by multi agent simulations. Whilst such simulations have been used to design real work settings the underlying theoretical grounding for the process is vague. The aim of this paper is to investigate whether the emergence of mutual knowledge (MK) in a group of co-located individuals can be explained as a percolation phenomenon. The followed methodology consists in coupling agent-based simulation with dynamic networks analysis to study information propagation phenomena: after using an agent-based simulation we generated and then analysed its traces as networks where agents met and exchanged knowledge. Deep analysis of the resulting networks clearly shows that the emergence of MK is comparable to a percolation process. We specifically focus on how changes at the microscopic level in our agent based simulator affect percolation and robustness. These results therefore provides theoretical basis for the analysis of social organizations. | cs.MA | cs | Coupling agent based simulation with dynamic networks analysis to study the
emergence of mutual knowledge as a percolation phenomenon
Julie Dugdale(1,2), Narjès Bellamine Ben Saoud(3,4), Fedia Zouai3, Bernard Pavard5
(1) Université Grenoble Alpes; LIG, Grenoble, France
(2) University of Agder, Norway
(3) Laboratoire RIADI -- ENSI Université de La Manouba, Tunisia
(4) Institut Supérieur d'Informatique, Université Tunis el Manar, Tunisia
(5) Université P. Sabatier, CNRS, IRIT, Toulouse, France
Abstract
The emergence of mutual knowledge is a major cognitive mechanism for the
robustness of complex socio technical systems. It has been extensively studied from
an ethnomethodological point of view and empirically reproduced by multi agent
simulations. Whilst such simulations have been used to design real work settings the
underlying theoretical grounding for the process is vague. The aim of this paper is to
investigate whether the emergence of mutual knowledge (MK) in a group of co-
located individuals can be explained as a percolation phenomenon. The followed
methodology consists in coupling agent-based simulation with dynamic networks
analysis to study information propagation phenomena: after using an agent-based
simulation we generated and then analysed its traces as networks where agents met
and exchanged knowledge. Deep analysis of the resulting networks clearly shows that
the emergence of MK is comparable to a percolation process. We specifically focus
on how changes at the microscopic level in our agent based simulator affect
percolation and robustness. These results therefore provides theoretical basis for the
analysis of social organizations.
Keywords:
agent-based simulation, complex network dynamics, Percolation, social networks,
mutual knowledge, emergence
Background
1.
The aim of this paper is to propose a theoretical model for emergent organisations
such as those often encountered in complex or degraded real-world cooperative
systems.
Usually, organisations such as those in emergency control rooms, space control
centres, and nuclear power plants are composed of a group of people interacting in a
proximal space (real or virtual). In addition to verbal interactions, people also interact
in an informal way through gestures and unobtrusive observations of the actions of
others.
Such informal organisations are highly paradoxical; whilst there are many rules and
procedures that constrain how the group should handle critical situations, the way that
the activities unfold in dealing with an event is mainly non deterministic and
unstructured. What is interesting is that this paradox is only apparent if we analyse
the situation from a systemic point of view. There are many examples of this apparent
paradox, for example in early field studies of Air Traffic Control (ATC) settings,
Mell observed that even if verbal exchanges between air traffic controllers are fully
1
constrained lexically and syntactically, in the real situation only 20% of exchanges
follow the rules [18]. In a study concerning incident management in ATC, non-verbal
communication and the use of informal artefacts were found to strongly structure
coordination processes within the team [14].
The intuition behind these mechanisms is that mutual knowledge emerges more
'easily' with informal organisations than with normalised exchanges. This is largely
due to the unobtrusive nature of an informal organisation where actors are gathering
information as they need it and when their interlocutor seems available. Likewise,
seemingly adhoc and informal broadcasting of information may also be selective and
modulated by the context of the situation [5].
We are particularly interested in social cooperation and understanding collective
behaviour where agents in a complex social system may rapidly share information.
We have seen that mutual knowledge emerges when people are spatially close
together and are willing to communicate. In such situations, information can
propagate very rapidly with a minimal perturbation to on-going cognitive processes.
Mutual knowledge has also been shown to contribute to the exceptional robustness of
socio-technical systems [17] [19]. However there are many factors that can adversely
affect emergent behaviours and are detrimental to cooperation. For example, if
interactions between agents are impaired by noise, or if there are too many
interlocutors, etc. the emergence of mutual knowledge may be drastically and rapidly
reduced without actors being aware of the situation. Likewise, if people working in a
group setting are too involved in their own individual activity, they are no longer able
to overhear broadcasted communications and, as a consequence, group efficiency is
drastically reduced (e.g. [17] [19]).
Despite some analytical studies that have examined information flows in
organisational settings, there is no particular theory that can explain the advantages of
such an informal organisation. Our own previous works in the design of such
organisations, e.g. Air Traffic Control, emergency medical control room, and space
control centres, used a multi agent simulation approach in order to empirically study
the advantages and disadvantages of complex emergent organisations [24]. From
these studies we were able to use these multi-agent models to design complex control
rooms [23]. Whilst these works employed the general properties of complex systems
theory, such as emergence and self-organisation, the underlying processes that led to
advantageous proximal cooperation remained unclear. In this paper we suggest the
percolation model as a good candidate for increasing our understanding of emergent
processes in complex organisations.
Therefore the aim of this paper is:
• To confirm that the percolation mechanism is a relevant model for explaining
the emergence of mutual knowledge in a group of people interacting locally in
a non coordinated way,
• To understand what are the most important factors that affect the percolation of
mutual knowledge (e.g. effect of population density, overhearing),
In section 2, we briefly introduce percolation as well as the two types of
percolation processes (site and link percolation). A more in-depth discussion on
mutual knowledge, as being generated as part of a complex socio-technical system, is
given in section 3. Section 4 describes the methodological approach that couples both
agent-based simulation and network analysis. The experiments performed and their
results are given in section 5. Here we specifically focus on how changes at the
microscopic level in our agent based simulator affect percolation and robustness.
2
Finally, section 6 draws some conclusions and discusses the implications of the work.
2. Percolation and its modelling process
2.1. Percolation and its types
The percolation mechanism was introduced by Broadbent and Hammersley and is
a long studied phenomenon in the domains of physics and mathematics [3]. A classic
example in physics is the study of porous material. Here the porosity of the material is
modelled in terms of a probability that an open space exists between two sites of
material. Above a certain probability threshold the material will be porous;
conversely, below the threshold, it will be impermeable. Percolation is a non-linear
process that can thus explain the emergence of connected clusters; clusters of empty
space in the case of porous material. What is interesting in this concept is its non
linear characteristic with singularities which explains how a process, above certain
threshold, can drastically change its characteristics following universal rules.
Furthermore these rules are independent of the domain of the process. Outside of
mathematics, physics and materials science, the percolation mechanism has been
observed in domains such as economics, ecology, biology, computer networks,
epidemiology, and social science [6]. In social science and economics this concept
has been used to better understand the interplay between local and global actors [30].
In the marketing domain, a social percolation model has been developed [28] where
agents represent consumers situated in a social network with the aim of understanding
potential markets for products.
The percolation problem considers a network in which each node is occupied with
probability p and links are present only between occupied nodes. As the probability p
increases, connected components, called clusters, emerge. Thus the percolation
problem studies the properties of the clusters, and in particular their sizes, as a
function of the occupation probability p. It is indeed intuitively clear that if p is small,
only small clusters can be formed, while a large p will eventually engulf the structure
of the original network. As a practical example to see the difference between site and
link percolation, consider a network of people (agents) that communicate in the
mountains through optical devices (such as a torch or laser). Here the agents will be
represented as nodes and the links will be the optical signals. If the weather is
perfectly clear and all agents are efficient, we will have a probability p of 100% of
communication and the network will be very efficient in propagating the information.
However, if the agents are totally efficient but with a low level of communication due
to bad weather, the information will not propagate very far (we have link
percolation). Alternatively, if the weather is perfectly fine, but the agents are
inefficient (e.g. not very attentive) then the lack of efficiency in the network is related
to the site's behaviour (not the links); we are then in a situation of site percolation.
Finally, if both the agents and the weather are not efficient, we will have a situation of
mixed percolation. From this simple example, we can better understand the difference
between site and link percolation [27].
2.2. Modelling site and link percolation processes
Let us first consider what happens in site percolation for a two-dimensional lattice
such as the one depicted in Figure 1. In Fig 1.a, the probability that a site is active is
0.4. We can see that some sites or nodes are isolated, whilst others are contiguous,
3
forming clusters1 yet there is no cluster that fills the entire network. In Fig 1b, we
reach the critical probability (p=0.6) when the network is almost filled by connected
sites (if a porous material were being modelled, we would say that we have reached
the probability threshold that makes the material porous). In Fig 1.c, the entire
network is connected.
Figure 1: An example of site percolation for a square network. Each site is occupied,
shown as black squares, with a probability p. Fig 1.a: when p is small (p=0.4), only
little clusters of occupied sites are formed. Fig 1.b: At percolation probability (p=
0.6), a large cluster starts to emerge. Fig 1.c: Above the percolation threshold, the
cluster invades the entire space. Note that if the dimension of the network is infinite,
the cluster will also be infinite above the threshold [22]
Figure 2 shows the same phenomena but for a link network. In this type of network,
only the links between sites are considered.
Figure 2: The same percolation phenomena in a link network [22]
To study percolation theoretically, we usually consider infinite networks that exhibit
interesting properties such as:
● There is a critical p (denoted by pc) below which the probability of having an
●
infinite cluster is always 0 and above which the probability is always 1
In networks of more than two dimensions, only simulation can approximate
the percolation threshold pc, it is not possible to calculate it
● The model has a singularity at the critical point p = pc believed to be of power-
law type
If P(p) is the probability of percolation, we could say that:
1The concept of clustering refers to the tendency observed in many natural networks to form cliques in the
neighborhood of any given link.
2 We adopt the usual network terminology where an 'edge' refers to a link in a non-directed network (i.e. the
meeting network) and where an 'arc' refers to a link in a directed network (i.e. the knowledge network). We will
4
P(p) =0 if p< pc (all clusters have a finite size)
P(p) >0 if p> pc (a giant cluster appears)
P(p) is a usually an increasing function with an exception point at the
percolation threshold
Figure 3: Probability P(p) for a node to belong to the infinite percolating cluster as a
function of the occupation probability p. We can see a singularity at the percolation
threshold Pc [15].
Above the percolation point, the system exhibits invariant behaviour. The exact
value of the critical exponents does not depend on the fine details of the percolation
model. In general, they just depend on the system's dimensionality and symmetries of
the order parameter. Thus, while the exact value of Pc depends on the lattice
geometry, the critical exponents do not. This universality also means that for the same
dimension independent of the type of the lattice or type of percolation (e.g., link or
site) the fractal dimension of the clusters at Pc is the same. If we call PG, the
probability for a node to belong to the infinite percolating cluster, we have [2, p. 125]:
This means that PG follows a power law scaling from close to Pc. The scaling
law expresses the insensibility of the characteristic quantities in a percolation process
to the local and microscopic details around the critical value Pc.
The study of the percolation transition as a function of the connectivity properties of
generalised random graphs finds a convenient formulation in the generating functions
technique [7], [20].
Barrat & al. [2] report conditions for a giant cluster to arise in graphs that have a local
tree structure with no cycles. They consider an uncorrelated network with degree
distribution P(k) and compute the probability q that a randomly chosen edge leads to a
vertex of degree k. This probability is written as the average over all possible degrees
k of the products of two probabilities: (i) the probability that the randomly picked
edge leads to a vertex of degree k and (ii) the probability that none of the remaining
edges lead to a vertex connected to a giant cluster. This leads to the condition for the
heterogeneity parameter (K) of the network [2] :
5
For a directed graph, the percolation phenomenon is also shown by the emergence of
a large cluster in a network. Such a network is a giant cluster if:
These are the criteria used to assess percolation in our generated dynamic
networks.
3. Mutual knowledge and complex socio-technical systems
Mutual knowledge (MK) is knowledge that communicating parties share in common
and that they know they share [16]. Mutual knowledge is also broadly referred to as
'common ground' and is an integral part in coordinating actions and collaborative
decision-making [8, 9, 10].
Whilst MK is a very important concept it is also an ambiguous notion because it
depends on the observer's capability and the richness of the media that people use to
communicate. If we consider people interacting through email messages, the MK will
depend, not only on the explicit messages exchanged, but also on the inferences that
each agent may make when they receive the message.
In more interactive situations like face to face communications, MK will not only
depend on verbal exchanges but also on non verbal ones such as gesture and posture.
Physical constraints (such as the distance between agents) and the artefacts that
mediate the communication can also affect MK.
It is important to understand that MK, like all emergent processes, needs an observer
to identify it [1]. Thus the emergence of social structures in animal societies needs a
human-being in order for it to be described. The actors themselves cannot see the
global picture. In addition, agents are usually not conscious of the richness of their
communication processes and an external observer is needed to assess the propagation
of information as well as the structure of the information itself.
MK is therefore an abstract concept that may be far removed from real agents real
representation. Nevertheless, in some ways it represents the knowledge that is
accessible at a certain time (analogous to potential energy in physical systems).
Extensive analysis of work situations may lead to an acceptable understanding of the
parameters that determine MK. If we consider a group of people interacting verbally
in a room, we may first imagine that communication content, distance, noise,
background knowledge or the goals of each agent are good candidates for MK
parameters. In real situations the efficiency of proximal cooperation may depend on
very subtle parameters such as how to direct information to a specific group of people
without bothering others, or how to broadcast information dynamically to different
groups of people without interfering with their on-going cognitive processes.
In real situations we can also observe that group efficiency behaves in some situations
like a percolation process. Below a certain threshold of activity (or noise, workload,
etc.), the group 'efficiency' is 'boosted' by its proximal local interaction. Above this
6
threshold the group may experience a drastic loss of performance, usually without
understanding the reason for it. In our previous works this situation was often
observed in emergency control rooms. If the workload is acceptable, the room is not
too noisy and overhearing process is efficient. We can then observe the 'percolation'
phenomena: everybody is aware of everything and the network is efficient. However,
when the workload increases, the noise increases and the overhearing capabilities are
low. The same group of people may no longer be aware of other activities in the
group and the efficiency of the group is low [4].
In order to understand better this process we previously developed empirically based
simulators using a multi agent approach. The approaches varied from using a very
detailed cognitive model of interaction [11] to more analytic approaches taking into
account a limited number of parameters (such as the distance between actors, the
radius of communication, the type of knowledge, and the rate of information
exchange).
In one of these models, agents where moving randomly in their environment and each
time they met they exchanged all their knowledge about the group to which they had
initially belonged. MK was defined as the total amount of knowledge that was shared
by each agent at each instant. We observed that the emergence of MK varied in a non-
linear way with the size of overhearing capabilities (Fig. 4). However we were not
able to formally identify this process as a percolation phenomenon. The aim of the
next section is to explain our methodology in order to identify this process as a
percolation process.
Figure 4: Evolution of MK (Time to reach 50% of Knowledge) in relation to the
overhearing capability. M= Mobility of agents, N = number of agents
4. Methodological approach
4.1. Overview
7
The approach to the study of percolation was conducted in 4 phases, as shown in
figure 5.
Figure 5: Workflow of our percolation study approach.
The agent-based model characterises the behaviour of the agents in their
environment. The model was implemented in the Repast toolkit [26] and data from
each time step was generated and stored in a database. Then, the data is used to
construct a series of networks using Pajek [21], which is then analysed and visualized
in R, which is a software package for statistical computation and graphics (R software
[25]).
We generate a set of networks where each network represents one iteration of
the simulation. The number of nodes in each network is equal to the number of agents
in the simulation run, which is kept constant for each experiment.
The changes from iteration to iteration of the simulation are thus reflected in
set of generated networks. Hence, we can see the dynamics of the network over time
and specifically how the links between the nodes change. Graphs showing the
evolution of the networks are generated using R [25].
4.2. Agent-based model and simulator
In this section we describe how the percolation mechanism has been modelled using
an agent-based approach. Agent-based simulation is based on the idea that it is
possible to represent computationally the behaviour of entities interacting in a world
and that emergent phenomena can emerge as a result of these interactions. This
approach therefore offers us a powerful tool in investigating macroscopic behaviours
that result from interactions at the microscopic level. The model simulates the case
where populations of agents, belonging to different groups, interact within an
environment. When an agent meets another agent within its perception radius, they
exchange information concerning their groups.
We are interested in measuring the level of MK over time. Intuitively this social
phenomenon arises through the propagation of information within a population. The
evolution of two artificial social networks (a Meeting network and an Exchanged
Knowledge network) was studied to see if these networks exhibit percolation. As
mentioned in section 1, the percolation phenomenon is characterised by the
emergence of a large cluster in a network with the following property:
8
K = (<k2>/2<k>) > 1
where k is the number of mean links between the nodes.
If a network exhibits this condition at time t, then it is in a percolation phase.
We also test how local factors impact this phenomenon at the macroscopic level. The
factors taken into account are: the size of population; the individual properties of the
agents (such as their ability to overhear, their propensity to forget information, and
their mobility); environmental properties such as its dimension, and the nature of the
exchanged information, such as the frequency of group changes are also considered.
The model consists of N agents (minimum 103) moving randomly in an environment
represented as a grid. Each cell of the grid is the same size and may contain one or
several agents. Agents are initially randomly assigned membership of a specific
group, which may change over time. Each time an agent meets another agent within
its neighbourhood, it provides information concerning its own group as well as the
group names of agents that it knows. An agent A can therefore know the agents that
agent B knows without having met them. If there is conflicting information, e.g. agent
A believes that agent C belongs to group 1, whereas agent B believes that it belongs
to group 2, then the most recent information is used for updating. Each agent is
characterized by the following information:
ID: Each agent is assigned a number which uniquely identifies that agent
●
● Group membership: Each agent is randomly assigned to a group at the start of
the simulation. An agent may change group randomly over time; a simulator
parameter is considered for this and is set to false by default.
● List of agents met: each agent keeps a list of all of the agents that it has met.
Each element of the list is composed of three fields: the ID of the agent that
has been met, the group membership of the agent that has been met, the
number of times that the two agents have met. Initially the list contains only
one item (the ID and the group membership of the agent itself and the number
1 - signifying that the agent has met itself once). If we assume that we have 2
agents: A and B, each time A meets B there are two possible cases: A is
meeting B for the first time, in this case A adds a new element to its list
containing the ID and group of B, and the number 1. If A has already met B, it
will simply increment the frequency.
● List of known agents: each agent keeps a list of all the agents that it knows
either directly or indirectly. Again each element of the list is composed of
three fields: the ID and group of the agent, and a timestamp indicting when
this information was obtained. As with the list of agents met, this list initially
contains information only about the agent itself. When the agent meets another
agent, say agent B, it adds details of the agents that B knows to its own list and
updates any incorrect information, such the change of group of an agent. Thus
if there is an element in B's list that is not in A's list, it will be added to A's
list (without changing the time). Otherwise, if the element is already in A's
list, A checks to see if the group membership is the same in both lists and if it
is, A uses the most recent time. If the group is different, A checks the time in
both lists and uses the group corresponding to the most recent time.
In addition each agent has the following abilities:
● Move: agents move randomly in the grid according to a certain speed that was
specified at the start of the simulation.
9
● Meet: Agent A meets Agent B if they are within each other's perception
radius. This radius is the same for all agents and is specified at the start of the
simulation. The radius represents the 'overhearing range' of an agent.
● Talk (to agents they meet): Agent A talks to Agent B means that A gives a list
of all the agents and their groups that it knows to B. An agent can talk to
several agents that it meets at the same time.
● Listen (to agents they meet): Agent A listens to Agent B means that A
receives the list of all the agents and their groups from B.
● Forget: A forgets B means that A forgets what group B is a part of. Practically
this corresponds to B being deleted from A's list of agents that it knows. The
probability of forgetting is a variable set at the beginning of the simulation.
● Update incorrect information: since agents can change groups over time it is
likely that an agent will, at some stage, have incorrect information about
another agent's group. Therefore each time two agents meet, they compare
lists and update incorrect information.
Table 1 summarizes the parameters and values that were used during the experiments
Abbreviation
Name
Variables
Evolution functions
d
ng
N
m
oh
pforg
chgr
Dimension/density
Number of Groups
Number of Agents
Mobility
Overhearing
Probability of
forgetting
Change
group/changing
information
Speed of diffusion
(knowledge network)
Meeting frequency
(meeting network)
Values
10002, 50002,
100002
2
103, 104, 105
100, 300, 600, 1000
0, 1, 2, 3, 4
0.001, 0.1, 0.5, 0.7
True, False
Table 1: Parameters and values of the agent-based percolation simulator
10
Figure 6 below shows the simulator interface.
Figure 6: Agent based simulator interface at time t. The tab 'Parameters' in the
bottom-left of the figure allows the user to enter the values for the parameters. The
icons at the top of the figure allow the user to drive the simulation (e.g. initialize the
model, start or stop the simulator or see a step-by-step iteration of the simulation,
etc.). Once the simulation is complete, other buttons allow the user to send the results
to external software packages, such as Pajek, R or even Excel to analyse the structural
properties of the two networks generated by the simulation.
4.3 Networks
The interaction between agents in the environment is represented by two networks: a
meeting network and an exchanged knowledge network. The meeting network is a
weighted non-oriented graph where the nodes represent the agents. An edge between
two agents with a weight x indicates that one agent has met the other x times since the
start of the simulation. Conversely, the knowledge network is a directed graph where
nodes represent the agents and where the arc2 from agent Ai to Aj indicates that Ai
knows Aj in the sense that Ai has received, through previously meeting Aj at time t,
the group membership information about Aj. Note that because it is a directed graph,
whilst Ai may know Aj, Aj does not necessarily know Ai.
2 We adopt the usual network terminology where an 'edge' refers to a link in a non-directed network (i.e. the
meeting network) and where an 'arc' refers to a link in a directed network (i.e. the knowledge network). We will
use the term 'link' when we are referring to both networks.
11
We are interested in studying the impact of local factors and the structure of the
network on the percolation phenomenon. At the beginning of the simulation we set
the parameter values, e.g. the size of the population, the overhearing range of the
agents, etc. At the end of each simulation we obtain two sets of networks: a set
showing how the meeting network has evolved over time, and another one, showing
how the knowledge network has evolved.
5. Experiments and Results
5.1 Knowledge network: analysing the percolation
The first objective is to look for a percolation process in knowledge network where
the arcs between the sites indicate that agent Ai knows agent Aj (directed graph). The
percolation phenomenon is shown by the emergence of a large cluster in a network.
Such a network is a giant cluster if:
If Ki = input arcs (number of agents that know Ai) and Ko = output arcs (number of
agents that Aj knows)
with the percolation condition :
PsG (p) = the probability that a site belongs to the infinite cluster of the network.
12
Figure 7: Evolution of the K parameter over time for a knowledge network of size 103
If the network verifies this condition at time t, we will say that it is in a percolation
phase. Figure 7 shows the evolution of K over time with a knowledge network of size
103. The figure shows that K > 1 from t =700.
We will now use condition 2 to compute if our site network shows percolation
characteristics. The expectation is that above a certain value of Ps (probability that one
node is connected i.e. it is linked to at least one other node) most of the sites will be
connected and a giant cluster will emerge; the network will percolate.
In order to visualize site percolation, we draw PsG, which is the probability that a site
belongs to the infinite cluster of the network, in relation to Ps. Our aim is to find the
point t that shows the percolation threshold (Fig. 8, A).
In figure 8, the values of PsG are largely greater than zero below the percolation
threshold (shown by the red line) and the meeting network does not exhibit site
percolation phenomenon. Concerning the knowledge network, the values of PbG are
close to zero below the percolation threshold, showing that arc percolation exists.
As we can see, in knowledge networks, the percolation condition is not reached for
site percolation (K>1 but Ps >0). Conversely we can observe arc percolation (K=1,
PbG ≈ 0). Figure 8, B shows that PbG begins to be positive at Pb=0.0006 (critical
point). Thus, in the knowledge network if arc percolation is considered, PbG (i.e. the
probability that an arc between agents belongs to a giant arc network) meets both the
first and second conditions (Fig. 8).
13
Figure 8 (A): Probability that a site belongs to the infinite cluster of a network (PsG)
in relation to Ps (probability that a site is active all over the network). The red line
shows Ps when K=1 (the condition for the existence of a cluster that percolates). (B):
Probability that an arc belongs to the infinite cluster of a network (PbG) in relation to
Pb(probability that an arc is active all over the network). The insert in figure 8, B,
shows a magnified view at the percolation threshold. The size of the network = 103.
Thus knowledge networks exhibit a arc percolation but no site percolation. In the next
section we perform the same analysis with meeting networks.
5.2 Meeting networks: Analysing the percolation phenomenon
The same reasoning can be applied to meeting networks and the calculation of
percolation condition becomes:
Figure 9 shows the graph Pbs = f(Ps) for a meeting network of 103 nodes. Following
the same steps, we can see that the meeting network exhibits a mixed percolation
phenomenon (node and edge percolation). This result holds for all of the studied
meeting networks (independent of initial variables).
14
Figure 9 (A): Probability that a site belongs to the infinite cluster of the meeting
network (PsG) in relation to Ps (probability that a site is active). (B): Probability that
an edge belongs to the infinite cluster of meeting network (PbG) in relationship to Pb
(probability that an edge is active). The red line shows Ps for where K=1 which is the
condition for the existence of a cluster that percolates. Size of the network = 103.
5.3 Comparison between the knowledge network and the meeting network
In order to see the similarities and differences between the knowledge network and
the meeting network we can compare the respective graphs shown in figure 10.
15
Figure 10: Evolution of Mutual Knowledge (red curve) and meeting rate (black curve)
over time (logarithmic scale). The vertical red line shows the time of percolation (Tc).
Percentage on the y-axis is the percentage of MK relation to the maximal MK where
every agent knows everything about all of the other agents.
As we can see, below the percolation point Tc, the knowledge is not shared; agents
may exchange some information but we do not have any propagation of this
information. Above the threshold (Tc), the percentage of shared knowledge grows
much more rapidly than the agents that meet. This means that even before all agents
have met, every agent knows everything and this process has the properties of a
percolation. This can be explained by the fact that when two agents meet, at
maximum one edge is added to the meeting network (if the two agents have not met
before). Conversely, up to 2(n-1) arcs can be added to the knowledge network since
the two agents exchange their information.
The common points of the two networks are:
● The two curves have the same appearance (the knowledge level and the
meeting proportion are close to zero at the start); only a few agents have met
and information has not been propagated. From Tc (the percolation threshold)
the curves rapidly increase and become stable and equal to 1 signifying that all
the agents have met and all the information has been diffused throughout the
population.
● This observation confirms our previous results: the propagation of information
and the frequency of meeting are governed by a percolation process; the link
percolation threshold is the same in the two networks, i.e. 0.001 for 103
agents.
The differences are:
● The speed of information propagation is quicker than the meetings. The time
necessary to reach full mutual knowledge is a lot shorter than the time for all
the agents to meet. As mentioned previously, when two agents meet, a
maximum of one edge is added to the meeting network (if the two agents have
not met before). Conversely, up to 2(n-1) arcs can be added to the knowledge
network since the two agents exchange their information.
● The time necessary to reach the percolation threshold is shorter in the meeting
network, but we cannot currently explain this difference, however the time
taken for the two networks to be totally connected is negligible.
● The link percolation threshold is the same in the two networks (knowledge
and meeting networks), i.e. 0.001 for 1000 agents.
In the following section we analyse how percolation is sensitive to network
parameters.
5.4 Communities structure
In order to detect the structure of communities in both studied series of networks (the
meeting and the knowledge network) (size 103) we used Pajek visualization. Figure
11 shows the result before, during and after the percolation threshold.
16
Figure 11: The knowledge network (up) / meeting network (down) (A) at the
beginning of the simulation, (B) at the percolation time Tc and (C) after the
percolation time. (N = 103).
For each of the networks, at the beginning of the simulation, the network is composed
of small clusters (of which the majority are isolated nodes). When we reach the
percolation threshold a large cluster appears, and in advancing the simulation all small
clusters disappear and the large cluster envelops the entire network. These results
show the presence of a percolation phenomenon.
5.5 Effect of the number of agents
In order to study the effect of the number of agents (N), we performed simulations
with N= 103, 104 and 105, d = 5000, m = 1, overhearing = 0 and forgetting = 0.
Following table 2, we note that for the knowledge, results show that the value for the
percolation threshold pc vary with 1/N. Indeed it is exactly the probability at which
the phase transition leads to the emergence of a large cluster appearing in random
graphs, as shown by Erdos and Renyi (1959).
The percolation threshold (link or node) is reached more rapidly when N increases.
Table 2 : Percolation threshold values (Nb) and the proportion of active arcs (Pbc)
and time (Tc) in knowledge networks of sizes 103, 104 and 105.
The emergence time of a large cluster (tc) is shorter when N is larger and we reach the
percolation threshold more quickly when N increases.
5.6 Effect of density
17
We note that the denser the population, the higher the meeting frequency (figure 12
B) and thus information propagation is more rapid (figure 12 A).
Figure 12: Information propagation (A) and meeting frequency (B) for d =
10002,50002 and 100002. Percentage = percentage of MK in relation to the maximal
MK.
We can see that density does not influence the percolation threshold value in the
meeting network (node and edge percolation) or in the knowledge network (node and
arc percolation). However it does influence the time needed to reach the threshold;
this being shorter for a higher density environment (Table 3).
Table 3: Percolation threshold values of the proportion of links (Pbc) and active nodes
(Psc) and time (Tc) in knowledge networks (RC) and in meeting networks (RR) for
different values of d (d=10002, d=50002 and d=100002)
5.7 Effect of overhearing
In order to observe this effect we ran a series of simulations, each time changing the
extent of the overhearing (overhearing = 0, 1, 2, 3, 4) whilst keeping the other
variables stable (N = 103, d = 50002, m = 1, forgetting = 0).
Figure 13 shows the effect of agents overhearing on the emergence of mutual
knowledge. We note that when overhearing is higher, the propagation of information
is faster and the agents produce mutual knowledge more quickly.
18
Figure 13: Evolution of mutual knowledge as a function of overhearing. Percentage of
MK = percentage of MK in relation to the maximal MK (ef = overhearing
distance).
Table 4: Percolation threshold values of the proportion of links (Pbc), active nodes
(Psc) and time (Tc) in knowledge networks (RC) and meeting networks (RR) for
different values of overhearing (overhearing = 0, 1, 2, 3 and 4).
In both networks the percolation threshold is reached more rapidly when overhearing
is higher, but the threshold value is always the same.
Concerning the structural properties of the networks, overhearing has no influence on
the shape of the representative curves. However, there is a difference in the number of
iterations that it takes to produce the curve, this shows that mutual knowledge
emerges more quickly when overhearing is higher.
In the two following sections, we try to assess the robustness properties of mutual
knowledge processes. We have already seen that MK follows a percolation process
with universal percolation singularities, as characterized by very fast information
propagation and threshold values, etc. but we found worthwhile to investigate how
19
such process is robust to destructive processes such as probability to forget, or
unexpected information modifications.
Figure 14: Meeting frequency as a function of overhearing distance (ef) and time.
5.8 Effect of the probability to forget on information propagation
Studying the effect of forgetting allows us to analyse network robustness against the
random destruction of some of its links (which is a way to topologically simulate the
probability to forget). The study is restricted to the knowledge network since
forgetting concerns knowledge and not meeting. Figure 15 shows that as the
probability to forget rises, the level of knowledge is weaker and the agents no longer
produce mutual knowledge.
20
Figure 15: Evolution of mutual knowledge as a function of the forgetting probability
(oub). Overhearing= 3, Density= 5000x5000, Size of population= 1000.
To better assess this result we calculated the knowledge level obtained in the same
time interval according to the probability of forgetting (figure 16). The graph shows a
rapid decrease as the probability of forgetting increases.
21
Figure 16 - Incidence of forgetting (Probability to forget) on the emergence of mutual
knowledge (MK). Overhearing= 3, Density= 5000x5000, Size of population= 1000.
These results allow us to conclude that the network is weakly resistant to the removal
of some arcs (low forgetting probability). It does not support a frequent removal of
arcs and becomes disconnected; the information remains local and does not
propagate. The robustness of the network is therefore not very strong.
5.9 Effect of changing information
Agents' group memberships are updated during the simulation. We have measured
the gap between the knowledge of agents and that of the 'real-world' ('real-world' in
this case is the list of true information at each instant. The agents' knowledge is the
information held by agents that may disagree with the real world).
We first studied the effect of the frequency of changing information. As intuitively
expected, the gap closes when the frequency of changing information is lower (figure
17).
Figure 17: Influence of the frequency of changing information (cght) on the
perception of the environment. Delta MK = the gap between agent MK and real MK.
Overhearing= 3, Density= 5000x5000, Size of population= 1000.
We can see that even a very weak probability of changing information induces a
permanent gap between the real knowledge held by agents and the mutual knowledge.
The process is therefore very fragile to loss of agent memory.
We then investigate if this process of loosing memory depends on the size of the
population. From figure 18, we can see that increasing the size of the population
increases the gap between agents' knowledge and that of the 'real-world'. Thus, the
22
higher number of individuals that an agent knows, the harder it is to update incorrect
information.
Figure 18 - Influence of population size (N) on on the gap between the knowledge of
agents and that of the 'real-world' (Delta MK). Overhearing= 3, Density= 5000x5000,
Size of population= 1000.
With a low population size (N=1000), the loss of individual memory has almost no
effect on the quality of mutual knowledge; agents update their knowledge more
rapidly than they forget their knowledge.
Finally we analyse how the population density affects mutual knowledge. Figure 19
demonstrates the importance of the environment where knowledge evolves over time.
In situations where information must be regularly updated, the simulation shows that
one of the most influential variables is population density; the denser the population,
the faster the propagation of false information.
23
Figure 19: Gap between agent's mutual knowledge and the real-world mutual
knowledge (Delta MK) for different environmental dimensions. Overhearing= 3, Size
of population= 1000)
Contrary to what was expected, overhearing increases the gap between agents'
knowledge and the real-world (figure 20). However, this gap starts to decrease when
the network becomes fully connected, i.e. when all agents are known. So instead of
improving the capacity of agents to update and correct false information, overhearing
favours the spread of false information, but after a certain time interval, which is
shorter when overhearing is higher. This effect will be reversed by playing an
important role in updating false knowledge.
24
Figure 20: Gap between mutual knowledge and that of the real world (Delta MK). A
small overhearing abilities (ef= 1) affects the mutual knowledge process more than an
larger overhearing ability (ef=4).
The study of robustness properties of social networks is very important because, as
we frequently see in real systems, security issues of complex socio-technical networks
are often related to the emergence of mutual knowledge.
These results show that the robustness of the social network depends on several
factors, such as the size of the population, radius of overhearing, and the density of
the population.
We investigate this process through two mechanisms: the probability of forgetting
agent knowledge and modifying agents' internal knowledge.
We have seen that:
● The network is robust to forgetting as soon as the probability of forgetting
does not exceed 0.1
● Randomly changing agents' internal information, if the probability of change
is low (<0.01) and the population density is small, does not impact mutual
knowledge
● Contrarily to expectations, a low overhearing ability (ef= 1) has a more
important effect than a high overhearing ability (ef-4), on mutual knowledge.
This result was interpreted by assuming that a higher overhearing ability more
greatly affects the propagation of false information than repairing it through
agent's meeting.
Following these results, it appears that social networks that rely on MK to share their
knowledge are robust but only if false information is marginal. We intend by deleter
information, the processes of forgetting or random agent information transformation.
25
If not, the process of sharing information through meeting is more favourable to the
propagation of false information than to the stabilisation of a true MK.
5.10 Summary of results
The meeting networks show a mixed (nodes and link) percolation phenomenon,
whereas the knowledge network only shows the arc percolation phenomenon.
At the start of their evolution, networks are composed of small isolated clusters and at
the percolation threshold a large cluster appears that eventually envelops the entire
network.
The value for the link percolation threshold is the same in the two networks and does
not depend on the size of the network.
The probability at the percolation threshold for link percolation is inversely
proportional to the size of the network; this result has also been demonstrated by
Erdos and Renyi for random graphs.
In both networks, the degree distribution follows a power law; this is one of the
principal characteristics of scale-free networks.
6 Conclusions
The aim of this paper was to question the theoretical nature of the emergence of MK.
Emergent behaviour usually refers to spontaneous outcome due to the interaction
between many actors in critical situations. Such emergent behaviours allow people to
efficiently cooperate in complex socio technical systems such as Air Traffic Control
and regulation centres, etc. Multi-agent simulations have been extensively used both
to reproduce such emergent behaviours and as tools to design robust social networks.
Nevertheless, identifying MK as a percolation process has never been strictly proved.
Percolation can be seen as a specific and interesting type of emergence because it
exhibits very specific properties that are independent of the network's characteristics.
As examples, we have seen that at the percolation threshold, the size of the connected
network (also called giant network) grows as a power law; that giant connected
networks show fractal dimensions, and that the robustness of percolated networks can
be assessed.
From a social theory point of view, considering percolation as a specific cooperative
property in social networks is very interesting since it provides a structural framework
of the emergence of global properties that go beyond individual representations. Thus,
it gives us a theoretical framework to understand emergence in real social networks.
We therefore investigate how percolation can also be seen as a mark of optimization.
Optimization is often considered as a driving force for the evolution of biological and
social structures. We can therefore consider that the emergence of MK through
communication processes is an evolution towards very efficient collective structures.
Our comparison was based on the work of Barratt and his colleagues [2] where they
propose criteria for the appearance of a giant cluster in graphs that have a local tree
structure with no cycles. Results clearly show that the dynamics of the emergence of
MK conform to the critical percolation condition. Conditions on heterogeneity
parameters are respected and non-linear behaviour is characteristic of emergent
systems.
26
Concerning the robustness, a network's robustness was tested with the random
destruction of some of its arcs (simulating the process of forgetting) as well as with
changing node information (simulating the process of cheating). Results confirm what
has often been observed in real situations; that the emergence of robust MK is more
easily obtained with networks of small population with no deterrent phenomena such
as forgetting or false information propagation. The emergence of MK appears to be
very sensitive to deterrent processes. If we view this in light of communication
theory, we could say that the emergence of MK (or efficient cooperation) may occur
very rapidly in a cooperative network, but this MK can drop quickly if actors do not
follow the Grice's maxims of good cooperation (specifically the sincerity condition)
[13]. These results are coherent with our empirical field observations that showed that
the efficiency of team cooperation drastically falls as soon as overhearing is reduced
(e.g. due to a noisy working environment) or as soon as actors hide information [23].
References
[1] Barr, D.J. (2004). Establishing conventional communication systems: Is
common knowledge necessary? Cognitive Science 28, 937-962.
[2] Barrat A., Barthelemy M. and Vespignani A., (2008) 'Dynamical Processes on
Complex Networks', Cambridge University Press.
[3] Broadbent S.R. and Hammersley J.M., (1957) 'Percolation processes. I. Crystals
and Mazes', Proceedings of the Cambridge Philosophical Society 53, 629-641.
[4] Benchekroun H., 1992. Modelisation des mecanismes d'interlocution. These de
doctorat. CNAM, Paris.
[5] Bressolle, M.C., Pavard, B. & Leroux, M., 1995. The role of multimodal
communication in cooperation and intention recognition : the case of air traffic
control, Proceedings of the International Conference on Cooperative Multimodal
Communication, Theory and Applications, Eindhoven, The Netherlands, 24-26
May 1995.
[6] Bunde A. & Havlin, S. (Eds) (1999). Percolation and disordered systems: theory
and applications? 266: Physica A.
[7] Callaway, D. S., Newman, M. E. J., Strogatz, S. H. and Watts, D. J. (2000),
Network robustness and fragility: Percolation on random graphs, Phys. Rev.
Lett. 85, 5468-5471.
[8] Clark, H. H., and Carlson, T. (1982). Hearers and speech acts. Language, 2, 332-
393.
[9] Clark, H. H., and Marshall, C. R. (1981). Definite reference and mutual
knowledge. In A. K. Joshi, B. Webber, & I. A. Sag (Eds.), Elements of discourse
understanding (pp. 10-63). Cambridge, United Kingdom: Cambridge University
Press.
[10] Crampton, C. (2001). The Mutual Knowledge Problem and its Consequences for
Dispersed Collaboration, Organization Science, Vol. 12, No.3, pp. 346-371.
[11] Dugdale J., Pavard B., Pallamine N., Bellamine N. (2008) Simulation and
Emergency Management. Information Systems for Emergency Management. Ed.
B. Van de Walle, M. Turoff and S. Starr Roxanne Hiltz.Advances in
Management Information Systems . Vladimir Swass Series Editor.
[12] Erdos, P.; Renyi, A. (1959). On Random Graphs. I. Publicationes Mathematicae
6: 290-297.
[13] Grice, Paul (1975). Logic and conversation. In Cole, P.; Morgan, J. Syntax and
semantics. New York: Academic Press. pp. 41-58.
[14] Hutchins, E., 1990. The technology of team navigation. In Intellectual
27
Teamwork. Eds. J. Galegher., R.E. Kraut and C. Edigo, Hillsdale, N.J. :LEA.
[15] Kesten H., (1987) 'Percolation Theory and First-Passage Percolation', Annals of
Probability.Royal Statistical Society, (1954) 'Symposium on Monte Carlo
Methods', Journal of the Royal Statistical Society B 16, 23-75.
[16] Krauss, R. and S. Fussell. 1990. Mutual knowledge and communicative
effectiveness. In Intellectual teamwork: The social and technological bases of
cooperative work, ed. J. Galegher, R.E. Kraut,and C. Egido. Hillsdale, NJ:
Erlbaum.
[17] Lalouette, C., Pavard, B. (2008, October). Enhancing interorganizational
resilience by loose coupling concept and complexity paradigm. In Proceedings of
the third resilience engineering symposium.
[18] Mell, J. (1992). Study of verbal communication between pilot and airtraffic
controller in standard and non-standard situations. Doctoral thesis. Paris: Ecole
Nationale de l'Aviation Civile.
[19] Moffat, J. (2007). Modelling human decision-making in simulation models of
conflict. The International C2 Journal, 1(1), 31-60.
[20] Newman MEJ. 2003. The structure and function of complex networks. SIAM
Rev. 45:167-256.
Nantes.
[21] Pajek. http://pajek.imfm.si/doku.php
[22] Pajot S., (2001) Percolation et economie, these de doctorat de lUniversite de
[23] Pavard, B., Dugdale, J., Saoud, N. B.-B., Darcy, S. & Salembier, P (2007)
Design of Robust Sociotechnical Systems. In : Resilience Engineering.:
Remaining sensitive to the possibility of failure. Ashgate.
[24] Rateb, F., Pavard, B., Bellamine-Bensaoud, N., Merelo, J. J., Arenas, M.G.
(2005) Modelling malaria with multi-agent systems. International Journal of
Intelligent Information Technologies, 1(2), 17-27, April-June 2005. ISSN: 1548-
3657, eISSN: 1548-3665.
[25] R software. http://www.r-project.org/
[26] Repast Suite http://repast.sourceforge.net/
[27] Roussenq J. (1992), Percolation. in Encyclopaedia Universalis, Paris, vol. 17, pp.
838-840.
[28] Solomon S. and WeisbuchG., (1999) 'Social Percolation Models'.
[29] Zouai, F. (2011) Etude de la dynamique des réseaux complexes. Cas de la
percolation dans les reseaux sociaux. Mémoire de Master Recherche. Université
de la Manouba, Ecole nationale des Sciences de l'Informatique
[30] S. Cantono and S. Solomon, (2010) New J. Phys.
28
|
1306.5166 | 1 | 1306 | 2013-06-21T15:15:22 | A variant of the multi-agent rendezvous problem | [
"cs.MA",
"cs.CG",
"cs.DS",
"cs.RO",
"math.PR"
] | The classical multi-agent rendezvous problem asks for a deterministic algorithm by which $n$ points scattered in a plane can move about at constant speed and merge at a single point, assuming each point can use only the locations of the others it sees when making decisions and that the visibility graph as a whole is connected. In time complexity analyses of such algorithms, only the number of rounds of computation required are usually considered, not the amount of computation done per round. In this paper, we consider $\Omega(n^2 \log n)$ points distributed independently and uniformly at random in a disc of radius $n$ and, assuming each point can not only see but also, in principle, communicate with others within unit distance, seek a randomised merging algorithm which asymptotically almost surely (a.a.s.) runs in time O(n), in other words in time linear in the radius of the disc rather than in the number of points. Under a precise set of assumptions concerning the communication capabilities of neighboring points, we describe an algorithm which a.a.s. runs in time O(n) provided the number of points is $o(n^3)$. Several questions are posed for future work. | cs.MA | cs |
A VARIANT OF THE MULTI-AGENT RENDEZVOUS PROBLEM
PETER HEGARTY, ANDERS MARTINSSON, DMITRY ZHELEZOV
Abstract. The classical multi-agent rendezvous problem asks for a deterministic algorithm
by which n points scattered in a plane can move about at constant speed and merge at a
single point, assuming each point can use only the locations of the others it sees when making
decisions and that the visibility graph as a whole is connected. In time complexity analyses
of such algorithms, only the number of rounds of computation required are usually considered,
not the amount of computation done per round. In this paper, we consider Ω(n2 log n) points
distributed independently and uniformly at random in a disc of radius n and, assuming each
point can not only see but also, in principle, communicate with others within unit distance,
seek a randomised merging algorithm which asymptotically almost surely (a.a.s.) runs in time
O(n), in other words in time linear in the radius of the disc rather than in the number of points.
Under a precise set of assumptions concerning the communication capabilities of neighboring
points, we describe an algorithm which a.a.s. runs in time O(n) provided the number of points
is o(n3). Several questions are posed for future work.
Let g, h : N → R+ be any two functions. We will employ the following notations throughout,
0. Notation
all of which are quite standard:
(i) g(n) ∼ h(n) means that limn→∞
g(n)
h(n) = 1.
g(n)
h(n) ≤ 1.
(ii) g(n) . h(n) means that lim supn→∞
(iii) g(n) & h(n) means that h(n) . g(n).
(iv) g(n) = O(h(n)) means that lim supn→∞
(v) g(n) = Ω(h(n)) means that h(n) = O(g(n)).
(vi) g(n) = Θ(h(n)) means that both g(n) = O(h(n)) and h(n) = O(g(n)) hold.
(vii) g(n) = o(h(n)) means that limn→∞
g(n)
h(n) < ∞.
g(n)
h(n) = 0.
n=1 are two sequences of random variables. We write
Now suppose instead that (g(n))∞
g(n) ∼ h(n) if, for all ε1, ε2 > 0 and n sufficiently large,
n=1, (h(n))∞
(0.1)
(0.2)
P(cid:18)1 − ε1 <
g(n)
h(n)
< 1 + ε1(cid:19) > 1 − ε2.
P(cid:18) g(n)
h(n)
< 1 + ε1(cid:19) > 1 − ε2.
Similarly, we write g(n) . h(n) if, for all ε1, ε2 > 0 and n sufficiently large,
We will employ the standard phrase "asymptotically almost surely (a.a.s.)" when considering
a sequence of events (An)∞
n=1 such that P(An) → 1 as n → ∞. We use the notation n ≫ 0 to
denote "n sufficiently large".
Finally, if p is a point in R2 and ε > 0, we denote by Bε(p) the open ball of radius ε about
p.
One day while performing the mathematical equivalent of sitting at a bar, drinking beer
and philosophising - that is, browsing the day's listings on arxiv.org - one of the three authors
1. Introduction
Date: October 16, 2018.
2000 Mathematics Subject Classification. 68W20, 68M14, 68T40, 60D05.
1
was somehow reminded of a classic scene from the movie Terminator 2 [CW]. At one point
toward the end of the film, the evil Terminator becomes liquified after an explosion. Blobs of
liquid metal are scattered on the ground and he appears to have been "terminated". However,
as if guided by some mysterious superior intelligence, the blobs suddenly start moving toward
one another, eventually merging and reforming the intact Terminator, who then sets off on the
rampage again.
At the time, we were completely unaware of the extensive literature on the subject of ren-
dezvous problems. The movie scene got us thinking, and the results presented in this article
are the result of those thoughts. Only afterwards, when we circulated a draft paper to a col-
league who is an expert in computational geometry, were we alerted to the existing literature.
Having examined the latter, it seems to us that the kind of multi-agent rendezvous problem
we ended up considering has two important differences from what is studied in the literature.
Firstly, we imagined that agents which could see each other could also, in principle, commu-
nicate and thus exchange information. Our agents are thus more anthropomorphised than the
robots normally imagined. Secondly, we took a "probabilist's approach", and sought randomised
algorithms which would work well, for generic configurations of agents, and in the limit as both
the number and the spatial distribution of agents tended to infinity. These facts probably render
our approach of less practical significance, though we will try to convince the reader that it is,
at the very least, still an interesting thought-experiment and involves some elegant mathematics.
Rendezvous problems (RPs) originated in work of Steven Alpern, who considered the prob-
lem faced by two people who are placed randomly in a known, bounded search region in R2
and move about at unit speed to find each other in the least expected time. Alpern originally
imagined that there were a finite number of possible meeting points, but later formalised the
continuous version of the problem, and even generalised it to other compact metric spaces [A].
A variety of models have been proposed in the case of an arbitrary number of agents, which are
then usually thought of as autonomous robots, rather than fully-fledged humans. An elegant,
and classical model, makes the following assumptions:
RP-1: Each robot is idealised as a mobile point in R2. Rendezvous means the merging of
all the robots at a single point. The robots are assumed to be identical.
RP-2: The robots can move at constant speed.
RP-3: There is some fixed constant ε > 0 such that, at all times, each robot can only see those
others which are within a distance ε of itself. For a given configuration of point robots, this
gives rise to a geometric graph G whose vertices are the points, and where an edge is placed
between any two points at distance at most ε from each other. Usually G is called the visibility
graph corresponding to the point configuration. Any two points connected by an edge we shall
refer to as neighbors.
RP-4: The initial configuration is such that the visibility graph is connected. However, the
robots have no a priori knowledge that they all lie within some specific region of R2. This is in
contrast to Alpern's original formulation, where the search region was known1.
RP-5: Robots do not communicate explicitly with each other, even when they are neighbors.
Each robot is equipped with a sensor which allows it to determine the exact locations of all its
neighbors at any instant. This is the only input it has from its neighbors when deciding how to
1Note that, given RP-3, any finite number of points can almost surely merge, even if the visibility graph G
consists entirely of isolated vertices and there is no a priori knowledge of a search region. For example, the points
could perform independent Brownian motions, and since, as is well-known (see [Du] for example), the trajectory
of a Brownian motion in R2 is almost surely dense, any finite number of points will most surely be able to merge
in this manner. However, as is also well-known, the expected running time for this procedure will always be
infinite, even in the case of two points whose initial separation is any positive constant greater than the visibility
range ε. Also note that the procedure will not work at all in higher dimensions, as Brownian motions are then
no longer almost surely dense.
2
move.
Under these assumptions, one seeks a deterministic algorithm which guarantees merging. It
is natural to seek an algorithm where the actions of the robots are synchronised. What this
means is that the algorithm should consist of rounds.
In each round, every robot notes the
locations of its neighbors, performs some computation and determines a point to which it then
moves, in a straight line, from its current location. All the robots execute these tasks before
the next round begins.
The classical solution to this problem was provided by Ando, Suzuki and Yamashita [ASY],
see also [AOSY], whose algorithm utilises a well-known concept in computational geometry. Let
k be a positive integer and let S be a set of k points in the plane, say S = {p1, . . . , pk}, where
pi = (xi,yi). For each point p ∈ R2, let
(1.1)
p − pi,
f (p) := max
1≤i≤k
where · denotes Euclidean distance. There is a unique point pS at which the function
f attains its minimum value, namely the center of the unique closed disc of minimum radius
which includes all the points of S. The point pS is called the center of the smallest enclosing
circle (CSEC), with respect to the points in S. In each round of the ASY-algorithm, a robot
moves in the direction of the CSEC of the points in its viewing range (including itself), but
not necessarily to exactly the CSEC because one must ensure that the visibility graph remains
connected. Indeed, the main technical challenge for the algorithm is to specify how to satisfy
this constraint. Note that, in the last round before merging, all the robots must be visible to
one another and then they all compute the same CSEC and move to it. By definition of the
CSEC, and given RP-2, this is the optimal solution to the multi-agent Rendezvous Problem
when there is global visibility.
We found a number of papers which investigated the time complexity of the ASY and similar
rendezvous algorithms, see for example [MBCF] and [De]. Two things struck us about these
analyses. Firstly, the complexity is defined in terms of the number of rounds. The time taken to
execute computations within each round is not considered. Secondly, the analyses tend to focus
on worst-case scenarios, in other words, the primary goal is to obtain an upper bound for the
number of rounds which is valid for any connected visibility graph. Hence the bounds tend to be
expressed in terms of the number of robots. A remark on page 2221 of [MBCF] hints that, if one
begins by assuming that the robots are confined to some fixed bounded region in R2, then for
"generic" initial configurations, the size and shape of the region may be what really determine
the time complexity rather than the number of points. This viewpoint means abandoning RP-4,
of course, though note that it is less restrictive than in Alpern's original formulation, since one
is not assuming that the initial search region is known to the agents, only that such a bounded
region exists. We have not seen any paper which follows up on this idea, however.
As already stated, we began thinking about Rendezvous-type problems well before we became
aware of the existing literature on the subject2. With hindsight, the point of view we took
addresses the two issues raised above about existing time-complexity analyses. However, to
do so and yet obtain the result we were looking for, we had to abandon RP-5, and allow our
robots certain abilities to communicate with their neighbors. Furthermore, we had to foucs on
asymptotic results and randomised rather than fully deterministic algorithms. We now describe
our viewpoint.
The simplest kind of closed, bounded region in R2 is a disc, so suppose the points are initially
confined to a disc of radius n. The disc radius is our primary parameter, and we will be
interested in asymptotic results, hence in letting n → ∞. We are also interested in "generic"
configurations, so we imagine that each point is placed uniformly at random in the disc, and
2Indeed, we didn't even know the term "rendezvous problem", and had instead invented our own term termi-
nator problem, in acknowledgement of our source of inspiration.
3
independently of all other points. It is well-known - see [P] for example - that if the number N
of points satisfies N = Ω(n2 log n), then the visibility graph will a.a.s. be connected, whereas
if N = o(n2 log n), then it will a.a.s. be disconnected. So we suppose the former holds. If we
run the ASY-algorithm, then in the last step each robot will have to compute the CSEC for
some set of N points in the plane. Trivially, any procedure for doing this will require time
Ω(N ) to execute. Hence, if we include this time in our analysis, we have a trivial lower bound
of Ω(n2 log n) for the time to rendezvous. Taking all the rounds of the procedure into account
will make things even worse, though perhaps not significantly since
(a) during earlier rounds, each robot will have far fewer than N neighbors, hence most of the
computations are done only towards the end. Since the average density of points is Ω(log n)
initially, and a robot moves distance O(1) per round, what we can be certain of is that the extra
contribution to the running time coming from earlier rounds is Ω(n log n).
(b) it is a classical result that computation of a CSEC can indeed be solved in time linear
in the number of points. Megiddo [M] was the first to describe a deterministic linear-time
algorithm. Later, Welzl [W] developed an alternative procedure which is considered the state-
of-the-art solution. His algorithm is probabilistic, with expected linear running time, but it is
much simpler to describe and to implement that Megiddo's.
To repeat, given that the initial visibility graph G is a generic, connected geometric graph
inside a disc of radius n, the actual running time of the ASY-algorithm under conditions RP 1-5
above is Ω(n2 log n). On the other hand, if the points somehow "knew" where to rendezvous,
then RP-2 means they could do so in time O(n). The goal of our study was to find a probabilistic
algorithm which a.a.s. led to rendezvous in time O(n). It seems clear, though we have not proven
it rigorously (see Section 3), that this is impossible under RP-5. So we are forced to allow some
kinds of communication between points, which is perhaps a more radical departure from the
classical multi-agent Rendezvous Problem than some a priori assumption about the geometry
of the point configuration. Part of the goal then becomes to find an algorithm which contains
as simple and as few as possible assumptions about how neighboring points can communicate.
This is admittedly a subjective requirement, but as long as we can make all assumptions precise
we will at least have a well-defined mathematical problem. There is also a conceptual problem
with, say, the ASY-algorithm as n → ∞, namely that each robot would need unlimited capacity
to store the results of the computation of a CSEC. We will thus also have a preference for
algorithms which overcome this problem.
Section 2 is the heart of the paper. Having formalised our assumptions about how points
can communicate we will prove our main result, Theorem 2.4. Informally, it asserts that there
is a randomised rendezvous algorithm which a.a.s.
runs in time O(n) provided the initial
configuration of points is not too dense. Significantly, our algorithm consists of two main steps,
the first and more difficult of which involves the choice of a leader amongst the points. Indeed,
it may make more sense to think of our algorithm as being for this purpose, rather than for
the purpose of rendezvous. We will have a lot more to say about this matter in Section 3,
which contains a critical analysis of our alternative assumptions, plus suggestions for further
developing the ideas of this paper. A sceptical reader may choose to only skim through Section
2, ignoring detailed proofs, before reading Section 3.
2. Rendezvous in a disc with local communication
We start with two lemmas which will be used in the proof of the main theorem below.
Lemma 2.1. Let f : N → N be an increasing function. For each n ∈ N suppose f (n) points are
placed uniformly and independently in the interior of a disc D in R2 of radius n. Let G = Gn
be the geometric graph whose vertices are these points, and with an edge between any two points
the Euclidean distance between whom is at most one. Set λn := f (n)
πn2 , the average density of
points in the disc. There is an absolute constant C1 > 0 such that, if λn > C1 log n, then a.a.s.
(i) the degree of every vertex in G lies between π
(ii) the graph diameter of G is at most 6n.
3 λn and 2πλn,
4
Proof. There is a rich literature on random geometric graphs - see [P] - and the above state-
ments certainly follow from well-known (stronger) ones. We sketch proofs here for the sake of
completeness. Choose n and let p be any point in D. The number of vertices of Gn which are at
distance less than one from p is distributed as Bin(f (n), aλn), where a is the area of B1(p) ∩ D.
Hence a ∈(cid:0) π
ability that the degree of any particular vertex of Gn lies outside the interval (cid:0) π
2 , π(cid:3), depending on how close p is to the boundary of D. It follows that the prob-
3 λn, 2λn(cid:3) is at
most e−C2λn, for some absolute constant C2 > 0. Thus (i) follows by a simple union bound.
Indeed, note that (i) holds if we replace the factors 1
2 and
greater than 1 respectively.
3 and 2 by any constants less than 1
To prove (ii), consider two vertices p and q in Gn. No matter where they are located in D,
we can connect them by a piecewise rectangular tube of width 1
3 and of length at most 2n (see
Figure 1). We can divide this tube into square sections of side-length 1
3 and observe that any
two vertices located in adjacent sections are joined by an edge in Gn. Hence, if no section of
the tube is empty, then the graph distance from p to q is at most 6n. The probability that any
section is completely empty of vertices is at most e−C3λn for some absolute constant C3 > 0.
The claim of part (ii) follows by some simple union bounds.
p
q
Figure 1. A tube connecting two points in the disc.
Lemma 2.2. Let n ∈ N and let k = k(n) be a positive integer such that k → ∞ as n → ∞.
Suppose n binary strings s1, . . . , sn, each of length k, are generated independently and uniformly
at random. Consider a fixed string si. Suppose it compares itself with each other string sj, j 6= i,
by reading each string bit-by-bit from left to right until it encounters a bit where the string differs
from itself (if sj = si then the whole of sj is read). Let Xi be the total number of bits read by
si and let X = maxi Xi. Then X . 2n.
Proof. Xi =Pj6=i Xij, where Xij is the number of bits of sj that are read by si. We have
k
k
Xt=1
Xt=1
5
(2.1)
E[Xij] =
t · 2−t ∼ 2,
(2.2)
E[X 2
ij] =
t2 · 2−t ∼ 6.
It
Hence E[Xi] ∼ 2n and since, for a fixed i, the Xij are independent, Var(Xi) ∼ 2n also.
follows from the Central Limit Theorem that, for any ε > 0, there exists cε > 0 such that
P(Xi > (2 + ε)n) < e−cεn. The lemma then follows from a union bound.
In order to be able to prove a rigorous mathematical result, Theorem 2.4 below, we need
to specify precisely our assumptions about the capabilities of the points that are tasked with
merging. A reader may variously complain that the following list of 11 assumptions is either
too long, too ad hoc or dubious from the point of view of physics. We accept such criticisms as
valid, but nevertheless think the thought experiment we are engaged in is worth pursuing, for
two reasons. On the one hand, we think our result is mathematically interesting since, as will
become clear from the proof of Theorem 2.4, our algorithm very much relies on the geometry
of the point configuration. On the other hand, none of our 11 assumptions seems completely
ridiculous from an anthropomorphic standpoint, that is, if we imagine our "points" as being
closer to intelligent human agents rather than much more primitive robots. The fact that our
algorithm involves choosing a leader gives greater justification for this anthropomorphic view-
point. One suggestion to a potential reader is to skip the list of assumptions and go directly to
Theorem 2.4 and its proof, returning to the list only if one encounters anything in our algorithm
that seems completely implausible. Otherwise, we shall return to this discussion in Section 3.
Henceforth, we assume a fixed choice of length and time units. All implicit constant factors in
the following list, and indeed throughout the remainder of Section 2, are positive and absolute.
We consider the first five assumptions to be minimal requirements if we wish to adhere to the
spirit of the classical multi-agent RP, as expressed by RP 1-5, but allow local communication
in principle.
Assumption 1: Each point can move at speed at most one. That this constant is the same
for all points corresponds, intuitively, to the assumption that all the points are identical, and in
particular have the same "mass". Consequently, we could further assume that if, at some stage
of the merging process, two points became stuck together, then this new "heavier point" could
subsequently only move at some maximum speed less than one. However, since in our algorithm
a leader is first chosen and then all the other points move to it, this last issue is not relevant.
See Questions 3.3 and 3.8 below, however.
Assumption 2: Each point can only directly communicate (by transmission or reception of
electromagnetic or chemical signals, say, though the precise details don't matter) with other
points that are within a distance one of itself. As in Section 1, two points which can commu-
nicate directly will be called neighbors. Henceforth, what was referred to in Section 1 as the
visibility graph will now be called the communication graph.
Assumption 3: All signals travel at a fixed speed, which we can think of as being determined
by the laws of physics (and maybe chemistry).
Assumption 4: If a point broadcasts a signal, then any neighbor which receives it can locate, in
time O(1), the exact point source of the broadcast.
Assumption 5: Each point is immortal
in the sense that it can carry on the various activi-
ties desired of it at the same constant rate indefinitely. Note that this assumption is necessary
to even make sense of any asymptotic result in which the diameter of the initial configuration
tends to infinity. Here it is natural to think of the points as representing biological agents, who
can "eat" to replenish their energy stores.
Our next two assumptions regard the types of signals that a point can broadcast and how
they are processed. We assume there are two types of signals.
6
Assumption 6: Each point can broadcast simoultaneously to all its neighbors. There are a
bounded, if perhaps large, number of such signals, which we will think of as being different
colors. Any color signal can be generated, turned off or recognised in time O(1). In addition,
we assume that every point can both isolate and filter received color signals. By isolating we
mean that, if a point is receiving signals in the same color from multiple sources, then it can
identify and locate individual sources at a rate of Ω(1) sites per time unit. By filtering we mean
that, if some set C of colors are presently being broadcast amongst a point's neighbors, but it
is only interested in some subset C′ of colors, then in time O(1) it can "put on an appropriate
pair of goggles" and henceforth scan only for colors in C′, and not be disturbed in any way by
interference from colors in C\C′.
Assumption 7: Each point can also generate single bits and send them to individual neigh-
bors. A bit can be generated, an individual neighbor identified, and the bit sent to the neighbor
all in time O(1). The receiving point can process the bit and (see Assumption 4) identify its
exact source in time O(1).
Remark 2.3. Assumption 6, together with the last sentence of Assumption 11 below, is a
powerful tool. To get a feeling for this, suppose all the points are initially inside an unspecified
disc of diameter one, so we have global visibility but no universal point of reference. Then, for
N points, a.a.s. they could rendezvous in time O(log N ). For example, suppose each point is red
by default. Each point can start generating random bits, and turn blue as soon as it generates a
zero. If, at some step, all the remaining red points turn blue, then these "last men standing" go
back to red and try again. It is clear that, almost surely, within O(log N ) rounds there will be
a time at which exactly one point is red. This point then becomes the leader. Since it is visible
to all other points, they can all merge at its location. By comparison, if we only assumed RP
1-5, then the computation of a CSEC would require time Ω(N ), and that is only if we ignore
the problem of storage capacity.
Finally, we make explicit four additional assumptions which our algorithm will exploit. They
concern the abilities of the points to perform certain tasks or manouevers. Note that Assump-
tion 9 will only be used to deal with the fact that each point has a finite storage capacity - see
Remark 2.7 below.
Assumption 8: If one point receives a color signal from another and then the sender, after
waiting time O(1) so that the receiver locates it, starts to move while still transmitting the
color, then the receiver can track its movements in real time, and hence follow it if necessary in
such a way that the vector separation between the two points remains constant. If the sender
remains stationary, then the receiver can move toward it in a straight line at speed one.
Assumption 9: Given three points p1, p2, p3 which are pairwise mutual neighbors, p1 can
point out p3 to p2 and p2 can process this information, all within time O(1). We can imagine
the actual "pointing" being done by, for example, p1 shining a laser at p3. A more "low-tech"
solution would be for p1 to walk toward p3, do a little dance around it and then walk back to
where it came from. This would work given Assumption 8, and also assuming each point can
leave a beacon (which transmits a color reserved for beacons), so that it can return to its exact
starting point after a walk within a radius of one.
Assumption 10: Given any fixed ε ∈ (0, 1], a point can sweep its ε-neighborhood in such a
way that it identifies the individual points in it at a rate of Ω(1) points per time unit and does
not miss any points. In fact, our algorithm will only require this capacity for some finite number
of predetermined values of ε, namely ε ∈n 1
10C15
will appear in the proof of Theorem 2.4. Note that, what we're assuming here is more than
just the ability to determine, in time O(1), whether an identified neighbor is within a distance
, 1
2 , 1o, where C15 is an absolute constant that
7
ε or not. Rather, we are assuming that all neighbors at distance greater than ε can be blotted
out as the scan is performed, perhaps by tuning a receiver so that it blocks out signals which
are weaker than a certain threshold. We're also assuming that a point can physically perform
a sweep, for example by rotating on an axis.
Assumption 11: Following on from the previous assumption, given a point p, a value of ε
and any ninety-degree sector of Bε(p), the point can determine in time O(1) whether that sec-
tor is empty of neighbors or not. Also, given a color, it can determine in time O(1) whether or
not anyone inside Bε(p) is currently broadcasting that color.
This completes our list of asssumptions. One thing that might strike the reader as surprising
is that we do not require the points to possess clocks, for example for the purpose of synchro-
nising their actions. In fact, as we shall see, our algorithm only requires a very rudimentary
synchronisation which can be achieved without perfect clocks, given the above assumptions. We
are ready to state our main result.
Theorem 2.4. There is a randomised algorithm A for which the following holds: There are
absolute positive constants C4, C5 such that, if f : N → N is a function satisfying C4n2 log n <
f (n) = o(n3) then, under Assumptions 1-11 above, if f (n) points are placed uniformly and
independently at random in the interior of a closed disc D = Dn of radius n in R2 and proceed
to execute the algorithm A, they will a.a.s. merge in time at most C5n.
Proof. Throughout the proof there will appear a sequence of absolute, positive constants which
will be denoted by Ci, for i = 6,7, . . . , 16. Some of these constants will be related to one another,
as well as to the constants C4, C5 appearing in the statement of the theorem. We will not bother
with making these dependencies explicit, however.
As in Lemma 2.1 we will denote the average point density by λn, i.e.: λn = f (n)
πn2 . We will
denote the boundary of the disc D by δD.
Our algorithm A will consist of two main steps:
Step 1: Choose a leader.
Step 2: Everyone move to the leader's location.
More precisely, the idea is that, in Step 1, the f (n) points should perform a sequence of
operations at the end of which, a.a.s. exactly one point will be in possession of information
which identifies it as "leader", whereas every other point will possess information which allows
them to rule themselves out as leader. In fact, Step 2 can begin once one point believes itself
to be the leader, the crucial thing being that a.a.s. no other point will subsequently reach the
same conclusion and muddy the waters. Step 1 is the trickier part of the algorithm and we will
go into the details below.
In Step 2, the leader signals its identity to all its neighbors by broadcasting the color red,
this being the color reserved specifically for the signal "I am the leader" (see Assumption 6).
Once a point receives a red signal and identifies its source, it broadcasts red in turn to all its
neighbors and then moves towards the location from which it received the signal. If it received
a red signal from multiple sources, it chooses one of these at random and follows it. Our various
assumptions (in particular, Assumption 8) guarantee that Step 2 will result in all f (n) points
merging at the leader's location in time O(n), if and only if the original red broadcast reaches
all f (n) points in time O(n). It follows from Lemma 2.1(ii) that this will a.a.s. be the case,
provided f (n) > C6n2 log n for a sufficiently large C6. Note, in particular, that Step 2 works for
arbitrarily dense configurations of points, the upper bound on f (n) in Theorem 2.4 is not needed.
It thus remains to describe the implementation of Step 1. There are two crucial ingredients
and both rely on the upper bound f (n) = o(n3).
8
The first ingredient is that, if f (n) = o(n3) then, by Lemma 2.1(i), a.a.s. every point has
o(n) neighbors. Thus, by Assumption 10, every point can count its neighbors in time o(n).
There is one subtlety here, however. Recall from Section 1 that we prefer to think of each point
as possessing a finite memory capacity. Hence, for n ≫ 0, it will in general not be able to
store internally the result of a count of its neighbors. We'll present our solution to this problem
further down, but it makes our algorithm a good deal more clunky than it would be otherwise.
See also Remark 2.7.
The second crucial ingredient is that the "boundary" of the communication graph will a.a.s.
contain O(n) points. We need a precise definition:
Definition 2.5. Let G be the communication graph of a collection of points distributed in a
disc. Let v be one of these points.
If v has a set of neighbors v1, . . . , vk such that, setting
vk+1 := v1,
(i) vi and vi+1 are also neighbors, for each i = 1, . . . , k,
(ii) reading clockwise, v1, . . . , vk are the vertices of a simple polygon in the plane which
encloses v,
then v is said to be a non-boundary point of G. If no such set of neighbors of v exists, we say
it is a boundary point of G. We denote by δG the set of boundary points of G.
Claim 2.6. Suppose f (n) satisfies the assumptions of Theorem 2.4. Then
(i) there are absolute positive constants C7, C8 such that, a.a.s., C7n ≤ δG . C8n.
(ii) a.a.s. every point in δG is at distance at most C9
from δD, where C9 is another
log n
λn
absolute constant.
Proof of Claim. These kinds of statements may also follow from "well-known facts" about
geometric graphs, but we choose to give complete proofs.
For a given n, let D1 denote the closed disc with the same center as D, but of radius n − 1
n .
Set D2 := D\D1 and, for each i = 1, 2, let δiG := δG ∩ Di.
Let p be a point of D, and let r denote its distance from δD. Let o denote the center of D
and let p1p2 be the chord through p which is perpendicular to the line through o and p (see
Figure 2 on page 10). It is easy to check that ppi ≥ 1
n and
n ≫ 0.
2 , for i = 1, 2, whenever r > 1
First consider points in D2. The total number of such points in the configuration is distributed
as Bin(cid:0)f (n), a
such points is o(n), and hence that δ2G = o(n) also.
πn2(cid:1), where a is the area of D2, hence a ∼ 2π. It follows that a.a.s. the number of
Next consider a point p in the configuration at distance r > 1
n from δD. If p is to lie in δG,
then at least one of the regions Ri, i = 1, . . . , 4, in Figure 3 on page 11 must be empty. By a
simple union bound, the probability of this is at most 4 · exp(−C10λn · max(cid:8)r, 1
C10 > 0. Part (ii) of the claim now follows. Note that this immediatetely implies in turn the
lower bound in part (i), since δG must be connected, as an induced subgraph of G. For the
upper bound in part (i), let us consider the random variable X which is the number of points
in the configuration at distance r > 1
n from δD for which at least one of the four regions in
Figure 3 is empty. By definition, X stochastically dominates δ1G, so it suffices to prove that
X . C11n. For the first moment we immediately have an upper bound
2(cid:9)), for some
(2.3)
2π(n − r)e−C10λnr dr ≤ C13n.
E[X] ≤ C12λnZ n
0
We will now show that Var(X) = O(n log n), which will suffice to complete the proof. We
can write X = P Xp, a sum of indicator variables, one for each of the f (n) points in the
configuration. Let Ap denote the event for which Xp = 1, that is, dist(p, δD) > 1
one of the four regions in Figure 3 is empty. There are O(n4λ2
so it suffices to prove that, for any pair of points,
n and at least
n) pairs of distinct points {p, q},
(2.4)
nn3(cid:19) .
P(Ap ∧ Aq) − P(Ap) · P(Aq) = O(cid:18) log n
λ2
9
p1
o
p
r
p2
Figure 2.
First of all, let us define three auxiliary events Wi, i = 1, 2, 3. Let W1 be the event that the
Euclidean distance between p and q is at least two. Let W2 be the event that at least one of p
from δD. Finally, let W3 be the complement of W1 ∪ W2. A priori,
and q is at least 10C9
we can decompose the left-hand side of (2.4) as
log n
λn
(2.5)
(cid:2)P(Ap ∧ AqW1) · P(W1) − P(Ap) · P(Aq)(cid:3) + P(Ap ∧ AqW2) · P(W2) + P(Ap ∧ AqW3) · P(W3).
The main idea here is that, if W1 occurs, then the events Ap and Aq are almost negatively
correlated. Intuitively, if say Ap occurs, what we then know is that at least one of the four
regions in Figure 3, for the point p, is empty. In other words, we just know that some region
of D is empty, which must make it less likely that any disjoint region is also empty. Since q
is at distance greater than two from p, each of its four associated regions is disjoint from the
corresponding regions for p. There is one small flaw with this reasoning, namely the knowledge
that q is far away from p slightly increases the probability of the event Ap to begin with,
and vice versa. Namely, we have already ruled out q as a close neighbor of p, and hence when
calculating the probability of Ap we can imagine starting with a configuration of f (n)− 1 rather
than f (n) points. In summary,
(2.6)
P(Ap ∧ AqW1) ≤ P(A′
p) · P(A′
q),
where A′
f (n) − 1 points (and similarly for q). Clearly,
p is the same event as Ap, but calculated with respect to an intitial configuration of
P(A′
p) ≤ P(Ap) ·(cid:16)1 −
a
πn2(cid:17)−1
,
(2.7)
where a is area of any of the four regions in Figure 3, hence a = π/4. Hence,
(2.8)
P(Ap ∧ AqW1) ≤ P(Ap) · P(Aq) ·(cid:18)1 +
10
1
2n2 + O(cid:18) 1
n4(cid:19)(cid:19) .
o
R1
R4
R2
p
R3
Figure 3. The diameter of the smaller disc equals one, so that any two points
inside it are neighbors.
But recall that W1 is the event that the distance between p and q is at least two, so clearly
P(W1) ≤ 1 − 2
n2 (the right constant is 2 rather than 22 = 4 because we have to consider points
close to δD).
It follows that the first square-bracketed term in (2.5) will be negative for all
n ≫ 0, and it just remains to bound the positive contributions coming from W2 or W3. The
definition of W2 immediately implies, however, by the same argument used to obtain part (ii)
of the Claim, that it contributes negligibly. Hence, it remains to show that
(2.9)
n(cid:19) .
P(Ap ∧ AqW3) · P(W3) = O(cid:18) log n
λn (cid:17) inside the disc boundary
But W3 occurs if and only if p is placed in a strip of width O(cid:16) log n
and then q is subsequently also placed inside this strip and at distance O(1) from q. Since the
points are placed independently and uniformly at random,
n3λ2
(2.10)
P(W3) = O n log n
n2 ×
λn
log n
λn
n2 ! = O(cid:18) log2 n
n (cid:19) .
n3λ2
This already suffices to prove that Var(X) = O(n log2 n) and hence to prove part (i) of the
Claim. To get rid of another log n factor, it suffices to show that P(Ap ∧ AqW3) = O(cid:16) 1
and hence to show that P(Ap) = O(cid:16) 1
log n(cid:17),
log n(cid:17), conditioned on the assumption that p is at distance
at most 10C9
this annulus and we already know that there are a.a.s. order n points in δG, so we are done.
from the disc boundary. But there are a.a.s. on the order of n log n points in
log n
λn
11
We now return to the description of Step 1, which we break down into three substeps:
Step 1A: Each point performs, in time O(1), some tests in an attempt to rule out that it
belongs to δG. If all these tests fail, then it turns blue to signal "I believe I might belong to δG".
The idea here is that each point does something very similar to checking the four regions as in
Figure 3, and turns blue if and only if at least one of those four regions is empty. A point cannot
test exactly this condition, since it does not know where the centre o of the disc is. However it
can, for example, fix some large number K, rotate a half turn and, at equal intervals of π/K,
scan the four quadrants in its 1
2 -neighborhood, as seen from its current orientation. Each time
it scans, it just wants to decide if each of the four quadrants is empty or not, and Assumption
11 implies that this can be done in time O(1). Hence all K tests can be performed in time
OK (1). It is clear that, for a sufficiently large but now fixed K, the set of blue points will have
the properties identified in Claim 2.6, that is, there will at least C14n and a.a.s. at most C15n
,
blue points. Furthermore, for C4 ≫ 0, the blue points will a.a.s. all lie within distance
say, of the disc boundary.
100C15
1
As preparation for Steps 1B and 1C, the points which do not turn blue also have to make a
choice:
(a) turn green if you can see at least one blue point at distance at most
10C15
Since C15 is an absolute constant, this can be incorporated into the algorithm.
1
from yourself.
(b) turn yellow if you see no blue point at distance less than
, but you do see a blue
1
10C15
point at distance less than 1
2 ,
(c) if you see no blue point at distance less than 1
2 , do nothing.
There is an issue of synchronisation here, but since there is an absolute constant time in which
every point can decide whether to turn blue or not, we can also ensure that every point has
decided blue/not blue before the choices (a)-(c) are made, without needing perfectly reliable
clocks. Assumptions 10 and 11 also guarantee that the latter choices can all be made within an
absolute constant time. Hence Step 1A runs in an absolute constant time.
To complete Step 1, the idea is that the leader should come from among the points which
are blue after Step 1A. Since there are a.a.s. Θ(n) such points, if each of them were to generate
at least C16 log n random bits, independent of all the others, then for C16 ≫ 0, a.a.s. there
would be a unique point which generates the largest binary number. There are basically three
problems to be solved in order to turn this into an algorithm:
(i) how do the blue points know how many bits to generate ?
(ii) how do they store the bits ?
(iii) how do they check their numbers against those generated by all other blue points ?
Step 1B will deal with (i) and (ii), whereas Step 1C will deal with (iii). As already mentioned,
problem (ii) could be ignored if we allowed each point to possess unlimited memory, and to
overcome this is where the algorithm becomes most technical.
C4
Step 1B: Each blue point scans all its green neighbors. Lemma 2.1 implies that, for C4 ≫ 0,
log n (say) and at most o(n) green neighbors.
a.a.s. every blue point will scan at least
As explained in Assumption 10, a blue point can actually scan its surroundings in such a way
that it doesn't miss any green points and identifies them at a rate of Ω(1) points per second,
and hence can complete the scan in time o(n). Now, since the total number of blue points is
. C15n, and C15 is an absolute constant, if every blue point generated as many random bits as
it had green neighbors, then a.a.s. there would be a unique largest binary number, provided C4
is sufficiently large.
1000C15
Each time a green point is scanned, the blue scanner generates a random bit. These bits now
need to be stored somewhere, and since their number a.a.s. goes to infinity, we cannot assume
the blue point stores them in its own internal memory. This is where the yellow points enter
the picture.
12
A blue point will choose points amongst its yellow neighbors to store two copies of the
random number it generates, which we call the S-copy (S for "stationary") and the M-copy (M
for "mobile"). The reason for two copies and for our curious terminology will become clear in
Step 1C. For present purposes, the basic idea is that, each time a green point is scanned then a
random bit is generated and two yellow points selected to store it. Each selected yellow point
receives two bits, the first is the random bit generated by blue, the second determines whether
it belongs to the S-copy or the M-copy (say 0 for S and 1 for M). As soon as a point accepts a
storage request then it turns purple, to indicate that it is no longer available for such requests.
Thus each storage point belongs unambiguously to either the S- or the M-copy of a unique
number. There is one more issue to deal with, namely the bits of (a copy of) a blue chief's
number must be stored in such a way that they form an unambiguous binary number, which
can be "read from left to right" unambiguously in Step 1C. This is achieved as follows. The first
time a blue point generates a random bit and finds two yellow points to store it, it also signals
to them that they are its headmen. The S-headman turns brown and the W-headman turns
orange. The headmen now also reposition themselves so that
(i) each remains within distance 1
(ii) the line segment from the headman to the chief is approximately at right angles to the
2 of its blue chief
boundary of the disc and directed outwards from the disc
(iii) the W-headman is to the right of the S-headman, seen from the center of the disc3.
Since all the points in question are a.a.s. within distance 3
5 of the disc boundary, they can make
good estimates of the various directions involved and, together with Assumptions 8 and 9 in
particular, it is clear that they can reposition themselves so as to satisfy (i)-(iii) in time O(1).
Their blue chief tracks their movements (Assumption 8) and so is aware of their exact resting
locations. Now it is time to scan the second green neighbor, generate the second random bit and
select two yellow points to store it. As well as storing their respective bits, these two hitherto
yellow (and now purple points) now walk toward their chief and reposition themselves somewhere
on the line segment between their chief and the appropriate headman. Using Assumptions 8-9,
this manoeuvre can be successfully accomplished in time O(1). Once in position, these points
take over the role of headman, whereas the previous headmen turn new colors to indicate that
they are now tailmen. In subsequent steps, the selected storage points reposition themselves
somehere on the line segment between their chief and the current appropriate headman. Once
they have done so, they take over the role of headman, and the previous headmen turn some
neutral color to indicate that they are now middlemen.
There is one final subtlety. Since the total number of blue points is . C15n, on average
any yellow point will be able to see no more than C15
π blue points. Putting it another way, on
average no more than C15
π blue points will be competing for the services of any yellow point as
storage space. Since the total number of yellow points will be close to (5C15)2 times the total
number of green ones, on average there will be enough storage space to go round. However,
we do not see how to rule out the possibility that there may locally be much denser clusters of
blue points. A priori, it follows from Claim 2.6 that up to O(log n) blue points may be visible
from any yellow point. In that case, it can happen that a blue point runs out of storage space
before it has scanned all its green neighbors. If that is the case, we shall let the blue point
"drop out of contention", i.e.: it abandons the generation of its random number. It then turns
pink to indicate that it is out of contention for leadership, and this information is relayed to all
the current members of its S- and W-strings, who subsequently revert to white (we let this be
the default color of a point that isn't doing anything). Note that, because the points in these
strings lie on a straight line from their chief, they cannot receive signals from the latter directly,
so information has to be relayed along the string. This can still be accomplished in time o(n),
3Actually, (ii) and (iii) are unnecessary requirements, they are merely of aesthetic value. We imagine the
points of each string lining up at right angles to the disc boundary. This "looks nice", but will not be needed for
the execution of Step 1C to follow. Indeed, all that's important is that each point along the string knows where
its two neighbors are, the shape of the string as a whole is irrelevant.
13
however.
Step 1C: Once a blue point has completed Step 1B, and if it has not turned pink, it signals to
the headman of its M-string to "start walking". The idea here is that the M-string will now walk
around the boundary of the disc such that the interior of the disc is on its right4. The headman
leads the way and the other points in the string follow (see Assumption 8). As it walks, it must
follow the trail of blue and pink points, and for each blue point it must consult the S-copy of
that point's number and compare it with its own. The two numbers are compared bit-by-bit
(we specify the exact protocol below) and the comparison is aborted once a bit is found where
the two numbers differ. If the S-number has a 1 in this position, the M-headman immediately
decides that its chief is out of contention for leadership. It stops walking once it returns to its
starting point, which it recognises by the fact that the S-string it just read is identical to its
own (it knows when it's finished reading a string since it recognises a tailman by the particular
color reserved for tailmen). If it returns without having encountered any S-number greater than
its own, it signals to its chief that it is leader. The following considerations ensure that this
procedure a.a.s. achieves what we want.
Firstly, because of Claim 2.6, a.a.s. all the chiefs are located very close to the disc boundary
and hence a wandering headman can make sure it doesn't miss any blue or pink points while
keeping the interior of the disc on its right. Secondly, the lower bound on f (n) ensures that
a.a.s. exactly one chief will be declared leader, provided C4 ≫ 0. Note that it doesn't matter
if some chiefs already dropped out by turning pink in Step 1B, since there will a.a.s. be Ω(n)
of them left in the game. Thirdly, by Lemma 2.2, a.a.s. each M-string will complete its walk
in time O(n), provided it does not does not waste any time "queueing". This is the only subtle
issue here. Since an M-string may read anything from one to Ω(λn) bits of an S-string, there
may be many M-strings reading the same S-string simoultaneously. To avoid queues developing,
we need to have an appropriate reading protocol. One solution is to have the M-headman read
the S-bits one by one:
it is no problem for it to follow the trail of S-bits, since they lie in a
straight line. Alternatively, each S-bit could keep a laser shining on its successor to point it
out, see Assumption 9. Simoultaneously, the M-headman requests bits from its own string to
compare with.
Imagine that each point in the string maintains two copies of its bit, one of
which is permanent and the other in short-term memory. The headman requests bits from the
first middleman. Whenever a middleman receives a request for a bit, it empties the contents
of its short-term memory and then requests the contents of the short-term memory of the next
point along the string. In this way, bits can be passed along the M-string to the headman at
a rate of Ω(1) bits per time unit, who can then compare them one-by-one with the trail of
S-bits. This protocol ensures that the S-strings are completely "passive" and hence can be read
simoultaneously by arbitrarily many M-headmen without any queues developing.
One final comment: In steps 1B and 1C we have not bothered about trying to synchronise the
activities of different actors. I think it is clear that this is not a problem. A.a.s. steps 1B and
1C will, as described, lead to one chief identifying itself as leader in time O(n). Once it does so,
it can initiate Step 2 by turning red. Note that, during Step 1, some points in the configuration
will have changed their locations. However, a.a.s. all these points were initially within distance
3
5 of the disc boundary and hence their movements will not affect the connectedness of the
communication graph. Hence, a.a.s. Step 2 will subsequently still succeed and the whole
algorithm terminate in time O(n).
Remark 2.6. The above algorithm would also work if f (n) ≤ κn3 and the constant κ were
somehow known in advance. When f (n) = Θ(n3), then the blue points will generate Θ(n)
random bits, but this is not a problem. If κ is not specified, then the difficulty is rather that the
constant we denoted C15 will now depend on κ, since it depends on the size of δG via C8 and
4or its left, it doesn't matter as long as each headman maintains a consistent orientation. Indeed, there is no
problem if different headmen head in opposite directions.
14
C11. In fact, it is only the size of δ2G that is problematic, as this will now be Ωκ(n), whereas
we can see from (2.3) that δ1G will a.a.s. remain bounded by an absolute constant times n.
As explained in the proof, the constant C15 determines the amount of competition, on average,
for yellow storage space. To be sure there is enough storage space to go round on average, the
ratio of yellow to green points must be sufficiently large, in other words the search radius for
green points must be sufficiently small, depending on κ. We do not see how to get around
this problem. Amusingly, though, if the points somehow knew that their average density was
Θ(n), then they would realise that generating Θ(n) random bits is extremely wasteful, since it
would suffice with Ω(log n) bits. The issue of whether there is a better alternative than counting
neighbors for the task of estimating how many random bits need to be generated by a blue point
will be taken up in Question 3.7 below.
Remark 2.7. If we allowed each point to possess unlimited storage capacity, then the descrip-
tion of Steps 1B and 1C could be simplified considerably. There would be no need for the colors
grren and yellow. Each blue point could just count all its neighbors, generate one random bit
per neighbor and store the entire binary number. It could then choose a headman from amongst
its neighbors and write a copy of the number to the headman's memory. The headman would
then walk around the boundary, comparing its number with that stored in each blue point. It
would a.a.s. know it has returned to its own blue chief when it reads a number equal to its own.
The fact that our algorithm does not utilise unlimited storage capacity per point seems
important, however. Another interesting observation is that the kinds of computations done in
Steps 1B and 1C are very primitive, basically only counting and comparing strings. Of course,
the sophistication of the procedure lies in the execution of "higher-level" tasks such as scanning,
pointing and following. These are the kinds of abilities a sceptical reader might object to.
3. Comments and Questions
In this section, whenever we formulate a precise question it should be understood that the
11 assumptions made in Section 2 are valid. However, just as important an issue going forward
is whether there are algorithms which work just as well under some weaker set of assumptions
about how points can interact. This should be kept in mind at all times.
For the task of merging at a single point to be strictly meaningful, one must assume from the
outset that one is dealing with point particles which are capable of locating other point particles
with infinite precision. This is true in both the classical multi-agent Rendezvous setting (RP-5)
and in ours (Assumption 4 in Section 2). The points are idealisations and each particle has,
in "reality" a fixed, if small size. In this paper, we have considered "generic" configurations of
points in a disc of radius n in R2. In order to ensure that the communication graph would be
a.a.s. connected, it was required that the average density of particles be Ω(log n), hence goes
to infinity with n. What this implies is that, once the idealisation of point particles is removed,
then generic configurations of particles will a.a.s. not be connected and RP-4 fails. This seems
like a serious obstacle to making practical sense of the whole project of studying Rendezvous
for generic configurations of agents. So, if you didn't already consider Theorem 2.4 just an
intellectual curiosity, then that seems even more apparent now. However, it is not clear that
all is in vain, especially if one is satisfied with a randomised algorithm that a.a.s. succeeds,
rather than a fully deterministic procedure. Suppose the average point density is Θ(1), this
obviously being the most "realistic" setting, from what we have just said. Almost surely, there
will be isolated points in this regime, but there will also be large connected components, so an
isolated point does not necessarily need to perform a Brownian motion in order to make contact
with other points, something which (as stated earlier) would on average take an infinite time to
succeed. This leads to our first and most important question:
Question 3.1. In the notation of Theorem 2.4, suppose f (n) = o(n2 log n). Does there exist
a randomised merging algorithm which a.a.s. runs in finite time and, if so, how quickly as a
function of n? In particular, consider these issues when f (n) = Θ(n2). The same questions
15
can be asked in the classical setting, i.e.: assuming RP 1-5, at least as long as we ignore the
problem of storage capacity.
We suspect nevertheless that the above questions have negative answers. Hence, perhaps a
better idea is to ask the same questions, but conditioned on the communcation graph being
connected:
Question 3.2. Same question as above, but conditioned on the initial communication graph
being chosen uniformly at random from among all connected such graphs on f (n) nodes. Even
more "realistically", one could condition on the graph being connected and the number of points
per unit area being bounded.
Once one accepts that each particle has a finite size, a second obvious conclusion is that,
in reality, Rendezvous means that all the particles come to occupy some sufficiently small
region of the plane, rather than merge at a single point5. An intermediate step, which is
more mathematically appealing, is to retain the assumption of point particles but redefine the
Rendezvous condition in this way, and see if it leads to a significant reduction in the run-times
of algorithms. In Theorem 2.4, the run-time is already linear in n, so relaxing the definition
of Rendezvous cannot improve performance significantly, at least as long as we want some
non-trivial convergence of the particle swarm, that is to a disc of radius o(n). We think this
emphasises the fact that our algorithm is really a procedure for choosing a leader, rather than for
merging. Our algorithm for choosing a leader can in turn be broken down into two main steps.
In the first step, the graph boundary δG is identified approximately. In the second step, the
points which think they might be in δG assign representatives to walk around the disc boundary
comparing randomly generated binary strings. This second step seems much more contrived
than the first. In addition, the eventual merging will take place at the leader's location, which
is a.a.s. very close to the disc boundary. In contrast, classical merging procedures involve some
kind of repeated "averaging", which means that, for a generic configuration of points in a disc,
those closer to the boundary will gradually move inwards, and the eventual merging will take
place somewhere close to the center of the disc.
This leads to at least two possible lines of further questioning. On the one hand, we can seek
an alternative to our algorithm which more resembles classical merging algorithms:
Question 3.3. In the setting of Section 2, is there a linear-time algorithm which does not have
the two-step character of ours, in which first a leader is chosen and then all the points move to
its location?
On the other hand, we can seek alternative linear-time algorithms after relaxing the definition
of Rendezvous so that the points are only required to come sufficiently close together. Since
part of the problem is to specify exactly how close, we will not formulate a precise question,
though at the very least, they should all be required to move inside some disc of radius o(n). In
this relaxation it is also natural to seek an algorithm which involves less sohpisticated communi-
cation between agents than that allowed by the assumptions of Section 2. Some communication
is probably still necessary, however. If we consider the ASY-algorithm, for example, and again
ignore storage issues, it was already noted in Section 1 that the time taken to execute any non-
trivial convergence is Ω(n log n). Indeed, extending a comment made in Section 1, we conjecture
that the same is true of any algorithm which assumes RP 1-5. It would also be interesting to
determine the actual (expected) rate of convergence for such algorithms, for example ASY, for
generic connected configurations. As already mentioned in Section 1, we are not aware of any
rigorous treatment of this problem in the literature.
If, instead of questioning either the assumptions made in Section 2 or the philosophy of con-
sidering generic configurations, we accept these principles, then there are still many unanswered
5The robustness of the ASY-algorithm to removal of unrealistic idealisations was tested numerically in [AOSY].
In their simulations, they replaced each point robot by a disc of fixed area and considered n robots to have
rendezvoued once they were all inside a disc of radius 3pn/2.
16
questions. What jumps out immediately is the upper bound on f (n) assumed in Theorem 2.4.
This played two crucial roles:
(i) it implied that the size of the graph boundary was a.a.s. O(n)
(ii) it implied that each point a.a.s. had O(n) neighbors, indeed o(n) neighbors, though see
also Remark 2.6. Thus, by counting its neighbors, a blue point had a way to estimate how many
random bits it needed to generate.
First consider (i). Here it is obviously crucial that the points are distributed in a disc, that
is, a bounded region of R2 with a "nice" boundary. This leads us to our next two questions.
The first is a bit vague:
n=1 of Lebesgue measurable compact regions in R2,
Question 3.4. For which sequences (Dn)∞
satisfying Area(Dn) → ∞ as n → ∞, does an analogue of Theorem 2.4 hold ? The analogue we
have in mind here is that f (n) points are distributed uniformly and independently in Dn, where
f (n) grows sufficiently fast so that the communication graph is a.a.s. connected, but not so fast
that any point is likely to have more than O(n) neighbors.
An alternative track, which we find more appealing, is to consider the multi-agent rendezvous
problem on a compact 2-dimensional manifold without boundary. The simplest question, to
which we have no answer, is the following:
Question 3.5. Let Sn denote the 2-sphere in R3 of radius n. Suppose f (n) points are distributed
uniformly and independently on Sn, where f (n) satisfies the same conditions as in Theorem 2.4.
Under Assumptions 1-11, does there exist a merging algorithm which a.a.s. runs successfully in
time O(n) ?
We have not seen the latter question asked even in the classical setting, where RP 1-5 are
assumed. Perhaps this is just our ignorance, or perhaps it is because, if one thinks of a sphere as
representing the Earth, then a compass provides a universal point of reference and rendezvous
can trivially be accomplished in time linear in the diameter. There are two further things to
note about the problem on a 2-sphere:
(a) A Brownian motion confined to Sn will a.a.s. arrive within distance one of any point in
finite time. Hence there is certainly some merging procedure on Sn which a.a.s. runs in finite
time. For example, two points could merge once they see each other, with a point choosing a
neighbor at random to merge with if it has more than one. Once points become isolated they
could perform a Brownian walk. The expected running time of even this rather stupid algorithm
is unclear to us, however, especially in light of Assumption 1 in Section 2.
(b) Something close to Question 3.5 arises if we try to extend our algorithm to higher dimen-
sions. Consider an initial configuration of points inside a ball in R3. Step 1A of our procedure
would identify a subset of blue points close to the boundary of the ball, hence close to a 2-sphere.
We are not quite left with Question 3.5, since these points can also move through the interior
of the ball now. However, these observations suggest that a different approach may also be
necesary in higher dimensions.
Now suppose we turn instead to point (ii). There is another obvious question:
Question 3.6. Can Theorem 2.4 be extended to even denser configurations of points ? In other
words, is there a linear-time merging algorithm on the disc which succeeds even when the average
point density is super-linear ?
An additional observation here is that the method for generating random numbers described
in Step 1B of our algorithm is very inefficient when f (n) ≫ n2 log n since, as long as f (n) grows
polynomially in n, it would suffice for each blue point to generate Ω(log n) bits, whereas in fact
each blue point generates on the order of f (n)/n2 bits. An alternative procedure we considered
is to have the blue points estimate n, the diameter of the disc. We considered a specific, and
rather complicated, signalling procedure, whereby the blue points send signals to their neighbors
17
which are relayed inwards in such a manner that, a.a.s., the signals bounce back only once they
have reached very close to the center of the disc. The details are not important as we could not
prove that our procedure worked, but we can state another question:
Question 3.7. As an alternative to Step 1B, is there a procedure by which the blue points can
estimate the diameter n of the disc, up to an absolute multiplicative factor say, a.a.s. in time
O(n) ?
We conclude with a couple more questions which seem interesting, though are perhaps less
important. One striking aspect of our "choose a leader then move to it" algorithm is that all
the points merge at one place. This is also true of classical rendezvous procedures. Noting also
Assumption 1, we can ask
Question 3.8. In the setting of Section 2, is there a linear-time algorithm which includes
merging of groups of points at intermediate steps and at different locations, rather than all
having all the points merge only at the end at a single location ?
We pose our last question as it may be interesting in its own right, as a problem in compu-
tational geometry:
Question 3.9. In the setting of Section 2, is there a linear-time algorithm for identifying the
graph boundary exactly?
Beyond the discussion above, several other possible lines of investigation present themselves
naturally, though they take us into even less well-defined territory. For example, one might
consider initial configurations of points which are not uniform i.i.d. Ultimately, one can question
the various assumptions listed in Section 2, and try to formulate an interesting problem under
some other set of assumptions. Last but not least, in the setting of Section 2, is there a really
simple linear-time algorithm which we have completely missed, which would render most of this
manuscript unnecessary ?
[A]
[AOSY] H. Ando, Y. Oasa, I. Suzuki and M. Yamashita, Distributed memoryless point convergence algotirhm for
S. Alpern, The rendezvous search problem, SIAM J. Control Optim. 33 (1995), No.3, 673 -- 683.
mobile robots with limited visibility, IEEE Trans. Robotics Automation 15 (1999), No.5, 818 -- 828.
References
[CW]
[De]
[ASY] H. Ando, Y. Suzuki and M. Yamashita, Formation and agreement problems for synchronous mobile
robots with limited visibility, in: Proceedings of the 1995 IEEE International Symposium on Intelligent
Control (ISIC 1995), pp. 453 -- 460.
J. Cameron and W. Wisher Jr., Terminator 2: Judgement Day, TriStar Pictures (1991).
B. Degener, B. Kempkes, T. Langner, F. Meyer auf der Heide, P. Pietrzyk and R. Wattenhofer, A tight
runtime bound for synchronous gathering of autonomous robots with limited visibility, in: Proceedings
of the 23rd ACM Symposium on Parallellism in Algorithms and Architectures (SPAA '11), pp. 139 -- 148
(2011).
R. Durrett, Probability: Theory and Examples (3rd edition), Belmont: Thomsen Brooks/Cole (2006).
N. Megiddo, Linear-time algorithms for linear programming in R3 and related problems, SIAM J. Com-
puting 12 (1983), No.4, 759 -- 776.
[Du]
[M]
[MBCF] S. Martinez, F. Bullo, J. Cortes and E. Frazzoli, On synchronous robotic networks - Part II: Time
complexity of rendezvous deployment algorithms, IEEE Trans. Automat. Control 52 (2007), No.12,
2214 -- 2226.
M. Penrose, Random Geometric Graphs, Oxford University Press (2007).
E. Welzl, Smallest enclosing discs (balls and ellipsoids), Lecture Notes Comp. Sci. 555 (1991), 359 -- 370.
[P]
[W]
Department of Mathematical Sciences, Chalmers University Of Technology and University of
Gothenburg, 41296 Gothenburg, Sweden
E-mail address: [email protected], [email protected], [email protected]
18
|
1601.07011 | 1 | 1601 | 2016-01-26T13:23:06 | Distributed Detection over Adaptive Networks: Refined Asymptotics and the Role of Connectivity | [
"cs.MA",
"cs.IT",
"cs.IT"
] | We consider distributed detection problems over adaptive networks, where dispersed agents learn continually from streaming data by means of local interactions. The simultaneous requirements of adaptation and cooperation are achieved by employing diffusion algorithms with constant step-size {\mu}. In [1], [2] some main features of adaptive distributed detection were revealed. By resorting to large deviations analysis, it was established that the Type-I and Type-II error probabilities of all agents vanish exponentially as functions of 1/{\mu}, and that all agents share the same Type-I and Type-II error exponents. However, numerical evidences presented in [1], [2] showed that the theory of large deviations does not capture the fundamental impact of network connectivity on performance, and that additional tools and efforts are required to obtain accurate predictions for the error probabilities. This work addresses these open issues and extends the results of [1], [2] in several directions. By conducting a refined asymptotic analysis based on the mathematical framework of exact asymptotics, we arrive at a revealing and powerful understanding of the universal behavior of distributed detection over adaptive networks: as functions of 1/{\mu}, the error (log-)probability curves corresponding to different agents stay nearly-parallel to each other (as already discovered in [1], [2]), however, these curves are ordered following a criterion reflecting the degree of connectivity of each agent. Depending on the combination weights, the more connected an agent is, the lower its error probability curve will be. Interesting and somehow unexpected behaviors emerge, in terms of the interplay between the network topology, the combination weights, and the inference performance. The lesson learned is that connectivity matters. | cs.MA | cs |
Distributed Detection over Adaptive Networks:
1
Refined Asymptotics
and the Role of Connectivity
Vincenzo Matta, Paolo Braca, Stefano Marano, Ali H. Sayed
Abstract
We consider distributed detection problems over adaptive networks, where dispersed agents learn
continually from streaming data by means of local interactions. The requirement of adaptation allows the
network of detectors to track drifts in the underlying hypothesis. The requirement of cooperation allows
each agent to deliver a performance superior to what would be obtained if it were acting individually.
The simultaneous requirements of adaptation and cooperation are achieved by employing diffusion
algorithms with constant step-size µ. In [1], [2] some main features of adaptive distributed detection
were revealed. By resorting to large deviations analysis, it was established that the Type-I and Type-II
error probabilities of all agents vanish exponentially as functions of 1/µ, and that all agents share the
same Type-I and Type-II error exponents. However, numerical evidences presented in [1], [2] showed
that the theory of large deviations does not capture the fundamental impact of network connectivity on
performance, and that additional tools and efforts are required to obtain accurate predictions for the
error probabilities. This work addresses these open issues and extends the results of [1], [2] in several
directions. By conducting a refined asymptotic analysis based on the mathematical framework of exact
asymptotics, we arrive at a revealing and powerful understanding of the universal behavior of distributed
detection over adaptive networks: as functions of 1/µ, the error (log-)probability curves corresponding
V. Matta and S. Marano are with DIEM, University of Salerno, via Giovanni Paolo II 132, I-84084, Fisciano (SA), Italy
(e-mail: [email protected]; [email protected]).
P. Braca is with NATO STO Centre for Maritime Research and Experimentation, La Spezia,
Italy (e-mail:
[email protected]).
A. H. Sayed is with the Electrical Engineering Department, University of California, Los Angeles, CA 90095 USA (e-mail:
[email protected]).
A short version of this work appeared in the conference publication [3].
October 28, 2018
DRAFT
2
to different agents stay nearly-parallel to each other (as already discovered in [1], [2]), however, these
curves are ordered following a criterion reflecting the degree of connectivity of each agent. Depending
on the combination weights, the more connected an agent is, the lower its error probability curve will
be. The analysis provides explicit analytical formulas for the detection error probabilities and these
expressions are also verified by means of extensive simulations. We further enlarge the reference setting
from the case of doubly-stochastic combination matrices considered in [1], [2], to the more general and
demanding setting of right-stochastic combination matrices; this extension poses new and interesting
questions in terms of the interplay between the network topology, the combination weights, and the
inference performance. The potential of the proposed methods is illustrated by application of the results
to canonical detection problems, to typical network topologies, for both doubly-stochastic and right-
stochastic combination matrices. Interesting and somehow unexpected behaviors emerge, and the lesson
learned is that connectivity matters.
Index Terms
Distributed detection, adaptive network, diffusion strategy, large deviations analysis, exact asymp-
totics.
I. INTRODUCTION AND MOTIVATION
Adaptive networks consist of spatially dispersed agents, learning continually from streaming
data by means of distributed processing algorithms. With a proper form of cooperation, each
agent is able to deliver an inference performance superior to what would be obtained if it
were acting individually. When called to operate in complex and dynamical scenarios (e.g., in
the presence of drifts in the statistical conditions, in the environmental conditions, and in the
network topology, among other possibilities), the agents must be further endowed with strong
adaptive capabilities, in order to respond in real-time to these variations.
Several useful distributed inference solutions are available that meet these requirements. Par-
ticularly relevant to our setting are the diffusion implementations based on adaptive algorithms
with constant step-size [4] -- [7]. The use of a constant step-size (as opposed to the decaying
step-size employed in the case of consensus algorithms [8] -- [19]) is key to enable continuous
learning, e.g., to meet the fundamental requirement of tracking. The interplay between adaptation
and learning is critical for guaranteeing the successful network operation and to produce reliable
inference.
October 28, 2018
DRAFT
3
Such an interplay has been examined rather deeply in the framework of estimation problems,
while less attention has been devoted to detection problems. Distributed detection over adaptive
networks can be succinctly described as follows. A network of dispersed agents, linked together
by a given topology, monitors a certain physical phenomenon. As time elapses, the agents gather
from the environment streaming data, whose statistical properties depend upon an unknown state
of nature, formally represented by a pair of hypotheses, say, H0 and H1. At each time instant,
each individual agent must produce a decision inferring the current state of nature, using its own
available observations and the local exchange of information obtained from consultation with its
neighbors. Due to the requirement of adaptation, the agents must be able to react promptly to
drifts in the current state of nature, while guaranteeing an adequate performance level (i.e., low
detection error probabilities) at the steady-state, namely, when a given hypothesis is in force for
sufficiently long time.
With reference to the above setting, in the recent works [1], [2] several interesting results were
derived that allowed the authors to establish fundamental scaling laws for adaptive distributed
detection by multi-agent networks. In particular, it was shown there that for diffusion strategies
with constant step-size µ, the steady-state error probabilities exhibit an exponential decay, as
µ → 0, as functions of 1/µ. By resorting to a detailed large deviations analysis, the exact
scaling law was fully characterized in [1], [2] in terms of a rate function Φ(γ) that can be
evaluated analytically. Notably, the results of [1], [2] showed that all network agents share the
same error exponent so that their error probabilities are asymptotically equivalent to the leading
exponential order as µ → 0. This is a remarkable feature of adaptive distributed detection
solutions. Readers may consult the introductory remarks of [1], [2], which contain a detailed
and motivated summary of these results.
One known limitation of large deviations analysis resides in the fact that it focuses only on the
leading exponential order, thus neglecting all sub-exponential terms. The practical implication of
this fact can be easily illustrated by a simple example. Assume network agents 1 and 2 exhibit
asymptotic error probabilities P1 and P2 of the form:
P1 = e
(1)
where o(1) stands for any correction such that o(1) → 0 as µ → 0. These two probabilities have
the same error exponent multiplying −1/µ (and is equal to one), but the error probability at
P2 = 2 e
− 1
µ = e
µ [1+o(1)],
− 1
µ ,
− 1
October 28, 2018
DRAFT
4
agent 2 is always twice that of agent 1. This is because the factor 2 is a sub-exponential term.
The above consideration is particularly relevant in the context of adaptive distributed detection
since numerical experiments reported in [1], [2] show that, depending on the particular network
structure, different agents have different error probabilities, while exhibiting the same scaling
law to the leading exponential order. Even more remarkably, the experimental results of [1], [2]
suggested that the differences in the error probabilities appear to be related to the degrees of
connectivity of the various agents. Accordingly, the main purpose of this paper is to perform a
refined asymptotic analysis overcoming the limitations of large deviations, in order to capture the
fundamental dependence of the error probabilities at different agents upon the specific network
connectivity. In this way, we will be able to show that the partial empirical evidences in [1], [2]
are in fact representative of a universal behavior of distributed detection over adaptive networks.
Another relevant contribution provided in this work is that, differently from what was done
in [1], [2], we shall not limit the analysis to doubly-stochastic combination matrices and will
consider instead the broader class of right-stochastic matrices. This latter class is relevant in
practical applications. Indeed, while doubly-stochastic matrices might look preferable in detection
applications in view of the asymptotic equipartition of the combination weights, they are not
always realizable in practice. This happens, for instance, due to physical limitations in the
topology and/or communication constraints. This generalization in the nature of the combination
policy poses new and interesting questions in terms of the interplay between the network
topology, the combination weights, and the inference performance.
In summary, as a result of the refined analysis presented in this work, we will not only be
able to evaluate in quantitative terms how network connectivity reflects on the performance of
different agents, but we will also be able to clarify how different network structures (i.e., different
combination matrices and network topologies) globally impact the network performance.
A. Related Work
Distributed detection is a classical topic, and the pertinent literature is therefore extensive. Any
set of references would be by-no-means exhaustive, so that we refer the reader to [20] -- [31] as
fundamental entry points on the subject. In this work we consider fully-decentralized detection
problems, i.e., fully-flat network architectures without fusion center, where the agents are only
allowed to interact locally. Several recent works address this scenario and, in particular, solutions
October 28, 2018
DRAFT
5
based on decentralized consensus strategies with decaying step-size have been successfully
proposed in [8] -- [13], and the detection performance of these algorithms has been accurately
characterized in different asymptotic frameworks [8] -- [13]. However, as already observed in
the introduction, a distinguishing feature of our work resides in the emphasis on adaptive
solutions. To enable adaptation, it has been shown that diffusion algorithms with constant step-
size are superior to consensus implementations due to an inherent asymmetry in the consensus
update that has been shown to be a source of potential instability in the consensus dynamics
when constant step-sizes are used [4] -- [7]. Several performance results are already available for
diffusion strategies in connection to their mean-square-error (MSE) estimation behavior [4] -- [7].
The corresponding results for detection applications are relatively limited. In [32] the problem
of using diffusion algorithms for detection purposes has been considered, with reference to
a Gaussian problem. More recently, the general problem of distributed detection over adaptive
networks has been posed in [1], [2]. By means of a large deviations analysis, the Type-I and Type-
II error exponents, for doubly-stochastic combination matrices, were characterized in closed-form
in these works. It was shown that the detection performance of the network solution is equivalent
to the fully-connected solution to the first-leading order in the exponent. However, numerical
and simulation results in [1], [2] showed that the network connectivity and the overall structure
of the combination weights do matter, and they influence the performance of the individual
agents. Analytical formulas for accurate evaluation of the detection error probabilities, and for
elucidating the relationship between the network structure and the expected performance, are
currently missing. Filling this important gap is the main theme of this work.
B. Summary of Main Result
In order to introduce the main result, we refer to the steady-state (i.e., as time goes to infinity)
output of the diffusion algorithm with constant step-size µ. The steady-state output at the k-
th agent is denoted by y(cid:63)
k,µ tends
to concentrate in the close proximity of the expected value of the local statistics, denoted by
E[x]. In order to characterize the error probabilities, we shall evaluate the probability that y(cid:63)
k,µ
deviates significantly from this expected behavior. Without loss of generality, we shall focus on
the following probability:
k,µ. It was shown in [1], [2] that, for small step-sizes, y(cid:63)
October 28, 2018
P[y(cid:63)
k,µ > γ],
γ > E[x].
(2)
DRAFT
With reference to the case of doubly-stochastic combination matrices, it was established in [1],
[2] that
µ ln P[y(cid:63)
k,µ > γ] = −Φ(γ),
lim
µ→0
where Φ(γ) is the so-called rate function. Equivalently, we can rewrite (3) as
P[y(cid:63)
k,µ > γ] = e
− 1
µ [Φ(γ)+o(1)].
6
(3)
(4)
This last form of representation highlights the fact that sub-exponential terms are neglected by
large deviations analysis.
A refined analysis can be pursued by seeking an asymptotic approximation, Pk,µ(γ), that
ensures the much stronger conclusion:
i.e.,
P[y(cid:63)
k,µ > γ] = Pk,µ(γ)[1 + o(1)],
P[y(cid:63)
k,µ > γ]
Pk,µ(γ)
lim
µ→0
= 1.
(5)
(6)
This framework is commonly referred to as exact asymptotics, and has been originally studied
in [33] with reference to the simplest case of normalized sums of independent and identically
distributed (i.i.d.) random variables -- see also [34], [35]. In the following, we shall often write
P[y(cid:63)
k,µ > γ] ∼ Pk,µ(γ)
to denote an asymptotic equivalence of the kind (6). It is immediately seen that (6) implies
lim
µ→0
µ ln P[y(cid:63)
k,µ > γ] = lim
µ→0
µ ln Pk,µ(γ),
(7)
(8)
which means that any exact asymptotic Pk,µ(γ) is able to reflect the leading exponential term.
The main result established in this paper can now be formally stated as follows:
Pk,µ(γ) =
µ
γ φ(cid:48)(cid:48)(θγ)
2πθ2
e
− 1
µ[Φ(γ)+k,µ(θγ )]
(9)
The main quantities necessary to evaluate the asymptotic approximation Pk,µ(γ) in the above
expression are now briefly introduced. The function φ(t) will be described in closed-form, and
depends on the underlying statistical model through the moment generating function of the local
October 28, 2018
DRAFT
(cid:114)
7
statistics, and on the network topology only through the limiting vector of combination weights
(aka the Perron eigenvector, see, e.g., [7]). The quantity θγ is the solution to the equation:
φ(cid:48)(θγ) = γ.
(10)
The function Φ(γ), which was the main object of study in [1], [2], is the so-called rate function,
and is computed directly from φ(t), namely, it is the Fenchel-Legendre transform of φ(t) [34],
[35].
Finally, the (sub-exponential) correction term k,µ(θγ) depends on the underlying statistical
model through the moment generating function of the local statistics, and on the network topology
through the actual (not only the limiting) network combination weights. This correction satisfies:
k,µ(θγ)
(11)
In particular, the ratio k,µ(θγ)/µ stays bounded as µ → 0. Loosely speaking, this means that the
overall correction appearing at the exponent in (9) plays, asymptotically, the role of a constant
correction.
µ→0−→ 0.
Despite its apparent complexity, Eq. (9) possesses a well defined structure, revealing important
connections with the physical behavior of the adaptive distributed system under consideration.
Let us elucidate some of these features. We start by the leading order in the exponent. It is easy
to see that:
µ ln Pk,µ(γ) = −Φ(γ) − k,µ(θγ) +
ln(cid:2)2πθ2
γ φ(cid:48)(cid:48)(θγ)(cid:3)
µ
2
ln µ − µ
2
µ→0−→ −Φ(γ),
(12)
which easily follows from (11). This means that the approximation Pk,µ in (9) can be regarded
as
Pk,µ(γ) = e
− 1
µ [Φ(γ)+o(1)].
(13)
This is consistent with result (3), which was established in [1], [2] for the case of doubly-
stochastic combination matrices. The terms of order o(1) in (13) collect all the sub-exponential
corrections that appear in (9). They can be separated into two categories. The first correction
in (9). This term is a typical sub-exponential refinement arising in
the framework of exact asymptotics, and is a consequence of a local Central Limit Theorem
is the term (cid:113) µ
γ φ(cid:48)(cid:48)(θγ )
2πθ2
October 28, 2018
DRAFT
8
(see [33] -- [35]). Observe that this correction, which is related to the network topology only
through the Perron eigenvector, is independent of the agent index k, and is therefore applicable
to all agents. In contrast, the second correction k,µ(θγ) depends on the agent index k and, as
it will be detailed in the exact statement of the main theorems, takes into account the entire
network topology and combination weights.
The above considerations lead to the important conclusion that (9) provides a detailed and
revealing assessment of the universal behavior of distributed detection over adaptive networks:
as functions of 1/µ, the error (log-)probability curves corresponding to different agents not only
stay nearly-parallel to each other, but they are also ordered following a criterion dictated by
the correction term k,µ(θγ). As we shall see later -- see Fig. 1 for an example, this criterion
reflects the degree of connectivity of each agent. Depending on the combination weights, the
more connected an agent is, the lower its error probability will be, and the correction term
k,µ(θγ) is sufficiently rich to capture this behavior.
Notation. We use boldface letters to denote random variables, and normal font letters for their
realizations. Capital letters refer to matrices, small letters to both vectors and scalars. Sometimes
we violate this latter convention, for instance, we denote the total number of sensors by S. The
symbols P and E are used to denote the probability and expectation operators, respectively. The
notation Ph and Eh, with h = 0, 1, means that the pertinent statistical distribution corresponds
to hypothesis H0 or H1.
II. PROBLEM FORMULATION
We consider a sensor network that collects observations about a physical phenomenon of inter-
est. Data are assumed to be spatially and temporally independent and identically distributed. From
the observation measured at time n, the k-th agent computes its local statistic (the observation
itself, or a suitable function thereof), which is denoted by xk(n), k = 1, 2, . . . , S. The mean and
variance of xk(n) will be denoted by E[x] and σ2
x, respectively. To avoid trivialities, throughout
the paper it is assumed that the random variable xk(n) is non-degenerate.
A. Diffusion Strategy
Following the framework developed in [1], [2], we focus on the class of diffusion strategies
for adaptation over networks [6], [7], [32], and in particular on the ATC (Adapt-Then-Combine)
October 28, 2018
DRAFT
implementation due to some inherent advantages in terms of a slightly improved mean-square-
error performance relative to other forms [6]. Extensions to other diffusion implementations, as
well as to consensus implementations, are certainly possible. In the ATC algorithm, each node
k updates its state from yk(n − 1) to yk(n) through local cooperation with its neighbors as
follows:
9
vk(n) = yk(n − 1) + µ[xk(n) − yk(n − 1)],
(14)
S(cid:88)
(cid:96)=1
ak,(cid:96)v(cid:96)(n),
yk(n) =
(15)
where 0 < µ (cid:28) 1 is a small step-size parameter. It is seen that node k first uses its locally
available statistic xk(n), to update its state from yk(n − 1) to an intermediate value vk(n).
The other network agents simultaneously perform similar updates using their local statistics.
Subsequently, node k aggregates the intermediate states of its neighbors using nonnegative convex
combination weights {ak,(cid:96)} that add up to one. Again, all other network agents perform a similar
calculation. Collecting the combination weights into a square matrix A = [ak,(cid:96)], then A is a right-
stochastic matrix, namely, the entries on each row add up to one. Formally:
ak,(cid:96) ≥ 0, A1 = 1,
(16)
with 1 being a column-vector with all entries equal to 1. We denote the n-th power of A by
Bn = [bk,(cid:96)(n)] (cid:44) An.
(17)
Throughout this article, we assume that A has second largest eigenvalue magnitude strictly less
than one, which yields [15], [36]:
where the limiting combination weights {p(cid:96)} satisfy, for all (cid:96) = 1, 2, . . . , S:
bk,(cid:96)(n) n→∞−→ p(cid:96) ⇔ Bn
S(cid:88)
pA = p,
p(cid:96) > 0,
n→∞−→ 1p
p(cid:96) = 1
(18)
(19)
and the row vector p = [p1, p2, . . . , pS] is usually referred to as the Perron eigenvector of A --
see, e.g., [7]. We remark that the condition on A is automatically satisfied by network topologies
that are strongly-connected [7], i.e., when there is always a path with nonzero weights between
(cid:96)=1
October 28, 2018
DRAFT
10
any pair of nodes, and at least one node in the network has a self-loop (ak,k > 0 for some agent
k).
B. Steady-State Distribution
In order to design and characterize an inference system based upon the sensor output yk(n),
knowledge of the distribution of yk(n) is crucial. However this knowledge is seldom available,
except for very special cases (e.g., Gaussian observations). A common and well-established
approach in the adaptation literature [6], [37] to address this difficulty is to focus on i) the
steady-state properties (as n → ∞), and ii) the small step-size regime (µ → 0). Accordingly,
throughout the paper, the term steady-state refers to the limit as the time-index n goes to infinity,
while the term asymptotic refers to the slow adaptation regime where µ → 0.
We start by considering the steady-state behavior of yk(n) for a given step-size µ. To this
aim, it is useful to recast the pair of equations given by (14) and (15) into the following single
equation, which is obtained by straightforward algebra:
yk(n) = (1 − µ)n
bk,(cid:96)(n)y(cid:96)(0)
+
S(cid:88)
(cid:123)(cid:122)
(cid:96)=1
transient term
(cid:124)
n(cid:88)
S(cid:88)
(cid:125)
µ(1 − µ)i−1bk,(cid:96)(i)x(cid:96)(n − i + 1).
i=1
(cid:96)=1
(20)
Since the random variables xk(n) are i.i.d. across time, and since we shall be only concerned
with the distribution of partial sums involving these terms, it is convenient to define the following
random variable:
µ(1 − µ)i−1bk,(cid:96)(i)x(cid:96)(i),
(21)
k(n) (cid:44) n(cid:88)
y(cid:63)
S(cid:88)
i=1
(cid:96)=1
n(cid:88)
S(cid:88)
which shares the same distribution of the second term on the RHS of (20) -- see also the
discussion in [1], [2]. Formally:
k(n) d=
y(cid:63)
µ(1 − µ)i−1bk,(cid:96)(i)x(cid:96)(n − i + 1),
where the notation d= denotes equality in distribution.
i=1
(cid:96)=1
October 28, 2018
(22)
DRAFT
In Theorem 1 of [1], [2], the existence of a steady-state random variable characterizing the
diffusion output has been established. This can be summarized by the following statement (the
symbol (cid:32) means convergence in distribution):
∞(cid:88)
S(cid:88)
yk(n) n→∞(cid:32) y(cid:63)
k,µ
(cid:44)
µ (1 − µ)i−1bk,(cid:96)(i)x(cid:96)(i)
11
(23)
(24)
(25)
where the first two moments of y(cid:63)
k,µ are given by
i=1
(cid:96)=1
E[y(cid:63)
k,µ] = E[x],
VAR[y(cid:63)
k,µ] = σ2
x
∞(cid:88)
S(cid:88)
µ2(1 − µ)2(i−1)b2
k,(cid:96)(i).
i=1
(cid:96)=1
Actually, the proof in [1], [2] focused on the case of doubly-stochastic combination matrices, but
it is immediate to verify that the same argument holds for right-stochastic combination matrices
and leads to (23) -- (25).
C. The Inference Problem
In the distributed detection formulation, each agent in the network must perform a binary
hypothesis test, in an adaptive and fully decentralized manner. In this setting, the agents in the
network continually collect an increasing amount of streaming data, whose statistical properties
depend upon an unknown binary state of nature, which is represented by a pair of hypotheses, say,
H0 and H1. The statistics xk(n) are spatially and temporally i.i.d., conditioned on the hypothesis
that gives rise to them. In what follows, we shall always assume that E0[x] (cid:54)= E1[x]. To get some
intuition about the meaning of this condition, consider that, when the local statistic x is chosen
as the log-likelihood of the local observation, the inequality E0[x] (cid:54)= E1[x] is simply a way to
state that the detection problem is identifiable, namely, that it is not singular [38]. Furthermore,
and without loss of generality, we assume that:
E0[x] < E1[x].
(26)
At time n, the k-th sensor needs to produce a decision about the state of nature, based upon
its state value yk(n). As discussed in [1], [2], the computation of yk(n) via algorithm (14)
and (15) is motivated by the fact that this implementation essentially results in a value yk(n) that
corresponds to a weighted average of the local statistics xk(n). Such additive constructions for
October 28, 2018
DRAFT
12
the state variable are not only convenient from an implementation point of view, but they are also
meaningful from a theoretical standpoint. Indeed, in the classical, centralized and non-adaptive
theory with i.i.d. data (where the optimal detection statistic is the sum of the log-likelihoods), as
well as in more general frameworks such as locally optimum detection or universal hypothesis
testing [38], [39], using additive detection statistics is the best choice. Moreover, the decision
regions employed by the detector are often in the form of single-threshold rules, which, for a
variety of detection problems, exhibit asymptotic optimality properties also in our adaptive and
distributed setting, as already shown in [1], [2]. Motivated by these considerations, in this work
the test implemented by agent k at time n is therefore chosen to be of the form:
yk(n)
(27)
The performance of the test is typically expressed in terms of the Type-I (choose H1 when
H0 is true) and Type-II (choose H0 when H1 is true) error probabilities. With reference to the
steady-state performance, the Type-I and Type-II error probabilities are defined as, respectively:
H0(cid:81)
H1
γ.
αk,µ (cid:44) P0[y(cid:63)
k,µ > γ],
βk,µ (cid:44) P1[y(cid:63)
k,µ ≤ γ].
(28)
III. MAIN THEOREMS
We now introduce the basic quantities necessary to characterize the system performance. As
we shall see, it is important to understand the behavior of the Logarithmic Moment Generating
Function (LMGF) of the steady-state variable y(cid:63)
k,µ. In Appendix A we state and prove a theorem
(Theorem 1) that establishes several useful properties of the LMGF, of its derivatives, and of
its limiting behavior as µ → 0. In the following description we shall mention and use some of
these properties, referring the reader to Appendix A for a more detailed explanation.
• Local LMGF.
The LMGF of the local statistics xk(n) is defined as:
ψ(t) (cid:44) ln E[etxk(n)],
(29)
and is independent of k. A fundamental role in our results will be played by the following
averaged version of the LMGF ψ(t), which has already been used in [1], [2]:
ω(t) (cid:44)
October 28, 2018
dτ.
(30)
DRAFT
(cid:90) t
ψ(τ )
τ
0
• Steady-state LMGF.
The LMGF of the steady-state variable y(cid:63)
k,µ is defined as:
φk,µ(t) (cid:44) ln E[ety(cid:63)
k,µ],
and admits the following representation [Theorem 1, Eq. (88)]:
∞(cid:88)
S(cid:88)
ψ(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)t(cid:1) .
φk,µ(t) =
• Limiting properties of the steady-state LMGF.
i=1
(cid:96)=1
We will be primarily concerned with the limiting normalized LMGF:
φ(t) (cid:44) lim
µ→0
µ φk,µ(t/µ),
where the above limit exists, and is given by [Theorem 1, Eq. (89)]:
S(cid:88)
(cid:90) p(cid:96)t
S(cid:88)
φ(t) =
ω(p(cid:96)t) =
(cid:96)=1
(cid:96)=1
0
ψ(τ )
τ
dτ
13
(31)
(32)
(33)
(34)
Equation (34) emphasizes that φ(t) depends on the underlying statistical model through
the LMGF ψ(t) of the local statistics xk(n), and on the network topology only through
the Perron eigenvector p, i.e., on the limiting combination weights. For doubly-stochastic
matrices we have p(cid:96) = 1/S, so that the above formula gives
φ(t) = Sω(t/S) = S
ψ(τ )
τ
dτ,
(35)
which is consistent with the results of [1], [2].
It is of interest to consider an alternative representation for (34). To this aim, we introduce
(cid:90) t/S
0
the LMGF of the averaged random variable(cid:80)S
¯ψ(t) (cid:44) S(cid:88)
(cid:90) t
and by straightforward calculations we get from (34):
(cid:96)=1
ψ(p(cid:96)t),
φ(t) =
¯ψ(τ )
dτ
τ
0
(cid:96)=1 p(cid:96) x(cid:96)(n), namely:
(36)
(37)
• Property of the LMGF derivatives.
In Theorem 1, part ii), we establish that the derivatives of φk,µ(t) and of the limiting
October 28, 2018
DRAFT
function φ(t) can be evaluated by interchanging the differential and limit operators, yielding,
in particular [Theorem 1, Eq. (90)]:
φ(cid:48)
k,µ(t) = µ
i=1
(cid:96)=1
(1 − µ)i−1bk,(cid:96)(i) ×
S(cid:88)
∞(cid:88)
ψ(cid:48)(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)t(cid:1) ,
S(cid:88)
1
t
(cid:96)=1
ψ(p(cid:96)t).
14
(38)
(39)
and [Theorem 1, Eq. (91)]
lim
µ→0
• Convergence errors.
φ(cid:48)
k,µ(t/µ) = φ(cid:48)(t) =
As we shall see, the convergence errors corresponding to (33) and (39) will play an important
role in characterizing the error probabilities. The following rates of convergence [Theorem
1, Eq. (93)] hold:
φ(t) − µ φk,µ(t/µ) = O(µ)
φ(cid:48)(t) − φ(cid:48)
k,µ(t/µ) = O(µ),
where the notation fµ = O(µ) means that the ratio fµ/µ stays bounded as µ → 0.
• Fenchel-Legendre transforms.
The Fenchel-Legendre transform of the function φ(t) is defined by [34], [35]:
Φ(γ) (cid:44) sup
t∈R
[γt − φ(t)].
(40)
(41)
(42)
We shall use capital letters to denote Fenchel-Legendre transforms, as done in (42). It is
further useful to introduce the domain where Φ(γ) is finite, namely:
DΦ = {γ ∈ R : Φ(γ) < ∞}.
(43)
The notation Do
Φ adopted in the sequel will denote the interior of the set DΦ.
We can now state our second theorem, which generalizes Theorem 3 from [1], [2] to handle the
case of right-stochastic matrices.
THEOREM 2 (Large deviations of y(cid:63)
that ψ(t) < +∞ for all t ∈ R. Then, for all k = 1, 2, . . . , S:
k,µ as µ → 0, for right-stochastic matrices). Assume
October 28, 2018
DRAFT
i) The steady-state variable y(cid:63)
k,µ obeys a Large Deviation Principle (LDP) with rate function
given by the Fenchel-Legendre transform Φ(γ) defined by (42), namely, for any Borel
measurable region Γ ∈ R:
15
µ ln P[y(cid:63)
k,µ ∈ Γ] = − inf
γ∈Γ
(44)
ii) φ(cid:48)(cid:48)(t) > 0 for all t ∈ R, implying that φ(t) is strictly convex. Moreover, the rate function
Φ, and attains its unique minimum at γ = E[x], with Φ(E[x]) =
Φ(γ).
lim
µ→0
Φ(γ) is strictly convex in Do
0.
Proof: By Theorem 1, part i), Appendix A, we can conclude that the function φ(t) in (33) exists
and is finite for all t ∈ R. Since, by definition, φ(t) is a normalized limiting LMGF, claim i) of
the present theorem follows by application of the Gartner-Ellis Theorem [34], [35].
Claim ii) is a direct extension to the case of right-stochastic combination matrices of the
results proved for doubly-stochastic combination matrices in Appendix C of [1].
(cid:3)
k,µ obeys a LDP with
According to property i) in Theorem 2, the steady-state random variable y(cid:63)
rate function Φ(γ). In view of the convexity properties of Φ(γ) stated in ii), for any γ ∈ Do
Φ,
with γ > E[x], the infimum over an interval of the form (γ,∞) always lies on the boundary
point γ. The same conclusion is reached for the infimum over an interval of the form (−∞, γ)
when γ < E[x]. Formally, for any γ ∈ Do
Φ, we have:
µ ln P[y(cid:63)
µ ln P[y(cid:63)
k,µ > γ] = −Φ(γ),
k,µ ≤ γ] = −Φ(γ),
lim
µ→0
lim
µ→0
γ > E[x]
γ < E[x].
(45)
(46)
Note also that, according to Theorem 2, part ii), the choice γ = E[x] yields Φ(γ) = Φ(E[x]) = 0,
that is, a null exponent. This is the uninteresting case where the error probability does not
vanish exponentially. Accordingly, this situation will be ruled out from the forthcoming analysis.
Relationships (45) and (46), along with the latter observation, have an immediate implication as
regards the choice of the detection threshold. Indeed, in the light of (28), one must ensure that
E0[x] < γ < E1[x],
(47)
DRAFT
October 28, 2018
16
in order to guarantee the exponential decay of both error probabilities αk,µ and βk,µ.
The exact asymptotics of y(cid:63)
k,µ are now characterized in the following theorem.
THEOREM 3 (Exact asymptotics of y(cid:63)
that ψ(t) < +∞ for all t ∈ R. Let γ ∈ Do
the stationary equation:
k,µ as µ → 0). Assume that xk(n) is not lattice, and
Φ, with γ > E[x], and let θγ be the unique solution to
Then, θγ > 0 and, for k = 1, 2, . . . , S:
φ(cid:48)(θγ) = γ.
P[y(cid:63)
k,µ > γ] ∼ Pk,µ(γ),
Pk,µ(γ) (cid:44)(cid:114)
µ
γ φ(cid:48)(cid:48)(θγ)
2πθ2
e
− 1
µ[Φ(γ)+k,µ(θγ )]
where
with
k,µ(t) = φ(t) − µφk,µ(t/µ).
The correction term k,µ(t) can be refined to:
k,µ(t) = [φ(t) − µφk,µ(t/µ)] +
[φ(cid:48)(t) − φ(cid:48)
k,µ(t/µ)]2
2 φ(cid:48)(cid:48)(t)
(48)
(49)
(50)
(51a)
(51b)
.
Proof: See Appendix B.
(cid:3)
REMARK I. The key ingredients to computing Pk,µ(γ) in (50) are the moment generating
function ψ(t) of the local statistics xk(n), and the combination matrix A. Indeed, the quantities
Φ(γ) and θγ depend on the function φ(t), which in turn depends on ψ(t) and on the limiting
combination weights p(cid:96). The correction term k,µ(t) (in both versions) depends on ψ(t) and on
the actual combination weights bk,(cid:96)(i).
(cid:3)
REMARK II. It is useful to comment on the reason for reporting two expressions for k,µ(t).
First, note that, from (41), the quantity
[φ(cid:48)(t) − φ(cid:48)
k,µ(t/µ)]2
2 φ(cid:48)(cid:48)(t)
=
µ
2 φ(cid:48)(cid:48)(t)
1
µ
October 28, 2018
(cid:20)φ(cid:48)(t) − φ(cid:48)
(cid:21)2
k,µ(t/µ)
µ
(52)
DRAFT
17
vanishes as µ → 0, implying that, asymptotically, using (51a) or (51b) in (50) makes no
difference. This equivalence is not surprising because, in principle, one can construct an ar-
bitrary number of distinct approximations that are asymptotically equivalent. The relevant fact
is that different forms for the correction correspond to different approximations of the true error
probability, which will perform differently for finite values of 1/µ. We shall see in the proof of
Theorem 3 why (51b) is a refined version of (51a). From now on, unless otherwise stated, we
shall make reference to (51b).
(cid:3)
REMARK III. Applying (40) and (41), it is immediately seen that:
(53)
meaning that the overall correction appearing at the exponent in (50) remains bounded as µ → 0.
(cid:3)
k,µ(θγ) = O(µ),
REMARK IV. The theorem is stated with reference to large deviations of the type P[y(cid:63)
k,µ > γ],
with γ > E[x]. We notice that the main formulas (50), (51a) and (51b), for the complementary
k,µ ≤ γ] remain unchanged. The only difference in the claim is that θγ < 0.
case γ < E[x] and P[y(cid:63)
These facts can be verified by repeating the proof of the theorem for this complementary case.
Alternatively, they can be verified by directly applying the claim of Theorem 3 to the random
variable −y(cid:63)
k,µ, and by observing that the difference between < and ≤ is immaterial in the proof
of the theorem.
(cid:3)
REMARK V. The proof of Theorem 3 assumes that the local statistic xk(n) is not of lattice
type. We recall that the distribution of a lattice random variable with span ∆ is concentrated
at the points d, d ± ∆, d ± 2∆, . . . for a certain real number d -- see, e.g., [40]. The use of
the non-lattice assumption in our proof is discussed in the Remark VI at the end of Appendix
B. It is useful to observe that, when dealing with the classical and simplest case of normalized
sums of i.i.d. random variables, the exact asymptotics for the lattice case must take into account
a further correction term that is related to the lattice span [34]. However, as already observed,
our asymptotic setting with vanishing step-size is markedly different from that addressed in the
literature for the problem of normalized sums of i.i.d. random variables. It is therefore not easy
to anticipate if the results of Theorem 3 hold as they are also for the lattice case, if they hold
October 28, 2018
DRAFT
as they are under additional conditions, or if further corrections are needed.
18
(cid:3)
Before ending this section it is useful to collect the steps needed to evaluate (50). Depending
on the particular application, the formulas in this listing might need to be evaluated numerically.
For practical purposes, the infinite summations appearing in k,µ(·) must be obviously truncated.
Evaluation of Pk,µ(γ) in (50).
1) Find the solution θγ to the stationary equation:
φ(cid:48)(θγ) =
1
θγ
2) Compute the rate function:
S(cid:88)
(cid:96)=1
ψ(p(cid:96)θγ) = γ.
Φ(γ) = sup
t∈R
[γt − φ(t)] = γθγ − φ(θγ).
3) Compute the second derivative:
φ(cid:48)(cid:48)(θγ) =
1
θ2
γ
4) Compute the convergence errors (40) and (41) by using (32) and (38):
(cid:96)=1
S(cid:88)
∞(cid:88)
∞(cid:88)
i=1
[(p(cid:96)θγ)ψ(cid:48)(p(cid:96)θγ) − ψ(p(cid:96)θγ)].
C1 = φ(θγ) − µ
C2 = φ(cid:48)(θγ) − µ
(cid:1) ,
S(cid:88)
S(cid:88)
ψ(cid:48)(cid:0)(1 − µ)i−1bk,(cid:96)(i)θγ
(cid:96)=1
ψ(cid:0)(1 − µ)i−1bk,(cid:96)(i)θγ
(cid:1) .
(1 − µ)i−1bk,(cid:96)(i) ×
(cid:96)=1
i=1
5) Compute the correction term in (51b):
k,µ(θγ) = C1 +
6) Using the above quantities, evaluate Pk,µ(γ).
C2
2
2 φ(cid:48)(cid:48)(θγ)
.
October 28, 2018
DRAFT
19
IV. IMPORTANCE SAMPLING AND CRAM ´ER'S TRANSFORM
Real-world detectors are usually required to be high-performing, i.e., they must exhibit very
low error probabilities. Unfortunately, standard Monte Carlo techniques to estimate their perfor-
mance become unfeasible for error probability values in the order of 10−6. To overcome this
issue, we shall resort to importance sampling techniques [42], which can dramatically reduce the
number of runs needed to reach a prescribed level of estimation accuracy. We start by illustrating
briefly the importance sampling philosophy. Then, in the next section, we show how it should
be applied to the adaptive distributed detection problem.
A. Cram´er's Transform
Let us refer to a random variable y, assumed continuous for ease of description. We shall
denote by f (y) its probability density function (pdf). Consider now another pdf f (y) that does
not vanish (except for zero-measure sets) when f (y) > 0, and introduce the following weighting
function:
w(y) =
,
(54)
(55)
where IE is the indicator of an event E, and E f [·] denotes expectation computed over the
transformed pdf f (y). The above equation shows that the quantity to be estimated can be regarded
as the expectation, under the transformed pdf, of the indicator of the event {y > γ}, weighted by
the function w(y). The rationale behind importance sampling is that, by an appropriate choice of
the weighting function, it is possible to map an event that is rare under the original sampling pdf
f (y), into an event that is not rare under the new sampling pdf f (y). In this way, the number of
Monte Carlo iterations needed to estimate the expectation is reduced, because (important) samples
are generated around the body (not the tail) of the new distribution. An accurate estimate of the
probability tails is enabled by the weighting function w(y).
When working with random variables obeying a LDP, there is a classical way to select the
transformed pdf f (y). This is usually referred to as exponential twisting of f (y), and amounts
October 28, 2018
DRAFT
that is, the likelihood ratio between f (y) and f (y). It then holds that:
P[y > γ] =
w(y) f (y)dy
f (y)
f (y)
(cid:90) ∞
(cid:90) ∞
(cid:2)w(y)I{y>γ}(cid:3) ,
f (y)dy =
γ
γ
= E f
to selecting [34], [35]:
f (y) = eηy−ln E[eηy] f (y) ⇔ w(y) = e−ηy+ln E[eηy].
20
(56)
When y is not a continuous random variable, the exponential twisting can be rephrased in terms
of probability measures m and m as:
m(dy) = eηy−ln E[eηy] m(dy).
(57)
The results presented in this section hold, mutatis mutandis, for this setting.
The aforementioned change of measure was originally proposed by Cram´er [41] to compute
large deviations exponents, and is accordingly also known as Cram´er's transform. The choice
of a given exponential twisting (i.e., the choice of the parameter η) is critical in determining
the accuracy of the estimates produced by the importance sampling algorithm. Interestingly,
theoretical studies suggest to use for importance sampling in general exactly the same exponential
twisting needed to compute the error exponents -- see, e.g., [42]. For the classical and simplest
case of summation of i.i.d. random variables, the latter kind of exponential twisting is well-known
-- see, e.g., [34], [35]. For our specific problem of adaptive distributed detection, the solution
is more involved, and will be discussed in the next section. In particular, it will be shown that
the theoretical results established in this paper are crucial to enable an accurate design of the
importance sampling simulations.
B. Importance Sampling for Adaptive Distributed Detection
Since we are interested in evaluating the steady-state performance of the adaptive distributed
algorithm, it is sufficient to examine the behavior, for n sufficiently large, of the random variable
k(n) introduced in (21), which, according to (22), has the same distribution of the diffusion
y(cid:63)
output with null transient.
In order to implement the importance sampling recipe, we need to perform the exponential
twisting (56). As already mentioned, the parameter η will be chosen as that value corresponding
to the change of measure used to compute the error exponents. For the specific case of adaptive
distributed detection, this exponential twisting is extensively discussed and employed in Appendix
B -- see, e.g., (143) and (178), and amounts to selecting η = θγ/µ, where θγ is the solution to
October 28, 2018
DRAFT
the stationary equation (48) in Theorem 3. Denoting by
φk,µ(t; n) (cid:44) ln E[ety(cid:63)
k(n)] =
n(cid:88)
S(cid:88)
i=1
(cid:96)=1
ψ(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)t(cid:1) ,
21
(58)
the LMGF of the random variable y(cid:63)
assumption on the {x(cid:96)(i)}), the choice η = θγ/µ applied in (56) then yields:
k(n) (the additive form above comes simply from the i.i.d.
(cid:16) θγ
(cid:17)
w(y) = e
− θγ
µ y+φk,µ
µ ;n
.
(59)
Unfortunately, in order to generate samples according to the transformed pdf f (y) in (56), the
weighting function w(y) is not sufficient, since one needs to know also the pdf f (y) of the
random variable y(cid:63)
k(n). As already observed, knowledge of this distribution is seldom available.
To overcome this issue, one may consider running first a Monte Carlo simulation by generating
random instances of the local statistics x(cid:96)(i), and then evaluating y(cid:63)
k(n) through (21). However,
recall that we are interested in implementing importance sampling. This means that we must
k(n)) that would produce
find the appropriate change of measure applied to x(cid:96)(i) (and not to y(cid:63)
the desired exponential change of measure (59) on the random variable y(cid:63)
k(n). To this aim, let
us denote by p(x) the pdf of x(cid:96)(i). We now show that the desired goal can be achieved by
drawing the (i, (cid:96))-th sample x(cid:96)(i) from the pdf
with
pi,(cid:96)(x) = eηi,(cid:96)x−ψ(ηi,(cid:96))p(x),
ηi,(cid:96) (cid:44) (1 − µ)i−1bk,(cid:96)(i) θγ.
(60)
(61)
All these samples are still generated independently, but now they are no longer identically
distributed. Note also that the above transformation depends upon the index k of the agent under
consideration, even if the subscript has been suppressed for ease of notation.
To see why the choice (60) will correspond to (59), let us introduce the joint ensemble
X = {x(cid:96)(i)}, for i = 1, 2, . . . , n and for (cid:96) = 1, 2, . . . , S. The expectation in (55) can then be
rewritten in terms of the distribution of X as follows:
k(n)>γ}] = Ep
k(n) > γ] = E[I{y(cid:63)
(cid:20)p(X)
k(n)>γ}
I{y(cid:63)
P[y(cid:63)
(62)
(cid:21)
.
p(X)
October 28, 2018
DRAFT
But since the x(cid:96)(i) are spatially and temporally i.i.d., we have from (60) and (59):
= e−(cid:80)n
i=1
p(X)
p(X)
(cid:80)S
(cid:80)S
(cid:96)=1 ηi,(cid:96)x(cid:96)(i)+(cid:80)n
(cid:17)
(cid:16) θγ
i=1
(cid:96)=1 ψ(ηi,(cid:96))
− θγ
µ y(cid:63)
k(n)+φk,µ
µ ;n
= e
= w(y(cid:63)
k(n)),
22
(63)
where, in the last equality, we applied definitions (21) and (58). This result shows that (62)
corresponds to the expectation in (55) with the pdf f (y) chosen as in (59).
Before ending this section, we stress that the above considerations show one further important
benefit: our large deviations results about distributed detection over adaptive networks are also
very useful in performing a careful design of a Monte Carlo simulator based on importance
sampling.
A. Network Topology and Combination Weights
V. ILLUSTRATIVE DESIGN
We consider a network made of S = 10 sensors, arranged so as to form the topology in the
inset of Fig. 1. Given the topology, two different combination matrices will be tested. The first
one is defined by the so-called Metropolis rule [7]. Denoting by Nk the neighborhood of the
k-th agent (including k itself), and by nk the cardinality Nk (aka the degree of the k-th agent),
the Metropolis rule is defined by:
ak,(cid:96) =
1/ max{nk, n(cid:96)},
1 − (cid:88)
ak,m,
m∈Nk\{k}
0,
(cid:96) ∈ Nk \ {k},
(cid:96) = k,
(cid:96) /∈ Nk.
(64)
This choice provides a doubly-stochastic A, and, hence, the corresponding Perron eigenvector
has uniform entries, p(cid:96) = 1/S for all (cid:96) = 1, 2, . . . , S.
The second combination policy is the uniform averaging rule [7]:
ak,(cid:96) =
October 28, 2018
1/nk,
0,
(cid:96) ∈ Nk,
(cid:96) /∈ Nk.
(65)
DRAFT
This choice provides a right-stochastic A, whose Perron eigenvector is available in closed form,
and has entries given by [7]:
p(cid:96) =
,
(cid:96) = 1, 2, . . . , S.
(66)
n(cid:96)(cid:80)S
m=1 nm
23
B. Hypothesis Test, Local Statistics and Simulation
We examine the following canonical shift-in-mean detection problem with noise distributed
according to a Laplace distribution, which is considered in [1], [2] as well. The Laplace pdf
(with scale parameter set to 1, without loss of generality) will be denoted by:
The hypothesis test can then be formulated as follows:
L (d) =
e−d.
1
2
H0
H1
: dk(n) ∼ L (d),
: dk(n) ∼ L (d − ρ),
(67)
(68)
(69)
where dk(n) denotes the measurement collected by agent k at time n, and ρ > 0 is the shift-in-
mean parameter. The local statistics xk(n) are chosen as the local log-likelihood ratios:
xk(n) = ln
= dk(n) − dk(n) − ρ.
(70)
Even if the above relationship shows clearly how to obtain the random variable xk(n) from the
knowledge of dk(n), it is useful, for later use, to evaluate explicitly the distribution of xk(n)
under the two hypotheses. To this aim, we rewrite (70) as:
(cid:19)
L (dk(n))
(cid:18)L (dk(n) − ρ)
−ρ,
+ρ,
2dk(n) − ρ,
xk(n) =
dk(n) < 0,
dk(n) > ρ,
dk(n) ∈ [0, ρ],
(71)
which is obtained by using, in the three ranges considered in (71), the explicit definition of the
absolute values appearing in (70). We see then that xk(n) is a random variable of mixed type,
taking values in the range [−ρ, ρ], and with two atoms located at ±ρ. For ease of description,
we find convenient to use, for random variables of mixed type, the generalized pdf written using
the Dirac-delta function δ(x).
Accordingly, let p0(x) and p1(x) denote the generalized pdfs of xk(n) under H0 and H1,
respectively. For a shift-in-mean with respect to a symmetric pdf (as L (d) is), it is well-known
October 28, 2018
DRAFT
that the log-likelihood ratio exhibits the symmetry property: p1(x) = p0(−x) -- see, e.g., [44].
Thus, it suffices to focus on p0(x). To this aim, we observe that, in view of (71), the cumulative
distribution function of xk(n) is
24
(cid:20)
dk(n) ≤ x + ρ
2
(cid:21)
0,
P0
1,
x < −ρ,
, x ∈ [−ρ, ρ),
(72)
x ≥ ρ,
P0[xk(n) ≤ x] =
which corresponds to the following generalized pdf:
p0(x) =
(cid:18) x
(cid:19)
(cid:18) x + ρ
(cid:19)
P0[dk(n) < 0]δ(x + ρ) + P0[dk(n) > ρ]δ(x − ρ) +
1
L
2
1
2
2ρ
δ(x − ρ) +
(cid:18) x
Π
e−ρ
2
δ(x + ρ) +
(cid:19)
e− x+ρ
2 Π
1
4
2ρ
2
,
=
(73)
where Π(x) is a unit-width rectangular window centered at 0. The corresponding LMGF of
xk(n) is computable in closed form:
(cid:18)(cid:90) ∞
(cid:19)
ψ0(t) = ln E0[etxk(n)] = ln
e(t−1)ρ
(cid:18) e−tρ
= ln
+
2
2
+
2
etxp0(x)dx
−∞
e−ρ/2ρ
sinch[ρ(t − 1/2)]
(cid:19)
,
(74)
where
sinch(x) (cid:44) sinh(x)
,
x
sinch(0) = 1.
(75)
Since, as observed, p1(x) = p0(−x), the LMGF under H1 is easily obtained as ψ1(t) = ψ0(−t).
We note in passing that the above explanation and the following derivations can be restated
in a more formal language by using, for the mixed-type random variable xk(n), a probability
measure made of the superposition of two singular, atomic measures (with masses located at
±ρ), and an absolutely continuous measure with density given by the third term in (73).
Before concluding this section, it remains to show how to implement, for the considered
Laplace example, the importance sampling method described in Sec. IV. We shall focus on
October 28, 2018
DRAFT
hypothesis H0, and, again, the results for H1 can be simply obtained from the relationship
p1(x) = p0(−x). Applying an exponential twisting with parameter η to the generalized pdf
in (73), we get:
25
p0(x) = eηx−ψ0(η)p0(x)
=
e−ηρ
2eψ0(η) δ(x + ρ) +
e(η−1)ρ
(cid:18) x
2eψ0(η) δ(x − ρ) +
e−ρ/2
4eψ0(η) ex(η−1/2)Π
2ρ
(cid:19)
.
Introducing the definitions:
p− =
e−ηρ
2eψ0(η) ,
p+ =
e(η−1)ρ
2eψ0(η) ,
and using (74), by simple algebra relation (76) becomes:
p0(x) = p−δ(x + ρ) + p+δ(x − ρ) + [1 − p− − p+] ×
(cid:18) x
(cid:19)
ex(η−1/2)
1
2ρ
sinch[ρ(η − 1/2)]
Π
2ρ
.
(76)
(77)
(78)
Generating a random variable xk(n) according to the distribution p0(x) is now an easy task. As
a matter of fact, (78) reveals that p0(x) is a mixture of three components such that xk(n) takes
on the value −ρ with probability p−, the value +ρ with probability p+, and that otherwise it
must be sampled from the pdf
(cid:18) x
(cid:19)
2ρ
.
(79)
ex(η−1/2)
sinch[ρ(η − 1/2)]
1
2ρ
Π
C. A Normal Approximation
In the following analysis, we shall compare the error probabilities estimated empirically, to
the refined asymptotic formulas obtained in the present work. As a further term of comparison,
we would like to add a normal approximation that will follow from the asymptotic normality
result proved in [1], [2]. Actually, this result was obtained there for doubly-stochastic connection
matrices, but the generalization to the case of right-stochastic matrices comes essentially at no
cost and can be stated as:
k,µ − E[x]
y(cid:63)
(cid:112)µ σ2
lim
µ→0(cid:32) N (0, 1)
October 28, 2018
(80)
DRAFT
where the limiting variance σ2
(i.e., the variance) of y(cid:63)
lim can be obtained by applying result (92) to the second cumulant
k,µ, which yields:
VAR[y(cid:63)
k,µ]
µ
S(cid:88)
(cid:96)=1
µ→0−→ σ2
x
2
(cid:44) σ2
lim.
p2
(cid:96)
(81)
26
We emphasize that this result is consistent with previous findings obtained in the context of
mean-square-error estimation -- see, e.g., [6].
We see from (81) that the ratio VAR[y(cid:63)
lim) converges to one as µ goes to zero. In
view of Slutsky's Theorem [47], this fact implies that the following alternative version of the
convergence in distribution in (80) holds:
k,µ]/(µσ2
(cid:113)
k,µ − E[x]
y(cid:63)
VAR[y(cid:63)
k,µ]
µ→0(cid:32) N (0, 1)
(82)
While the two formulations (80) and (82) are asymptotically equivalent, it is expected that (82)
offers a better performance since it replaces the asymptotic variance with the actual one. The
relationship shown in (82) suggests the following way to approximate the probability P[y(cid:63)
k,µ > γ]:
γ − E[x]
(cid:113)
VAR[y(cid:63)
k,µ]
,
P[y(cid:63)
k,µ > γ] ≈ Q
(83)
where Q(·) denotes the complementary cumulative distribution function of a standard normal
distribution.
D. Analysis of the Results
Let us now examine the adaptive distributed network of detectors in operation, by reporting
the evidence arising from our Monte Carlo analysis. We refer to a sufficiently large time horizon,
such that the steady-state assumption applies, and evaluate the error probabilities for different
values of the step-size.
The exact asymptotics provided by Theorem 3 will be computed by implementing the six-
steps recipe described at the end of Sec. III. Clearly, in doing so, we must use the LMGF ψ0(t)
if we are working under H0, and the LMGF ψ1(t) if we are working under H1. The normal
approximation will be instead obtained as described in the previous section -- see (83).
In the two examples that we are going to discuss we choose the detection threshold as detailed
in [1], [2], obtaining γ = 0. This implies, by the symmetry property p1(x) = p0(−x), that the
October 28, 2018
DRAFT
27
Fig. 1.
Laplace example discussed in Sec. V-B, with ρ = 0.6. The network topology is depicted in the inset plot, and the
combination weights ak,(cid:96) follow the Metropolis rule (64). The performance of agents 2, 4 and 5 is displayed. Dots refer to the
empirical steady-state error probabilities αk,µ = βk,µ at different sensors, obtained via Monte Carlo simulation with importance
sampling, as described in IV. Solid curves refer to the exact asymptotics provided by Theorem 3. Dashed curves refer to the
normal approximation (83).
error probabilities of first and second kind defined by (28) are equal, namely:
αk,µ = βk,µ.
(84)
Consistently, in the following description, the terminologies "error probability" and "error ex-
ponent" refer to any of these errors.
We start by considering the doubly-stochastic combination matrix obtained with the Metropolis
rule (64). In Fig. 1, the performance of the agents is displayed as a function of 1/µ, and
different agents are marked with different colors. The main system features, which were already
commented in [1], [2], are here summarized.
October 28, 2018
DRAFT
1020304050!14!12!10!8!6!4!21ΜErrorprobabilitySolid:Pk,ΜΓDashed:GaussianDots:Simulation12345678910 28
Fig. 2.
Laplace example discussed in Sec. V-B, with ρ = 0.6. The network topology is depicted in the inset plot, and the
combination weights ak,(cid:96) follow the averaging rule (65). The performance of agents 2, 4 and 5 is displayed. Dots refer to the
empirical steady-state error probabilities αk,µ = βk,µ at different sensors, obtained via Monte Carlo simulation with importance
sampling, as described in IV. Solid curves refer to the exact asymptotics provided by Theorem 3. Dashed curves refer to the
normal approximation (83).
The first evidence is that the different curves pertaining to different agents stay nearly parallel
for sufficiently small values of the step-size µ, i.e., the detection error probabilities at different
sensors vanish exponentially as functions of 1/µ, sharing the same detection error exponent.
This was the main result revealed by the large deviations analysis performed in [1], [2].
However, as already noticed in [1], [2], the large deviations tool is not powerful enough to
capture an important feature of the distributed behavior. Indeed, the second evidence emerging
from the simulations is that the error probability curves in Fig. 1 are basically ordered, and
the ordering is closely related to the network connection structure. Comparing the detection
performance of three specific agents, namely, agents 2, 4, 5, we see that the ordering reflects the
October 28, 2018
DRAFT
1020304050!14!12!10!8!6!4!21ΜErrorprobabilitySolid:Pk,ΜΓDashed:GaussianDots:Simulation12345678910 29
degree of connectivity of each agent. For instance, agent 4 has the highest number of neighbors,
and its performance is the best one, while agent 2 is the most isolated, and its error probability
curve appears consistently as the highest one. According to what one expects, agent 5 is in an
intermediate position. The new fact here is that, using the results of the current manuscript, we
are now able to provide a systematic analysis of the above features, as well as of the exact
interplay with network connectivity.
First of all, what in [1], [2] was only a partial evidence arising from a particular numerical
experiment, emerges now, thanks to the refined asymptotic analysis, as the universal behavior
of adaptive distributed detection.
Moreover, the refined asymptotic approximations provided by Theorem 3 can be used to obtain
quantitative predictions of the actual system performance, as we proceed to explain. The refined
formulas are represented by the solid curves in Fig. 1. We see that the empirical probability
points (the dots) converge toward the theoretical solid curves as the step-size µ decreases (i.e.,
as we move to the right in the plot). Remarkably, the theoretical formulas provided by our
theorems are able to embody the dependencies between the network connection structure and
the detection performance at different sensors, as it is witnessed by the correct ordering of
the curves. In addition, a gap between the exact asymptotics (solid curves) and the normal
approximation (dashed curves) is clearly observed. This should come as no surprise, since the
normal approximation is expected to be accurate when working with small deviations, and must
accordingly provide a wrong prediction in the large deviations regime -- see also the discussion
in [1], [2].
We now switch to the analysis of the right-stochastic combination matrix provided by the
uniform averaging rule (65). The corresponding results are reported in Fig. 2. First, we note
that, as in the doubly-stochastic case, the error probability curves vanish exponentially fast as
functions of 1/µ, and stay nearly parallel as µ goes to zero. This is consistent with the prediction
that they share the same error exponent, as dictated by Theorem 2 for the general case of right-
stochastic combination matrices. Also in this example, we are able to appreciate the goodness
of the refined approximations obtained with Theorem 3, and the fact that the empirical points
depart from the normal approximation.
To get further insights, we now compare the detection performance of the two aforementioned
October 28, 2018
DRAFT
combination matrices. To begin with, we compute the error exponents pertaining to the two
systems. For the particular example considered, we obtain (recall that the threshold is set to
γ = 0):
ΦDS(0) ≈ 0.75 > ΦRS(0) ≈ 0.7,
30
(85)
for the doubly-stochastic and the right-stochastic case, respectively. These values appear to
suggest that that the doubly-stochastic combination policy asymptotically outperforms the right-
stochastic combination policy. This conclusion may somehow be expected because, asymptoti-
cally, a doubly-stochastic combination policy weights the local statistics equally, while a right-
stochastic combination policy does not. In the presence of i.i.d. observations, the former strategy
seems to be preferable. To see if this first-order analysis suffices, let us now apply the refined
asymptotic formulas.
Accordingly, in Fig. 3 we display the theoretical error probability curves obtained with
Theorem 3, for the Metropolis combination matrix (doubly-stochastic case, line and markers) and
for the uniform averaging combination matrix (right-stochastic case, solid curves). An interesting
behavior arises. We observe that the relative performance of the two different combination policies
depends strongly on the connectivity of the individual agent. For instance, for the well-connected
agent 4, the doubly-stochastic combination policy delivers superior performance, while exactly
the converse is true for the scarcely connected agent 2. Moreover, if we consider as network
performance the arithmetic average of the error probabilities (black curves), we see that the
right-stochastic combination policy is globally superior.
An explanation for this behavior is as follows. Denoting by λDS
2
the second largest
magnitude eigenvalues of the doubly-stochastic and right-stochastic combination matrices, re-
spectively, we see that
and λRS
2
2 ≈ 0.83 > λRS
λDS
2 ≈ 0.7,
(86)
implying that the right-stochastic weights will converge to the corresponding Perron eigenvector
faster than the doubly-stochastic weights. Examining the detection performance, we see from
here that the slower convergence of the doubly-stochastic combination matrix has a detrimental
effect on the less connected agents. We could say that the benefits of the higher (doubly-stochastic
case) exponent are more than compensated by the faster (right-stochastic case) convergence to
the steady-state behavior.
October 28, 2018
DRAFT
31
Fig. 3.
Laplace example discussed in Sec. V-B, with ρ = 0.6. The network topology is depicted in the inset plot, and the
performance corresponding to two combination matrices is compared, namely, the doubly-stochastic matrix obtained with the
Metropolis rule (64), and the right-stochastic matrix obtained with the uniform averaging rule (65). The performance of agents
2 and 4 is displayed, along with the average network performance, namely, the arithmetic mean of the error probabilities of all
agents (black). All curves are computed by using the exact asymptotics provided by Theorem 3.
Notice that the observed behavior is not in contrast with what is predicted by (85). Based on an
analysis at the first leading order in the exponent, the doubly-stochastic combination policy are
asymptotically superior to the right-stochastic one. This means that the two curves corresponding
to sensor 2 in the figure must cross for a certain vanishingly small µ. What the refined analysis
is able to tell is that this value might be too small for a given regime of analysis.
Before concluding, we would like to stress that the above comparison between combination
policies should be considered preliminary, in that: i) the comparison has been made for two
systems operating with the same value of the step-size µ, and, for a more complete view,
the analysis should be complemented by examining also the transient behavior of the two
October 28, 2018
DRAFT
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!1020304050!14!12!10!8!6!4!21ΜErrorprobability!!!:DoublyStochasticSolid:RightStochastic12345678910 32
combination policies; ii) the evidences are obtained with reference to a particular doubly-
stochastic matrix and a particular right-stochastic matrix. This notwithstanding, while no general
conclusions can be drawn at this stage regarding the relative advantages of the two kinds of
combination strategies for individual agents, a general trend seems to emerge: connectivity
matters. Since the connectivity features are embodied in the higher-order corrections, the simplest
large deviations analysis is not sufficient, and the refined exact asymptotics provided by Theorem
3 are crucial in assessing the performance of adaptive distributed detection over networks.
APPENDIX A
In the following, the r-th derivative of a function f (t) will be denoted by f (r)(t), with the
convention that f (0)(t) = f (t). When convenient, the first three derivatives will be alternatively
denoted by f(cid:48)(t), f(cid:48)(cid:48)(t), and f(cid:48)(cid:48)(cid:48)(t). Moreover, the notation fµ = O(µ) means that the ratio fµ/µ
stays bounded as µ → 0.
THEOREM 1 (Fundamental properties of φ(r)
introduce the quantity:
k,µ(t)). Assume that ψ(t) < ∞ for all t ∈ R, and
ξi,(cid:96) (cid:44) µ(1 − µ)i−1bk,(cid:96)(i).
Then, the following facts hold:
i) The LMGF of y(cid:63)
k,µ can be computed as:
φk,µ(t) =
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
ψ(ξi,(cid:96)t)
and the limiting LMGF φ(t) is:
S(cid:88)
(cid:96)=1
ω(p(cid:96)t)
φ(t) =
ii) For r = 1, 2, . . . , it holds that:
φ(r)
k,µ(t) =
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
ξr
i,(cid:96)ψ(r)(ξi,(cid:96)t)
and
October 28, 2018
φ(r)
k,µ(t/µ)
µr−1 = φ(r)(t)
lim
µ→0
(87)
(88)
(89)
(90)
(91)
DRAFT
In particular, the r-th cumulant, φ(r)
k,µ(0), of the steady-state random variable y(cid:63)
k,µ satisfies:
φ(r)
k,µ(0)
µr−1 = φ(r)(0) =
lim
µ→0
ψ(r)(0)
r
pr
(cid:96)
(92)
S(cid:88)
(cid:96)=1
33
where ψ(r)(0) is the r-th cumulant of the local statistic xk(n).
iii) With reference to the convergence in (89) and (91), the following refined estimate of the
convergence error holds for r = 0, 1, . . . :
φ(r)
k,µ(t/µ)
µr−1 = φ(r)(t) + O(µ).
(93)
(cid:3)
In order to establish the validity of Theorem 1, we start by proving a couple of useful lemmas.
Our first lemma is a simple generalization of Lemmas 1 and 2 from [1], [2]. We shall focus on
a function f (t) twice differentiable in R, with f (0) = 0. For such a function, we have that:
• The function f (t)/t is continuous for all t ∈ R. For any t (cid:54)= 0, the result follows from
the continuity of f (t). For t = 0, the result is easily verified by recalling that f (0) = 0,
yielding:
f (t)
= f(cid:48)(0).
lim
t→0
(94)
• The derivative (d/dt)(f (t)/t) is continuous for all t ∈ R. For any t (cid:54)= 0, this result follows
immediately from the assumed smoothness properties of f (t). For t = 0, the result is easily
verified by recalling that f (0) = 0, and observing that
f(cid:48)(t)t − f (t)
f(cid:48)(cid:48)(0)
f (t)
t
lim
t→0
d
dt
= lim
t→0
t
=
,
2
(95)
where we used L'Hospital's rule [43].
We introduce the auxiliary functions:
h1(t) =
×
t2
2
t2
(cid:12)(cid:12)(cid:12)(cid:12) ,
(cid:12)(cid:12)(cid:12)(cid:12) ,
f (τ )
τ
f (τ )
τ
f(cid:48)(τ ),
f(cid:48)(τ ),
dτ
max
τ∈[0,t]
(cid:12)(cid:12)(cid:12)(cid:12) d
(cid:12)(cid:12)(cid:12)(cid:12) d
max
τ∈[0,t]
max
τ∈[t,0]
max
τ∈[t,0]
dτ
h2(t) = t ×
and
October 28, 2018
t ≥ 0,
t < 0.
t ≥ 0,
t < 0.
(96)
(97)
DRAFT
We can easily show that:
34
∀t ∈ R.
0 ≤ h1(t) < ∞,
0 ≤ h2(t) < ∞,
(98)
Indeed, h1(t) ≥ 0 and h2(t) ≥ 0 by definition. Finiteness of both functions follows from
Weierstrass extreme value theorem [43] since, by the properties for f (t) and f (t)/t discussed
above, the maxima appearing in (96) and (97) are maxima of continuous functions over compact
sets for any finite t.
LEMMA 1. Let f (t) be twice continuously differentiable in R, with f (0) = 0. Let, for all i ∈ N:
0 ≤ bi ≤ 1,
i→∞−→ p > 0,
bi
with bi − p ≤ Cλi
for some C > 0 and 0 < λ < 1. Then, we have, for all t ∈ R:
(99)
(100)
(101)
∞(cid:88)
n(cid:88)
where:
i=1
(cid:90) µpt
0
f (µ(1 − µ)i−1bit) =
1
µ
f (τ )
τ
dτ + R(t)
R(t) ≤ h1(µpt)
2 − µ
+
Cλh2(µt)
1 − λ(1 − µ)
Proof. First, note that the integral in (100) is well-posed, since the function f (t)/t is continuous
for all t ∈ R. Then, for the case t = 0, Eq. (100) is trivially verified, since f (0) = 0 by
assumption, and h1(0) = h2(0) = 0 by definitions (96) and (97). We now prove the lemma for
the case t > 0, and the proof for t < 0 will follow similarly.
Let us introduce the infinite partition of the interval [0, µpt]:
τi = µ(1 − µ)i−1pt,
i = 1, 2, . . .
(102)
By using a first-order Taylor expansion we can write:
i=1
=
October 28, 2018
n(cid:88)
n(cid:88)
f (µ(1 − µ)i−1bit)
f (τi) +
i=1
i=1
µ(1 − µ)i−1f(cid:48)(ti)(bi − p)t.
(103)
DRAFT
We then obtain from (104) that
(cid:90) µpt
µ(1−µ)npt
δi = τi − τi+1 = µ τi.
g(τ )dτ =
=
i=1
τi+1
g(τ )dτ
(cid:90) τi
n(cid:88)
n(cid:88)
g(τi)δi − n(cid:88)
f (τi) − n(cid:88)
n(cid:88)
i=1
i=1
i=1
i=1
= µ
g(cid:48)(¯ti)
δ2
i
2
g(cid:48)(¯ti)
δ2
i
2
,
(104)
(105)
(106)
(107)
for a value ti that is certainly contained in the interval [0, µt], because so are the points µ(1 −
µ)i−1bit and τi. Let us focus on the first term on the RHS of (103), and let g(t) = f (t)/t. A
second-order Taylor expansion of the function G(t) =(cid:82) τi
t g(τ )dτ around the point τi gives:
35
G(τi+1) = G(τi) + g(τi)δi − g(cid:48)(¯ti)
δ2
i
2
for a certain ¯ti ∈ (τi+1, τi) and where
= g(τi)δi − g(cid:48)(¯ti)
δ2
i
2
,
Computing now(cid:80)n
n(cid:88)
where
i=1
i=1 f (τi) from (106), and substituting into (103), we have
f (µ(1 − µ)i−1bit) =
1
µ
f (τ )
dτ + rn(t),
µ(1−µ)npt
τ
(cid:90) µpt
rn(t) (cid:44) 1
µ
g(cid:48)(¯ti)
δ2
i
2
+
n(cid:88)
n(cid:88)
i=1
(108)
Both series on the RHS of (108) are absolutely convergent as n → ∞. Indeed, using the definition
of the auxiliary function h1(t) in (96), we see that the first series satisfies:
i=1
µ(1 − µ)i−1f(cid:48)(ti)(bi − p)t.
n(cid:88)
i=1
1
µ
g(cid:48)(¯ti) δ2
i
2
=
µ
2
≤ µ
(µpt)2
∞(cid:88)
n(cid:88)
g(cid:48)(¯ti)(1 − µ)2(i−1)
i=1
October 28, 2018
i=1
(1 − µ)2(i−1)h1(µpt) =
h1(µpt)
2 − µ
.
(109)
DRAFT
n(cid:88)
µ(1 − µ)i−1f(cid:48)(ti)(bi − p)t
i=1
≤ h2(µt)
(1 − µ)i−1bi − p
n(cid:88)
∞(cid:88)
i=1
[λ(1 − µ)]i−1
≤ Cλh2(µt)
Cλh2(µt)
1 − λ(1 − µ)
i=1
=
.
(110)
Convergence of the first term on the RHS in (107) follows by the definition of integration. If
we now denote by R(t) the limit of rn(t) as n → ∞, we have in fact proved (100) along with
the upper bound in (101).
(cid:3)
The next lemma establishes some useful properties of the function ω(t) defined in (30).
LEMMA 2 (Derivatives of ω(t)). Let ψ(t) be the LMGF of the local statistic xk(n), and assume
that ψ(t) < ∞ for all t ∈ R. Then, the function ω(t) in (30) is infinitely differentiable in R,
and, for all r ∈ N:
ω(r)(t) =
ψ(r)(τ )τ r−1dτ
(111)
(112)
Similarly, using the definition of h2(t) in (97), and the assumption bi − p ≤ Cλi, we see that
the second series obeys:
36
(cid:90) t
0
1
tr
(cid:90) t
0
where, for t = 0, the above equation should be read as:
ω(r)(0) = lim
t→0
1
tr
ψ(r)(τ )τ r−1dτ =
ψ(r)(0)
r
.
Proof. We prove the lemma by induction. Let us consider first the case t (cid:54)= 0. Property (111)
holds for r = 1:
ω(cid:48)(t) =
ψ(t)
t
=
1
t
ψ(cid:48)(τ )dτ,
(113)
having used ψ(0) = 0 since ψ(t) is a LMGF. Now we show that if (111) holds for some r, then
it must hold for r + 1. Indeed, note that
(cid:90) t
0
(cid:18) 1
(cid:90) t
tr
0
(cid:19)
ψ(r)(τ )τ r−1dτ
ω(r+1)(t) =
d
dt
ω(r)(t) =
d
dt
October 28, 2018
.
(114)
DRAFT
Applying the rule of integration by parts we have:
(cid:90) t
0
1
tr
ψ(r)(τ )τ r−1dτ
(cid:20)
=
ψ(r)(τ )
τ r
r
− 1
rtr
Differentiating the above expression yields:
1
tr
ψ(r)(t)
=
r
(cid:90) t
0
ψ(r+1)(τ )
τ r
r
dτ
− 1
tr
ψ(r+1)(τ )τ rdτ.
(cid:21)t
(cid:90) t
0
0
ω(r+1)(t) =
ψ(r+1)(t)
− ψ(r+1)(t)
r
(cid:90) t
r
0
+
1
tr+1
ψ(r+1)(τ )τ rdτ.
This proves the claim for t (cid:54)= 0. Now, since ω(r)(t) exists for all t (cid:54)= 0, we have [43]:
provided that the above limit exists. By applying L'Hospital's rule [43] we have in fact:
ω(r)(0) = lim
t→0
ω(r)(t),
(cid:90) t
0
ψ(r)(τ )τ r−1dτ =
ψ(r)(0)
r
ω(r)(0) = lim
t→0
1
tr
and the proof is complete.
We are now ready to prove Theorem 1.
37
(115)
(116)
(117)
(118)
(cid:3)
(119)
(120)
Proof of Theorem 1. Recalling the definition of ξi,(cid:96) in (87), the finite-horizon random variable
k(n) in (21), and its LMGF φk,µ(t; n) in (58) can be written as, respectively:
y(cid:63)
and
φk,µ(t; n) =
S(cid:88)
(cid:96)=1
y(cid:63)
k(n) (cid:44) n(cid:88)
S(cid:88)
n(cid:88)
i=1
ψ(ξi,(cid:96)t) =
i=1
(cid:96)=1
ξi,(cid:96)x(cid:96)(i),
S(cid:88)
n(cid:88)
(cid:96)=1
i=1
ψ(ξi,(cid:96)t).
We now apply Lemma 1 to each of the S inner summations, with the choices f (t) = ψ(t),
bi = bk,(cid:96)(i), and p = p(cid:96). Note that the hypotheses of the lemma are satisfied. Indeed ψ(t) is a
LMGF, implying that ψ(0) = 0 and that ψ(t) is infinitely differentiable in R. By assumption, we
October 28, 2018
DRAFT
have ψ(t) < ∞ for all t ∈ R. As to the conditions in (99), by Perron's theorem [36, Theorem
8.5.1], there exist C > 0 and 0 < λ < 1 such that:
38
bk,(cid:96)(i) − p(cid:96) ≤ Cλi.
By Lemma 1 we conclude that
lim
n→∞ φk,µ(t; n) =
=
ψ(ξi,(cid:96)t)
∞(cid:88)
S(cid:88)
(cid:90) p(cid:96)µt
S(cid:88)
(cid:96)=1
i=1
1
µ
(cid:96)=1
0
ψ(τ )
τ
dτ +
(121)
(122)
S(cid:88)
(cid:96)=1
R(cid:96)(t),
where we made explicit the dependence of the remainder terms upon (cid:96) = 1, 2, . . . , S.
In view of the continuity theorem for the moment generating functions [45], and since by
k,µ, the limit of φk,µ(t; n) (i.e., the limit of the
definition y(cid:63)
LMGF of y(cid:63)
k(n) converges in distribution to y(cid:63)
k(n)) represents φk,µ(t) (i.e., the LMGF of y(cid:63)
k,µ). Formally, we have:
φk,µ(t) = lim
n→∞ φk,µ(t; n) =
ψ(ξi,(cid:96)t),
and claim (88) is proved. For the same reason we can rewrite (122) as:
and
φk,µ(t) =
1
µ
µ φk,µ(t/µ) =
ψ(τ )
τ
dτ +
ψ(τ )
dτ + µ
R(cid:96)(t),
R(cid:96)(t/µ),
where, from (101), the remainder term is bounded by:
S(cid:88)
(cid:96)=1
µ
R(cid:96)(t/µ) ≤ µ
2 − µ
h1(p(cid:96)t) +
µ SCλh2(t)
1 − λ(1 − µ)
.
∞(cid:88)
i=1
(cid:96)=1
S(cid:88)
S(cid:88)
S(cid:88)
(cid:96)=1
(cid:96)=1
(cid:90) p(cid:96)µt
S(cid:88)
(cid:90) p(cid:96)t
S(cid:88)
(cid:96)=1
0
(cid:96)=1
0
τ
S(cid:88)
(cid:96)=1
n(cid:88)
S(cid:88)
(123)
(124)
(125)
(126)
(127)
(128)
DRAFT
We conclude from (124) that claim (89) and claim (93) for the case r = 0 hold.
To establish the remaining claims, we start by noting from (120) that, for all r = 1, 2, . . . :
φ(r)
k,µ(t; n) =
i,(cid:96)ψ(r)(ξi,(cid:96)t).
ξr
Now, recalling the definition of ξi,(cid:96) from (87), we have:
i=1
(cid:96)=1
0 ≤ ξi,(cid:96) ≤ µ(1 − µ)i−1,
t ∈ [−a, a] ⇒ ξi,(cid:96)t ∈ [−a, a].
October 28, 2018
Thus, for all t ∈ [−a, a] we can write:
ξr
i,(cid:96)ψ(r)(ξi,(cid:96)t) ≤ µr(1 − µ)r(i−1) max
τ∈[−a,a]
ψ(r)(τ ).
But since
µr
(1 − µ)r(i−1) =
µr
1 − (1 − µ)r ,
∞(cid:88)
i=1
39
(129)
(130)
we conclude that φ(r)
k,µ(t; n) is majorized by an absolutely convergent series. In view of Weierstrass
theorem about uniform convergence [43, Theorem 7.10, p. 148], this result implies that φ(r)
k,µ(t; n)
converges uniformly as n → ∞ in any compact interval [−a, a]. Since this holds for all r =
1, 2, . . . , and since we already showed that φk,µ(t; n) → φk,µ(t) in (88), the results about uniform
convergence and differentiability [43, Theorem 7.17, p. 152] allow us to conclude that the
derivative of the limit function equals the limit of the series of derivatives, namely, that (90)
holds true. Moreover, using (127) for n → ∞, we can write for any t (cid:54)= 0:
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
φ(r)
k,µ(t) =
1
tr
(ξi,(cid:96)t)rψ(r)(ξi,(cid:96)t).
(131)
Applying Lemma 1 to the function f (t) = trψ(r)(t), with the same choices done before for bi
and p, Eq. (100) gives:
which implies:
φ(r)
k,µ(t) =
1
µtr
φ(r)
k,µ(t/µ)
µr−1 =
1
tr
(cid:90) p(cid:96)µt
(cid:90) p(cid:96)t
0
S(cid:88)
S(cid:88)
(cid:96)=1
(cid:96)=1
0
ψ(r)(τ )τ r−1dτ +
R(cid:96)(t),
ψ(r)(τ )τ r−1dτ +
R(cid:96)(t/µ),
1
tr
S(cid:88)
S(cid:88)
(cid:96)=1
(cid:96)=1
µ
tr
(cid:90) p(cid:96)t
S(cid:88)
(cid:96)=1
0
(cid:90) p(cid:96)t
S(cid:88)
and, using (101), the remainder term vanishes as µ → 0, implying that
φ(r)
k,µ(t/µ)
µr−1
µ→0−→ 1
tr
ψ(r)(τ )τ r−1dτ.
S(cid:88)
Now, in view of Lemma 2, and by the definition of φ(t) in (89), we have:
φ(r)(t) =
(135)
Comparing (134) and (135) we arrive at (91) for all t (cid:54)= 0. It remains to consider the case t = 0,
namely, to show that:
(cid:96) ω(r)(p(cid:96)t) =
pr
(cid:96)=1
(cid:96)=1
0
ψ(r)(τ )τ r−1dτ.
1
tr
φ(r)
k,µ(0)
µr−1
µ→0−→ φ(r)(0) =
ψ(r)(0)
S(cid:88)
r
(cid:96)=1
pr
(cid:96),
(136)
DRAFT
October 28, 2018
(132)
(133)
(134)
ξr
i,(cid:96)
i=1
(cid:96)=1
∞(cid:88)
S(cid:88)
S(cid:88)
∞(cid:88)
∞(cid:88)
S(cid:88)
∞(cid:88)
S(cid:88)
(cid:96)=1
(cid:96)=1
i=1
i=1
i=1
(cid:96)=1
φ(r)
k,µ(0)
(90)
= ψ(r)(0)
(87)
= ψ(r)(0)µr
(1 − µ)r(i−1)br
k,(cid:96)(i)
= ψ(r)(0)µr
ψ(r)(0)µr
(cid:124)
= ψ(r)(0)
µr
1 − (1 − µ)r
k,(cid:96)(i) − pr
(cid:125)
(cid:96))
(1 − µ)r(i−1)(br
(cid:96)(1 − µ)r(i−1) +
pr
(cid:123)(cid:122)
S(cid:88)
S(cid:88)
pr
(cid:96) + R,
(cid:44)R
(cid:96)=1
φ(r)
k,µ(0)
µr−1 = ψ(r)(0)
µ
1 − (1 − µ)r
pr
(cid:96) +
R
µr−1 .
(cid:96)=1
br
k,(cid:96)(i) − pr
(cid:96) ≤ rbk,(cid:96)(i) − p(cid:96),
40
(137)
(138)
(139)
where the last equality follows from (112). This equality is established as follows:
or
Using the inequality1
and (121) we can write
R
µr−1 ≤ rSCλψ(r)(0)
µ
(140)
Therefore, the second term on the RHS of (138) vanishes as µ → 0, and by application of
L'Hospital's rule [43] to the first term we get (136). Finally, examining (133) and (140) shows
that the O(µ) error estimate in (93) holds true.
1 − [λ(1 − µ)r]
.
1The inequality follows from the factorization, holding for r ∈ N:
r−1(cid:88)
ar − br = (a − b)
akbr−1−k,
and from the fact that bk,(cid:96)(i) and p(cid:96) are not greater than one.
k=0
October 28, 2018
DRAFT
APPENDIX B
41
The proof relies on some properties of moment generating functions, which are collected in
Sec. B-A, and on a technical lemma, whose proof is deferred to Sec. B-B.
Proof of Theorem 3. By definition, θγ solves the stationary equation
(141)
Since φ(cid:48)(cid:48)(t) > 0, it follows that φ(cid:48)(t) is strictly increasing, and, hence, θγ is unique. Moreover,
from (92) used with r = 1 we obtain:
φ(cid:48)(θγ) = γ.
φ(cid:48)(0) = ψ(cid:48)(0) = E[x],
(142)
recalling that the first derivative of the LMGF, evaluated at the origin, gives the expectation
of the random variable [34], [35]. The equality φ(cid:48)(0) = E[x] implies that θγ > 0, because by
assumption γ > E[x], and φ(cid:48)(t) is strictly increasing.
For ease of notation, in the following we shall suppress the subscript γ, and we shall use
the symbol θ to denote the unique solution of (141). Let us denote by m(dy) the probability
measure associated with y(cid:63)
k,µ, and introduce the following measure transformation, namely, an
exponential translation in the form:
m(dy) = e
θ
µ y−φk,µ( θ
µ)m(dy).
(143)
The use of a similar transformation for large deviations analysis was originally proposed by
Cram´er [41], and then used, e.g., in [33], to find the exact asymptotics in the simplest case of
normalized sums of i.i.d. random variables -- see also [34], [35]. We will explain how this
transformation is also helpful in our more general case, where the asymptotic parameter µ is
not simply the index of a summation of i.i.d. random variables.
Denoting by IE the indicator of an event E, and by E m[·] the expectation operator under the
October 28, 2018
DRAFT
measure m(dy), some straightforward algebra allows us to write:
(cid:90) ∞
γ
k,µ>γ}] =
m(dy)
P[y(cid:63)
(cid:90) ∞
k,µ > γ] = E[I{y(cid:63)
− θ
e
µ y+φk,µ( θ
=
γ
µ y(cid:63)
k,µ+φk,µ( θ
− θ
= E m[e
− 1
µ [γθ−φ(θ)] e
k,µ−φ(cid:48)
= e
µ [y(cid:63)
− θ
k,µ( θ
µ) m(dy)
µ) I{y(cid:63)
E m[e
− 1
E m[e
= e
µ [Φ(γ)+k,µ(θ)] e
− θ
k,µ−φ(cid:48)
µ [y(cid:63)
k,µ( θ
k,µ>γ}]
µ)] ×
θ
µ)] I{y(cid:63)
µ[γ−φ(cid:48)
k,µ( θ
µ)] I{y(cid:63)
k,µ>γ}],
k,µ>γ}]
µ)] e
− 1
µ[φ(θ)−µφk,µ( θ
θ
µ[γ−φ(cid:48)
k,µ( θ
µ)] ×
where we used definition (51a) for k,µ(θ), and the fact that
Φ(γ) = sup
t∈R
[γt − φ(t)] ⇔ Φ(γ) = γθ − φ(θ),
which holds true because φ(t) is strictly convex and θ solves the stationary equation (141). It is
now useful to introduce the normalized random variable
For ease of notation, we introduce the quantity
where the convergence follows from (91). With this definition, the expression for w(cid:63)
rewritten as:
k,µ can be
It is also convenient to introduce the normalized threshold:
42
(144)
(145)
(146)
(147)
(cid:44) y(cid:63)
w(cid:63)
k,µ
(cid:113)
k,µ − φ(cid:48)
φ(cid:48)(cid:48)
k,µ(θ/µ)
k,µ(θ/µ)
.
(cid:115)
sk,µ (cid:44)
φ(cid:48)(cid:48)
k,µ(θ/µ)
µ
µ→0−→(cid:112)φ(cid:48)(cid:48)(θ),
w(cid:63)
k,µ =
k,µ − φ(cid:48)
y(cid:63)
√
µsk,µ
k,µ(θ/µ)
.
k,µ(θ/µ)
µsk,µ
,
γk,µ (cid:44) γ − φ(cid:48)
√
(cid:20)
E m
− θsk,µ√
e
µ (w(cid:63)
k,µ−γk,µ) I{w(cid:63)
k,µ>γk,µ}
(cid:21)
.
(148)
(149)
(150)
DRAFT
and the normalized expectation:
Ck,µ (cid:44) 1√
µ
October 28, 2018
Using the above definitions, Eq. (144) can be rewritten as:
P[y(cid:63)
k,µ > γ] =
µ Ck,µ e
− 1
µ [Φ(γ)+k,µ(θ)].
√
43
(151)
Examining expression (50) for Pk,µ(γ), it is immediate to recognize that the theorem will be
established if we are able to show that:
µ→0−→
Ck,µ
1(cid:112)2πθ2φ(cid:48)(cid:48)(θ)
(152)
The key for proving the above claim is the asymptotic normality of the random variable w(cid:63)
k,µ
under the transformed measure m(dy). To see this, let us exploit the fundamental properties of
k,µ. First, according to property P2 in Sec. B-A, and since the first three central moments of
w(cid:63)
a random variable coincide with its first three cumulants [48], under the considered transformed
measure we have:
k,µ] = φ(cid:48)
E m[y(cid:63)
k,µ(θ/µ))2] = φ(cid:48)(cid:48)
k,µ(θ/µ))3] = φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ),
k,µ(θ/µ),
k,µ(θ/µ).
E m[(y(cid:63)
E m[(y(cid:63)
k,µ − φ(cid:48)
k,µ − φ(cid:48)
(153)
(154)
(155)
Accordingly, from (146) and (147), we conclude that w(cid:63)
third moment is:
k,µ is zero-mean, unit-variance, and its
E m[(w(cid:63)
k,µ)3] =
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
µ3/2s3
k,µ
.
(156)
A crucial step consists now in finding an asymptotic expansion for the cumulative distribution
function of the random variable w(cid:63)
k,µ under the transformed measure. To this aim, we follow the
procedure employed for deriving the Berry-Esseen and Edgeworth expansions -- see, e.g., [40],
and we introduce the function:
Gk,µ(w) (cid:44) N (w) − E m[(w(cid:63)
k,µ)3]
(w2 − 1) n(w)
= N (w) − φ(cid:48)(cid:48)(cid:48)
6
k,µ(θ/µ)
6 µ3/2s3
k,µ
(w2 − 1) n(w),
(157)
where N (w) and n(w) = N(cid:48)(w) are, respectively, the cumulative distribution function and
the probability density function of a standard normal. Denoting by Fk,µ(w) the cumulative
October 28, 2018
DRAFT
distribution function of the random variable w(cid:63)
establish that (Lemma 3):
k,µ under the transformed measure, in B-B we
1√
µ
sup
w∈R
Fk,µ(w) − Gk,µ(w) µ→0−→ 0.
(158)
44
This implies that w(cid:63)
k,µ is asymptotically normal under the transformed measure m(dy). More
importantly, Eq. (158) provides a refined, uniform estimate of the convergence error, a property
that we shall exploit soon to prove (152). To this aim, we evaluate explicitly the expectation
in (150). Using integration by parts, we obtain:
(cid:90) ∞
Ck,µ =
1√
µ
(cid:20)
=
1√
µ
θsk,µ
µ
1√
µ
=
γk,µ
− θsk,µ√
e
(cid:90) ∞
(cid:90) ∞
γk,µ
0
− θsk,µ√
e
µ (w−γk,µ)dFk,µ(w)
µ (w−γk,µ) [Fk,µ(w) − Fk,µ(γk,µ)]
(cid:21)∞
+
γk,µ
− θsk,µ√
µ (w−γk,µ)[Fk,µ(w) − Fk,µ(γk,µ)]dw
e
e−x [Fk,µ(pk,µ(x)) − Fk,µ(γk,µ)] dx,
(159)
(160)
(161)
(162)
(163)
(164)
DRAFT
where the last equality follows from the change of variable
(w − γk,µ) → x,
θsk,µ√
µ
and from the definition:
pk,µ(x) (cid:44)
√
µ
θsk,µ
x + γk,µ.
Let us also introduce the analogue of (159) for the normal approximation Gk,µ(w) in (157),
namely,
(cid:90) ∞
e−x [Gk,µ(pk,µ(x)) − Gk,µ(γk,µ)] dx.
Dk,µ (cid:44) 1√
µ
By using Lemma 3, we can write:
0
(cid:90) ∞
Ck,µ − Dk,µ ≤ 2√
µ
2√
µ
=
sup
w∈R
sup
w∈R
Fk,µ(w) − Gk,µ(w)
Fk,µ(w) − Gk,µ(w) µ→0−→ 0.
0
e−xdx
By virtue of this result, and in order to prove (152), it is enough to show that
µ→0−→
Dk,µ
October 28, 2018
1(cid:112)2πθ2φ(cid:48)(cid:48)(θ)
.
Substituting definition (157) for Gk,µ(w) into (162), we have:
(cid:90) ∞
Dk,µ =
0
e−x [N (pk,µ(x)) − N (γk,µ)] dx +
1√
µ
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
6 µ2s3
[n(γk,µ)(γ2
(cid:90) ∞
k,µ − 1) − n(pk,µ(x))(p2
e−x ×
k,µ
0
k,µ(x) − 1)]dx.
Consider first the second term on the RHS. By (93) we can write (recall from (141) that φ(cid:48)(θ) =
γ):
k,µ(θ/µ) = φ(cid:48)(θ) + O(µ) = γ + O(µ),
φ(cid:48)
45
(165)
(166)
(167)
which, in view of (149), implies that
lim
µ→0
γk,µ = 0.
We therefore conclude that limµ→0 pk,µ(x) = 0 from (161) and (147). Accordingly, we have:
k,µ − 1) − n(pk,µ(x))(p2
k,µ(x) − 1)] = 0,
lim
µ→0
[n(γk,µ)(γ2
(168)
and, hence, by dominated convergence [46] (note that the function n(w)(w2 − 1) is bounded),
the second term on the RHS in (165) vanishes as µ → 0, having further used the fact that the
ratio φ(cid:48)(cid:48)(cid:48)
converges in view of (91) and (147).
k,µ(θ/µ)
6 µ2s3
k,µ
With regards to the first term on the RHS of (165), by using a second-order Taylor expansion
of N (w) around the point γk,µ we obtain, for an intermediate point qk,µ(x):
N (pk,µ(x)) = N (γk,µ) + n(γk,µ) (pk,µ(x) − γk,µ) +
n(cid:48)(qk,µ(x))
(pk,µ(x) − γk,µ)2
√
2
µ
θsk,µ
x +
= N (γk,µ) + n(γk,µ)
n(cid:48)(qk,µ(x))
µx2
2(θsk,µ)2 ,
(169)
DRAFT
October 28, 2018
where we used (161). The above expansion then gives:
The second term on the RHS of (170) vanishes because
(cid:90) ∞
1√
µ
=
0
n(γk,µ)
θsk,µ
√
µ
2(θsk,µ)2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:90) ∞
0
x2e−xn(cid:48)(qk,µ(x))dx
(cid:90) ∞
n(γk,µ)
θsk,µ
e−x [N (pk,µ(x)) − N (γk,µ)] dx
0
xe−xdx +
(cid:90) ∞
(cid:90) ∞
(cid:12)(cid:12)(cid:12)(cid:12) ≤ sup
w∈R
0
= 2 sup
w∈R
x2e−xn(cid:48)(qk,µ(x))dx.
(cid:90) ∞
x2e−xdx
n(cid:48)(w)
n(cid:48)(w) < ∞.
0
46
(170)
(171)
On the other hand, the first term on the RHS of (170) satisfies:
1(cid:112)2πθ2φ(cid:48)(cid:48)(θ)
xe−xdx =
µ→0−→
n(γk,µ)
θsk,µ
0 xe−xdx = 1, and sk,µ converges to (cid:112)φ(cid:48)(cid:48)(θ)
2π, (cid:82) ∞
(172)
,
√
since γk,µ vanishes by (167), n(0) = 1/
by (147). The proof of (51a) is now complete.
0
It remains to justify the alternative expression (51b). The rationale behind our choice is as
follows. Examining (172), it is seen that the term n(γk,µ) is replaced by its limiting constant
√
2π. In order to get a better approximation for finite values of 1/µ, we can decide to
value of 1/
retain the dependence on µ by including n(γk,µ) in the main formula. This amounts to replacing
Ck,µ in (151), not with the limiting value in (152), but rather with
√
2π n(γk,µ)
1(cid:112)2πθ2φ(cid:48)(cid:48)(θ)
.
(173)
By making explicit the definition of γk,µ in (149), and of the Gaussian pdf, the additional term
is:
√
2π n(γk,µ) = e
− [γ−φ(cid:48)
k,µ(θ/µ)]2
2µs2
k,µ
∼ e
− 1
µ
[φ(cid:48)(θ)−φ(cid:48)
k,µ(θ/µ)]2
2φ(cid:48)(cid:48)(θ)
,
(174)
where, in the last step, we have used γ = φ(cid:48)(θ), and, to be consistent with the approximations
k,µ with its limiting value φ(cid:48)(cid:48)(θ). Examining the correction at
adopted so far, we have replaced s2
the exponent of (174) we recognize exactly the second term in (51b).
October 28, 2018
(cid:3)
DRAFT
47
A. Some general properties of moment generating functions
Consider a random variable x with Moment Generating Function (MGF), Logarithmic Moment
Generating Function (LMGF), and Characteristic Function (CF) defined respectively, as:
ν(t) (cid:44) E[etx], ψ(t) (cid:44) ln ν(t), ϕ(u) (cid:44) E[ejux].
(175)
We stress that, in this specific subsection, x is intended to be a generic random variable, and
there is no need here to give it any particular meaning in terms of a local or a global network
statistic. We assume that ν(t) exists for all t ∈ R. Under this assumption, we can also define
the MGF relative to a complex argument z, namely:
ν(z) (cid:44) E[ezx], z ∈ C,
(176)
which is the analytic continuation of ν(t) over the complex plane [46]. Note that, for z = t + ju:
ν(z) = E[ezx] ≤ E[ezx] = ν(t).
(177)
In this section, we focus on the following measure transformation (exponential measure trans-
lation):
m(dx) = eηx−ψ(η)m(dx) =
eηx
ν(η)
m(dx),
(178)
and establish several important properties characterizing the statistical behavior of x under the
transformed measure m(dx).
• Property P1. The MGF under the transformation is:
ν(t) =
ν(η + t)
ν(η)
Indeed, by definition:
ν(t) (cid:44) E m[etx] =
E[e(η+t)x]
ν(η)
=
ν(η + t)
ν(η)
.
• Property P2. The r-th cumulant χ(r) of the variable x under the transformation is
χ(r) = ψ(r)(η)
Again, from property P1, the LMGF of x under the transformed measure is:
ψ(t) (cid:44) ln ν(t) = ln ν(η + t) − ln ν(η) = ψ(η + t) − ψ(η).
October 28, 2018
(179)
(180)
(181)
(182)
DRAFT
Accordingly, the r-th cumulant of the variable x under the transformation is
(cid:12)(cid:12)(cid:12)(cid:12)t=0
χ(r) =
dr
dtr
ψ(t)
= ψ(r)(η).
• Property P3. The CF of the random variable x under the transformation is
ϕ(u) =
ν(η + ju)
ν(η)
48
(183)
(184)
where ν(z), for a complex z, was defined by (180). This result follows from an argument
similar to the proof of Property P1.
• Property P4. Consider the shifted random variable
x = x − ψ(cid:48)(η).
Since the CF of the shifted-and-scaled random variable ax + b is
E[ejt(ax+b)] = ϕ(at)ejbt,
(185)
(186)
the CF of the shifted variable x under the transformed measure is readily computed as:
ϕ(u) = ϕ(u)e−jψ(cid:48)(η)u =
ν(η + ju)
ν(η)
e−jψ(cid:48)(η)u.
(187)
Assume that the parameter η ruling the measure transformation lies in a bounded interval,
say, η ≤ ηmax. Then, for sufficiently small δ > 0, a positive constant M = M (δ, ηmax)
exists such that for u ≤ δ, and for any choice of η ≤ ηmax, the following expansion
holds for the CF of the shifted random variable x:
log ϕ(u) = −ψ(cid:48)(cid:48)(η)
u2
2
+ ψ(cid:48)(cid:48)(cid:48)(η)
(ju)3
6
+ R(u)
with the remainder bounded as:
R(u) ≤ M
u4
24
.
(188)
(189)
In order to prove the above property, let us introduce the logarithm of a complex number
z:
For z < 1, it can be alternatively defined by the Taylor series [46]:
log z = lnz + j arg(z).
∞(cid:88)
n
n=1
(−1)n+1
zn.
log(1 + z) =
October 28, 2018
(190)
(191)
DRAFT
49
We shall now consider conditions under which the function log ϕ(u) can be represented by
a Taylor series.
Since ϕ(u) is the CF of a zero-mean random variable, the following known bound is obtained
from the Taylor expansion of the complex exponential [40, Eq. (4.14), p. 514]:
ϕ(u) − 1 ≤ E m[ x2]
u2
2
= ψ(cid:48)(cid:48)(η)
u2
2
,
(192)
where in the last equality we used property P2. We now select a parameter δ > 0 such that
δ2 <
2
ψ(cid:48)(cid:48)(η)
max
η≤ηmax
,
(193)
where the above choice is meaningful because ψ(η) is a LMGF, which implies that ψ(cid:48)(cid:48)(η) >
0, and that ψ(cid:48)(cid:48)(η) is a continuous function, and thus admits a maximum within the interval
[−ηmax, ηmax]. Therefore, over the range u ≤ δ, Eqs. (192) and (193) give:
ϕ(u) − 1 < 1.
(194)
This allows representing, in the considered range for u, the function log ϕ(u) via its Taylor
series. By the definition of cumulants, this series can be written as [48, Eq. (6.8), p. 45]:
n(cid:88)
r=2
log ϕ(u) =
χ(r) (ju)r
r!
+ o(un),
(u → 0).
(195)
Note that, in the above formula, the series coefficients should be given by the cumulants
of x (computed under the transformed measure). As it can be seen, we used instead the
cumulants χ(r) of x. This is because x is a shifted version of x, and so its cumulants are
(but for the first one, which is zero) the same as those of x.
Consider now the three-term series. Using the explicit formulas for the cumulants under the
transformed measure provided by (181), we get:
log ϕ(u) = −ψ(cid:48)(cid:48)(η)
+ ψ(cid:48)(cid:48)(cid:48)(η)
(ju)3
u2
2
+ r(¯u)
6
where ¯u ∈ (−δ, δ), and the function r(u) can be expressed as:
(cid:20)ν(η + ju)
d4
du4 log ϕ(u)
d4
du4 log
ν(η)
(cid:21)
,
e−jψ(cid:48)(η)u
r(u) =
=
October 28, 2018
u4
24
,
(196)
(197)
DRAFT
50
which leads to (the argument η + ju is omitted for ease of notation):
r(u) =
(cid:2)−6(ν(cid:48))4 + 12ν(ν(cid:48))2ν(cid:48)(cid:48) − 4ν2ν(cid:48)ν(cid:48)(cid:48)(cid:48)−
3ν2(ν(cid:48)(cid:48))2 + ν3ν(4)(cid:3) .
1
ν4
(198)
We now show that r(u) is bounded for η ≤ ηmax and u ≤ δ. To this aim, it is useful to
introduce the region of the complex plane:
A = {Re(z) ∈ [−ηmax, ηmax], Im(z) ∈ [−δ, δ]}.
(199)
First, note that, since all the derivatives of ν(z) are analytic (and, hence, continuous) in
C [46], the term between brackets in (198) is bounded within the bounded and closed set
A. In addition, for z ∈ A, Eq. (194) implies that ν(z) > 0. Indeed, if we had ν(z) = 0
for a certain z ∈ A, then there would exist η and u, with η ≤ ηmax and u ≤ δ, such that
ν(η + ju) = 0, implying:
ϕ(u) − 1 =
e−jψ(cid:48)(η)u − 1
(200)
and this condition would violate (194). We can accordingly write, for all η ≤ ηmax and
u ≤ δ:
ν(η)
(cid:12)(cid:12)(cid:12)(cid:12) ν(η + ju)
(cid:12)(cid:12)(cid:12)(cid:12) ≤ max
z∈A
(cid:12)(cid:12)(cid:12)(cid:12)
1
ν(η + ju)
(cid:12)(cid:12)(cid:12)(cid:12) = 1,
(cid:12)(cid:12)(cid:12)(cid:12) < ∞,
(cid:12)(cid:12)(cid:12)(cid:12) 1
ν(z)
(201)
where finiteness follows since ν(z) is analytic (and, hence, continuous) in C, and it does
not vanish within the set A. We have in fact shown that r(u) is bounded for η ≤ ηmax
and u ≤ δ, namely, that a constant M = M (δ, ηmax) exists such that r(¯u) ≤ M. In the
light of (196), this completes the proof of (188) and (189).
B. Technical lemma relevant to Theorem 3
Let us introduce the main quantities involved in the forthcoming derivation.
• The measure transformation:
m(dy) = e
θ
µ y−φk,µ( θ
µ)m(dy).
• The random variable introduced in (148):
k,µ − φ(cid:48)
y(cid:63)
√
w(cid:63)
k,µ =
k,µ(θ/µ)
,
µsk,µ
October 28, 2018
(202)
(203)
DRAFT
51
whose cumulative distribution function under the transformed measure has been denoted by
Fk,µ(w). Using property P3 and applying (186), it is easily seen that the CF of w(cid:63)
k,µ under
the transformed measure is:
(cid:16) θ
(cid:17)
−ju
e
φ(cid:48)
k,µ(θ/µ)
√
µsk,µ ,
(204)
νk,µ
ϕk,µ(u) =
µ + j
µsk,µ
u√
(cid:16) θ
(cid:17)
νk,µ
µ
where νk,µ(·) denotes the MGF (of real or complex argument) of the random variable y(cid:63)
under the original measure.
k,µ
• The normal approximation Gk,µ(w) defined by (157):
k,µ(θ/µ)
6 µ3/2s3
Gk,µ(w) = N (w) − φ(cid:48)(cid:48)(cid:48)
k,µ
(w2 − 1) n(w).
(205)
The following lemma shows that Fk,µ(w) converges to Gk,µ(w) as µ goes to zero. From (91)
and (147), we conclude that the second term on the RHS of (205) vanishes as µ → 0. This
result will imply that w(cid:63)
k,µ is asymptotically normal. In addition, regarding Gk,µ(w) as a higher-
order approximation for Fk,µ(w) (namely, including a correction with respect to the crudest
approximation N (w)) the lemma provides the useful information that the (worst-case) rate of
µ. This observation is critical for
convergence of the approximation error is in the order of
the proof of Theorem 3.
√
LEMMA 3 (Asymptotic normality of w(cid:63)
of the convergence error). Under the same hypotheses of Theorem 3:
k,µ under the transformed measure (202): Estimate
1√
µ
sup
w∈R
Fk,µ(w) − Gk,µ(w) µ→0−→ 0.
(206)
Proof. The following classical result [40, page 538] is used in our proof.
Let F be a probability distribution with zero mean and characteristic function ϕ. Suppose that
F − G vanishes at ±∞ and that G has a derivative g such that g ≤ m. Finally, suppose that
g has a continuously differentiable Fourier transform ζ such that ζ(0) = 1 and ζ(cid:48)(0) = 0. Then,
for all w and for all T > 0:
(cid:90) T
−T
(cid:12)(cid:12)(cid:12)(cid:12) ϕ(u) − ζ(u)
u
(cid:12)(cid:12)(cid:12)(cid:12) du +
F (w) − G(w) ≤ 1
π
October 28, 2018
24m
πT
.
(207)
DRAFT
(cid:34)
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
6 µ3/2s3
k,µ
(cid:34)
(cid:35)
(cid:35)
It is now immediate to verify that the functions Fk,µ(w) and Gk,µ(w) introduced above meet the
conditions required for (207) to hold. Indeed, Gk,µ(w) → 0 as w → −∞, and Gk,µ(w) → 1 as
w → +∞, implying that the difference Fk,µ(w) − Gk,µ(w) vanishes at ±∞ because Fk,µ(w) is
a cumulative distribution function. Moreover,
d
dw
Gk,µ(w) = n(w)
1 − w
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
3 µ3/2s3
k,µ
−
(w2 − 1) n(cid:48)(w).
(208)
From (91) used with r = 3, we conclude that the quantity φ(cid:48)(cid:48)(cid:48)
vanishes as µ → 0 and,
hence, it is bounded for sufficiently small values of µ. The required boundedness condition on
dw Gk,µ(w)(cid:12)(cid:12) then follows by the boundedness of the functions n(w), w n(w), and (w2−1)n(cid:48)(w).
(cid:12)(cid:12) d
k,µ(θ/µ)
µ3/2s3
k,µ
Finally, the Fourier transform of (208) is given by [40]:
ζk,µ(u) (cid:44) e− u2
2
1 +
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
6 µ3/2s3
k,µ
(ju)3
.
(209)
It is seen that ζk,µ(u) is continuously differentiable, and that ζk,µ(0) = 1, ζ(cid:48)
k,µ(0) = 0.
In order to use the above lemma, let us select, for a given > 0, a constant a such that:
and set
a >
24m
,
T =
a√
µ
⇒ 24m
πT
√
≤
µ.
With these particular choices, application of (207) yields:
52
(210)
(211)
(212)
(213)
DRAFT
Due to arbitrariness of , the claim (206) will be proved if we show that the first term on the
RHS of (212) vanishes as µ → 0. To get this result, we split the integral as follows:
1√
µ
≤ 1√
µ
(cid:90) a√
µ
− a√
µ
Fk,µ(w) − Gk,µ(w)
u
µ
µ
− a√
(cid:90) a√
(cid:12)(cid:12)(cid:12)(cid:12) ϕk,µ(u) − ζk,µ(u)
(cid:12)(cid:12)(cid:12)(cid:12) ϕk,µ(u) − ζk,µ(u)
(cid:12)(cid:12)(cid:12)(cid:12) du + .
(cid:12)(cid:12)(cid:12)(cid:12) du = I1 + I2,
u
1√
µ
October 28, 2018
where
and
I1 =
(cid:90) δ√
µ
− δ√
µ
1√
µ
(cid:90)
(cid:12)(cid:12)(cid:12)(cid:12) ϕk,µ(u) − ζk,µ(u)
(cid:12)(cid:12)(cid:12)(cid:12) du,
(cid:12)(cid:12)(cid:12)(cid:12) ϕk,µ(u) − ζk,µ(u)
(cid:12)(cid:12)(cid:12)(cid:12) du,
u
u
I2 =
1√
µ
µ <u< a√
δ√
µ
53
(214)
(215)
for a sufficiently small δ whose choice will be carefully addressed in the following. We start by
proving that I1 vanishes as µ → 0.
Part i (I1 → 0 as µ → 0). From (88) we know that the MGF (with real argument t) of the
random variable y(cid:63)
k,µ under the original measure can be represented as:
∞(cid:89)
S(cid:89)
ν(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)t(cid:1) ,
νk,µ(t) =
where ν(t) = E[etxk(n)] is the MGF of the local statistic xk(n). We also know that the MGF
νk,µ(t) can be extended to the complex plane by considering
i=1
(cid:96)=1
νk,µ(z) = E[ezy(cid:63)
k,µ]
(217)
for complex z. It would be useful to have the same infinite-product representation as in (216)
for νk,µ(z), namely:
∞(cid:89)
S(cid:89)
i=1
(cid:96)=1
ν(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)z(cid:1) .
νk,µ(z) =
This turns out to be true, and can be justified by an argument similar to that used in [45], which
is now illustrated in detail. Consider the sequence of functions, for n = 1, 2, . . .
(216)
(218)
n(cid:89)
S(cid:89)
ν(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)z(cid:1) .
fn(z) =
(219)
First, fn(z) is analytic in C, since it consists of the product of functions ν(·) that are analytic
in C. Moreover, by (177), we can write, for z = t + ju:
(cid:96)=1
i=1
fn(z) ≤ n(cid:89)
S(cid:89)
ν(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)t(cid:1) = fn(t).
(220)
i=1
(cid:96)=1
For any n, the RHS is a MGF and, hence, its second derivative over t is strictly positive, which
implies that the maximum of the RHS within an interval t ≤ tmax is attained at either one or
both endpoints of the interval, namely, that:
October 28, 2018
fn(z) ≤ max{fn(−tmax), fn(tmax)} .
(221)
DRAFT
But since fn(t) converges to νk,µ(t) < ∞ as n → ∞, we have that, for any > 0, a certain n0
exists such that, for all n ≥ n0:
54
fn(−tmax) ≤ νk,µ(−tmax) + , fn(tmax) ≤ νk,µ(tmax) + ,
(222)
which shows that the sequence fn(z) is uniformly bounded, at least for n ≥ n0, in the closed
strip {z ∈ C : Re(z) ≤ tmax}. Moreover, since the sequence fn(t) (t ∈ R) admits a limit, we
can say that fn(z) (z ∈ C) admits a limit at each point of the real-axis, namely, of the interval
{z ∈ C : Re(z) ≤ tmax, Im(z) = 0}. This allows invoking Vitali's convergence theorem [49],
which asserts that fn(z) converges, namely:
fn(z) n→∞−→ f (z) =
∞(cid:89)
S(cid:89)
i=1
(cid:96)=1
ν(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)z(cid:1) ,
uniformly in each bounded closed subregion of the open strip {z ∈ C : Re(z) < tmax}, and
the limit function f (z) is therein analytic. Since we are free to choose tmax, we conclude that
f (z) is analytic in C. Moreover, on the real axis
∞(cid:89)
S(cid:89)
ν(cid:0)µ(1 − µ)i−1bk,(cid:96)(i)t(cid:1) = νk,µ(t),
f (t) =
i=1
(cid:96)=1
and since νk,µ(z) is analytic and equals νk,µ(t) on the real axis, by [46, Corollary, p. 209] we
finally get (218).
ηi,(cid:96) (cid:44) (1 − µ)i−1bk,(cid:96)(i) θ,
√
ui,(cid:96) (cid:44) (1 − µ)i−1bk,(cid:96)(i)
µu
sk,µ
,
and used (90) to compute φ(cid:48)
k,µ(θ/µ). Now, from property P4 it is seen that the function
ν(ηi,(cid:96) + jui,(cid:96))
ν(ηi,(cid:96))
e−jψ(cid:48)(ηi,(cid:96))ui,(cid:96)
October 28, 2018
In the light of the above finding, we are allowed to use (218) into (204). This yields, after
some manipulations:
ϕk,µ(u) =
where we introduced
i=1
(cid:96)=1
∞(cid:89)
S(cid:89)
ν(ηi,(cid:96) + jui,(cid:96))
ν(ηi,(cid:96))
e−jψ(cid:48)(ηi,(cid:96))ui,(cid:96),
(223)
(224)
(225)
(226)
(227)
(228)
DRAFT
can be regarded as the CF, evaluated at the point ui,(cid:96), of the shifted random variable x(cid:96)(i)−ψ(cid:48)(ηi,(cid:96))
under the measure transformation
m(dx) =
eηi,(cid:96)x
ν(ηi,(cid:96))
m(dx).
(229)
55
We shall denote this CF by ϕηi,(cid:96)(ui,(cid:96)), having made explicit the dependence upon ηi,(cid:96), thus
getting:
ϕk,µ(u) =
ϕηi,(cid:96)(ui,(cid:96)).
(230)
∞(cid:89)
S(cid:89)
i=1
(cid:96)=1
√
µu
sk,µ
From (227) we get:
We also know from (147) that limµ→0 sk,µ =(cid:112)φ(cid:48)(cid:48)(θ). By applying the definition of limit, this
ui,(cid:96) ≤
(231)
.
implies that, for any > 0, and for sufficiently small values of µ:
ui,(cid:96) ≤ √
µu
(cid:32)
1(cid:112)φ(cid:48)(cid:48)(θ)
(cid:33)
(cid:32)
+
≤ δ
1(cid:112)φ(cid:48)(cid:48)(θ)
(cid:33)
+
(cid:44) ¯δ,
(232)
where the last inequality holds because we are focusing on the integral I1 in (214) and, hence,
we consider the range
µu ≤ δ.
√
Noting now that ηi,(cid:96) ≤ θ by (226), the choice
¯δ2 <
2
ψ(cid:48)(cid:48)(η)
max
η≤θ
(233)
corresponds to (193), and allows invoking property P4, so that it is legitimate to use the Taylor
expansion:
log ϕηi,(cid:96)(ui,(cid:96)) = −ψ(cid:48)(cid:48)(ηi,(cid:96))
u2
i,(cid:96)
2
+ ψ(cid:48)(cid:48)(cid:48)(ηi,(cid:96))
(jui,(cid:96))3
6
+
R(ui,(cid:96)).
(234)
This leads to
October 28, 2018
n(cid:89)
S(cid:89)
i=1
(cid:96)=1
elog ϕηi,(cid:96) (ui,(cid:96))
i,(cid:96)
u2
2 +ψ(cid:48)(cid:48)(cid:48)(ηi,(cid:96))
(jui,(cid:96))3
6 +R(ui,(cid:96))
i=1
(cid:96)=1
ϕηi,(cid:96)(ui,(cid:96)) =
n(cid:89)
S(cid:89)
n(cid:89)
S(cid:89)
e−ψ(cid:48)(cid:48)(ηi,(cid:96))
(cid:80)S
(cid:80)n
(cid:96)=1 −ψ(cid:48)(cid:48)(ηi,(cid:96))
(cid:96)=1
i=1
=
= e
i=1
i,(cid:96)
u2
2 +ψ(cid:48)(cid:48)(cid:48)(ηi,(cid:96))
(jui,(cid:96))3
6 +R(ui,(cid:96)).
(235)
DRAFT
56
We now show that all terms in the exponent are convergent as n → ∞. By using the expressions
for ηi,(cid:96), ui,(cid:96) in (226) and (227), and recalling the definition of ξi,(cid:96) in (87), straightforward
manipulations yield:
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
ψ(cid:48)(cid:48)(ηi,(cid:96))
u2
i,(cid:96)
2
=
=
∞(cid:88)
S(cid:88)
1
µs2
k,µ
i=1
φ(cid:48)(cid:48)
k,µ(θ/µ)
µs2
k,µ
(cid:96)=1
u2
2
=
u2
2
,
i,(cid:96)ψ(cid:48)(cid:48)(ξi,(cid:96)θ/µ)
ξ2
u2
2
(236)
having used, in the last two equalities, Eq. (90) and the definition of sk,µ in (147). Similarly,
(jui,(cid:96))3
6
i,(cid:96)ψ(cid:48)(cid:48)(cid:48)(ξi,(cid:96)θ/µ)
ξ3
(ju)3
6
A1 (cid:44)
=
=
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
1
k,µ
µ3/2s3
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
µ3/2s3
k,µ
ψ(cid:48)(cid:48)(cid:48)(ηi,(cid:96))
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
(ju)3
.
6
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
(237)
(238)
(239)
Finally, by defining
A2 (cid:44)
R(ui,(cid:96)),
and using (189), we have (recall that bk,(cid:96)(i) ≤ 1):
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
A2 ≤
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
ui,(cid:96)4
R(ui,(cid:96)) ≤ M
24
∞(cid:88)
S(cid:88)
i=1
(cid:96)=1
≤ M
24s4
k,µ
µ2
(1 − µ)4(i−1)u4
=
M S
24s4
k,µ
µ2
1 − (1 − µ)4u4.
We stress that the constant M appearing in the above formulas bounds uniformly all terms
R(ui,(cid:96)). Indeed, as already observed, in the considered range of analysis we have ui,(cid:96) ≤ ¯δ
and ηi,(cid:96) ≤ θ. Under this assumption, from property P4 we know that the bounding constant M
depends only on the pair (¯δ, θ), and, hence, not on the particular pair (i, (cid:96)).
Using (230) and (235), it is now legitimate to rewrite
ϕk,µ(u) = e− u2
2 +A1+A2.
(240)
DRAFT
October 28, 2018
Making explicit the definition of ζk,µ(u) appearing in (214), we have:
(cid:33)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
6 µ3/2s3
k,µ
(ju)3
k,µ(θ/µ)
µ3/2s3
k,µ
(ju)3
6
=
ϕk,µ(u) − ζk,µ(u)
2
1 +
(cid:32)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕk,µ(u) − e− u2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕk,µ(u)e
2 − 1 − φ(cid:48)(cid:48)(cid:48)
(cid:12)(cid:12)
2 (cid:12)(cid:12)eA1+A2 − 1 − A1
(cid:18)
(cid:19)
u2
A2 +
A22
2
= e− u2
2
= e− u2
≤ e− u2
2
where, in the last inequality, we resorted to the following result, which is known to hold for any
pair of complex numbers z and w -- see, e.g., [40, Eq. (2.8), p. 534]:
emax(z,w).
ez − 1 − w ≤
z − w +
w2
(cid:18)
(242)
Consider first the term A1. Using (147) and (91) with r = 3, we can write:
Making explicit the definition of limit, for any > 0, and for µ small enough, it holds that:
eA1+A2,
1
2
(cid:19)
(cid:12)(cid:12)(cid:12)(cid:12) φ(cid:48)(cid:48)(cid:48)(θ)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)√
(cid:12)(cid:12)(cid:12)(cid:12) +
(cid:19)
µu3
[φ(cid:48)(cid:48)(θ)]3/2
(cid:12)(cid:12)(cid:12)(cid:12) .
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) φ(cid:48)(cid:48)(cid:48)
lim
µ→0
k,µ(θ/µ)
µ2s3
k,µ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) φ(cid:48)(cid:48)(cid:48)
(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) φ(cid:48)(cid:48)(cid:48)(θ)
k,µ(θ/µ)
µ2s3
k,µ
[φ(cid:48)(cid:48)(θ)]3/2
µu3,
a1√
1
6µ
≤ 1
6µ
1
µ
=
A1 =
(cid:32)
(cid:33)
1
s4
k,µ
57
(241)
(243)
(245)
(246)
DRAFT
√
µu3
(244)
where the constant a1 has been implicitly defined. We switch to the term A2. Applying L'Hospital's
rule [43], and using again (147), we see that:
4[φ(cid:48)(cid:48)(θ)]2 ,
implying that, for any > 0, and for sufficiently small values of µ:
1 − (1 − µ)4
lim
µ→0
=
µ
1
A2 ≤ 1
µ
≤ 1
µ
1
µ
=
October 28, 2018
M S
24s4
k,µ
(cid:18)
M S
24
√
1 − (1 − µ)4 (
µ
1
4[φ(cid:48)(cid:48)(θ)]2 +
(cid:19)
µu)4
√
(
µu)4
√
µu)4,
a2 (
where the constant a2 has been implicitly defined. Observe now that, for any a > 0, any 0 ≤
x ≤ δ, and for any integers m and n such that m < n:
58
δn−m ≤ 1/a ⇒ axn ≤ xm.
µu, and recalling that in the considered range √
Setting x = √
implies that the parameter δ can be certainly chosen to satisfy:
µu)4 ≤ √
µu3 ≤ √
√
a2(
a1√
,
µu2
8
µu2
8
,
yielding (in the considered range):
(247)
µu ≤ δ, the above relationship
Finally, using these inequalities in (241), the integrand in (214) can be upper bounded as:
A1 ≤ u2
8
,
A2 ≤ u2
8
.
(cid:18)√
(cid:12)(cid:12)(cid:12)(cid:12) ≤ e− u2
4
(cid:19)
2u7
2
µ a2u3 + µ3/2 a2
,
(250)
u
implying that I1 → 0 as µ → 0.
Part ii (I2 → 0 as µ → 0). The following chain of inequalities holds true:
(248)
(249)
1√
µ
(cid:12)(cid:12)(cid:12)(cid:12) ϕk,µ(u) − ζk,µ(u)
(cid:90)
(cid:90)
I2 ≤ 1√
µ
(cid:12)(cid:12)(cid:12)(cid:12) du +
u
(cid:12)(cid:12)(cid:12)(cid:12) ϕk,µ(u)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 +
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
6 µ3/2s3
u
k,µ
(ju)3
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) du
≤ 1
δ
ϕk,µ(u)du +
µ <u< a√
δ√
µ
e− u2
2
µ
1√
µ
u> δ√
(cid:90)
(cid:90)
1
u> δ√
δ
φ(cid:48)(cid:48)(cid:48)
k,µ(θ/µ)
6 µ2s3
µ
µ
µ <u< a√
δ√
e− u2
(cid:90)
k,µ
2 du +
u> δ√
µ
(cid:90) ∞
(cid:90) ∞
e− u2
2 u2du,
(251)
which follows by repeated application of the triangle inequality for complex numbers z + w ≤
z + w. Since the integrals
e− u2
are convergent, and since the quantities φ(cid:48)(cid:48)(cid:48)
the last two integrals on the RHS of (251) vanish as µ → 0.
e− u2
2 du,
−∞
−∞
2 u2du
k,µ(θ/µ)/µ2 and sk,µ converge in view of (91) and (147),
(252)
October 28, 2018
DRAFT
It remains to prove that the first integral on the RHS of (251) vanishes as µ → 0. To this aim,
observe that, by (225):
ϕk,µ(u) =
∞(cid:89)
S(cid:89)
i=1
(cid:96)=1
ν(ηi,(cid:96) + jui,(cid:96))
ν(ηi,(cid:96))
.
Since, from (177), we know that
ν(ηi,(cid:96) + jui,(cid:96))
ν(ηi,(cid:96))
≤ 1,
all the terms of the product in (253) are not greater than one, implying that, in particular:
ϕk,µ(u) ≤ 2(cid:98)1/µ(cid:99)(cid:89)
S(cid:89)
ν(ηi,(cid:96) + jui,(cid:96))
i=(cid:100)1/µ(cid:101)
ν(ηi,(cid:96))
(cid:96)=1
,
(255)
where (cid:100)x(cid:101) (resp., (cid:98)x(cid:99)) denotes the smallest (resp., the largest) integer not smaller (resp., not
greater) than x. Since in the above product we have 1/µ ≤ (cid:100)1/µ(cid:101) ≤ i ≤ 2(cid:98)1/µ(cid:99) ≤ 2/µ, and
since, in the range of interest δ < √
(1 − µ)2/µ−1bk,(cid:96)(i)
µu < a, for ui,(cid:96) defined in (227) we can write:
(1 − µ)1/µ−1bk,(cid:96)(i)
δ < ui,(cid:96) <
a.
sk,µ
The following convergences hold:
59
(253)
(254)
(256)
(257)
(258)
(259)
(261)
(262)
DRAFT
lim
µ→0
(1 − µ)1/µ = 1/e,
lim
i→∞ bk,(cid:96)(i) = p(cid:96),
lim
µ→0
sk,µ(i) = (cid:112)φ(cid:48)(cid:48)(θ).
(cid:32)
(cid:33)
−
δ < ui,(cid:96) <
(cid:32)
p(cid:96)/e2(cid:112)φ(cid:48)(cid:48)(θ)
v1 (cid:44)
sk,µ
[known limit]
[Eq. (18)]
[Eq. (147)]
(cid:33)
p(cid:96)/e(cid:112)φ(cid:48)(cid:48)(θ)
Using the explicit definition of limit, the above three relationships imply that, for any > 0, it is
possible to choose µ sufficiently small so as to ensure that, in the considered range i ≥ (cid:100)1/µ(cid:101):
+
a (cid:44) v2,
(260)
where, due to arbitrariness of , we can safely assume that v1 > 0.
Moreover, for ηi,(cid:96) defined in (226), we have 0 ≤ ηi,(cid:96) ≤ θ. Accordingly, by defining the set:
S(cid:96) = {(η, u) ∈ R2 : 0 ≤ η ≤ θ, v1 ≤ u ≤ v2},
(cid:90)
ϕk,µ(u)du ≤ 2
a − δ√
µ
c1/µ+1,
µ <u< a√
δ√
µ
we have:
October 28, 2018
where
S(cid:89)
S(cid:89)
(cid:96)=1
c =
=
(cid:18)
ν(η + ju)
ν(η)
max
(η,u)∈S(cid:96)
ν(¯η(cid:96) + j ¯u(cid:96))
ν(¯η(cid:96))
(cid:96)=1
60
(cid:19)
.
(263)
In the above equation, (¯η(cid:96), ¯u(cid:96)) are the pairs where the maxima are attained, which belong to S(cid:96)
since ν(η+ju)
, regarded as a function of the pair (η, u), is continuous.
ν(η)
ν(η)
In view of property P3, ν(η+ju)
is (as a function of u) nothing but the CF of xk(n) under
an exponential measure translation. If xk(n) is not lattice [40], the magnitude of the CF attains
the value one only at the origin. This implies that all the maxima appearing in (263) are strictly
less than 1, because u is bounded away from zero in S(cid:96). Since, for any c < 1, the function c1/µ
vanishes faster than any power of µ as µ → 0, we conclude that the integral in (262) vanishes
as µ goes to zero, and the proof of the lemma is complete.
(cid:3)
REMARK VI. The proof of Lemma 3 assumes that the local statistic xk(n) is not lattice. The
distribution of a lattice random variable with span ∆ is concentrated at the points d, d ± ∆, d ±
2∆, . . . for a certain real number d. The pertinent CF becomes:
∞(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∞(cid:88)
m=−∞
ν(η + ju)
m=−∞
=
ν(η)
where pm = P[xk(n) = d + m∆]. Accordingly, we have
ν(η)
pmeη(d+m∆)ejt(d+m∆)
ν(η + jt)
ν(η)
=
pmeη(d+m∆)ejtm∆
ν(η)
,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
,
(264)
(265)
which shows that the magnitude of the CF has period 2π/∆ and, in particular, it assumes the
value one at all points t = 2πh/∆, for h = 0,±1,±2, . . . , which would violate the last step in
our proof above.
October 28, 2018
(cid:3)
DRAFT
REFERENCES
61
[1] P. Braca, S. Marano, V. Matta, and A. H. Sayed, "Asymptotic performance of adaptive distributed detection over networks,"
submitted for publication. Available as arXiv:1401.5742v2 [cs.IT], Jan. 2014.
[2] P. Braca, S. Marano, V. Matta, and A. H. Sayed, "Large deviations analysis of adaptive distributed detection," in Proc.
IEEE ICASSP, Florence, Italy, May 2014, pp. 6112 -- 6116.
[3] V. Matta, P. Braca, S. Marano, and A. H. Sayed, "Exact asymptotics of distributed detection over adaptive networks," in
Proc. IEEE ICASSP, Brisbane, Australia, Apr. 2015, pp. 3377 -- 3381.
[4] C. G. Lopes and A. H. Sayed, "Diffusion least-mean squares over adaptive networks: Formulation and performance
analysis," IEEE Trans. Signal Process., vol. 56, no. 7, pp. 3122 -- 3136, Jul. 2008.
[5] F. S. Cattivelli and A. H. Sayed, "Diffusion LMS strategies for distributed estimation," IEEE Trans. Signal Process., vol. 58,
no. 3, pp. 1035 -- 1048, Mar. 2010.
[6] A. H. Sayed, S.-Y. Tu, J. Chen, X. Zhao, and Z. J. Towfic, "Diffusion strategies for adaptation and learning over networks,"
IEEE Signal Process. Mag., vol. 30, no. 3, pp. 155 -- 171, May 2013.
[7] A. H. Sayed, "Adaptive networks," Proceedings of the IEEE, vol. 102, no. 4, pp. 460 -- 497, Apr. 2014.
[8] P. Braca, S. Marano, and V. Matta, "Enforcing consensus while monitoring the environment in wireless sensor networks,"
IEEE Trans. Signal Process., vol. 56, no. 7, pp. 3375 -- 3380, Jul. 2008.
[9] P. Braca, S. Marano, V. Matta, and P. Willett, "Asymptotic optimality of running consensus in testing statistical hypotheses,"
IEEE Trans. Signal Process., vol. 58, no. 2, pp. 814 -- 825, Feb. 2010.
[10] -- -- , "Consensus-based Page's test in sensor networks," Signal Processing, vol. 91, no. 4, pp. 919 -- 930, Apr. 2011.
[11] D. Bajovic, D. Jakovetic, J. Xavier, B. Sinopoli, and J. M. F. Moura, "Distributed detection via Gaussian running consensus:
Large deviations asymptotic analysis," IEEE Trans. Signal Process., vol. 59, no. 9, pp. 4381 -- 4396, Sep. 2011.
[12] D. Bajovic, D. Jakovetic, J. M. F. Moura, J. Xavier, and B. Sinopoli, "Large deviations performance of consen-
sus+innovations distributed detection with non-Gaussian observations," IEEE Trans. Signal Process., vol. 60, no. 11,
pp. 5987 -- 6002, Nov. 2012.
[13] D. Jakovetic, J. M. F. Moura, and J. Xavier, "Distributed detection over noisy networks: Large deviations analysis," IEEE
Trans. Signal Process., vol. 60, no. 8, pp. 4306 -- 4320, Aug. 2012.
[14] J. Tsitsiklis, D. Bertsekas, and M. Athans, "Distributed asynchronous deterministic and stochastic gradient optimization
algorithms," IEEE Trans. Autom. Control, vol. 31, no. 9, pp. 803 -- 812, Sep. 1986.
[15] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging," Systems and Control Letters, vol. 53, no. 1, pp.
65 -- 78, Sep. 2004.
[16] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah "Randomized gossip algorithms," IEEE Trans. Inf. Theory, vol. 52, no. 6,
pp. 2508 -- 2530, Jun. 2006.
[17] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multiagent optimization," IEEE Trans. Autom. Control,
vol. 54, no. 1, pp. 48 -- 61, Jan. 2009.
[18] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione, "Gossip algorithms for distributed signal
processing," in Proc. IEEE, vol. 98, no. 11, pp. 1847 -- 1864, Nov. 2010.
[19] S. Kar and J. M. F. Moura, "Convergence rate analysis of distributed gossip (linear parameter) estimation: Fundamental
limits and tradeoffs," IEEE J. Sel. Topics Signal Process., vol. 5, no. 4, pp. 674 -- 690, Aug. 2011.
[20] M. Longo, T. D. Lookabaugh, and R. M. Gray, "Quantization for decentralized hypothesis testing under communication
constraints," IEEE Trans. Inf. Theory, vol. 36, no. 2, pp. 241 -- 255, Mar. 1990.
October 28, 2018
DRAFT
62
[21] P. K. Varshney, Distributed Detection and Data Fusion.
[22] R. Viswanathan and P. K. Varshney, "Distributed detection with multiple sensors: Part I -- Fundamentals," in Proc. IEEE,
Springer-Verlag, New York, 1997.
vol. 85, no. 1, pp. 54 -- 63, Jan. 1997.
[23] R. S. Blum, S. A. Kassam, and H. V. Poor, "Distributed detection with multiple sensors: Part II -- Advanced topics," in
Proc. IEEE, vol. 85, no. 1, pp. 64 -- 79, Jan. 1997.
[24] J. F. Chamberland and V. V. Veeravalli, "Decentralized detection in sensor networks," IEEE Trans. Signal Process., vol. 51,
no. 2, pp. 407 -- 416, Feb. 2003.
[25] J. F. Chamberland and V. V. Veeravalli, "Wireless sensors in distributed detection applications," IEEE Signal Process.
Mag., vol. 24, no. 3, pp. 16 -- 25, May 2007.
[26] B. Chen, L. Tong, and P. K. Varshney, "Channel-aware distributed detection in wireless sensor networks," IEEE Signal
Process. Mag., vol. 23, no. 4, pp. 16 -- 26, Jul. 2006.
[27] V. Saligrama, M. Alanyali, and O. Savas, "Distributed detection in sensor networks with packet losses and finite capacity
links," IEEE Trans. Signal Process., vol. 54, no. 11, pp. 4118 -- 4132, Nov. 2006.
[28] J. N. Tsitsiklis, "Decentralized detection," Advs. Statistical Signal Processing, vol. 2, pp. 297 -- 344, 1993.
[29] P. Willett, P. F. Swaszek, and R. S. Blum, "The good, bad and ugly: Distributed detection of a known signal in dependent
Gaussian noise," IEEE Trans. Signal Process., vol. 48, no. 12, pp. 3266 -- 3279, Dec. 2000.
[30] K. Liu and A. M. Sayeed, "Asymptotically optimal decentralized type-based detection in wireless sensor networks," in
Proc. IEEE ICASSP, Montreal, Canada, May 2004, vol. 3, pp. 873 -- 876.
[31] S. Barbarossa, S. Sardellitti, and P. Di Lorenzo "Distributed detection and estimation in wireless sensor networks," in
Academic Press Library in Signal Processing, vol. 2, R. Chellapa and S. Theodoridis, Eds., pp. 329 -- 408, Academic Press,
Elsevier, 2014.
[32] F. S. Cattivelli and A. H. Sayed, "Distributed detection over adaptive networks using diffusion adaptation," IEEE Trans.
Signal Process., vol. 59, no. 5, pp. 1917 -- 1932, May 2011.
[33] R. R. Bahadur, and R. R. Rao, "On deviations of the sample mean," The Annals of Mathematical Statistics, vol. 31, no. 4,
pp. 1015 -- 1027, Dec. 1960.
[34] A. Dembo and O. Zeitouni, Large Deviations Techniques and Applications. Springer, 1998.
[35] F. den Hollander, Large Deviations. American Mathematical Society, 2008.
[36] R. Horn and C. Johnson, Matrix Analysis. Cambridge University Press, Cambridge, UK, 1985.
[37] A. H. Sayed, Adaptive Filters. Wiley, NJ, 2008.
[38] H. V. Poor, An Introduction to Signal Detection and Estimation. Springer-Verlag, 1988.
[39] S. A. Kassam, Signal Detection in Non-Gaussian Noise. Springer-Verlag, 1987.
[40] W. Feller, An Introduction to Probability and Its Applications, vol. 2, Wiley, NY, 1971.
[41] H. Cram´er, "Sur un nouveau th´eor`eme-limite de la th´eorie des probabilit´es," in: Actualit´es Scientifiques et Industrielles,
Hermann, Paris, no. 736, pp. 5 -- 23, 1938.
[42] J. A. Bucklew, Large Deviations Techniques in Decision, Simulation and Estimation. Wiley, 1990.
[43] W. Rudin, Principles of Mathematical Analysis. McGraw-Hill, 1976.
[44] H. L. Van Trees, Detection, Estimation, and Modulation Theory. Part I. Wiley, 1968, (reprinted, 2001).
[45] J. H. Curtiss, "A note on the theory of moment generating functions," The Annals of Statistics, vol. 13, no. 4, pp. 430 -- 433,
1942.
[46] W. Rudin, Real and Complex Analysis. McGraw-Hill, 1987.
October 28, 2018
DRAFT
[47] H. Shao, Mathematical Statistics, 2nd ed., Springer, 2003.
[48] R. N. Bhattacharya and R. R. Rao, Normal Approximation and Asymptotic Expansions. Robert E. Krieger Publishing
Company, 1986.
[49] E. C. Titchsmarsh, The Theory of Functions. Oxford University Press, 1939.
63
October 28, 2018
DRAFT
|
1105.0510 | 1 | 1105 | 2011-05-03T08:27:26 | Voting in a Stochastic Environment: The Case of Two Groups | [
"cs.MA",
"cs.SI",
"eess.SY",
"math.OC",
"physics.soc-ph"
] | Social dynamics determined by voting in a stochastic environment is analyzed for a society composed of two cohesive groups of similar size. Within the model of random walks determined by voting, explicit formulas are derived for the capital increments of the groups against the parameters of the environment and "claim thresholds" of the groups. The "unanimous acceptance" and "unanimous rejection" group rules are considered as the voting procedures. Claim thresholds are evaluated that are most beneficial to the participants of the groups and to the society as a whole. | cs.MA | cs |
ISSN 0005-1179, Automation and Remote Control, 2011, Vol. 72, No. 6, p. 345. c(cid:13) Pleiades Publishing, Ltd., 2011.
Original Russian Text c(cid:13) P.Yu. Chebotarev, A.K. Loginov, Ya.Yu. Tsodikova, Z.M. Lezina, and V.I. Borzenko, 2010,
published in Problemy Upravleniya, 2010, No. 1, pp. 18 -- 25.
CONTROL SCIENCES
Voting in a Stochastic Environment: The Case of Two Groups
P. Yu. Chebotarev, A. K. Loginov, Ya. Yu. Tsodikova,
Z. M. Lezina, and V. I. Borzenko
Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia
Received April 6, 2009
Abstract -- Social dynamics determined by voting in a stochastic environment is analyzed for
a society composed of two cohesive groups of similar size. Within the model of random walks
determined by voting, explicit formulas are derived for the capital increments of the groups
against the parameters of the environment and "claim thresholds" of the groups. The "unan-
imous acceptance" and "unanimous rejection" group rules are considered as the voting proce-
dures. Claim thresholds are evaluated that are most beneficial to the participants of the groups
and to the society as a whole.
1. INTRODUCTION
This work was carried out within the framework of analyzing social dynamics determined by
democratic decisions in a stochastic environment. The subject of investigation is the relationship
between the parameters of social dynamics and social strategies of the participants including egoism,
collectivism (corporatism) and altruism. The results for the situation where a cohesive group
competes with egoistic participants are reported in [1, 2]; competition between egoists and two
cohesive groups (parties) was studied in [3]. The latter paper also considers the so-called "snowball
of cooperation" mechanism which, under certain conditions, motivates those participants initially
having egoistic strategies to act altruistically.
The present paper investigates a special case of the situation considered in [3], namely, compe-
tition of two groups. The relative simplicity of this case enables us to obtain explicit formulas for
the most interesting relationships.
Let us recall the basic features of our model of voting in a stochastic environment. Being
composed of a finite number of participants, the society sequentially votes for/against proposals that
are regularly generated by the "environment" according to a random law. A proposal is identified
with a vector of capital increments of the participants; an alternative interpretation involves the
utility values of the participants. According to the model, the capital/utility increments entering
the proposal are realizations of independent and identically distributed random variables. Here,
the normal distribution is mainly considered. During the vote, an egoistic participant supports any
proposal that increases his/her capital. In contrast to egoists, the members of a group vote jointly
for those proposals that would be favorable for the whole group (the latter is characterized by a
certain index). In particular, a group may support proposals that increase either the capital of the
majority of its members or the total group capital. Each proposal is accepted (and implemented)
or rejected under a particular voting procedure. Generally, the procedures of α-majority are used.
The parameter α ∈ [0, 1) determines the share of votes necessary and sufficient for the acceptance
of a proposal.
The presence of a stochastic environment is the focus of the model under consideration;
it
preserves many basic phenomena of social reality. At the same time, it enables the investigation of
these phenomena by analytical methods.
345
346
CHEBOTAREV, LOGINOV, TSODIKOVA, LEZINA, AND BORZENKO
In voting theory, the stochasticity assumption is often applied to choosing the voters' positions
(see, e.g., [4]). In cooperative game theory, it is widely adopted with respect to the payments [5].
A distinctive feature of our model is a stochastic mechanism for generating proposals for voting. In
other words, we study a model of random walks controlled by voting. In this context, some papers
on the dynamical correction of tax rates are relevant (see, e.g., [6]); note, however, that these
works focus on selecting the rate itself (which is optimized within special models of production and
consumption) and not on the random walks in the space of individual utilities. We also mention a
recent survey on modelling political competition processes that involve voting [7].
2. EXPECTED CAPITAL OF GROUP MEMBERS
Suppose that a society consists of two groups of similar size; every group votes jointly. A natural
voting rule requires that the accepted proposals are supported by both groups. In the sequel, we
use the term unanimous acceptance group rule for this voting procedure. Suppose that the first
(second) group supports a proposal if and only if this proposal leads to the increase in the mean
capital of the group members by at least t1 (respectively, t2). Here, the proposal support thresholds
t1 and t2 are variable parameters which may take positive, zero, or negative values. It is interesting
to investigate the dependence of the future capital of the groups and the whole society on the
proposal support thresholds t1 and t2 (we will also call them claim thresholds).
In addition to
the unanimous acceptance group rule, we will consider the rule according to which a proposal is
accepted if and only if it is supported by at least one group. This voting procedure will be called
the unanimous rejection group rule, since a proposal is rejected iff it is rejected by both groups.
In the following theorem, we derive expressions for the average capital increments under the above
voting rules.
Theorem 1. Suppose a proposal is accepted if and only if it is supported by at least one group.
Then the mathematical expectation of the one-step capital increment 1 of a member of group i
(i = 1, 2) is given by the formula
M ( di) = µF3−i + (µFi + σifi)F3−i.
(1)
Here, Fi = F (µi/σi), Fi = 1 − Fi, fi = f (µi/σi), f (·) and F (·) denote the density and distribution
function of the standard normal distribution, respectively,2 µi = µ − ti, σi = σ.√gi, ti is the claim
threshold of group i, gi is the size of group i, while µ and σ are the parameters of the normal
distribution N (µ, σ2) that describes the independent capital increments forming proposals.
Suppose a proposal is accepted if and only if it is supported by both groups. Then the mathematical
expectation of the one-step capital increment of a member of group i (i = 1, 2) is given by the formula
M ( di) = (µFi + σifi)F3−i.
(2)
Proof. For a given proposal, let G1 (G2) be the event that the first (respectively, second) group
supports the proposal. Denote by G1G2 the simultaneous implementation of G1 and G2. Similarly,
G1G2 and G1G2 are the simultaneous implementation of G1 and the complement of G2, and of the
complements of G1 and G2, respectively. Let P (·) denote the event probability.
To prove the first statement of the theorem, we first suppose that the unanimous rejection group
rule is used. Using the total probability formula one has
M ( di) = M ( di G3−i) P (G3−i) + M ( di GiG3−i) P (GiG3−i) + M ( di GiG3−i) P (GiG3−i).
(3)
1 Notation with tilde serves to indicate actual (i.e, taking into account the acceptance of proposals) capital increments,
as distinct from the proposed ones.
2 f (t) = 1
√2π
e−t2/2 and F (t) = 1
√2π R t
−∞
e−x2/2dx.
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
VOTING IN A STOCHASTIC ENVIRONMENT
347
On the assumption of GiG3−i, the proposal is rejected; hence, M ( di GiG3−i) = 0. Un-
der G3−i, the proposal is accepted and independence of the proposal components implies that
M ( di G3−i) = µ. Similarly, the proposal is accepted under GiG3−i, and independence of the com-
ponents di yields M ( di GiG3−i) = M (di GiG3−i) = M (di Gi). Next, M (di Gi) = M (d ave
Gi)
with d ave
standing for the result of averaging the components of the proposal (these components
correspond to the ith group). Observe that d ave
is a random variable having the distribution
N (µ, σ2
i > ti. Using the formula for the
conditional mean of a normal random variable (it can be proved by integration) one finally obtains
i ), where σi = σ.√gi, and that Gi is tantamount to d ave
i
i
i
M ( di GiG3−i) = M (d ave
i
i > ti) = µ +
d ave
σifi
Fi
,
(4)
where Fi = F (µi/σi), fi = f (µi/σi), and µi = µ − ti. It is also easy to demonstrate that
P (G3−i) = F3−i. Using independence of the proposal components we have the expression
P (GiG3−i) = P (Gi) P (G3−i) = FiF3−i.
(5)
Substituting all this into (3) yields
M ( di) = µF3−i + (cid:18)µ +
σifi
Fi (cid:19) Fi F3−i = µF3−i + (µFi + σifi)F3−i.
This completes the proof of the first part.
Now prove the second statement of the theorem. Suppose that the unanimous acceptance group
rule is used. In this case,
Similarly to the derivation of (4), one obtains
M ( di) = M ( di GiG3−i) P (GiG3−i).
M ( di GiG3−i) = M (d ave
i
i > ti) = µ +
d ave
σifi
Fi
and similarly to (5),
P (GiG3−i) = P (Gi) P (G3−i) = FiF3−i.
(6)
(7)
(8)
Substitute (7) -- (8) in (6) to get M ( di) = (µFi + σifi) F3−i. This completes the proof. ♦
Of special interest are the comparative capital increments of the groups. Indeed, it is important
to know the capital-related performance of each group against the background of the results achieved
by the other one. The expected one-step capital increment of Group 1 as compared with the one
of Group 2 is expressed as M ( d1 − d2). Theorem 1 enables one to obtain simple formulas for
M ( d1 − d2).
Corollary 1. In the notation of Theorem 1, if the unanimous rejection group rule is adopted,
then
Under the unanimous acceptance group rule, one has
M ( d1 − d2) = σ1 f1 F2 − σ2 f2 F1.
M ( d1 − d2) = σ1 f1 F2 − σ2 f2 F1.
The proof of Corollary 1 is straightforward.
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
(9)
(10)
348
CHEBOTAREV, LOGINOV, TSODIKOVA, LEZINA, AND BORZENKO
3. ANALYSIS OF SOCIAL DYNAMICS
In order to apply and interpret the results obtained, let us first consider the case where the
mean proposed capital increment is zero, i.e., the environment is neutral; suppose that the standard
deviation of the proposed capital increment is 10: µ = 0, σ = 10. Suppose that each group includes
300 participants and Group 1 supports a proposal whenever it increases the total capital of the
group (t1 = 0). The claim threshold t2 of Group 2 varies.
In the case of the unanimous acceptance group rule (sometimes, we will call it the (G1∧G2)-rule),
the expected capital increments of the groups and the whole society versus the claim threshold t2
of Group 2, as stated by Theorem 1, are depicted in Fig. 1.
Fig. 1. Expected capital increments of the groups and the society provided that the unanimous acceptance
group rule is adopted (in this example, it coincides with the majority rule): 1 -- Group 1; 2 -- Group 2; 3 --
the difference between Group 1 and Group 2; 4 -- the whole society. The parameters are: µ = 0; σ = 10;
g1 = g2 = 300.
In particular, the average capital increment of Group 2 represents a symmetrical bell-shaped
curve. This can be explained as follows. Setting a positive Group 2 threshold (high claims) leads to
the rejection of some favorable to Group 2 (on the average) proposals, whereas setting a negative
threshold (low claims) leads to the acceptance, along with the favorable to Group 2 proposals, of
some proposals unfavorable to it. So both strategies decrease the expected capital increment of
Group 2 as against the case of t2 = 0. A noteworthy fact is that positive and negative thresholds
of the same absolute value are equivalent. Let us prove it.
Corollary 2. Let µ = 0. Then for fixed values of the model parameters except for t2, any t ∈ IR,
and under either group voting rule, M ( d2 t2 = t) = M ( d2 t2 = −t) holds.
Proof. Applying Theorem 1 to the case µ = 0 one obtains
M ( d2) = σ2f2 F1
for the unanimous rejection group rule and
M ( d2) = σ2f2 F1
(11)
(12)
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
for the unanimous acceptance group rule.
VOTING IN A STOCHASTIC ENVIRONMENT
349
The required statement follows from the facts that σ2 and F1 are independent of t2 and that
f2 = f (cid:18) µ − t2
σ2 (cid:19) is an even function of t2 provided that µ = 0. ♦
Remark 1. Due to (11) -- (12), for both voting rules under consideration, the expression of M ( d2)
in terms of t2 is proportional to the normal density function, which is a bell-shaped curve with zero
limits as t2 → −∞ and t2 → ∞. Finally, if t1 = 0, then under both voting rules, M ( d2) = 1
2 σ2f2
holds. ♦
Next we study the characteristics of social dynamics for each voting rule separately.
3.1. The Unanimous Acceptance Group Rule
If the unanimous acceptance group rule is adopted, then Group 1 benefits from low (negative)
claim threshold t2 of Group 2. Indeed, in this case Group 2 rarely vetoes proposals beneficial to
Group 1 and so Group 1 can maximize its benefits. Contrariwise, if t2 is high, then Group 1 can
rarely expect approval of proposals beneficial to it.
Now, suppose that Group 2 aims not at maximizing the average capital increment, but at
ensuring maximum advantage over Group 1. In other words, Group 2 maximizes M ( d2 − d1) rather
than M ( d2). A tool Group 2 can use for this is its claim threshold t2. For instance, in the example
illustrated by Fig. 1, M ( d2 − d1) as a function of t2 attains its maximum when M ( d1 − d2) attains
its minimum, i.e., at t2 ≈ 0.46. In the proposition below, we find the "claim threshold" t2 that
maximizes M ( d2 − d1) in the general case.
Proposition 1. Suppose the unanimous acceptance group rule is adopted. Then, in the notation
of Theorem 1, the expected advantage M ( d2 − d1) of a Group 2 member over a Group 1 member
in the sense of capital increment attains its unique maximum at the claim threshold t2 of Group 2
defined by
Proof. Differentiating (10) with respect to t2 we obtain
t+
2 = µ +
σ1f1
F1
.
dM ( d1 − d2)
d t2
=
d (σ1 f1 F2 − σ2 f2 F1)
d t2
= σ1 f1f2·(cid:0)−σ−1
2 (cid:1) − σ2f2−(t2 − µ)
σ2
2
(13)
F1.
Setting the derivative equal to zero results in the desired expression (13). Finding the second
derivative of the function M ( d1− d2) ensures that a minimum has been found. Hence, (13) provides
the unique maximum of M ( d2 − d1). ♦
Remark 2. It should be noted that the optimal claim threshold t+
2 found in Proposition 1 is
independent of the group size g2. It is also interesting that expression (13) coincides with (4) in
which i = 1.
In turn, (4) with i = 1 expresses the expected capital increment of a member
of Group 1 provided it exceeds the threshold t1. This observation suggests the following simple
algorithm of estimating the optimal "claim threshold" t+
2 by Group 2: (1) find the average capital
increments of Group 1 members according to the proposals that Group 1 supports; (2) set the
threshold t2 equal to the value found on Step (1) averaged over the proposals.
Thus, to achieve an advantage over Group 1, Group 2 should set its claim threshold at a higher
level than Group 1 does. As has been noted before, the optimal threshold t2 is equal to the mean
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
350
CHEBOTAREV, LOGINOV, TSODIKOVA, LEZINA, AND BORZENKO
2 = q 2
"above threshold" value of Group 1 capital increment. In particular, for the example of Fig. 1, (13)
gives t+
3π . Now, if Group 1 wishes to act in the same way (i.e., to maximize its advantage
M ( d1 − d2) with respect to t2 = t+
2 . If this
process continues, it leads to an infinite increase in the claims (the process is divergent and the
corresponding game possesses no Nash equilibrium). ♦
2 ), it should, in turn, set a higher threshold: t1 > t+
How does the maximum comparative gain M ( d2 − d1) of Group 2 depend on the fixed claim
threshold t1 of Group 1? It increases as t1 → −∞, since in this case Group 1 supports all proposals,
and thus, the decisions are made by Group 2. If t1 grows, then this maximum gain decreases and
tends to zero as t1 → ∞ since in this case, no proposals are accepted due to the veto of Group 1.
Now consider the following question which seems quite interesting. What claim threshold t2 of
Group 2 is optimal for the whole society (thus, leading to the maximum expected capital increment
of the society)? The answer is provided by the following proposition.
Proposition 2. Suppose the unanimous acceptance group rule is adopted. Then the expected
capital increment of the whole society attains its maximum value at the claim threshold t2 of Group 2
given by
t0
2 = −
g1
g2 (cid:18)µ +
σ1f1
F1 (cid:19) .
(14)
Proof. Theorem 1 (see (2)) implies that the expected capital increment of the whole society is
g1M ( d1) + g2M ( d2) = g1(µF1 + σ1f1)F2 + g2(µF2 + σ2f2)F1
= ((g1 + g2)µF1 + g1σ1f1) F2 + g2σ2f2F1.
The derivative of this function with respect to t2 is
d (g1M ( d1) + g2M ( d2))
d t2
= ((g1 + g2)µF1 + g1σ1f1)f2·(cid:0)−σ−1
2 (cid:1) + g2σ2 f2F1−(t2 − µ)
σ2
2
.
Setting it equal to zero one obtains g2 F1 (t0
2 − µ) = −(g1 + g2)µF1 − g1σ1f1, which leads to (14).
Finally, the second derivative test confirms that this point provides a unique maximum value. ♦
Remark 3. Comparing Propositions 1 and 2 produces a somewhat unexpected result. Namely,
at g1 = g2 those values of claim threshold t2 of Group 2 leading to the maximum advantage over
Group 1 (on the one hand) and to the best results for the whole society (on the other hand) turn
out to be opposite. In particular, this is the case for the example illustrated by Fig. 1. Note that
even if g1 6= g2, the thresholds t+
2 always have different signs (or both are zero), since by
(13) and (14),
2 and t0
t0
2 = −
g1
g2
t+
2 . ♦
Proposition 2 determines t2 (as a function of t1) that guarantees the maximum capital increment
of the whole society. Let us maximize this maximum over t1 to find the global optimal point of the
society. Obviously, the corresponding solutions t1 and t2 satisfy the system of equations
t1 = −
g2
g1 (cid:18)µ +
σ2f2
F2 (cid:19) ;
t2 = −
g1
g2 (cid:18)µ +
σ1f1
F1 (cid:19)
(15)
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
VOTING IN A STOCHASTIC ENVIRONMENT
351
with σ1, σ2, f1, f2, F1, and F2 defined in Theorem 1. Most likely, the solution of (15) cannot generally
be expressed in elementary functions. In particular, this applies even to the simplest case of g1 = g2
(two groups of equal size) in which the system (15) reduces to
t1 = −µ −
σ1f1
F1
;
t2 = t1.
Nevertheless, the solution can be obtained numerically. For instance, the dependence on µ of
the claim threshold t = t1 = t2 optimal for the society is shown (in the case of g1 = g2 = 300 and
σ = 10) in Fig. 2.
Fig. 2. Relationship between µ and the claim threshold t = t1 = t2 that maximizes the capital increment of
the whole society provided that g1 = g2 = 300 and σ = 10.
In particular, for g1 = g2 = g and µ = 0, the optimal threshold is t = −σy0(cid:14)√g, where y0 ≈ 0.506
is the unique solution of the equation y = f (y)/F (y) with f (y) and F (y) designating the density
and distribution function of the standard normal distribution.
One of the most instructive results of this study is that under the unanimous acceptance group
rule, the claim threshold t of the groups that is optimal for the whole society is negative (see Fig. 2).
In the case g1 = g2, it tends to zero from below as µ → −∞ and asymptotically approaches to the
line t = −µ as µ → ∞. Negative claim thresholds of the groups can be interpreted as a willingness
to a moderately negative result, that is, as a certain appetite for risk.
3.2. The Unanimous Rejection Group Rule
Now consider the unanimous rejection group rule (we will also call it the (G1 ∨ G2)-rule). With
this procedure, for acceptance of a proposal, the support of either one group is sufficient. The
analytical results obtained in Section 2 are illustrated by the diagram in Fig. 3.
It can be easily observed that this diagram mirrors the one shown in Fig. 1. Indeed, the following
statement holds.
Corollary 3. Let µ = 0. If the model parameters, except for t2, are fixed, then for any t ∈ IR,
MG1∨G2( d1 t2 = t) = MG1∧G2( d1 t2 = −t) holds, where MG1∧G2 (MG1∨G2) is the expectation, pro-
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
352
CHEBOTAREV, LOGINOV, TSODIKOVA, LEZINA, AND BORZENKO
Fig. 3. Expected capital increments of the members of Groups 1 and 2 and the society under the unanimous
rejection group rule: 1 -- Group 1; 2 -- Group 2; 3 -- the difference between Group 1 and Group 2; 4 -- the whole
society; the parameters are: µ = 0; σ = 10; g1 = g2 = 300.
vided the unanimous acceptance group rule (resp., the unanimous rejection group rule) is adopted.
Moreover, at t1 = 0 one has MG1∨G2( d2 t2 = t) = MG1∧G2( d2 t2 = t) for any t ∈ IR.
Proof. According to Theorem 1, for µ = 0 we have:
MG1∧G2( d1 t2 = −t) = σ1f1 F (t/σ1),
MG1∨G2( d1 t2 = t) = σ1f1 (1 − F (−t/σ1)) = σ1f1 F (t/σ1) = MG1∧G2( d1 t2 = −t).
Moreover, t1 = 0 implies that MG1∧G2( d2 t2 = t) = σ2 f (−t/σ2)F (0) = MG1∨G2( d2 t2 = t).
This completes the proof. ♦
Remark 4. Due to Corollary 2, µ = 0 implies that M ( d2 t2 = t) = M ( d2 t2 = −t) for either
voting rule under study. Consequently, if µ = 0 and t1 = 0, then every relationship under (G1∨G2)-
rule can be derived from the corresponding relationship obtained for the (G1 ∧ G2)-rule by means
of the y-axis reflection. Thus, in this case, the situations where voting is organized according to
the (G1 ∧ G2)-rule and according to the (G1 ∨ G2)-rule are, in a certain sense, dual. ♦
Note that under the (G1 ∨ G2)-rule, Group 1 benefits from excessive claims of Group 2 (i.e.,
from high threshold t2). In this case, if threshold t1 is positive, but not so high, then most decisions
are made for the benefit of Group 1. On the other hand, at a low threshold t2, Group 2 ensures
implementation of almost all proposals, and so most proposals unfavorable to Group 1 pass through
the "riddle" of voting.
Now let us consider the maximization problem for the Group 2 advantage M ( d2 − d1) over
Group 1 provided that the (G1 ∨ G2)-rule is adopted.
Proposition 3. Under the (G1∨G2)-rule, the expected advantage M ( d2− d1) of a Group 2 member
over a Group 1 member attains its maximum value at the claim threshold t2 defined by
t+
2 = µ −
σ1f1
F1
.
(16)
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
VOTING IN A STOCHASTIC ENVIRONMENT
353
Proof. Differentiating (9) with respect to t2 yields:
dM ( d1 − d2)
d t2
=
d (σ1 f1 F2 − σ2 f2 F1)
d t2
= σ1 f1f2·(cid:0)σ−1
2 (cid:1) − σ2f2−(t2 − µ)
σ2
2
F1.
Setting this derivative equal to zero one arrives at (16). The second derivative test confirms
that this unique extremum is a minimum. Thus, (16) provides a maximum value to the function
M ( d2 − d1). ♦
The maximum comparative gain M ( d2− d1) of Group 2 attained at the claim threshold t+
2 tends
to zero as t1 → −∞ (in the limit, all proposals are accepted by the votes of Group 1) and tends to
its maximum value when t1 → ∞ (in the limit, Group 1 does not support any proposals).
As well as in the case of the (G1 ∧ G2)-rule, consider the problem of maximizing, by choosing
t2, the capital of the whole society. The solution is provided by the following proposition.
Proposition 4. Suppose that the (G1 ∨ G2)-rule is adopted. Then the expected capital increment
of the whole society attains its maximum value at the Group 2 claim threshold given by
t0
2 = −
g1
g2 (cid:18)µ −
σ1f1
F1 (cid:19) .
(17)
Proof. By Theorem 1 (see Eq. (1)) the expected capital increment of the whole society is
g1M ( d1) + g2M ( d2) = g1·(cid:0)µF2 + (µF1 + σ1f1)F2(cid:1) + g2·(cid:0)µF1 + (µF2 + σ2f2)F1(cid:1)
= (g1 + g2)µF1 + g1σ1f1 + g2σ2F1f2 + (cid:0)g1·(µ − µF1 − σ1f1) + g2µF1(cid:1)F2.
The derivative of this function w. r. t. t2 is:
d (g1M ( d1) + g2M ( d2))
d t2
= g2σ2F1f2−(t2 − µ)
σ2
2
+ (cid:0)g1·(µ − µF1 − σ1f1) + g2µF1(cid:1)f2·(cid:0)−σ−1
2 (cid:1).
Setting it equal to 0 we have: −g2 F1 (t0
2 − µ) = g1·(µF1 − σ1f1) + g2µF1, which leads to (17).
Finally, using the second derivative test we verify that the point in question provides the maximum
value. ♦
Remark 5. Comparing Propositions 3 and 4 we obtain the following (cf. Remark 3). At g1 = g2,
the value t+
2 of the claim threshold t2 of Group 2 that leads to the maximum advantage over Group 1
and the value t0
2 that maximizes the capital of the whole society are opposite. In particular, this
is the case for the example in Fig. 3. Moreover, even if g1 6= g2, the thresholds t+
2 have
different signs (or both are equal to zero), since, as well as in the case of the (G1 ∧ G2)-rule,
2 = −(g1/g2)t+
t0
2 and t0
2 . ♦
Now consider the issue of maximizing the capital increment of the whole society when the
(G1 ∨ G2)-rule is employed.
In a certain sense, the result is dual to the one derived for the
(G1 ∧ G2)-rule. Namely, Proposition 4 implies that the claim thresholds of the groups, t1 and t2,
that maximize the capital increment of the whole society satisfy the following system of equations:
t1 = −
g2
g1 (cid:18)µ −
σ2f2
F2 (cid:19) ;
t2 = −
g1
g2 (cid:18)µ −
σ1f1
F1 (cid:19) ,
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
354
CHEBOTAREV, LOGINOV, TSODIKOVA, LEZINA, AND BORZENKO
where σ1, σ2, f1, f2, F1, and F2 are defined in Theorem 1. In the simplest case g1 = g2, the above
system of equations reduces to
t1 = −µ +
σ1f1
F1
;
t2 = t1.
The dependence on µ of the claim threshold t of both groups that maximizes the capital of
the society is illustrated (for the case of g1 = g2 = 300 and σ = 10) in Fig. 4. In particular, for
g1 = g2 = g and µ = 0, the corresponding threshold is t = σy0(cid:14)√g, where y0 ≈ 0.506 is the solution
of the equation y = f (y)/F (y), f (y) and F (y) being the density and distribution function of the
standard normal distribution, respectively.
Fig. 4. Relationship between the parameter µ and the claim threshold t = t1 = t2 of both groups that
maximizes the capital increment of the whole society in the case of g1 = g2 = 300 and σ = 10.
In contrast to voting based on the (G1 ∧ G2)-rule, the claim threshold of both groups that
maximizes the capital of the society under the (G1 ∨ G2)-rule is positive (if the voting rule is not
conservative, then the voters have to be conservative); for g1 = g2 it tends to zero from above as
µ → ∞ and asymptotically approaches to the line t = −µ as µ → −∞ (see Fig. 4).
4. CONCLUSION
This paper has studied the social dynamics determined by voting of two cohesive groups in a
stochastic environment. Two groups of similar size (a kind of a "two-party system") have been
considered; this makes reasonable the application of the voting procedure which requires the support
of both groups for the acceptance of any proposal (the unanimous acceptance group rule).
Another procedure in which the support of either group suffices for the acceptance of a proposal
(the unanimous rejection group rule) looks a little less natural. However, using this procedure does
not lead to an infinite sequence of contradictory decisions suggested by different groups since the
proposals in the model are generated by the stochastic environment.
Each group supports only those proposals that increase the average capital of its members by an
amount exceeding the claim threshold set by the group. The relationship between social dynamics
and the claim thresholds of the groups was the main subject of the present study.
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
VOTING IN A STOCHASTIC ENVIRONMENT
355
In particular, we have derived analytical expressions for the capital increments of the groups in
terms of their claim thresholds. It has been shown that under the unanimous acceptance group
rule and the equal size of the groups, the claim threshold of a group that ensures the maximum
advantage over the other group and the claim threshold that maximizes the capital of the whole
society are opposite. Moreover, the first above threshold is equal to the average capital increment of
the competing group members over the proposals supported by the competing group. Furthermore,
it turns out that the claim thresholds of the groups that maximize the capital of the whole society
are negative, which can be interpreted as the profitability of a certain appetite for risk in the case
of a conservative voting rule; the dependence of these thresholds on the environment favorability
has been established. Similar (in fact, dual) results have been obtained for the unanimous rejection
group rule.
It should be emphasized that the phenomena observed are not only inherent in the model
under consideration; the most important ones have real prototypes widely discussed in the relevant
literature.
REFERENCES
1. Borzenko, V.I., Lezina, Z.M., Loginov, A.K., Tsodikova, Ya.Yu., and Chebotarev, P.Yu., Strategies of
Voting in Stochastic Environment: Egoism and Collectivism, Autom. Remote Control , 2006, vol. 67,
no. 2, pp. 311 -- 328.
2. Chebotarev, P.Yu., Analytical Expression of the Expected Values of Capital at Voting in the Stochastic
Environment, Autom. Remote Control , 2006, vol. 67, no. 3, pp. 480 -- 492.
3. Chebotarev, P.Yu., Loginov, A.K., Tsodikova, Ya.Yu., Lezina, Z.M., and Borzenko, V.I., Analysis of
Collectivism and Egoism Phenomena within the Context of Social Welfare, Autom. Remote Control ,
2010, vol. 71, no. 6, pp. 1196 -- 1207.
4. Anderson, S.P., Kats, A., and Thisse, J.-F., Probabilistic Voting and Platform Selection in Multi-Party
Elections, Soc. Choice Welfare, 1994, vol. 11, no. 4, pp. 305 -- 322.
5. Suijs, J., Cooperative Decision Making in a Stochastic Environment, PhD Dissertation, Tilburg: Tilburg
Univ., 1998.
6. Aiyagari, S.R. and Peled, D., Social Insurance and Taxation under Sequential Majority Voting and
Utilitarian Regimes, J. Econom. Dynam. Control , 1995, vol. 18, no. 8, pp. 1511 -- 1528.
7. Zakharov, A.V., The Models of Political Competition: Literature Review, Ekonom. Mat. Metody, 2009,
vol. 45, no. 1, pp. 110 -- 128.
AUTOMATION AND REMOTE CONTROL Vol. 72 No. 6
2011
|
1811.01458 | 3 | 1811 | 2019-09-10T21:21:00 | Bayesian Action Decoder for Deep Multi-Agent Reinforcement Learning | [
"cs.MA",
"cs.AI",
"cs.LG"
] | When observing the actions of others, humans make inferences about why they acted as they did, and what this implies about the world; humans also use the fact that their actions will be interpreted in this manner, allowing them to act informatively and thereby communicate efficiently with others. Although learning algorithms have recently achieved superhuman performance in a number of two-player, zero-sum games, scalable multi-agent reinforcement learning algorithms that can discover effective strategies and conventions in complex, partially observable settings have proven elusive. We present the Bayesian action decoder (BAD), a new multi-agent learning method that uses an approximate Bayesian update to obtain a public belief that conditions on the actions taken by all agents in the environment. BAD introduces a new Markov decision process, the public belief MDP, in which the action space consists of all deterministic partial policies, and exploits the fact that an agent acting only on this public belief state can still learn to use its private information if the action space is augmented to be over all partial policies mapping private information into environment actions. The Bayesian update is closely related to the theory of mind reasoning that humans carry out when observing others' actions. We first validate BAD on a proof-of-principle two-step matrix game, where it outperforms policy gradient methods; we then evaluate BAD on the challenging, cooperative partial-information card game Hanabi, where, in the two-player setting, it surpasses all previously published learning and hand-coded approaches, establishing a new state of the art. | cs.MA | cs | Bayesian Action Decoder for Deep Multi-Agent Reinforcement Learning
Jakob N. Foerster * 1 2 H. Francis Song * 3 Edward Hughes 3 Neil Burch 3 Iain Dunning 3 Shimon Whiteson 1
Matthew M. Botvinick 3 Michael Bowling 3
9
1
0
2
p
e
S
0
1
]
A
M
.
s
c
[
3
v
8
5
4
1
0
.
1
1
8
1
:
v
i
X
r
a
Abstract
When observing the actions of others, humans
make inferences about why they acted as they
did, and what this implies about the world; hu-
mans also use the fact that their actions will be
interpreted in this manner, allowing them to act in-
formatively and thereby communicate efficiently
with others. Although learning algorithms have
recently achieved superhuman performance in a
number of two-player, zero-sum games, scalable
multi-agent reinforcement learning algorithms
that can discover effective strategies and conven-
tions in complex, partially observable settings
have proven elusive. We present the Bayesian
action decoder (BAD), a new multi-agent learn-
ing method that uses an approximate Bayesian
update to obtain a public belief that conditions on
the actions taken by all agents in the environment.
BAD introduces a new Markov decision process,
the public belief MDP, in which the action space
consists of all deterministic partial policies, and
exploits the fact that an agent acting only on this
public belief state can still learn to use its private
information if the action space is augmented to
be over all partial policies mapping private infor-
mation into environment actions. The Bayesian
update is closely related to the theory of mind
reasoning that humans carry out when observ-
ing others' actions. We first validate BAD on a
proof-of-principle two-step matrix game, where
it outperforms policy gradient methods; we then
evaluate BAD on the challenging, cooperative
partial-information card game Hanabi, where, in
the two-player setting, it surpasses all previously
published learning and hand-coded approaches,
establishing a new state of the art.
*Equal contribution 1University of Oxford, UK 2Work done at
DeepMind. JF has since moved to Facebook AI Research, Menlo
Park, USA. 3DeepMind, London, UK. Correspondence to: Jakob
Foerster <[email protected]>, Francis Song <[email protected]>.
Proceedings of the 36 th International Conference on Machine
Learning, Long Beach, California, PMLR 97, 2019. Copyright
2019 by the author(s).
1. Introduction
In multi-agent reinforcement learning (RL), agents must
learn to act in an environment that contains multiple learn-
ing agents, often under partial observability (Littman, 1994).
In recent years, a variety of deep RL methods have been
adapted to this setting (Foerster et al., 2016a; Lowe et al.,
2017; Perolat et al., 2017; Jaderberg et al., 2018). In the par-
ticular case of cooperative, partially observable multi-agent
settings, a key challenge is to discover communication proto-
cols while simultaneously learning policies. Such protocols
are essential for many real-world tasks where agents must
interact and communicate seamlessly with other agents.
State-of-the-art deep RL methods for learning communica-
tion protocols mostly use backpropagation across a com-
munication channel (Sukhbaatar et al., 2016; Foerster et al.,
2016a). This approach has two limitations. First, it can only
be applied to cheap-talk channels in which the communi-
cation action has no effect on the environment. Second, it
misses the conceptual connection between communication
and reasoning over the beliefs of others, which is known to
be important to how humans learn to communicate (Grice,
1975; Frank & Goodman, 2012).
A well-known domain that highlights these challenges is
Hanabi, a popular, fully cooperative card game of incom-
plete information that is difficult even for humans (Hanabi
won the Spiel des Jahres award in 2013). A distinguishing
feature of the game is that players see everyone's hands but
their own and must find effective conventions for commu-
nication to succeed. Since there is no cheap-talk channel,
most recent methods for emergent communication are inap-
plicable, necessitating a novel approach. Because of these
unique features, Hanabi has recently been proposed as a
new benchmark for multi-agent learning (?).
The goal in Hanabi is to play a legal sequence of cards and,
to aid this process, players are allowed to give each other
hints indicating which cards are of a specific rank or colour.
These hints have two levels of semantics. The first level is
the surface-level content of the hint, which is grounded in
the properties of the cards that they describe. This level of
semantics is independent of any possible intent of the agent
in providing the hint, and would be equally meaningful if
provided by a random agent. For example, knowing which
Bayesian Action Decoder
Figure 1: a) In an MDP the action u is sampled from a policy π that conditions on the state features (here separated into f pub
and f a). The next state is sampled from P (s(cid:48)s, u). b) In a PuB-MDP, public features f pub generated by the environment and
the public belief together constitute the Markov state sBAD. The 'action' sampled by the BAD agent is in fact a deterministic
partial policy ∆π ∼ πBAD( ∆πsBAD) that maps from private observations f a to actions. Only the acting agent observes f a and
deterministically computes u = ∆π(f a). u is provided to the environment, which transitions to state s(cid:48) and produces the new
. BAD then uses the public belief update to compute a new belief B(cid:48) conditioned on u and ∆π (Equation 1),
observation f pub(cid:48)
thereby completing the state transition.
cards are of a specific colour often does not indicate whether
they can be safely played or discarded.
A second level of semantics arises from information con-
tained in the actions themselves, i.e., the very fact that an
agent decided to take a particular action and not another,
rather than the information resulting from the state transition
induced by the action. This is essential to the formation of
conventions and to discovering good strategies in Hanabi.
To address these challenges, we propose the Bayesian ac-
tion decoder (BAD), a novel multi-agent RL algorithm for
discovering effective communication protocols and policies
in cooperative, partially observable multi-agent settings. In-
spired by the work of Nayyar et al. (2013), BAD uses all
publicly observable features in the environment to compute
a public belief over the players' private features. This ef-
fectively defines a new Markov process, the public belief
Markov decision process (PuB-MDP), in which the action
space is the set of deterministic partial policies, parame-
terised by deep neural networks, that can be sampled for a
given public state. By acting in the space of deterministic
partial policies that map from private observations into envi-
ronment actions, an agent acting only on this public belief
state can still learn an optimal policy. Using approximate,
factorised Bayesian updates and deep neural networks, we
show for the first time how a method using the public belief
of Nayyar et al. (2013), can scale to large state spaces and
allow agents to carry out a form of counterfactual reasoning.
When an agent observes the action of another agent, the
public belief is updated by sampling a set of possible private
states from the public belief and filtering for those states in
which the teammate chose the observed action. This process
is closely related to the kind of theory of mind reasoning
that humans routinely undertake (Baker et al., 2017). Such
reasoning seeks to understand why a person took a specific
action among several, and what information this contains
about the distribution over private observations.
We experimentally validate an exact version of BAD on a
simple two-step matrix game, showing that it outperforms
policy gradient methods. We then apply an approximate
version to Hanabi, where BAD achieves an average score of
24.174 points in the two-player setting, surpassing the best
previously published results for learning agents by around
9 points and approaching the best known performance of
24.9 points for (cheating) open-hand gameplay. BAD thus
establishes a current state-of-the-art on the Hanabi Learning
Environment (?) for the two player self-play setting. We
further show that the beliefs obtained via Bayesian reasoning
have 40% less uncertainty over possible hands than those
using only grounded information.
t ), where τ a
t
2. Background and Setting
Consider a partially observable multi-agent environment
with A agents. At time t each agent a takes action ua
sampled from policy πa(uaτ a
t
is its action-
t = {oa
observation history τ a
t are the
0, ua
observations of agent a at time t, which is given by the obser-
vation function O(a, st) in state st. The next Markov state
st+1 of the environment is produced by the transition func-
tion P (st+1st, ut), which conditions on the joint action
ut = {u1
t ∈ U. In the fully cooperative
setting considered here, each agent receives a per-timestep
team reward rt+1(st, ut) that depends on the last state and
t }, where ua
0, .., oa
t }. Here oa
t , .., uA
Bayesian Action Decoder
t γtrt
(cid:2)(cid:80)
t in the hidden activations.
(cid:3), where γ is the discount
last joint action. We allow centralised training but require de-
centralised execution, from which follows that the policies
πa are known to all agents. This setting can be formalised
as a Dec-POMDP (Oliehoek, 2012).
The goal of multi-agent RL is to find a set of agent policies
{πa}a=1,...,A that maximise the total expected return per
episode J = Eτ∼P (τπa)
factor. In deep RL, optimisation involves training neural
networks that represent policies and value functions. In par-
tially observable settings, the networks are typically recur-
rent, e.g., LSTMs (Wierstra et al., 2009), as they can learn
to represent a sufficient statistic of the action-observation
history τ a
Here we consider a setting where the Markov state st con-
sists of a set of discrete features ft, composed of public
features f pub
. The public features
are common knowledge to all agents, while private features
are observable by at least one, but not all, of the agents.
t are the private features observable by agent a. We use
f a
the notation f pri
[i] to indicate the i-th private state feature.
For example, in a typical card game the cards being played
openly on the table are part of f pub
, the cards held by each
player are in f pri
t contains the cards held by agent a,
and f pri
[i] corresponds to a specific card held by a specific
player. We assume that this separation of state features is
common knowledge to all agents. An example of this sep-
aration for the case of an MDP is illustrated in Figure 1a.
Furthermore, while all our formalisms and methods can be
extended to synchronous action settings, for simplicity we
assume a turn-based setting with one agent acting per step.
and private features f pri
, f a
t
t
t
t
t
t
3. Method
Below we introduce the Bayesian Action Decoder (BAD).
BAD scales the public belief of Nayyar et al. (2013) to
large state spaces using factorised beliefs, an approximate
Bayesian update, and sampled deterministic policies param-
eterised by deep neural networks.
t
t
0
Fortunately, in our setting the common knowledge described
above makes it possible to compute a public belief, (Nay-
yar et al., 2013), that makes the recursion of I-POMDPs
unnecessary. In our case the public belief Bt is the posterior
over all of the private state features given only the public
features, i.e., Bt = P (f pri
f pub≤t ), where ≤ t indicates his-
). Because Bt conditions only
, .., f pub
tory: f pub≤t = (f pub
on publicly available information, it can be computed inde-
pendently by every agent via a common algorithm, yielding
the same result for all agents. Furthermore, since all agents
know f pub, it suffices for Bt to be a posterior over f pri, not
ft = {f pri, f pub}.
While the public belief avoids recursive reasoning, it is not
obvious how it can be used to guide behaviour: agents that
condition their actions only on the public belief will never
exploit their private observations. As Nayyar et al. (2013)
propose, we can construct a special public agent whose
policy πBAD conditions on the public observation and the
public belief but which nonetheless can generate optimal
behaviour.1 This is possible because an action selected by
πBAD specifies a partial policy, ∆π : {f a} → U, for the
acting agent, deterministically mapping private observations
to environment actions. The sampling of a deterministic
partial policy also addresses a fundamental tension in using
policy gradients to learn communication protocols, namely,
differentiation and exploration require high-entropy policies,
while communication requires low-entropy policies. By
sampling in the space of deterministic policies, both can be
achieved.
Intuitively, the public agent can be viewed as a third party
that can observe only the public observation and belief.
While πBAD cannot observe the private state, it can tell each
agent what to do for any private observation it might receive.
Thus at each timestep, the public agent selects ∆π based on Bt
and f pub
t = ∆π(f a)
by supplying the private observation hidden from the public
agent; the public agent then uses the observed action ua
t to
construct the new belief Bt+1.
; the acting agent then selects the action ua
t
3.1. Public belief
3.2. Public Belief MDP
In single-agent partially observable settings, it is clearly
useful for an agent to maintain beliefs about the hidden
environment state, since this is a sufficient statistic for its
action-observation history (Kaelbling et al., 1998). In multi-
agent settings, however, it is not obvious what the beliefs
should be over. It is not enough to maintain beliefs over the
environment state alone, as other agents also have unobserv-
able internal states. In interactive POMDPs (I-POMDPs;
Gmytrasiewicz & Doshi 2005), agents model each other's
beliefs, beliefs over these beliefs, and so on, but this is often
computationally intractable.
t are public information, observing ua
t induces
Since ∆π and ua
a posterior belief over the possible private state features f pri
t
given by the public belief update:
P (f a
t ua
t ,Bt, f pub
t
, ∆π) =
P (ua
∝ 1( ∆π(f a
t f a
P (ua
t ), ua
t , ∆π)P (f a
t Bt, f pub
t )P (f a
t
t Bt, f pub
t
)
, ∆π)
t Bt, f pub
t
(1)
).
(2)
1πBAD conditions on the public observation because the public
belief is a sufficient statistic for the public observation, but only
over the private features.
Bayesian Action Decoder
Using this Bayesian belief update, we can define a new
Markov process, the public belief MDP (PuB-MDP), as
illustrated in Figure 1b. The state sBAD = {B, f pub} of
the PuB-MDP consists of the public observation and pub-
lic belief; the action space is the set of deterministic par-
tial policies that map from private observations to envi-
ronment actions; and the transition function is given by
BADsBAD, ∆π). The next state contains the new public
P (s(cid:48)
belief calculated using the public belief update. The reward
function marginalises over the private state features:
B(f pri)r(s, ∆π(f pri)).
(cid:88)
rBAD(sBAD, ∆π) =
(3)
f pri
Since s(cid:48)
BAD includes the new public belief, and that belief
is computed via an update that conditions on ∆π, the PuB-
MDP transition function conditions on all of ∆π, not just the
selected action ua
t . Thus the state transition depends not just
on the executed action, but on the counterfactual actions:
t .
those specified by ∆π for private observations other than f a
In the remainder of this section, we describe how factorised
beliefs and policies can be used to efficiently learn a public
policy πBAD for the PuB-MDP.
3.3. Sampling Deterministic Partial Policies
For each public state, πBAD must select a distribution
πBAD( ∆πsBAD) over deterministic partial policies. The size
of this space is exponential in the number of possible private
observations f a, but we can reduce this to a linear depen-
dence by assuming a distribution across ∆π that factorises
across the different private observations, i.e., for all ∆π,
πBAD( ∆πBt, f pub) :=
πBAD( ∆π(f a)Bt, f pub, f a).
(cid:89)
(4)
Figure 2: Payoffs for the toy matrix-like game. The two
outer dimensions correspond to the card held by each player,
the two inner dimensions to the action chosen by each player.
Payouts are structured such that Player 1 must encode in-
formation about their card in the action they chose in order
to obtain maximum payoffs. Although presented here in
matrix form for compactness, this is a two-step, turn-based
game, with Player 1 always taking the first action and Player
2 taking an action after observing Player 1's action.
the number of cards held by all players. To avoid this un-
favourable scaling, we can instead represent an approximate
factorised belief state
f pub≤t ) ≈(cid:89)
P (f pri
t
P (f pri
t
[i]f pub≤t ) =: Bfact
t
.
(5)
f a
i
With this restriction, we can parameterise πBAD with factors
BAD(uaBt, f pub, f a) using a function approx-
of the form πθ
imator such as a deep neural network. In order for all of
the agents to perform the public belief update, the sampled
∆π must be public. We resolve this by having ∆π sampled
deterministically from a given Bt and f pub
, using a common
knowledge random seed ξt. The seeds are shared prior to
the game so that all agents sample the same ∆π. This re-
sembles the way humans share common ways of reasoning
in card games and allows the agents to explore alternative
policies jointly as a team. Further details on the mechanics
of parameterising and sampling from πBAD are provided in
the Supplemental Material.
t
3.4. Factorised Belief Updates.
In general, representing exact beliefs is intractable in all
but the smallest state spaces. For example, in card games
the number of possible hands is typically exponential in
From here on we drop the superscript and use B exclu-
sively to refer to the factorised belief. In a card game each
factor represents per-card probability distributions, assum-
ing approximate independence across the different cards
both within a hand and across players. This approximation
makes it possible to represent and reason over the otherwise
intractably large state spaces that commonly occur in many
settings, including card games.
To carry out the public belief update with a factorised repre-
sentation we maintain factorised likelihood terms Lt[f [i]]
for each private feature that we update recursively:
Lt[f [i]] := P (ua≤tf [i],B≤t, f pub≤t , ∆π≤t)
t f [i],Bt, f pub
(6)
(7)
, ∆πt)
≈ Lt−1[f [i]] · P (ua
= Lt−1[f [i]] · Eft∼Bt
(cid:2)1(ft[i], f [i])1( ∆π(f a
(cid:2)1(ft[i], f [i])(cid:3)
Eft∼Bt
t
t )(cid:3)
(8)
t ), ua
,
1000484100000104840010001048400010004841000Player 2 (acts second)Player 1 (acts first)Card 1Card 2Card 2Card 1Player 2 actionPlayer 1 action"Tiny Hanabi"●Each player knows their card but can't see the other player's card.●Player 1 acts first.●Player 2 observes Player 1's action and acts second.ABCABC Bayesian Action Decoder
where (7) assumes that actions are (approximately) condi-
tionally independent of the future given the past. As indi-
cated, these likelihood terms are calculated by sampling,
and the more samples the better.
3.5. Self-Consistent Beliefs
f [−i]
∝ E
f [−i]∼Bk
(cid:2)Lt(f [i])P (f [i]f [−i], f pub
)(cid:3),
t
(9)
This factorisation is only an approximation, even in simple
card games: knowledge that a player is holding a specific
card clearly influences the probability that another player
is holding that same card. Furthermore, our approximation
can yield beliefs that are not even self-consistent, i.e., they
are not the marginalisation of any belief over joint features.
While not central to the key ideas behind BAD, we introduce
a general iterative procedure that can account for feature
interactions in factorised models. Starting with a public
belief B0 = Bt we can iteratively update the belief to make
it more self-consistent through re-marginalisation:
Bk+1(f [i]) =
Bk(f [−i])P (f [i]f [−i], f pub≤t , ua≤t, ∆π≤t)
(cid:88)
(10)
where f [−i] denotes all features excluding f [i]. In the last
step we used the factorised likelihoods from above to con-
vert to an expectation, so that we can use samples to ap-
proximate the intractable sum across features. The notion
of refining the distribution over one feature while keeping
the distribution across all other features fixed is similar to
Expectation Propagation for factor graphs (Minka, 2001).
However, the card counts constitute a global factor, making
the factor graph formulation less useful. While this iterative
update can in principle be carried out until convergence, in
practice we terminate after a fixed number of iterations.
4. Experiments and Results
4.1. Matrix Game
We first present proof-of-principle results for a two-player,
two-step partially observable matrix-like game (Figure 2).
The state consists of 2 random bits (the cards for Player 1
and 2) and the action space consists of 3 discrete actions.
Each player observes its own card, with Player 2 also ob-
serving Player 1's action before acting, which in princi-
ple allows Player 1 to encode information about its card
with its action. The reward is specified by a payoff tensor,
r = Payoff[card1][card2][u1][u2], where carda and ua are
the card and action of the two players, respectively. The
payout tensor is structured such that the optimal reward can
only be achieved if the two players establish a convention,
in particular if Player 1 chooses informative actions that can
be decoded by Player 2.
Figure 3: BAD, both with and without counterfactual (CF)
gradients, outperforms vanilla policy gradient on the matrix-
like game. Each line is the mean over 1000 games, and the
shade indicates the standard error of the mean (s.e.m.).
As shown in Figure 3, BAD clearly outperforms the base-
line policy-gradient method on the toy matrix game. In this
small, exact setting, it is also possible to estimate counterfac-
tual (CF) policy gradients that reinforce not only the action
taken, but also these counterfactual actions. This can be
t τ a) with log P ( ∆πBt, f pub)
achieved by replacing log πa(ua
in the estimation of the policy gradient. However, the addi-
tional improvement in performance from using CF gradients
is minor compared to the initial performance gain from us-
ing a counterfactual belief state.
Code for the matrix game with a proof-of-principle im-
plementation of BAD is available at https://bit.ly/
2P3YOyd.
4.2. Hanabi
Here we briefly describe the rules of 2-player Hanabi.
There are 5 cards in a hand. For each of the 5 colours there
are three 1s, one 5, and two each of all other ranks, i.e., 10
cards per colour for a total of 50 = 5 × 10 cards in the deck.
While this is a modestly large number of cards, even for
2 players it leads to 6.2 × 1013 possible joint hands at the
beginning of the game.
4.3. Observations and Actions
Each player observes the hands of all other players, but
not their own. The action space consists of 2 × 5 options
for discarding and playing cards, and 5 + 5 options per
teammate for hinting colours and ranks. Hints reveal all
cards of a specific rank or colour to one of the teammates,
Bayesian Action Decoder
Equation 8). As derived in the Supplemental Material,
Bk+1(f [i]) ∝
Bk(f [j])
×HM(f [i]). (12)
(cid:32)
C(f )−(cid:88)
j(cid:54)=i
(cid:33)
e.g., 'Player 2's card 3 and 5 are red'. Hinting for colours
and ranks not present in the hand of the teammate (so-called
'empty hints') is not allowed.
Each hint costs one hint token. The game starts with 8
hint tokens, which can be recovered by discarding cards.
After a player has played or discarded a card, she draws a
new card from the deck. When a player picks up the last
card, everyone (including that player) takes one more action
before the game terminates. Legal gameplay consists of
building 5 fireworks, which are piles of ascending numbers,
starting at 1, for each colour. When the 5 has been added to
a pile the firework is complete and the team obtains another
hint token (unless they already have 8). A life token is lost
each time a player plays an illegal card; after three mistakes
the game terminates. Players receive 1 point after playing
any playable card, with a perfect score being 25 = 5 × 5.
The number of hint and life tokens at any time are observed
by all players, as are the played and discarded cards, the last
action of the acting player and any hints provided.
4.4. Beliefs in Hanabi
The basic belief calculation in Hanabi is straightforward:
f pub
consists of a vector of 'candidates' C containing counts
for all remaining cards, and a 'hint mask' HM, an ANh ×
t
(NcolorNrank + 1) binary matrix that is 1 if in a given 'slot'
the player could be holding a specific card according to the
hints so far, and 0 otherwise; the additional 1 accounts for
the possibility that the card may not exist in the final round
of play. Slots correspond to the features of the private state
space f [i], for example the 3rd card of the second player.
Hints contain both positive and negative information: for
example, the statement 'the 2nd and 4th cards are red' also
implies that all other cards are not red.
The basic belief B0 can be calculated as
B0(f [i]) = P (f [i]f pub) ∝ C(f ) × HM(f [i]).
(11)
We call this the 'V0 belief', in which the belief for each
card depends only on publicly available information for that
card. In our experiments, we focus on baseline agents that
receive this basic belief, rather than the raw hints, as public
observation inputs; while the problem of simply remember-
ing all hints and their most immediate implication for card
counts is potentially challenging for humans in recreational
play, we are here more interested in the problem of forming
effective conventions for high-level play.
As noted above, this basic belief misses an important inter-
action between the hints for different slots. We can calculate
an approximate version of the self-consistent beliefs that
avoids the potentially expensive and noisy sampling step
in Equation 10 (note that this sampling is distinct from the
sampling required to compute the marginal likelihood in
We call the resulting belief at convergence (or after a max-
imum number of iterations) the 'V1 belief'. It does not
condition on the Bayesian probabilities but considers inter-
actions between hints for different cards. In essence, at each
iteration the belief for a given slot is updated by reducing
the candidate count by the number of cards believed to be
held across all other slots.
By running the same algorithm but including L, we obtain
the Bayesian beliefs BB that lie at the core of BAD:
BB0(f [i]) ∝ C(f ) × HM(f [i]) × L(f [i]),
(13)
(cid:33)
BBk+1(f [i]) ∝
Bk(f [j])
(cid:32)
C(f ) −(cid:88)
j(cid:54)=i
× HM(f [i]) × L(f [i]).
(14)
In practice, to ensure stability, the final 'V2 belief' that we
use is an interpolation between the Bayesian belief and the
V1 belief: V2 = (1 − α)BB + αV1 with α = 0.01 (we
found α = 0.1 to also work). For the Bayesian update we
sampled S = 3, 000 hands during training and S = 20, 000
hands for the final test games.
4.5. Architecture Details for Baselines and Method
Advantage actor-critic agents were trained using the
Importance-Weighted Actor-Learner Architecture (Espeholt
et al., 2018), in particular the multi-agent implementation
described in Jaderberg et al. (2018). In this framework,
'actors' continually generate trajectories of experience (se-
quences of states, actions, and rewards) by having agents
(self-)playing the game, which are then used by 'learners' to
perform batched gradient updates (batch size was 32 for all
agents). Because the policy used to generate the trajectory
can be several gradient updates behind the policy at the time
of the gradient update, V-trace was applied to correct for
the off-policy trajectories. The length of the trajectories, or
rollouts, was 65, the maximum length of a winning game.
Further details for the hyperparameters, architecture, and
training are given in the Supplemental Material.
4.6. Results on Hanabi
The BAD agent achieves a new state-of-the-art mean per-
formance of 24.174 points on two-player Hanabi. In Fig-
ure 4a we show training curves and test performance for
BAD and two LSTM-based baseline methods, as well as the
performance of FireFlower (https://github.com/
lightvector/fireflower), the best known hand-
coded bot for two-player Hanabi. For the LSTM agents,
Bayesian Action Decoder
(a)
(b)
(c)
Figure 4: a) Hanabi training curves for BAD and the V0 and V1 baseline methods using LSTMs rather than the Bayesian
belief. Thick lines indicate the final evaluated agent for each agent type, with the dots showing the final test score. Error
bars (standard error of the mean, s.e.m.) are smaller than the dots. Upward kinks in the curves are generally due to agents
'evolving' in PBT by copying its weights and hyperparameters (plus perturbations) from a superior agent. b) Distribution of
game scores for BAD on Hanabi under testing conditions. BAD achieves a perfect score in almost 60% of the games. The
dashed line shows the proportion of perfect games reported for FireFlower, the best known hard-coded bot for two-player
Hanabi. c) Per-card cross entropy with the true hand for different belief mechanisms during BAD play. V0 is the basic belief
based on hints and card counts, V1 is the self-consistent belief, and V2 is the BAD belief which also includes the Bayesian
update. The BAD agent conveys around 40% of the information via conventions, rather than grounded information.
test performance was obtained by using the greedy version
of the trained policy, resulting in slightly higher scores than
during training. To select the agent, we first performed a
sweep over all agents for 10,000 games, then carried out a
final test run of 100,000 games on the best agent from the
sweep to obtain an unbiased score. For the BAD agent we
also increased the number of sampled hands. The results for
other learning methods from the literature perform below
the range of the y-axis (far below 20 points) and are omitted
for readability. We note that, under a strict interpretation of
the rules of Hanabi, games in which all three error tokens
are exhausted should be awarded a score of 0. Under these
rules the same BAD agent achieves 23.917 ± 0.009 s.e.m,
the best known score, even though it was not trained under
these conditions. SmartBot and FireFlower achieve average
scores of 22.99 and 22.56 respectively.
While not all of the game play BAD learns is easy to follow,
some conventions can be understood simply from inspecting
the game. Printouts of 100 random games can be found at
https://bit.ly/2zeEShh. One convention stands
out: Hinting for 'red' or 'yellow' indicates that the newest
card of the other player is playable. We found that in over
80% of cases when an agent hints 'red' or 'yellow', the next
action of the other agent is to play the newest card. This
convention is very powerful: Typically agents know the
least about the newest card, so by hinting 'red' or 'yellow',
agents can use a single hint to tell the other agent that the
card is playable. Indeed, the use of two colours to indicate
Agent
Learning steps Mean ± s.e.m.
Prop. perfect
SmartBot
FireFlower
V0-LSTM
V1-LSTM
BAD
-
-
20.2B
21.1B
16.3B
23.09
23.37 ± 0.0002
23.622 ± 0.005
23.919 ± 0.004
24.174± 0.004
29.52%
52.6%
36.5%
47.5%
58.6%
Table 1: Test scores on 100K games. The LSTM agents
were tested with a greedy version of the trained policy, while
the final BAD agent was evaluated with V1 mix-in α = 0.01,
20K sampled hands, and inverse softmax temperature 100.0.
The FireFlower bot was evaluated over 25K games.
'play newest card' was present all of the highest-performing
agents we studied. Hinting 'white' and 'blue' are followed
by a discard of the newest card in over 25% of cases. We
also found that the agent sometimes attempts to play cards
which are not playable in order to convey information to
their team mate. In general, unlike human players, agents
play and discard predominantly from the last card. In the
supplementary material we also include a written analysis
of our bot by the creator of FireFlower.
Figure 4c shows the quality of the different beliefs. While
the iterated belief update leads to a reduction in cross en-
tropy compared to the basic belief, a much greater reduction
in cross entropy is obtained using counterfactual beliefs.
Bayesian Action Decoder
This clearly demonstrates the importance of learning con-
ventions for successful gameplay in Hanabi: Roughly 40%
of the information is obtained through conventions rather
than through the grounded information and card counting.
5. Related Work
5.1. Learning to Communicate
Many works have addressed problem settings where agents
must learn to communicate in order to cooperatively solve
a toy problem. These tasks typically involve a cheap-talk
communication channel that can be modeled as a continu-
ous variable during training, which allows differentiation
through the channel as first proposed by Foerster et al.
(2016a) and Sukhbaatar et al. (2016). In this work we fo-
cused on the case where agents must learn to communicate
via grounded hinting actions and observable environment
actions rather than a cheap-talk channel. This is closest to
the "hat game" of Foerster et al. (2016b), who proposed a
simple extension to recurrent deep Q-networks rather than
explicitly modeling action-conditioned Bayesian beliefs. An
idea similar to the Pub-MDP was introduced in the context
of decentralised stochastic control by Nayyar et al. (2013),
who also formulated a coordinator that uses "common in-
formation" to map local controller information to actions.
However, they did not provide a concrete solution method
that can scale to a high-dimensional problem like Hanabi.
5.2. Hanabi
A number of papers have been published on Hanabi. Baffier
et al. (2016) showed that optimal gameplay in Hanabi is
NP-hard even when players can observe their own cards.
Encoding schemes similar to the hat game essentially solve
the five-player case (Cox et al., 2015), but only achieve
17.8 points in the two-player setting (Bouzy, 2017). Walton-
Rivers et al. (2017) developed a variety of Monte Carlo tree
search and rule-based methods for Hanabi, but the reported
scores were roughly 50% lower than BAD. Osawa (2015)
defined a number of heuristics for the two-player case that
reason over possible hands given the other player's action.
While this is similar in spirit to our approach, the work was
limited to hand-coded heuristics, and the reported scores
were around 8 points lower than BAD. Eger et al. (2017)
investigated humans playing with hand-coded agents, but no
pairing resulted in scores higher than 15 points on average.
The best result for two-player Hanabi we could find
was for the FireFlower described at github.com/
lightvector/fireflower, which has been reported
to achieve an average of 23.37 points (52.6% perfect games).
While FireFlower uses the same game rules as those used in
our work, it is entirely hand-coded and involves no learning.
5.3. Belief State Methods
The continual re-solving (nested solving) algorithm used by
DeepStack (Moravc´ık et al., 2017) and Libratus (Brown &
Sandholm, 2018) for poker also use a belief state space. Like
BAD, when making a decision in a player state, continual re-
solving considers the belief state associated with the current
player and generates a joint policy across all player states
consistent with this belief. The policy for the actual player
is then selected from this joint policy. Continual re-solving
also does a Bayesian update of the beliefs after an action.
There are key differences, however. Continual re-solving
performa exact belief updates, which requires a joint policy
space small enough to enumerate; belief states are also
augmented with opponent values; continual re-solving is
a value-based method, where the training process consists
of learning the values of belief states under optimal play;
finally, the algorithm is designed for two-player, zero-sum
games, where it can independently consider player state
values while guaranteeing that an optimal choice for the
joint action policy can be found.
6. Conclusion and Future Work
We presented the Bayesian action decoder (BAD), a novel
algorithm for multi-agent reinforcement learning in cooper-
ative partially observable settings. BAD uses a factorised,
approximate belief state that allows agents to efficiently
learn informative actions, leading to the discovery of con-
ventions. We showed that BAD outperforms policy gradi-
ents in a proof-of-principle matrix game, and achieves a
state-of-the-art performance of 24.174 points on average
in the card game Hanabi. We also showed that using the
Bayesian update leads to a reduction in uncertainty across
the private hands in Hanabi by around 40%. To the best of
our knowledge, this is the first instance in which deep RL
has been successfully applied to a problem setting that both
requires the discovery of communication protocols and was
originally designed to be challenging for humans. BAD also
illustrates clearly that using an explicit belief computation
achieves better performance in such settings than current
state-of-the-art RL methods using implicit beliefs, such as
recurrent neural networks.
In the future, we aim to apply BAD to games with more
players and further generalise BAD by learning more of its
components, e.g., the V0-belief. While the belief update
necessarily involves a sampling step, most of the other com-
ponents can likely be learned end-to-end. We also plan to
extend the BAD mechanism to value-based methods and
further investigate the relevance of counterfactual gradients.
Similar to what was suggested as next steps in (?), we hope
to extend the setting to a point where our bots can learn to
collaborate with human players.
Bayesian Action Decoder
Acknowledgements
We thank Marc Lantot, Shibl Mourad, Angeliki Lazaridou,
Jelena Luketina, Anuj Mahajan, Gregory Farquhar, Kelsey
Allen, Thore Graepel, Nando de Freitas, and Nolan Bard for
valuable discussions.
This project has received funding from the European Re-
search Council (ERC) under the European Unions Horizon
2020 research and innovation programme (grant agreement
#637713).
References
Baffier, J.-F., Chiu, M.-K., Diez, Y., Korman, M., and
Mitsou, V. Hanabi is NP-complete, even for cheaters
who look at their cards. arXiv:1603.01911, 2016. URL
https://arxiv.org/abs/1603.01911.
Baker, C. L., Jara-Ettinger, J., Saxe, R., and Tenen-
baum, J. B.
Rational quantitative attribution of
beliefs, desires and percepts in human mentalizing.
Nat. Hum. Behav., 1(4):1 -- 10, 2017.
doi: 10.1038/
s41562-017-0064. URL http://dx.doi.org/10.
1038/s41562-017-0064.
Bouzy, B. Playing Hanabi near-optimally. In Advances in
Computer Games, pp. 51 -- 62. Springer, 2017.
Brown, N. and Sandholm, T. Superhuman ai for heads-up
no-limit poker: Libratus beats top professionals. Science,
359(6374):418 -- 424, 2018.
Cox, C., De Silva, J., Deorsey, P., Kenter, F. H. J., Ret-
ter, T., and Tobin, J. How to Make the Perfect Fire-
works Display : Two Strategies for Hanabi. Math.
Mag., 88:323, 2015.
doi: 10.4169/math.mag.88.5.
323. URL http://www.jstor.org/stable/10.
4169/math.mag.88.5.323.
Eger, M., Martens, C., and Cordoba, M. A. An intentional
AI for hanabi. 2017 IEEE Conf. Comput. Intell. Games,
CIG 2017, pp. 68 -- 75, 2017. doi: 10.1109/CIG.2017.
8080417.
Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih,
V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning,
I., Legg, S., and Kavukcuoglu, K. IMPALA: Scalable
Distributed Deep-RL with Importance Weighted Actor-
Learner Architectures. arXiv:1802.01561, 2018. URL
http://arxiv.org/abs/1802.01561.
Foerster, J. N., Assael, Y. M., de Freitas, N., and Whiteson,
S. Learning to communicate to solve riddles with deep
distributed recurrent Q-networks. arXiv:1602.02672,
2016b. URL https://arxiv.org/abs/1602.
02672.
Frank, M. C. and Goodman, N. D. Predicting pragmatic
reasoning in language games. Science, 336(6084):998,
2012. doi: 10.1126/science.1218633.
Gmytrasiewicz, P. J. and Doshi, P. A framework for sequen-
tial planning in multi-agent settings. J. Artif. Intell. Res.,
24:49 -- 79, 2005. doi: 10.1613/jair.1579.
Grice, H. P.
Logic and conversation.
In Cole, P.
and Morgan, J. L. (eds.), Syntax and Semantics: Vol.
3: Speech Acts, pp. 41 -- 58. Academic Press, New
York, 1975. URL http://www.ucl.ac.uk/ls/
studypacks/Grice-Logic.pdf.
Jaderberg, M., Czarnecki, W. M., Dunning, I., Marris, L.,
Lever, G., Castaneda, A. G., Beattie, C., Rabinowitz,
N. C., Morcos, A. S., Ruderman, A., Sonnerat, N., Green,
T., Deason, L., Leibo, J. Z., Silver, D., Hassabis, D.,
Kavukcuoglu, K., and Graepel, T. Human-level perfor-
mance in first-person multiplayer games with population-
based deep reinforcement learning. arXiv:1807.01281,
2018. doi: arXiv:1807.01281. URL http://arxiv.
org/abs/1807.01281.
Kaelbling, L. P., Littman, M. L., and Cassandra, A. R. Plan-
ning and acting in partially observable stochastic domains.
Artif. Intell., 101(1-2):99 -- 134, 1998. doi: 10.1016/
S0004-3702(98)00023-X. URL http://dx.doi.
org/10.1016/S0004-3702(98)00023-X.
Littman, M. L. Markov games as a framework for multi-
agent reinforcement learning. Mach. Learn. Proc. 1994,
pp. 157 -- 163, 1994. doi: 10.1016/B978-1-55860-335-6.
50027-1. URL http://linkinghub.elsevier.
com/retrieve/pii/B9781558603356500271.
Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, O. P.,
and Mordatch, I. Multi-agent actor-critic for mixed
cooperative-competitive environments. In Advances in
Neural Information Processing Systems, pp. 6379 -- 6390,
2017.
Minka, T. P. Expectation Propagation for Approximate
Bayesian Inference. Uncertain. Artif. Intell., 17(2):362 --
369, 2001. URL https://arxiv.org/abs/1301.
2294.
Foerster, J., Assael, Y. M., de Freitas, N., and Whiteson,
S. Learning to communicate with deep multi-agent re-
inforcement learning. In Advances in Neural Informa-
tion Processing Systems, pp. 2137 -- 2145, 2016a. URL
https://arxiv.org/abs/1605.06676.
Moravc´ık, M., Schmid, M., Burch, N., Lis`y, V., Morrill, D.,
Bard, N., Davis, T., Waugh, K., Johanson, M., and Bowl-
ing, M. Deepstack: Expert-level artificial intelligence in
heads-up no-limit poker. Science, 356(6337):508 -- 513,
2017.
Bayesian Action Decoder
Nayyar, A., Mahajan, A., and Teneketzis, D. Decentral-
ized stochastic control with partial history sharing: A
common information approach. IEEE Trans. Automat.
Contr., 58(7):1644 -- 1658, 2013. ISSN 00189286. doi:
10.1109/TAC.2013.2239000. URL https://arxiv.
org/abs/1209.1695.
Oliehoek, F. A. Decentralized pomdps. In Reinforcement
Learning, pp. 471 -- 503. Springer, 2012.
Osawa, H. Solving hanabi: Estimating hands by opponent's
actions in cooperative game with incomplete informa-
tion. In AAAI workshop: Computer Poker and Imperfect
Information, pp. 37 -- 43, 2015.
Perolat, J., Leibo, J. Z., Zambaldi, V., Beattie, C., Tuyls,
K., and Graepel, T. A multi-agent reinforcement
learning model of common-pool resource appropria-
tion. arXiv:1707.06600, 2017. ISSN 10495258. URL
http://arxiv.org/abs/1707.06600.
Sukhbaatar, S., Szlam, A., and Fergus, R. Learning Multia-
gent Communication with Backpropagation. Advances
in Neural Information Processing Systems, 2016. ISSN
10495258. URL http://arxiv.org/abs/1605.
07736.
Walton-Rivers, J., Williams, P. R., Bartle, R., Perez-Liebana,
D., and Lucas, S. M. Evaluating and modelling Hanabi-
In Evolutionary Computation (CEC),
playing agents.
2017 IEEE Congress on, pp. 1382 -- 1389. IEEE, 2017.
Wierstra, D., Forster, A., Peters, J., and Schmidhuber, J.
Recurrent policy gradients. Log. J. IGPL, 18(5):620 -- 634,
2009. doi: 10.1093/jigpal/jzp049.
Bayesian Action Decoder
A. Parameterising and Sampling from the
Distribution over Partial Policies
BAD requires us to parameterise a probability distribution
over partial policies using a deep neural network:
P ( ∆πsBAD) = πθ
BAD( ∆πsBAD).
(15)
The first insight is that we can trivially use a neural network
to map from public states, sBAD, into probabilistic partial
policies. To do so, we simply start with a feedforward
policy that takes as input both sBAD and f a and produces a
distributions over actions:
πθ(sBAD, f a) → P (usBAD, f a).
(16)
Next, we note that if we fix a given sBAD, we now have a
probabilistic partial policy which maps each private obser-
vation f a into a probability distribution over actions. This
partial policy is produced deterministically as a function of
sBAD via the parameters θ:
π(uf a) : {f a} → {P (U)} sBAD,
π(uf a) = πθ(usBAD, f a).
(17)
(18)
Now, this is close, but not quite what we want. Above we
have a deterministic map from sBAD into probabilistic par-
tial policies, π(uf a). Instead, we require a differentiable
distribution over deterministic partial policies.
Perhaps surprisingly, this can be accomplished by condition-
ing the sampling from π(uf a) on a common knowledge
random seed, ξ:
∆π : {f a} → U sBAD,
{f a} → u ∼ π(uf a) ξ,
{f a} → u ∼ πθ(usBAD, f a) ξ.
(19)
(20)
(21)
(22)
Thus, when we sample ξ we are effectively sampling an
entire deterministic partial policy.
B. Hyperparameters and Training Details
For the toy matrix game, we used a batch size of 32 and the
Adam optimiser with all default TensorFlow settings; we
did not tune hyperparameters for any runs.
In the V0-LSTM and V1-LSTM BAD agents, all observa-
tions were first processed by an MLP with a single 256-unit
hidden layer and ReLU activations, then fed into a 2-layer
LSTM with 256 units in each layer. The policy π was a
softmax readout of the LSTM output. The baseline network
was an MLP with a single 256-unit hidden layer and ReLU
activations, which then projected linearly to a single value.
Since the baseline network is only used to compute gradient
updates, we followed ? (?) in feeding each agent's own
hand (i.e., the other agent's private observation) into the
baseline by concatenating it with the LSTM output; thus we
make the common assumption of centralised training and
decentralised execution. We note that the V0 and V1-LSTM
agents differed only in their public belief inputs.
The Hanabi BAD agent consisted of an MLP with two 384-
unit hidden layers and ReLU activations that processed all
observations, followed by a linear softmax policy readout.
To compute the baseline, we used the same MLP as the
policy but included the agent's own hand in the input (this
input was present but zeroed out for the computation of the
policy).
For all agents, illegal actions (such as hint for a red card
when there are no red cards) were masked out by setting the
corresponding policy logits to a large negative value before
sampling an action. In particular, for the non-acting agent at
each turn the only allowed action was the 'no-action'. For
Hanabi, we used the RMSProp optimiser with = 10−10,
momentum 0, and decay 0.99. The RL discounting factor γ
was set to 0.999. The baseline loss was multiplied by 0.25
and added to the policy-gradient loss. We used population-
based training (PBT) (?Jaderberg et al., 2018) to 'evolve' the
learning rate and entropy regularisation parameter during the
course of training, with each training run consisting of a pop-
ulation of 30 agents. For the LSTM agents, learning rates
were sampled log-uniformly from the interval [1, 4) × 10−4
while the entropy regularisation parameter was sampled log-
uniformly from the interval [1, 5) × 10−2. For the BAD
agents, learning rates were sampled log-uniformly from the
interval [9 × 10−5, 3 × 10−4) while the entropy regularisa-
tion parameter was sampled log-uniformly from the interval
[3, 7)×10−2. Agents evolved within the PBT framework by
copying weights and hyperparameters (plus perturbations)
according to each agent's rating, which was an exponen-
tially moving average of the episode rewards with factor
0.01. An agent was considered for copying roughly every
200M steps if a randomly chosen copy-to agent had a rating
at least 0.5 points higher. To allow the best hyperparameters
to manifest sufficiently, PBT was turned off for the first 1B
steps of training.
The BAD agent was trained with 100 self-consistent iter-
ations, a V1 mix-in of α = 0.01, BAD discount factor
γBAD = 1, inverse temperature 1.0, and 3000 sampled hands.
Since sampling from card-factorised beliefs can result in
hands that are not compatible with the deck, we sampled 5
times the number of hands and accepted the first 3000 legal
hands, zeroing out any hands that were illegal.
Bayesian Action Decoder
C. Self-Consistent Belief Approximation for
Hanabi
t
We will use the same notation as in the main text: f pub
consists of a vector of 'candidates' C containing counts
for all remaining cards, and a 'hint mask' HM, an ANh ×
NcolorNrank binary matrix that is 1 if in a given 'slot' the
player could be holding a specific card according to the
hints given so far, and 0 otherwise". Furthermore, L(f [i]),
is the marginal likelyhood.
Then the basic per-card belief is simply:
B0(f [i]) ∝ C(f ) × HM(f [i]) × L(f [i]),
(cid:80)
C(f ) × HM(f [i]) × L(f [i])
(cid:0)C(f ) × HM(f [i]) × L(f [i])(cid:1).
g C(g) × HM(g[i]) × L(g[i])
B0(f [i]) =
= βi
(23)
(24)
(25)
In the last two lines we are normalising the probability, since
the probability of the i-th feature being one of the possible
values must sum to 1. For convenience we also introduced
the notation βi for the normalisation factor.
Next we apply the same logic to the iterative belief update.
The key insight here is to note that conditioning on the
features f [−i], i.e., the other cards in the slots, corresponds
to reducing the card counts in the candidates. Below we use
M (f [i]) = HM(f [i]) × L(f [i]) for notational convenience:
Bk+1(f [i])
=
Bk(f [−i])P (f [i]f [−i], f pub≤t , ua≤t, ∆π≤t)
(cid:19)
f [−i]
(cid:18)
(26)
(cid:88)
(cid:88)
C(f ) −(cid:88)
Bk(g[−i])βi
1(g[j] = f )
M (f [i]).
=
g[−i]
j(cid:54)=i
(27)
In the last line we relabelled the dummy index f [−i] to g[−i]
for clarity and used the result from above. Next we sub-
stitute the factorised belief assumption across the features,
(cid:19)
1(g[j] = f )
M (f [i])
1(g[j] = f )
M (f [i])
(cid:19)
(cid:19)
(28)
(29)
M (f [i]).
(30)
Bk(g[−i])βi
=
Bk+1(f [i])
Bk(g[−i]) =(cid:81)
(cid:88)
(cid:88)
(cid:89)
(cid:89)
(cid:88)
g[−i]
g[−i]
j(cid:54)=i
=
(cid:39) βi
g[−i]
j(cid:54)=i
j(cid:54)=i
j(cid:54)=i Bk(g[j]) :
(cid:18)
C(f ) −(cid:88)
(cid:18)
C(f ) −(cid:88)
(cid:18)
C(f ) −(cid:88)
j(cid:54)=i
Bk(g[j])
Bk(g[j])βi
1(g[j] = f )
j(cid:54)=i
In the last line we have ommited the dependency of βi on
the sampled hands f [−i]. It corresponds to calculating the
average across sampled hands first and then normalising
(which is approximate but tractable) rather than normalising
and then averaging (which is exact but intractable). We can
now use product-sum rules to simplify the expression.
Bk+1(f [i])
(cid:39) βi
Bk(g[j])
(cid:88)
1(g[j] = f )
(cid:19)
M (f [i])
j(cid:54)=i
(cid:89)
(cid:88)
j(cid:54)=i
g
g[−i]
(cid:18)
C(f ) −(cid:88)
(cid:18)
C(f ) −(cid:88)
(cid:18)
C(f ) −(cid:88)
(cid:18)
C(f ) −(cid:88)
j(cid:54)=i
j(cid:54)=i
= βi
= βi
∝
Bk(g[j])1(g[j] = f )
(cid:19)
(cid:19)
Bk(f [j])
M (f [i])
j(cid:54)=i
Bk(f [j])
M (f [i]).
(cid:19)
(31)
M (f [i])
(32)
(33)
(34)
This concludes the proof.
D. Anecdotal Analysis
Below we present commentary from David Wu (https:
//github.com/lightvector/), the creator of the
FireFlower bot, on our BAD agent. While this is anecdotal
evidence, we believe it provides some interesting insights
into the gameplay that our BAD agent discovers. The com-
ments are taking verbatim from an email exchange with
David:
D.1. Communicating Playables
• As you observed before, the bot uses R and Y often to
hint newest-card-playability.
• In addition to the R and Y hints, it also often uses direct
hints to the newest card to indicate playability, in the
way that natural human conventions do, and I think
these include both color and number hints.
• When the R and Y hints or direct hints to the newest
card hit multiple cards, the bot often was indicating
multiple plays. In the small sample size of cases we
looked over, it tended to be the case that the R/Y hints
were more often "play in the order from newest to
oldest" while the direct hints were more often in the
order of "play from oldest to newest". I think this
was not 100% consistent though, but in all cases when
looking at the direct beliefs, it was clear that in each
case there was a strong ordering convention was in
force for that hint, it's just that we didn't see enough
Bayesian Action Decoder
cases to be able to determine the precise rules for which
one when. Generally though, it makes a lot of sense to
vary the ordering convention in different parts of the
hint space to add flexibility in hinting.
• The bot uses certain other kinds of direct hints to older
cards to suggest that those cards are one step away
from playable, or something of that nature. Sometimes
the belief state shows that this is not absolutely certain,
but over time as other things happen the probability
mass sometimes gradually updates and concentrates
on the card on the truth, such that once the preceding
card is played, the bot may then play the formerly-one-
step-away card without any further suggestion.
• For these "delayed" one-step-removed hints to older
cards, there is also a similar variation in ordering con-
ventions in the case those hints hit more than one
card, sometimes they're in "age-order" and sometimes
they're in "reverse-age-order".
• Commonly the R and Y hints also indicate other plays
or delayed plays besides the play of the newest card.
The bot chooses the manner of hinting the first card as
playable (R vs Y vs direct hint) to try to communicate
other useful information at the same time, if possible.
• I think occasionally the bot seems to "single out" a card
by directly hinting all other cards in the hand *besides*
that card over successive turns, and sometimes this
implies that the singled-out unhinted card is playable.
I'm not sure on this one though, I'd need to see more
cases.
• I think there seems to be some interesting other con-
ventions that seem to function to give information to
allow play of older red and yellow cards, which are
necessary since direct hints of R and Y mean to play
the newest card rather than the card hinted.
D.2. Communicating Protection
• As you observed before, the bot discards its newest
card by default.
• G hints that do not directly hit the first card appear
to mean that the newest card is dangerous and should
not be discarded. Possibly it is more specific, and
actually just means that it's a 5, the examples I recall
all involved 5s. The bot also can just directly hint the
newest card in various ways.
• The bot is very aggressive about protecting the newest
card if the newest card is a 5 or otherwise dangerous
(the last copy in the deck), whether by giving a G hint,
or a direct hint, or otherwise. This is so consistent
that pretty much any action other than an immediate
protection causes the other bot to infer that the newest
card is NOT a 5 or the last copy of a card whose first
copy has been lost.
• However, the bot does *not* do this any longer if there
is a common-knowledge-extremely-safe discard in that
player's hand (e.g. a redundant copy of a card already
played). In that case, it is understood that the bot will
prefer to discard that instead. Then, protection of the
newest card is not necessarily urgent any more, and
neither will a player necessarily infer that the newest
card is safe from a failure by their partner to protect it
immediately.
• There seems to be some interesting dictionary of hints
that we haven't worked out yet about ways to signal
to discard cards besides the newest, which prevents
junk from accumulating in the hand as non-playable
but useful-to-hold-on-to cards enter the hand.
Miscellaneous Communication
• Often the hints, and sometimes its other actions just
come "attached" with miscellaneous information. The
most extreme example is I observed one game where
as a result of the bots discarding, it was immediately
implied that a particular card in the other player's hand
was almost certainly red. This information was not
immediately useful (the red card was not yet playable,
nor was it likely to have been discarded soon), it was
simply just extra information attached to the action of
discarding in that particular case. Presumably the bot
was by convention constrained to almost certainly do
some other action in that situation had that partner's
card counterfactually not been red.
• This kind of extra not-immediately-useful "attached"
information is perhaps the most non-human part of the
bot's convention set. But actually it doesn't happen as
often as one might expect from a "nonhuman" agent.
For the most part I didn't see this all that much for
plays and discards (that one extreme example notwith-
standing). This makes sense, as having too many such
conventions would overly constrain the ability of the
players to act, as discard/play are both critical actions
you need to take very frequently regardless of the other
player's hand.
• Even for hint actions, most hints were very sensible
and humanly explainable, or clearly appeared that they
would be humanly explainable had we had a larger
sample size so that we could be surer about the general-
ity of its meaning and exactly how the bot had packed
different meanings into the hint space. There were only
a few hint actions that I found particularly "weird" in
what inference was made.
Bayesian Action Decoder
thereafter (i.e. getting plays out in a timely man-
ner, collectively never discarding any card that
could be useful thereafter, optimizing who draws
the next card for parity, etc), and by my under-
standing of their conventions, nothing stopping
them from doing so.
-- But instead of playing, the bot wasted a turn giv-
ing their partner a hint. When you inspected the
V2 belief state, it gave no useful information - the
dominant effect of the hint was actually to con-
centrate probability mass *away* from the truth
giving the partner a misleading belief about a card,
and had almost no other effects.
-- Their partner then proceeded to also not play and
instead discarded their newest card, which unnec-
essarily lost one of the copies of a useful 4. There
was a copy of the 4 left in the deck, but such a dis-
card is still bad. If the remaining copy of that card
is the very bottom card of the deck, it guarantees
that you cannot get 25 points, so every unneces-
sary discard of the first copy of any card loses you
EV due to the chance for the other copy to be the
last card.
Speculating a little here - perhaps something about the bot's
policy or convention set hasn't converged as sharply in the
endgame? It's certainly the case that the gradient there is
much smaller - even a clear mistake near the end tends to
cost you only a little in EV if you're measuring by score,
whereas near the start of the game it can cost you a lot.
And the expected penalty for discarding the first copy of a
useful 4 when otherwise well ahead is slight, since it then
usually only harms you when that 4 is precisely the last card
in the deck which only happens 1/N times, so one might
imagine the average gradient there for good behavior to be
very small.
• A priori, there's no particular reason why a bot's con-
ventions couldn't, for example, completely change
depending on the turn number modulo 3, and be ex-
tremely hard for humans to comprehend. But for the
most part, the conventions of this bot weren't like that
- they were pretty understandable, or at least seemed
consistent and sensible even if we didn't have all the
exact meanings mapped out.
Overall Quality of Play and Game Flow
• The bot is *very* strong in the early game, and there
its convention set is overall far more efficient than "nat-
ural" human convention sets (although not-necessarily
human convention sets that were constructed to be
more artificial and encoding-like).
It's really quite
beautiful.
• The bot is superhumanish at tracking inferred informa-
tion over time, e.g. on the one hand inferring that a
card is not scary in first position, then as it drifts back
later in the hand, inferring this or that other property
incidentally, and inferring based on the "aging" of the
card that it is probably not this or that, and so on, until
only a couple possibilities remain. It's not uncommon
that in the midgame, both players know almost all the
relevant things about their hands.
• The bot might be tactically weak in occasional situa-
tions on or near 0 hints, where the ensuing sequence of
actions is heavily constrained. It seems to have a very
strong preference to discard and get away from 0 hints,
even when as far as we can tell based on its convention
set it should be possible to just stay at 0 hints and play
out some cards, and where discarding at that moment
is suboptimal. For example if the ensuing sequence
of plays would result in a few 5s being played thereby
recovering some hints for free, and the partner's im-
mediate discard is also completely safe in the event
that the partner wants to discard, whereas one's own
discard unnecessarily loses a copy of a card that could
be useful in the future. (If I read the paper right, there
is no explicit lookahead in this bot?).
• The bot makes a few seemingly-clear mistakes in the
endgame (as far as we can tell), although only slight
ones. For example, one of the games we looked at:
-- The players were in a close-to-winning state - they
both knew all the playable cards in their hands or
had inferred them with high confidence, and all
they needed to do was play those cards and wait
to draw the few remaining cards to play.
-- They had plenty enough hints and headroom
to theoretically execute essentially-perfect play
|
1104.1279 | 1 | 1104 | 2011-04-07T09:32:00 | Context Aware Multisensor Image Fusion for Military Sensor Networks using Multi Agent System | [
"cs.MA"
] | This paper proposes a Context Aware Agent based Military Sensor Network (CAMSN) to form an improved infrastructure for multi-sensor image fusion. It considers contexts driven by a node and sink. The contexts such as general and critical object detection are node driven where as sensing time (such as day or night) is sink driven. The agencies used in the scheme are categorized as node and sink agency. Each agency employs a set of static and mobile agents to perform dedicated tasks. Node agency performs context sensing and context interpretation based on the sensed image and sensing time. Node agency comprises of node manager agent, context agent and node blackboard (NBB). Context agent gathers the context from the target and updates the NBB, Node manager agent interprets the context and passes the context information to sink node by using flooding mechanism. Sink agency mainly comprises of sink manager agent, fusing agent, and sink black board. A context at the sensor node triggers the fusion process at the sink. Based on the context, sink manager agent triggers the fusing agent. Fusing agent roams around the network, visits active sensor node, fuses the relevant images and sends the fused image to sink. The fusing agent uses wavelet transform for fusion. The scheme is simulated for testing its operation effectiveness in terms of fusion time, mean square error, throughput, dropping rate, bandwidth requirement, node battery usage and agent overhead. | cs.MA | cs | International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
Context Aware Multisensor Image Fusion for
Military Sensor Networks using Multi-Agent
system
Ashok V. Sutagundar*, S. S. Manvi**
*Department of Electronics and Communication Engineering, Basaveshwar Engineering
College, Bagalkot-587102, INDIA.
**Wireless Information Systems Research Lab, Department of Electronics and
Communication Engineering, REVA , Institute of Technology and Management,
Yelahanka, Bangalore
[email protected], [email protected].
Abstract
This paper proposes a Context Aware Agent based Military Sensor Network (CAMSN) to form an improved
infrastructure for multi-sensor image fusion. It considers contexts driven by a node and sink. The contexts such as
general and critical object detection are node driven where as sensing time (such as day or night) is sink driven. The
agencies used in the scheme are categorized as node and sink agency. Each agency employs a set of static and mobile
agents to perform dedicated tasks. Node agency performs context sensing and context interpretation based on the
sensed image and sensing time. Node agency comprises of node manager agent, context agent and node blackboard
(NBB). Context agent gathers the context from the target and updates the NBB, Node manager agent interprets the
context and passes the context information to sink node by using flooding mechanism. Sink agency mainly comprises of
sink manager agent, fusing agent, and sink black board. A context at the sensor node triggers the fusion process at the
sink. Based on the context, sink manager agent triggers the fusing agent. Fusing agent roams around the network, visits
active sensor node, fuses the relevant images and sends the fused image to sink. The fusing agent uses wavelet
transform for fusion. The scheme is simulated for testing its operation effectiveness in terms of fusion time, mean
square error, throughput, dropping rate, bandwidth requirement, node battery usage and agent overhead.
Keywords
Context Aware, Image fusion, Discrete Wavelet transform (DWT), Agent, Military sensor Networks (MSN).
1.Introduction
Recent advancement in wireless communications and sensor technology has enabled the development of
low-cost Wireless Sensor Networks (WSNs). The sensor networks can be used for various application areas
such as health, military, environmental monitoring, home, etc,. A typical WSN consists of a large number
of low-cost and low-energy sensors, which are scattered in an area of interest to collect observations and
pre-process the observations. Each sensor node has its own communication capability to communicate with
other sensor nodes or the central node (fusion center) via a wireless channel. More recently, the production
of cheap CMOS cameras and microphones, which can acquire rich media content from the environment,
created a new wave into the evolution of wireless sensor networks. For instance, the Cyclops imaging
module is a light-weight imaging module which can be adapted to MICA21 or MICAz sensor nodes. Thus,
a new class of WSNs came to the scene that can be applied to military applications and are known as
Wireless Military Sensor Networks (WMSNs).
DOI : 10.5121/ijasuc.2011.2113 147
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
Lot of papers have been published in this research area [1]. Most of this research has, however, focused on
wireless networks of sensor nodes that collect scalar data such as temperature, pressure, and humidity
sensors. Such sensors generate a limited amount of information, which can be insufficient for many
applications, even if a large number of sensors are deployed. As VSNs offer new opportunities for many
promising applications compared to scalar sensor networks, they also raise new challenges that are not fully
addressed by current research on WSNs. Camera sensors generate a huge amount of data compared to
scalar sensors. Processing and transmitting such data by generally low-power sensor nodes is challenging
due to their computational and bandwidth requirements. It is emphasized in [2] that these applications
demand a re-consideration of the computation communication paradigm of traditional WSNs, which has
mainly focused on reducing the energy consumption, targeting to prolong the longevity of the sensor
network. Though, the applications implemented by WMSNs have a second goal, as important as the energy
consumption, to be pursued. This goal is the delivery of application-level quality of service (QoS) and the
mapping of this requirement to network layer metrics, like latency. A VSN with overlapped field of views
could exploit the redundancy between the field of view of each visual sensor to avoid inconsistencies and
obtain more accurate results. Therefore, a key challenge in visual sensor network contexts is how to get the
most relevant information from the environment and fuse it in the most efficient way. However getting the
most relevant information from each visual sensor is not a simple task. Multiple factors could affect the
visual sensor information, for example in tracking activities, occlusions of static objects could affect the
tracking positions.
Context awareness can then be defined as detecting a user’s internal or external state. Context-aware
computing describes the situation of a wearable or mobile computer being aware of the user’s state,
surroundings, and modifying its behavior based on this information [3]. Context awareness plays a
significant role in MSNs because it allows for interpreting various contexts such as temporal, emergency
and computational contexts coming from the battlefield based on information regarding the current state of
the object movement and the state of the environment. Context-aware sensing is an integral part of the
MSNs design to achieve the ultimate goal of long-term pervasive enemy battle field monitoring.
This paper proposes a Context Aware Agent based Military Sensor Network (CAMSN) to form an
improved infrastructure for multi-sensor image fusion. It considers contexts driven by a node and sink. The
contexts such as general and critical object detection are node driven where as sensing time (such as day or
night) is sink driven. The agencies used in the scheme are categorized as node and sink agency. Each
agency employs a set of static and mobile agents to perform dedicated tasks. Node agency performs
context sensing and context interpretation based on the sensed image and sensing time. Node agency
comprises of node manager agent, context agent and node blackboard (NBB). Context agent gathers the
context from the target and updates the NBB, Node manager agent interprets the context and passes the
context information to sink node by using flooding mechanism. Sink agency mainly comprises of sink
manager agent, fusing agent, and sink black board. A context at the sensor node triggers the fusion process
at the sink. Based on the context, sink manager agent triggers the fusing agent. Fusing agent roams around
the network, visits active sensor node, fuses the relevant images and sends the fused image to sink. The
fusing agent uses wavelet transform for fusion.
1.1Related works
Some of the related works are as follows. The focus of article [4] is on the military requirements for
flexible wireless sensor networks. Based on the main networking characteristics and military use-cases,
insights into specific military requirements for flexible wireless sensor networks are discussed. The article
structures the evolution of military sensor networking devices by identifying three generations of sensors
along with their capabilities. Existing developer solutions are presented and an overview of some existing
tailored products for the military environment is given. The work presented in [5] investigates the design
tradeoffs for using WSN for implementing a system, which is capable of detecting and tracking military
148
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
targets such as tanks and vehicles. Such a system has the potential to reduce the casualties incurred in
surveillance of hostile environments. The system estimates and tracks the target based on the spatial
differences of the target object signal strength detected by the sensors at different locations. The work
depicted in paper [6] proposes to use the mobile agent paradigm for reducing and aggregating data in planar
sensor network architecture. Mobile agents can be used to greatly reduce the communication cost,
especially over low bandwidth links, by moving the processing function to the data rather than bringing the
data to a central processor.
The work given in paper [7] describes the use of the mobile agent paradigm to design an improved
infrastructure for data integration in distributed sensor network (DSN). Instead of moving data to
processing elements for data integration, as is typical of a client/server paradigm, it moves the processing
code to the data locations. This saves network bandwidth and provides an effective means for overcoming
network latency, since large data transfers are avoided. The work given in [8], describes the image
transmission problem in sensor networks. It pre-processes the images in the sensors before sending them to
the server, but this preprocessing requires extra energy in the sensors. In [9], a hierarchical multi-quality
image fusion method based on region-mapping is proposed to improve quality of the image. Camera view
is divided into regions according to mapping relation and the structured deployment of nodes. The methods
of motion attention analysis and nearest neighbor sampling are also adopted to optimize the local attention
region of the image. The fusion method is analyzed and verified to improve the quality of the monitor
image with no need for more energy and bandwidth. The work presented in [10] gives architecture to
implement scene understanding algorithms in the visual surveillance domain. The main objective is to
obtain a high level description of the events observed by multiple cameras not to fuse the tracking
information. In [11], authors present a visual sensor network system with overlapped field of views,
modeled as a network of software agents. The communication of each software agent allows the use of
feedback information in the visual sensors, called active fusion. The work limits only to indoor scenario
only.
1.2Our Contributions
However most of the related works focus on how to solve different visual sensor problems, there are few
works that focus on building a software architecture which allows a context aware wireless visual sensor
network, where agents and wavelet transform can be applied to take active part in the fusion process. Our
contributions in this paper are as follows: 1) contexts are gathered from the target (sensor node) as well as
generated at the sink, 2) several static agents are defined for gathering and interpretation of the context, 3)
militant activities can be monitored using the proposed MSNs, 4) used mobile agents to perform several
tasks that aid information processing, fusing, etc., in asynchronous fashion, 5) based on the context,
low/high resolution image fusion is employed by changing agent code, 6) embedding wavelet code in the
agents for fusing images, 7) fuses only images from active sensor nodes, and 8) reduces communication
overheads and energy.
The rest of the paper is organized as follows. In section 2, we present context aware computing. Section 3
depicts the wavelet based multi-sensor image fusion. Section 4 describes context aware based information
fusion. Section 5 presents the simulation model and performance metrics. Section 6 explains results and
section 7 concludes the work.
2.Context Aware Computing
Context is any information that can be used to characterize the situation of an entity. An entity is a person,
place, or object that is considered relevant to the interaction between a user and an application, including
the user and applications themselves. In other words, any information that depicts the situation of a user can
be entitled context. This would include the number of people in the area, the time of day, and any devices
the user may employ. One can however distinguish between those contextual characteristics that are critical
149
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
to the application versus those that may be relevant but are not critical. Within computing applications,
there are three major context categories of interest: user context, computing resources, and environmental
aspects. Orthogonal to this view, context can be explicit (that is, information provided directly by the user)
or implicit (derived from on one hand sensors, on the other hand from an analysis of user behavior).
In previous decades, a narrow aspect of user context, the user preferences in terms of search and retrieval of
data was denoted in the database community with the concept of views. Today, user context is considered
far more broadly and includes user interests, goals, social situation, prior knowledge and history of
interaction with the system. Thus, the dimension of time may also be included in the user’s context.
``Context information may be utilized in an immediate, just in time, way or may be processed from a
historical perspective’’ [12, 13]. Environmental contextual aspects include (but are not limited to) location,
time, temperature, lighting conditions, and other persons present. Note that there is some ambiguity in the
literature in the use of the term environment, which can refer to the computing environment as well as the
actual physical environment [14,15,16,17] of the user. Context information is often acquired from
unconventional heterogeneous sources, such as motion detectors or GPS receivers. Such sources are likely
to be distributed and inhomogeneous. The information from these sensors must often be abstracted in order
to be used by an application; for example, GPS data may need to be converted to street addresses. Finally,
environmental context information must be detected in real time and applications must adapt to changes
dynamically.
To monitor the environment for changes and to adapt the behavior of the node to the current environment
CAC can be used. The applications and protocols need not be aware of the environment at all, but rather
focus on taking care of the tasks they have been designed for in the first place. As the network environment
changes, the node must be able to rapidly adapt to the new situation. In this work, context image sensor
nodes are configured based on the context. In order to reduce the communication overhead, we consider the
wavelet based image fusion in WSNs.
3.Wavelet Based Multi Sensor Image Fusion
In this section, we describe the wavelet transform and wavelet based image fusion.
3.1Wavelet Transform
Wavelets are mathematical functions that cut up data into different frequency components, and then study
each component with a resolution matched to its scale. They have advantages over traditional Fourier
methods in analyzing physical situations where the signal contains discontinuities and sharp spikes.
Wavelets were developed independently in the fields of mathematics, quantum physics, electrical
engineering, and seismic geology. Interchanges between these fields during the last decade has led to many
new wavelet applications such as image compression, turbulence, human vision, radar, and earthquake
prediction. The wavelet transform has become a useful computational tool for a variety of signal and image
processing applications. For example, the wavelet transform is useful for the compression of digital image
files, smaller files are important for storing images using less memory and for transmitting images faster
and more reliably. Wavelets are functions that satisfy certain mathematical requirements and are used in
representing data or other functions. Wavelet algorithms process data at different scales or resolutions [20,
21, 22, 23].
3.2Image Fusion
Image fusion is the processing of images about a given region obtained from different sensors by a specific
algorithm, so that the resultant image is more reliable, clear and more intelligible. Image fusion can take
150
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
place on pixel-level, feature-level, and decision-level. Pixel-level image fusion is the basic for other levels
and multiresolution image fusion based on multi-scale decomposition is a main research branch in it.
Figure 1: Sub band coding, Figure 2: Image Fusion using DWT
The 1-D wavelet transform can be extended to a two-dimensional (2-D) wavelet transform using separable
wavelet filters. With separable filters the 2-D transform can be computed by applying a 1-D transform to all
the rows of the input, and then repeating on all of the columns. The wavelet transforms of the input images
are appropriately combined, and the new image is obtained by taking the inverse wavelet transform of the
fused wavelet coefficients. An area-based maximum selection rule and a consistency verification step are
used for feature selection. The original image of a one-level (K=1), 2-D wavelet transform, with
corresponding notation is shown in figure 1.
The 2-D subband decomposition is just an extension of 1-D subband decomposition. The entire process is
carried out by executing 1-D subband decomposition twice, first in one direction (horizontal), then in the
orthogonal (vertical) direction. For example, the low-pass subbands (Li) resulting from the horizontal
direction is further decomposed in the vertical direction, leading to LLi and LHi subbands. To obtain a two-
dimensional wavelet transform, the one-dimensional transform is applied first along the rows and then
along the columns to produce four subbands: low-resolution, horizontal, vertical, and diagonal. (The
vertical subband is created by applying a horizontal high-pass, which yields vertical edges.) At each level,
the wavelet transform can be reapplied to the low-resolution subband to further decorrelate the image.
Figure 3 depicts the image fusion using DWT. DWT is first performed on each source images, then a
fusion decision map is generated based on a set of fusion rules. The fused wavelet coefficient map can be
constructed from the wavelet coefficients of the source images according to the fusion decision map.
Finally the fused image is obtained by performing the inverse wavelet transform. When constructing each
wavelet coefficient for the fused image, we have to determine which source image describes this coefficient
better. This information will be kept in the fusion decision map. The fusion decision map has the same size
as the original image. Each value is the index of the source image which may be more informative on the
corresponding wavelet coefficient and thus, make decision on each coefficient.
Assume that node i and node j have the common information to be sent to the sink node. In order eliminate
the common information between the neighboring nodes, a fusion process is adopted. Multi sensor image
fusion using wavelet transform is described as follows (1) and (2).
The fused image is given by (3)
151
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
where, fi(x,y): Image of sensor node i, f j(x,y): Image of sensor node j, fij(x,y): Fused Image of sensor node
i and sensor node j, Fi(u,v): Node i image in transform domain, and Fj(u,v): Node j image in transform
domain.
4.Context Aware Based Image Fusion
The scheme comprises of three phases: context gathering, context interpretation, and image fusion. Context
gathering - contexts are gathered from the target, i.e., sensed image and time from the target, and stored in
the node for a short period until its interpretation is done. Context interpretation sensed images are
compared with previous image and set of critical image features (weapons, explosives, enemy, etc.) stored
at the node. If image analysis yields some general or critical object feature existence, information fusion
process is invoked. Information fusion- relevant images from active sensor nodes corresponding to object
existence are fused to get a clear picture of the object and make some decisions.
Image fusion can classify into two types namely low resolution image fusion and high resolution image
fusion. Low resolution image fusion is used for contexts like general object detection, and general image
gathering by sink. High resolution image fusion is done for contexts such as critical object detection, image
gathering in night time by sink. During nights, it is better to monitor the target periodically by the deployed
MSN since lighting condition is poor and possibility of enemy attack, militant activities, etc,. are high. Sink
driven image fusion is based on the time of sensing, available network bandwidth and sensor node battery.
Static and mobile agents are employed to perform the fusion process. The scheme assumes that an agent
platform is available in the nodes of WSNs. However, if an agent platform is unavailable, the agent
communicates by traditional message exchange mechanisms such as message passing method. Several
agencies exist at each of the nodes based on their role in MSN, which will be discussed in this section.
Following are the assumptions made in proposed work.
1. Context Interpretation is based on the image detection. Image detection is not in the scope of this
work.
2. Based on the lighting condition, camera of the sensor node automatically configures itself to get
the better quality images.
3. Flooding protocol is used to pass the context information to the sink node [25].
4. Geographical routing protocol is used by mobile agent to reach the active sensor nodes [26].
5. Low and high resolution image fusion code is used to fuse the images.
This section presents system environment and the agencies at the sensor and sinks nodes.
1.1 System Environment
Sensor node comprises of static sensor nodes and sink nodes (also called as end nodes that require
information). A sensor node is said to form a cluster around it based on the communication range. Each
cluster will have cluster head node. Figure 3 depicts the WSN environment. A sensor node may have
several channels with different sensory devices connected to each of them. Sensor nodes are geographically
distributed and periodically collect measurements of different modalities such as acoustic, seismic, and
infrared from the environment. Every sensor node of WSN has predetermined value of the signal strength.
Once the information is sensed by sensor node, it compares the information signal strength with
predetermined value, if it is greater (if the deviation is more), a message is sent to its cluster head saying
that it is an active node.
The signal energy from each channel can be detected individually and processed in the analog front end.
The amount of signal energy that reaches an individual sensor is an effective indicator of how close the
node is to a potential target. Once the signal is captured and pre-processed by a sensor node, the strength
152
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
level of the detected signal is broadcast (through an omni directional antenna) to its cluster nodes if it
greater than some predefined value.
Figure 3: Distributed Sensor Network with Three Clusters
1.2 Agent Technology
The traditional programming paradigm uses functions, procedures, structures and objects to develop
software for performing a given task. This paradigm does not support development of flexible, intelligent
and adaptable software’s, and also does not facilitate all the requirements of Component Based Software
Engineering (CBSE) [27,28]. In recent developments, agent technology is making its way as a new
paradigm in the areas of artificial intelligence and computing, which facilitates sophisticated software
development with features like flexibility, scalability and CBSE requirements [29,30,31].
Agents are the autonomous programs situated within an environment, which sense the environment and act
upon it to achieve the goals. Agents have special properties such as mandatory and orthogonal properties.
The orthogonal properties provide strong notion of the agents [32, 33, 34]. The mandatory properties are as
follows: autonomy, decision-making, temporal continuity, goal oriented. The orthogonal properties are as
follows: mobility, collaborative, learning. A mobile agent platform comprises of agents, agent server,
interpreter and transport mechanisms. We classify the agent technologies as single-agent systems and
muti-agent systems. In the context of single-agent systems Local or user interface agents and Networked
agents can be identified, while in the area of multi-agent systems DAI (Distributed Artificial Intelligence)-
based agents and Mobile agents can be distinguished. In single-agent systems, an agent performs a task on
behalf of a user or some process, while performing a task; the agent may communicate with the user as well
with local or remote system resources. In contrast, the agents in multi-agent systems (MAS) may
extensively cooperate with each other to achieve their individual goals, and also may interact with users
and system resources.
1.3 Image Signal Strength
Image signal strength is measured by estimating the difference between previous image and present image
stored at the sensor node. To measure the change in the image, entropy of the difference image between the
present and previous image is taken. The notion of entropy may be used to estimate the information
content. For an image, the simplest idea is to create these states that correspond to the possible values in
which all pixels are involved. The image entropy would be given by [35]:
153
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
Where Pk is the probability of gray level k, k=0,1,2,3,…,255, assuming an 8-bit image. Equation (4)
represents the information content of an image. If a change in an image takes place, the entropy, H, changes
as well. Otherwise no change is detected. Firstly, subtract the previous image from a sensed image to obtain
the features which is given by (5).
Where Ij is the present image, I0 is the previous image of the node. Entropy is then applied to measure the
change in information. This step provides the image signal strength estimation, which is expressed by
equation (6).
Where Pj is difference image and H denotes entropy. Only the gray scale information of the images is used
to calculate the image entropy. A threshold signal of the image is considered to find out the required signal
strength where it is defined as the amount of information change from present to the previous image. This
facilitates to decide the active mode of a sensor node. If Pj is greater than threshold, then that node is
considered to be active otherwise inactive.
1.4 Fused Image Transmission Model
Let R1, R2, R3,....... Rk be the sensor nodes deployed in the field, and
is Fusion Factor, and N1 is the Fused
Image at R1 and later fuses with image from other nodes, i.e., shown in equation (7).
Figure 4 shows the network path representation. We assume that most of the sensor nodes carry the similar
images therefore instead of sending the individual images from each of the sensor nodes, the images are
fused in order to remove the redundancy and fused image fij(x,y) is being sent to the sink node.
Figure 4: Network Path Representation
Agent carries the fused images from node to node depending on its given itinerary or chosen itinerary.
Total transmission load, TLoad can be given by (8).
where, hc: Number of hops required to transmit the fused image to sink node, mal: Total number of packets
required to transmit the image and fusion code and TLoad: Transmission Load.
154
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
1.5 Agencies
Each node in MSN comprises of an agent platform and the proposed agent information fusion model. The
sensor nodes and sink node comprises of the node agency and sink agency respectively. Here we describe
both the agencies.
1.5.1 Node Agency
Node agency comprises of static agents and a node blackboard (NBB) for inter agent communication.
Agents are Node Manager Agent (NMA) and Context agent (CA). The agency is depicted in figure 5.
Node Manager Agent (NMA): This agent resides in all the sensor nodes of MSN. It creates context
agent and NBB and is responsible for synchronizing the actions of the agents within themselves and outside
world/agents. CA senses the image and updates the NBB periodically with image, capture time and image
signal strength. NMA compares the sensed image with previous image and set of critical images residing
in NBB and interprets the context. In sleep mode, NMA does not transmit any information to sink. The
agent also monitors the battery life; if battery is exhausting, sends the status of the battery to its sink node.
NMA sends information of the node such as node id, geographical information, context, and signal strength
information to the sink node, if and only if it has a signal strength above the predefined threshold set by the
sink (varies with time and critical situations).
Node black Board (NBB): This knowledge base is read and updated by the agents. NBB comprises of
node id, active mode /sleep mode, sensing time, previously sensed image, set of critical images, bandwidth
required to transmit in the available bandwidth, image signal strength and geographical status, and location
of the node. Critical images comprises of images of weapons, explosives, enemy dresses, etc. Figure 6
depicts the sample NBB at 11.13AM, when the CA senses the image and updates the NBB. The object is
detected to be a context-general object (Cgo).
Figure 5: Node Agency Figure 6: Node Black Board
Context Agent (CA): It gathers the context from the target such as time of sensing the image and
updates the NBB periodically. It interacts with NMA regarding the updating of the NBB. The agent is in
direct contact with the sensor. It can be also triggered as and when required (aperiodic) based on the need
of information by the sink.
155
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
1.5.2 Sink Agency
The sink agency comprises of static, mobile agents and sink black board (SBB). Agents employed are
Fusing agent (FA), and Sink Manager Agent (SMA) which are mobile and static, respectively. Sink agency
is shown in figure 7.
Sink Manager Agent (SMA): This agent resides in the sink node of MSN. It creates fusing agent and
SBB and is responsible for synchronizing the actions of the agents within themselves and outside
world/agents and monitors and updates the SBB continuously. It is responsible for the image fusion
process. SMA gets fused images from the target in two ways: 1) based on the sensor node context (context
driven) and 2) whenever user seeks (sink driven) fused information. SMA retrieves active sensor node's
geographical locations information from the SBB, generates the optimum routes using standard WSN
geographical routing protocol as and when required. Based on the context or as and when user needs the
information, SMA triggers FA with the necessary fusion code, forward and reverse route. The fused image
is given to the user monitoring the network either by alert alarm or updated in the user database of images.
Fusing Agent (FA): It is a mobile agent equipped with image fusion code (low/high resolution based on
the context) that migrates from one to another active node depending on the routing information provided
by the SMA. Whenever it visits active nodes, fuses the image, and moves to another active node along
with the fused image. The agent repeats fusion until it visits all active nodes. After completion of the fusion
process, FA sends the fused image to the sink node using reverse route specified by SMA.
Sink Black Board (SBB): It is the knowledge base that can read and updated by SMA. It stores the
information about the node id, signal strength, context information, time of sensing, image signal strength,
bandwidth required to transmit the image of each active node, available network bandwidth, and
geographical locations of the active nodes. Sample contents of SBB is shown in the figure 8 (using 1 as the
sink node).
Figure 7: Sink Agency
Figure 8: Sink Black Board
1.6 Agency Algorithm
In this section we discuss the node and sink agency algorithms.
1.6.1 Node Agency Algorithm
Nomenclature:x,y: index value of the image, D(x; y) = Difference image,C(x; y)= Critical image, P(x; y)=
Present image, B(x; y)= Previous image, Cgo= General object context, Cco= Critical object context, i: index
value of the critical images, , N _N: Size of the image,and H: Entropy operator, H1;H2=Entropy of present
and previous images, Tth= Threshold of image signal strength, Pth =Percentage of similarity between
previous and present image, and Nstatus= Node status, Ncriim= Number of critical images.
156
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
Begin
1. CA gathers the image and time of sensing
2. NMA gets the context information and does the interpretation
For x= 1 to N do,
For y= 1 to N do,
H1=H(P(x,y));
H2=H(B(x,y));
endfor
endfor
Pth = (H1/H2)X100;
NMA updates the Pth to NBB;
if(Pth > Tth)
Nstatus = Active;
NMA updates the node status to NBB;
1. else
Nstatus = Inactive;
Delete the image from NBB;
NMA updates the node status to NBB;
Go to (4);
For i= 1 to Ncriim,
For x= 1 to N do,
For y= 1 to N do,
D(x,y)= Pim(x,y) - Bim(x,y);
if(Di(x,y) = =Ci(x,y)) Then,
Context= Cco;
else
endfor
context =Cgo;
endfor
endfor
3. NMA updates the context information to NBB;
4. Stop
End
CA gathers the context such as image and sensing time and updates the NBB. NMA finds the similarity
between the previous and present image by taking the entropy ratio of them. The percentage of similarity
ratio compared with Tth, if it is greater than the threshold value then node is assigned with active status
otherwise inactive status is assigned. If the node is inactive then present image is deleted from NBB. The
D(x,y) is compared with set of critical images, C(x,y) of the node and if any critical object is found, then it
is assigned with Cco, else Cgo. NMA also finds the entropy of the difference image and compares with
threshold signal strength of an image, if it is greater; than status of the node is stored as active else status of
the node is stored as inactive. NMA updates the context information to NBB and sends it to the sink node
by using the flooding mechanism. In flooding mechanism message is sent from node to node until it
reaches the sink node.
4.6.2 Sink Node Agency Algorithm
Nomenclature:F(x,y) = Fused Image, N(x,y)= Active node Image, Nactive= Number of active nodes Cgo=
General Object Context, Cco= Critical Object Context, Rtable= Routing table to reach active sensor node, i=
157
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
index value of the active nodes, x,y= index value of the node image, Fgocode= General objection detection
image fusion code, Fcocode= Critical objection detection image fusion code NXN= Size of the image
.
Begin
1. SMA gets the context information from SBB;
2. SMA creates FA, and provides the context information;
3. Based on the Context FA gets the fusion code and routing table
if (context= = Cgo) Then,
FA gets the Fgocode and Rtable;
else
FA gets the Fcocode and Rtable;
4. Along with fusion code and routing information, FA visits active sensor nodes for fusion
F(x,y)=0;
For i=1 to Nactive
For x= 1 to N do,
For y= 1 to N do,
F(x,y)= F(x,y) + Ni(x,y);
endfor
endfor
endfor
5. FA sends the fused image to the sink node;
6. FA is disposed;
7. Go to (1) as and when the user seeks the information and repeat the process;
8. stop;
End
Whenever the SMA gets the context information or user seeks information, it creates the FA, based on the
context; FA retrieves low/ high resolution fusion code. With fusion code FA visits all the active sensor
nodes to fuse the images, and returns with fused image to the sink node. Fusion of images is done using
DWT, which is discussed in the section 3.2. For simplicity, in the sink agency algorithm fusion has
considered to be an addition operation.
4.7 Agent Interaction
Context based image fusion system agent interaction sequence is depicted in figure 9, which provides a
detailed view of the image fusion by agents in a MSNs. The numbers shown on the directed arcs denotes
the action number in the sequence of interactions that takes place.
158
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
Figure 9: Information Fusion System Agent Interaction
The interaction sequence is as follows:
1. CA of the active sensor nodes gathers the context namely sensing time and image. NMA compares
the sensed image with the previous and set of critical images of the node and interprets the
context.
2. NMA sends the context information into the sink by using the flooding protocol.
3. NMA updates the context information to SBB.
4. SMA gets the context information, creates the FA, along with fusion code (low/high resolution)
FA visits the first active node and fuses the image and migrate to the next active node and
continue the process till it visits all the active nodes.
5. FA returns to sink node along with fused image.
6. During night time, as and when user requires fused information, SMA generates the FA agent and
sends it for the fusion.
7. FA returns with fused image to sink node. Fused image will be sent to the military base camp
through the Internet or satellite.
5. Simulation model and performance metrics
We have carried out the simulation of the context aware image fusion using multi-agents for MSNs. The
proposed model has been simulated in various network scenarios on a Pentium-4 machine by using
MATLAB tool for the performance and effectiveness of the approach.
Simulation Model
Proposed scheme is simulated in various network scenarios to access the performance and effective of an
approach. Simulation environment comprises of six models namely network model, Channel Model,
Propagation model, Battery model, CAC model, and information fusion model. These models are described
as follows.
1. Network Model:
We considered simulation area of A X B sq meters with LOC scenario. A MSN consists of num static
nodes that are placed randomly within the given area. MSN have communication radius, r mts, and
network bandwidth, net_BW. Fcode kbytes wavelet based image fusion code roams around the network.
159
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
2. Channel Model
The communication environment is assumed to be contention-free (e.g., a media access scheme such as
time division media access (TDMA) may be assumed). The transmission of packets is assumed to occur in
discrete time. A node receives all packets heading to it during receiving interval unless the sender node is in
off state. For simplicity, we have considered the channel to be error free. The characteristics of sensor
networks and applications motivate a MAC that is different from traditional wireless MACs such as IEEE
802.11 in almost every way: energy conservation and self-configuration are primary goals, while per-node
fairness and latency are less important. Sensor MAC protocol (S-MAC) uses different techniques to reduce
energy consumption and support self-configuration. To reduce energy consumption in listening to an idle
channel, nodes periodically sleep. Neighboring nodes form virtual clusters to auto-synchronize on sleep
schedules. Simulation environment uses the S-MAC protocol.
3. Propagation model
Free space propagation model is used with propagation constant β. Transmission range of WMSN node
communication radius is r for a single-hop distance d. It is assumed that at any given time, the value of
transmitted power is NPow milliwatts for every node.
4. Battery model
In MSN, image sensor nodes are deployed in the battle field, recharging of the nodes at the target is
difficult. So, we have used a solar cell recharging model [36] and a layered clustering model to deal with
the restrict energy consumption under the consideration of visual quality. The system lifetime can be
prolonged by rechargeable solar cell that can be recharged by solar panel in daytime. Image sensor nodes
consumes node-batt millivolts (day, non critical information= less, day critical information= medium, and
night=more) to sense an image.
5. CAC Model
Various contexts considered are: General object detection context, Critical object detection context and
night detection context. The images and their respective contexts are stored in the knowledge base of each
sensor. Contexts are randomly generated by assigning a number such that 1= General object detection, 2 =
Critical object detection, 3 = Time (night or poor lighting condition) context.
6. Information Fusion Model
We are using two information fusion methods namely context driven image fusion and user driven image
fusion. Context driven fusion uses low resolution image fusion or high resolution image fusion. User/sink
driven image fusion uses the high resolution image fusion. Each of the sensor node is associated with
battery of node-batt millivolts. It is assumed that 1 millivolt is decremented for every usage
(transmission/processing). k number of active nodes are randomly chosen as active nodes from num nodes.
Set of critical images, present and previous images are stored in the node. Size of the gray images stored at
each of the sensor node is fixed. The fixed gray scale image of size rows X columns, 8 bits/pixel is assigned
to each of the active sensor node. Th, is the percentage of the threshold image signal strength.
5.1 Simulation Procedure
To illustrate some results of the simulation, we have taken A = 100, B =200 sq meters and N=1 to 5, num=5
to 15, netBW=40mbps, node_batt=90 millivolts, Th= (50%, 60 %, 70 %), Gray scale image of varying size
rows_columns=(32X32; 64X64; 128X128; 256X256), (8,12,16,24) bits/pixel, time of sensing (Tsensing=
(8AM, 11.30AM,5PM, 7PM), Pth= Present Signal strength (30%,50%,60%,70%), communication radius
r=10 mts , netBW= 4MBPS, Propagation Constant β =3.5, and Fcode = (4 , 8, 12) Kbytes. Each of the
simulation executes in varying seconds.
160
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
Begin
• Generate the WMSN for the given radius and number nodes.
• Apply the proposed context aware fusion model.
• Compute the performance of the system.
End
Performance parameters considered in the simulation are as follows.
1. Node Battery Usage: It is defined as the battery depleted with the usage of a node.
2. Dropping rate: It is defined as the number of packets missed during the transmission of packets
(Equation 9), i.e., it is a ratio of packets dropped to packets sent.
1. Fusion time: It is the total time required by the FA to fuse the images from active sensor nodes.
2. Bandwidth Requirement: It is the amount of bandwidth required to transmit the image to sink
node, i.e., it is a ratio of image size to available bandwidth. It is given by (Equation 10)
(9)
3. Throughput: It is the ratio of Image packets (data) received to the image packets sent and is given
by (Equation 11).
=
(11)
4. Mean Square Error: It is defined as the standard deviation of the difference image between the
ideal and standard image.
5. Agent Overhead: It is the additional code which acquires the communication channel. It is given
by (Equation 12).
(12)
6. Results
In this section, we discuss the various results obtained through the simulations. The results include node
battery usage, dropping rate, throughput, fusion time, bandwidth requirement, agent overhead and mean
square error.
6.1 Node Battery Usage
The life time of the sensor node mainly depends on the battery life time and its power. Sensor nodes will be
active whenever the sensor nodes have the information otherwise nodes status will be inactive (sleep
mode). In active mode of the sensor node, sensor nodes consume more power and in inactive mode, a node
consumes less power. This is evaluated by choosing the one sensor node of WSN for repeated simulation.
For each of the simulation, sensor node sends the varying number of packets. We observe from the figure
10 that the battery life decreases as the number of packets sent by the sensor node increases.
161
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
Sensor nodes require more power to transmit the critical and dark images which are of large size (better
resolution) as compared to day time image. During day time, only low resolution images are to be sent to
the sink node that needs less power. From figure 11, we can notice that power consumed by each of the
sensor node to transmit the day time, non critical image is less as compared to the critical dark images.
Figure 10: Battery in Milli Volts Vs. Number of times node used, Figure 11: Power in Milli watts Vs.
Number of times node used
6.2 Dropping Rate
During the transmission of the image packets in the network, some packets may not reach the sink node.
Dropping rate depends on the number of packets dropped. From figure 12, we can notice that dropping rate
increases as there is increase in the number of nodes and decrease in signal strength value (Th).
Figure 12: Dropping Rate Vs No. of Active Nodes, Figure 13: Fusion Time in milliseconds Vs. No. of
Active Nodes
6.3 Fusion Time
Since the sensor nodes have less processing capability, mobile agent requires more time to fuse the image
from all the active nodes.
162
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
If there is day time and non critical images in all the active nodes, then all the active nodes information has
to be fused and sent to the sink node which needs less time as we use low resolution image fusion
technique and better lightning condition. In case of critical object detection context, if the lighting condition
is poor, we use high resolution image fusion method which needs more time for fusion. From the figure 13,
we depict that for critical images as the number of nodes and Th increases, the fusion time also increases.
We can notice that fusion time required for the noncritical object detection context is less compared to the
noncritical object detection context.
6.4 Bandwidth Requirements
We allocate 8 or 12 bits/pixel for the low resolution image. For high resolution image, we allocate 15 or 24
bits/pixel depending on the availability of the bandwidth. For critical and dark images, we allocate more
number of bits compared to the general object detection context images.
6.5 Throughput
Throughput depends on number of image packets sent and received. In this paper, we have considered the
gray scale image. While transmitting the image, it is divided into packets and sent. The figure 16 shows
that it decreases with increase in the image size
Figure 15: BW Vs. Size of the image, Figure 16: Throghput Vs. Image size.
6.6 Agent Overhead
Agent overhead depends on the size of agent. Fusing agent has to carry image along with the wavelet based
image fusion code. Agent code is extra overhead to the communication channel. Figure 17 shows that as
the image size increases, the overhead decreases, as the same size agent is used for handling small as well
as large size image. We can also observe, as the size of the agent increases agent overhead also increases.
163
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
Figure 17: Agent Overhead Vs. Image Size, Table 1: MSE for different Wavelet Bases
6.7 Mean Square Error
We have also analyzed the performance of the image fusion in terms of mean square error (MSE) ρ for
different wavelet bases such as biorthogonal, db, haar, and Mayer. ρ is defined as the standard deviation of
the difference image between the ideal and standard image (See table 1).
7. Conclusion
This paper proposed a context aware agent based distributed sensor network (CADSN) to form an
improved infrastructure for multi-sensor image fusion to monitor the militant activities. The proposed work
is based on context aware computing which uses software agents for image fusion in WMSN. In an
environment where source nodes are close to each other, and considerable redundancy exists in the sensed
data, the source nodes generate a large amount of traffic on the wireless channel, which not only wastes the
scarce wireless bandwidth, but also consumes a lot of battery energy. Instead of each source node sending
sensed images to the sink node, images from the different active nodes are fused and sent to sink node by
using mobile agent. The use of agents facilitates the following: 1) asynchronous operation, i.e., does not
require a continuous connectivity between source and sink, 2) flexibility to change the embedded code to
perform context/user driven fusion, 3) adaptability to varying network conditions and the environment for
image fusion, 4) ease of maintenance since the code can be debugged and upgraded independent of other
agents in the system, 5) reusing of the code is possible by other applications with slight modifications and
put in the system, thus enables dynamic software architecture.
However, there are some issues to be questioned in the proposed scheme, which can be taken up as further
work: security in information fusion by mobile agent, standard agent framework supporting persistence and
security to agents, tackling the active node failures during agent's fusion process?, positioning of the
cameras concerned to images to be fused, and security to the sensors. Furthermore, we are planning to
employ cognitive agents to tackle the aforesaid questions.
164
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
ACKNOWLEDGMENTS
We are thankful to Visvesvaraya Technological University (VTU), Karnataka, INDIA for sponsoring the
part of the project under VTU Research Grant Scheme, grant no. VTU/Aca/2009-10/A-9/11624, Dated:
January 4, 2009.
References
[1] M. Rahimi, “Cyclops: In Situ Image Sensing and Interpretation in Wireless Sensor Networks’’, Proc. ACM Conf.
Embedded Networked Sensor Sys., San Diego, CA, Nov. 2005.
[2] I. Akyildiz, T. Melodia, and K. R. Chowdhury,” A survey of wireless multimedia sensor networks”, The
International Journal of Computer and Telecommunications Networking, Vol.51, No.4, pp. 921-960, 2007.
[3] A. Krause, D. Siewiorek, A. Smailagic, and J. Farringdon, “ Unsupervised, dynamic identification of physiological
and activity context in wearable computing”, IEEE, Proc. 7th International Symposium on Wearable Computers, pp.
88-97, 2003.
[4] MichaelWinkler, Graeme Barclay,Kester Hughes , “Theoretical and practical aspects of military wireless sensor
networks”, Journal of Telecommunications and Information Technology (JTIT), February, 2008
[5] Tatiana Bokareva,Wen Hu, Salil Kanhere, “Wireless Sensor Networks for Battle field Surveillance”, Land Warfare
Conference, Brisbane, 2006
[6] Taekyoung Kwon, Yong Yuan, and Victor C.M. Leung, ”Mobile Agent Based Wireless Sensor Networks”, Journal
of Computers, Academy Publishers, Vol. 1, No. 1, April,2006
[7] H. Qi, S.S. Iyengar and K. Charabarty, “ Multi-Resolution Data Integration using Mobile Agents in Distributed
Sensor Networks”, IEEE-Systems Man Cybernetics, Vol. 31, No. 3, pp. 383-390, August 2001
[8] King-Shan Lui and Edmund Y. Lam, “.Image transmission in sensor networks”, IEEE Workshop on Signal
Processing Systems, pp. 726-730, November 2005.
[9] Minghua Yang,Yuanda Cao ,Li Tan ,Changyou Zhang ,Jiong Yu, “A new multi-quality image fusion method in
visual sensor network”, Third International Conference on Intelligent Information Hiding and Multimedia Signal
Processing, IIHMSP, Vol. 2, pp.667 - 670, 2007.
[10] J. Orwell, S. Massey, P. Remagnino, D. Greenhill, and G. A. Jones,” A multi-agent framework for visual
surveillance”, IEEE, Proc. 10th International Conference on Image Analysis and Processing ICIAP 99, Washington,
DC, USA, 1999.
[11] Federico Castanedo, Miguel A, Patricio , M. Molina,” A Multi-Agent Architecture to Support Active Fusion In A
Visual Sensor Network”, Second ACM/IEEE International Conference on distributed smart cameras , Stanford
University, California, USA, pp. 1-8, sept., 2008
[12] G. Chen and D. Kotz,.”A survey of context-aware mobile computing”, Technical report 2000-381 , Department of
Computer Science, Dartmouth College, 2000.
[13] A. Dey and G. Abowd,”Towards a better understanding of context and context-awareness”, Technical report,
Georgia Institute of Technology College of Computing 1999.
[14] M. Gaber, S. Krishnaswamy, and A. Zaslavsky,” Ubiquitous data stream mining”, Proc. Eighth Pacific-Asia
Conference on Knowledge Discovery and Data Mining, Sydney, Australia,2004.
165
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
[15] J. Gwizdka, “Whats in the context?”, Proc. CHI 2000 Workshop on the What, Who, Where, When, Why and How
of Context-Awareness, Netherlands, April 1-6,2000.
[16] J. Heidemann, “ Building efficient wireless sensor networks with low-level naming”, Proc. 18th ACM Symposium
on Operating Systems Principles ,Chateau Lake Louise, Banff, Canada, 2001.
[17] J. Hill, R. Szewczyk, A.Woo, S. Hollar, D. Culler, and K.Pister, “System architecture directions for networked
sensors”, Proc. of ASPLOS , pp.93 - 104, Boston, MA, USA, 2000.
[18] W.M.Merrill, K.Sohrabi, ”Open standard Development platforms for distributed sensor networks”, Proc. SPIE,
Unattended Ground Sensor Technologies and Applications Iv (AeroSense), Vol. 4743, pp. 327 - 337,Orlando, FL, April
2002
[19] P.K. Willett, P.F. Swaszek, and R.S. Blum,” The good, bad, and ugly: distributed detection of a known signal in
dependent Gaussian noise”, IEEE Trans. Signal Processing, Vol. no.48, pp. 3266-3279 , Dec. 2000.
[20] H. Li, B.S. Manjunath, and S.K. Mitra, "Multisensor image fusion using the wavelet transforms. Graphical Models
and Image Processing , Vol.57, pp.235-245, 1995.
[21] J.L. Moigne and R.F. Cromp, “ The use of wavelets for remote sensing image registration and fusion”, Technical
Report TR-96-171, NASA , 1996.
[22] N.G. Kingsbury, “The dual-tree complex wavelet transform with improved orthogonality and symmetry
properties”, IEEE International Conference on Image Processing, pp. 375-378, Vancouver, BC, Canada, Sept. 10-13,
2000.
[23] A. V. Sutagundar, S. S. Manvi, S. R. Bharmagoudar,” Wavelet Based Image Indexing and Retrieval”, Proc. IEEE
ICETET, Nagapur,India, 2007.
[24] Sajid Hussain, Abdul w. Martin, “Hierarchical cluster based routing in wireless routing in wireless sensor
Networks”, available at www.cs.virginia.edu/ ipsn06/wip/Hussain-1568986753.pdf.
[25] A. Qayyum, L. Viennot, and A. Laouiti. “Multipoint relaying: An efficient technique for flooding in mobile
wireless networks”, Technical Report 3898, INRIA -Rapport de recherche, 2000.
[26] Yan Yu, R. Govindan and D. Estrin., “Geographical and Energy Aware Routing: A Recursive Data Dissemination
Protocol for Wireless Sensor Networks”, Technical Report UCLA/CSD-TR-01-0023, May 2001
[27] C. Castelfranchi, and E. Lorini, “Cognitive anatomy and functions of expectations”, In Proc. IJCAI03 Workshop
on Cognitive Modeling, Acapulco, Mexico, Aug. 9-11, 2003
[28] M. L. Griss, and G. Pour, “ Development with agent components”, IEEE Computer Magazine, Vol.34, pp.37-43,
2001.
[29] M. S. Kakkasageri, S. S. Manvi, and A. K. Sinha,” Agent based multicast routing in MANETs”, Proc. Thirteenth
National conference on communications, IIT Kanpur, India, pp. 188-192, 2007.
[30] S. S. Manvi, and P. Venkataram,” Applications of agent technology in communications: A review”, Computer
Communications, Vol.27 pp.1493-1508, 2004.
[31] A. V. Sutagundar, S. S. Manvi, M. N. Birje, “ Agent based Location Aware Services in Wireless Mobile
Networks”, Proc. IEEE, AICT, Mauritius, 2007.
[32] T. Magedanz, K.Rothermel and S. Krause, “Intelligent agents: An emerging technology for next generation
telecommunications”, IEEE Proc. Globecom, pp. 464-472, San Francisco, CA, 1996.
[33] V. Oham, and A. Karmouch,” Mobile software agents: An overview”, IEEE Communications Magazine 36, pp.
25-37, 1998.
166
International Journal of Ad hoc, Sensor & Ubiquitous Computing (IJASUC) Vol.2, No.1, March 2011
[34] A. V. Sutagundar, S. S. Manvi, “Agent based Information Fusion in Wireless Sensor Networks”, Proc. IEEE
TENCON, Hydrabad, India, 2008.
[35] M. Sonka, V. Hlavac, R. Boyle, ” Image Processing Analysis, and Machine Vision” , Brooks/Cole Public cashing
company, second edition, pp.15-23, 1999.
[36] Xiaoming Fan, William Shaw, Ivan Lee,” Layered Clustering for Solar Powered Wireless Visual Sensor
Network”, Ninth IEEE International Symposium on Multimedia (ISM 2007), pp.237-244, Taichung, Taiwan, Taichung,
Taiwan, Dec.10-12, 2007.
Authors
A. V. SUTAGUNDAR completed his M. Tech from Visvesvaraya Technological
University, Belgaum, Karnataka. He is pursing his PhD in the area of Cognitive
Agent based Information Management in wireless Networks. Presently he is
serving as an Assistant Professor of Department of Electronics and
Communication Engineering Bagalkot, Karnataka. His areas of interest include
Signal and system, Digital Signal Processing, Digital Image Processing,
Multimedia Networks, Computer communication networks, Wireless networks,
Mobile ad-hoc networks, Agent technology. He has published 16 papers in
referred National/International Conferences and 4 papers in international journals.
Sunilkumar S. Manvi received M.E. degree in Electronics from the University of
Visweshwariah College of Engineering, Bangalore and Ph.D degree in Electrical
Communication Engineering, Indian Institute of Science, Bangalore, India. He is
currently working as a Professor and Head of Department of Electronics and
Communication Engineering, Reva Institute of Technology and Management,
Bangalore, India. He is involved in research of Agent based applications in
Multimedia Communications, Grid computing, Vehicular Ad-hoc networks, E-
commerce and Mobile computing. He has published about 100 papers in national and
international conferences and 40 papers in national and international journals. He has
published 3 books. He is a Fellow IETE (FIETE, India), Fellow IE (FIE, India) and
member ISTE (MISTE, India ), member of IEEE (MIEEE, USA),
He has been listed in Marqui’s Whos Who in the World.
167
|
cs/9905003 | 1 | 9905 | 1999-05-10T17:07:11 | Collective Choice Theory in Collaborative Computing | [
"cs.MA",
"cs.DC"
] | This paper presents some fundamental collective choice theory for information system designers, particularly those working in the field of computer-supported cooperative work. This paper is focused on a presentation of Arrow's Possibility and Impossibility theorems which form the fundamental boundary on the efficacy of collective choice: voting and selection procedures. It restates the conditions that Arrow placed on collective choice functions in more rigorous second-order logic, which could be used as a set of test conditions for implementations, and a useful probabilistic result for analyzing votes on issue pairs. It also describes some simple collective choice functions. There is also some discussion of how enterprises should approach putting their resources under collective control: giving an outline of a superstructure of performative agents to carry out this function and what distributing processing technology would be needed. | cs.MA | cs |
Collective Choice Theory in Collaborative
Computing
Walter D Eaves
June 20, 2018
Abstract
This paper presents some fundamental collective choice theory for
information system designers, particularly those working in the field of
computer -- supported cooperative work. This paper is focused on a pre-
sentation of Arrow's Possibility and Impossibility theorems which form
the fundamental boundary on the efficacy of collective choice: voting and
selection procedures. It restates the conditions that Arrow placed on col-
lective choice functions in more rigorous second -- order logic, which could
be used as a set of test conditions for implementations, and a useful prob-
abilistic result for analyzing votes on issue pairs. It also describes some
simple collective choice functions. There is also some discussion of how
enterprises should approach putting their resources under collective con-
trol: giving an outline of a superstructure of performative agents to carry
out this function and what distributing processing technology would be
needed.
1 Collective Choice in Information Systems
1.1 Naming Services
1. Windows NT
If one uses a system then, from time to time, one might receive an event
message saying that "The Browser Has Forced an Election..." [MSK99a,
MSK99b].
2. The Internet and the Domain Naming Service DNS
Internet connected systems could not function without the Domain Nam-
ing Service and this too relies upon elections: individual system adminis-
trators choose when their name -- server is to authoritative or not [Wel99,
Ricly] and which name -- servers it will rely upon.
The difference between the two systems is that Windows NT is designed to
manage the naming of relatively small domains and can use a direct election
amongst all its naming components, the browsers. This is represented in figure
1, where the master browser is elected by the itself and the other browsers.
The browsers apply the same fitness criteria in choosing their
master. There can be no conflict in policy.
1
elect
Master
Browser
Browser
Browser
elect
elect
Names
Names
Names
Figure 1: Browser programs choosing a master browser
The Internet's DNS has to rely upon a loosely co -- ordinated database of
name servers. This is represented in figure 2. Here the Administrator of each
DNS chooses which other DNS it will use to resolve names. In this example,
for all names other than their own, the Superior Administrator is chosen as
authoritative.
The Administrators need not apply the same fitness criteria in
choosing their superior DNS. There may be conflicts in policy.
They can therefore attempt malevolent actions collectively if they so wish,
for example:
Consider an electronic commerce web site. The user's web browser
makes a secure connection to the site, providing a protected chan-
nel. If the DNS entry for the server's address was replaced by one
indicating an attacker's address, the browser will connect to the ma-
licious site, possibly without the user's knowledge. In this scenario,
the DNS spoofer could monitor the traffic over the "secure" connec-
tion, since the secure connection would actually be to the spoofer,
and forward the transaction data to the real website or process the
traffic itself [Wel99].
Without human intervention, computer programs have wholly predictable
behaviour and cannot possess ulterior motives: people can. (The problems that
can arise from badly managed networks are described to a much greater extent
in [EFL+97].)
2
Administrator
Superior
Administrator
Administrator
elect
elect
own
own
own
DNS
Superior
DNS
DNS
Names
Names
Names
Figure 2: DNS Administrators choosing a superior DNS
3
But even if computer programs have predictable behaviour they
may not apply the same fitness criteria in making choices. This may
lead to conflicts in policy.
1.2 Groups choosing policies
Most people will have come across moderated newsgroups and mailing lists.
Potentially these services could be made self -- managing and would be simple
examples of a computer -- supported working environments.
1. Joining the Group
A policy decision is needed to determine whether or not an individual
should be allowed to join a particular group and take the rights and priv-
ileges enjoyed by its members. A collective choice, probably by a mem-
bership committee, is made based upon the applicant's credentials.
The procedure is usually carried out using a set of recommendations. The
individual requesting the rights fills in a form stating his credentials. Peo-
ple are assigned to check the applicant's trustworthiness, qualifications and
so forth. If the membership committee is satisfied, someone is instructed
to assign the applicant to the group.
2. Expulsion from the Group
Should the membership committee decide to expel an individual from the
group, that, too, would be a collective policy decision.
A situation that could clearly arise is for a number of individuals to infiltrate
a group, subvert it by having themselves elected to the membership committe
and then expelling all the members of the group who are not sympathetic to
the infiltrators. Of course, those individuals might also do this legitimately, if
the selection of the membership committee reflected the views of the current
membership. What legitimates actions is a wider consensus.
1.3 Lattices and Access Control
Organizing the membership of groups is a sub -- process needed for the formation
of lattices of membership classes for an access control system, first put forward
by Denning in [Den76]. The only provably safe access control systems are those
that are based on Mandatory Access Control, MAC, schemes, also as described
by Denning [Den82].
1.3.1 Operating Systems
Discretionary Access Control, DAC, schemes are used in most multi -- user operat-
ing systems such as VAX -- VMS [Cor84] and Unix [Cur90] and resource -- sharing
operating systems such as Windows NT [Jum98]. These are not as discretionary
as one might think:
• Individual users are allocated to groups
• Privileged User(s): the "super -- user" or "Administrator" set group mem-
berships.
4
It could be described as a dictatorial Discretionary Access Control scheme.
The only latitude that individuals have is to be able to grant or deny access
rights to members of their own group or to everyone. The privileged user can
undo any access control operations performed by any individual.
To use some better terminology: subjects are entities who may possess access
rights and objects are those entities to which subjects have rights to use. An
operating system that uses a MAC scheme is represented in figure 3 and one
that uses a DAC scheme is represented in figure 4.
1. MAC Scheme
This is a simple scheme. The administrator classifies all the subjects and
the objects and it classifies some subjects above other subjects so that
higher subjects can access everything that lower subjects can.
Administrator
classify
classify
classify
classify
Subject
access
Objects
access
Subject
access
Objects
Figure 3: Mandatory Access Control Scheme
2. DAC Scheme
5
This is more sophisticated. The administrator classifies all the subjects
and can denote that they belong to certain a Group. Every subject belongs
to the group of Everyone.
Individual subjects own some objects and can choose to grant access to
• Either: their groups (or groups)
• Or: to Everyone
In
Groups do not own anything and neither can the group Everyone.
figure 4, the object A is owned by one of the subjects and has allowed
access to the group its owner belongs to and to Everyone. The object B
can be accessed by the group, but not by Everyone.
There is degree of autonomny granted to the subjects in that they may
classify objects to be accessible to the members of the groups they belong
to, but they may not choose which group, or groups, they belong to.
Neither may they change ownership of an object they own1.
1.3.2 Database Management Systems DBMS
Some DBMSs have more flexible DACs which allow some individuals to be more
privileged than others by granting them the right to grant rights. This feature
is available in some DBMSs that support the Structured Query Language, SQL,
which was based upon System -- R, which is descibed by Denning in [Den89].
In effect, this is the same as the DAC scheme for operating systems, see
figure 4, but the owner can grant access to subjects other than those in its
group and it can grant to others the right to grant access, but not to those to
whom access has been explicitly denied by the owner.
1.4 General Resource Management Systems
Information processing systems can be thought of as general resource manage-
ment systems and the Open Distributed Processing ODP standards, in par-
ticular the prescriptive model, [ISO95], describe how an enterprise modelling
language could be used to state the relationships between resource owners and
users as behavioural contracts given in terms of a set of permissions, prohibi-
tions and obligations. This can be seen as a generalization of the MAC and
DAC schemes, but the ODP standards are only reference models and each in-
formation system should contain some component that embodies its enterprise
model. This is a wholly new superstructure to an information processing sys-
tem: there are some simple class relationships describing the ODP model set
out in appendix A, these rather vaguely state the information model for the
superstructure. Some mechanisms that could be used for applying policy, im-
plementing the administrators, has already been proposed [Eav99b].
Recently, the Java programming environment has provided a powerful lan-
guage for expressing permissions [SUN98]. However, it assumes the existence
of some agent which would negotiate the contracts between resource owner and
user. There have been only a few efforts made to develop arbitrator agents
which could generate such contracts.
1A POSIX [Lew91] requirement, Unix typically does allow ownerships to be changed.
6
Group
Subject
classify
classify
Administrator
classify
Everyone
belongs-to
own
Objects
classify
belongs-to
classify
Objects
access
A
belongs-to
Group
Subject
belongs-to
classify
own
Objects
access
Objects
B
A
belongs-to
Subject
classify
Objects
own
B
A
e
m
e
h
c
S
l
o
r
t
n
o
C
s
s
e
c
c
A
y
r
a
n
o
i
t
e
r
c
s
i
D
:
4
e
r
u
g
i
F
7
There are some system proposals which attempt to apply abstract behavioural
rules in terms of concrete permissions:
in papers by Minsky [Min89, Min95].
Minsky's treatment is for information processing systems in general, but a pa-
per by Rabitti et al. [RBKW91] describes authorization generation mechanisms
which support a lattice model of authorization policy for an object -- oriented
database. The innovation of the system is that it generates authorization policy
as it operates. Authorization is viewed as having three dimensions:
Expression Authorizations specified by users, which are known as explicit and
those that are derived by the system as known as implicit.
Direction An authorization can be positive, stating what may be done, or
negative stating what may not be done.
Strength An authorization may be strong, in which case it may not be over-
ridden, or weak, in which it can.
This model has been extended [BW94] and a recent contribution by Castano
[Cas97] introduces metrics that can be used to generate concrete permissions
from more abstract specifications, including:
• Operation compatibility
• Individuality similarity co -- efficient
• Authorization compatability
• Semantic correspondence
• Clustering of Individuals
Although Bertino et al. attempt to produce mechanical means of generating
authorization policy no -- one would seriously expect a system to be driven wholly
by mechanical recommendation, it would require choices to be made by people
and, if that is the case, then a suitable collective choice procedure must be
found.
1.5 Summary
Information processing systems are not mechanical systems. They represent
the interests of people and the agents that comprise an information processing
system will require policies that represent people's interests which will be used
in formulating behavioural contracts between agents that own resources and the
agents that use them.
2
Issues in Collective Choice Theory
The principal difficulty in collective choice theory is that if a group of people
have to choose between more than two issues, there is no choice procedure they
can adopt that is not open to abuse. By abuse, it is meant:
• Denial of service: some agent can exercise a veto on any policy proposed.
8
• Enforced service: some agent can force a policy upon others.
An abuse takes the form of insincere behaviour: an agent acts not to fulfil his
own interests, but to prevent others from fulfilling theirs.
This simple example, known as the Voting Paradox might help:
Scenario 2.1 (Denial of Service by Policy Cycle). Three agents, x, y and
z, have to choose between three services: A, B and C. x and y rank sincerely, but
z ranks the policies to prevent x and y from reaching a compromise, i.e. policy
A, see table 1.
Agent
x
y
z
Ranking
A > B > C
C > A > B
B > C > A
Table 1: Policy cycle used to deny service
Just to clarify terminology: an election is an expression of collective choice
and the policy chosen by an election is the outcome of what statisticians might
call a trial. In a trial, there are a number of choices available to each voter, or
individual, taking part. A policy is usually chosen with regard to an issue; the
proposal that a policy should be followed regarding an issue is called a motion.
2.1 Two Policy Issues
This could also be described as a a two outcome trials. These arise when there
are only two policies which can result: the choice is to accept a policy or not.
1. Number of Choices is always three
Although there are two policies, there are three choices: one can vote For
or Against. One may also be given the explicit right to Abstain, and, by
doing so, state that one cannot vote for or against. One may also choose
Not to Vote.
2. Abstentions and Not Voting
Usually in referenda, there is no option to abstain and those who do not
vote are considered to have abstained. This assumption is legitimate if one
is sure that all individuals who are entitled to vote have been informed
that they may do so and have made a decision not to. In most business
processes, this would not be the case.
In what follows, it is assumed that all abstentions are explicitly made and
that individuals who do not vote have excluded themselves.
3. Choice Function
Simple majority rule is the usual method for resolving two policy elections,
but it by no means the only one. Two policy collective choice is dealt with
in some detail in appendix B.
9
What is required from a choice function is that it is not open to abuse and
it is decisive. Two policy issues can not be abused because it can be shown
to be the case that the majority have chosen the policy. The only difficulty is
resolving ties.
2.2 Three policy issues
A three -- outcome trial, for example, one of A, B or C must be chosen. The
voting procedures described here are explained at greater length in [Saa94] and
and there is some documentation on voting methods at [VOT98]. What follows
illustrates some of the problems that arise from using them.
1. Number of Choices can vary
(a) Four Choices
One can organize the election so that are four choices: any of A, B
or C and to abstain.
(b) Seven (or Eleven) Choices
One could also allow voters to express a choice between their first
two preferences and to abstain. So a vote might be: (A, B) which
implies that A > B > C.
One can also allow voters to state they are indifferent between their
first two choices, but prefer them to the third; an example of this
kind of vote is: (A = B) > C.
(c) Thirteen (or Seven) Choices
Also one can allow voters to express their choices as a ranking over
all three policies, in two ways:
• Strong ordering no statements of indifference allowed.
• Weak ordering statements of indifference are allowed.
The former allows seven choices of rankings, the latter thirteen.
2. Choice Functions
(a) Simple Majority Rule
This could only be used when the voters are presented with four
choices and clearly would not work:
• Three voters: a tie can result from a policy cycle and can be
used by one voter to deny service, see table 1.
• Seven voters:
if three vote A, two B and two C, then, even
though a majority did not want A, A is chosen.
(b) Single Transferable Vote
This could be used when the voters are presented with seven or eleven
choices. It suffers from the same problem as the next procedure.
(c) Hare Voting System and Borda Preferendum
These two can be used if one presents to the voters thirteen or seven
choices; they, and the Single Transferable Vote procedure, all suffer
from the same fault, [Dor79, Fis73], which is that voting is affected
10
Voter
i
j
k
Policy and Ranking
w
4(3)
4(3)
1(1)
9(7)
y
2(2)
2(2)
4(3)
8(7)
x
3(-)
3(-)
2(-)
8(-)
z
1(1)
1(1)
3(2)
5(4)
Table 2: The Borda "Preferendum"
by irrelevant alternatives. This is best illustrated by using the results
of a Borda Preferendum which has each voter rank their alternatives
in order, see table 2.
If the voters i, j and k are asked to rank the four policies w, x, y and
z, then the order is w > x > y > z, but if asked to choose between
w, y and z then (w = y) > z, but it was clear that w was preferred
over y. The anomaly being that an irrelevant policy, x, serves to
differentiate between relevant ones.
(d) Condorcet Procedure
This procedure is often employed in committees but also suffers from
irrelevant alternatives affecting the selection of a final choice. (There
are some good examples of how a Condorcet procedure can be abused
in [Saa94].) It is simply a series of two policy elections: (A, B), (A, C)
and (B, C). If any policy beats the other two, then it is chosen. If
there is a policy cycle then there is no Condorcet winner. It may also
allow an irrelevant alternative to beat a potential Condorcet winner.
It will be shown that there is no satisfactory choice function for more than
two policy issues. The next note shows that the number of different orderings
for a given number of issues increases dramatically.
2.3 n policy issues
The total number of different weak orderings for n policies can be calculated as
follows:
1. Generate all the partitions[Ski90, p. 56] of n.
2. Calculate the number of permutations for each partition, call this N (partitions).
3. For each partition find the number of number of ways in which the policies
could be allocated to the elements of the partition, N (policies).
4. Multiply N (partitions) by N (policies) for each partition and sum them
together.
ΣpartitionsN (policies) N (partitions)
(1)
A Mathematica[Ins99] package is available[Eav99a] that performs the calcu-
lation. Table 3 lists the total number of different preference orders for up to 6
policies and clearly shows how large the search space becomes.
11
Policies Orderings
1
3
13
75
541
4683
1
2
3
4
5
6
Table 3: Number of Different Preference Orderings for n policies
2.4 Sincere and Sophisticated Voting
Sophisticated voting utilizes some strategy whereby a voter does not vote for
their first choice to ensure that their least -- preferred policy is not chosen. For
example, an electorate of seven votes sincerely for three policies x, y and z thus:
4 x, 3 y and 2 z, then x would be chosen. However, the z voters may prefer y
to x so their sophisticated vote is for y.
An interesting example of a sophisticated vote is global abstention.
If a
motion is formulated which requires a choice between A or B, but all voters
prefer C which is not proposed, a sophisticated response is a global abstention.
Unfortunately, sophisticated voters enjoy an advantage over sincere voters,
but, to do so, they must formulate their own voting policy, which usually re-
quires that they have some information as to the relative strengths of the dif-
ferent coalitions within an electorate and their choice of voting policy would,
presumably, be decided by a sincere vote amongst them. One of the attractions
of presenting an electorate with a complex agenda -- of more than two issues --
is that they are less able to formulate strategies amongst themselves, so that
complex agendas should elicit more sincere voting, but the likelihood of a policy
cycle arising is greater, as is made clear in a later section, §5 and by equation
(2). Sincere voting would allow voters' underlying values to more precisely de-
termined, which one would hope, would in the long -- run be a more stable basis
for decision -- making.
3 Collective Choice Mathematical Model
This is more rigorous presentation of collective choice theory. This following
section introduces the notation that will be used to formalize the conditions
that are placed on collective choice functions.
3.1 Some Notation
Notation 3.1 (Relations, Preferences and Their Ordering: ≻). is a pref-
erence ordering over a set of objects in a finite set X = {x1, . . . , xn} constructed
thus:
R is an instance of a class of binary relations between any two objects. To
state that x1 is related to x2 in some way, one would write: x1 R x2. The
particular relation might be any of the following >, ≥, ≤, =. At this stage,
R is taken to be transitive and connected.
12
≻ is a statement of an individual's preference order or preferences over the
elements of X.
It is a tuple, i.e. a vector, with exactly #X elements,
with each element being of the form xi R xj, where R is instantiated to
one of the values that the class might take. The ordering is assumed to
be consistent for whatever qualities R possesses. There must be at least
one statement of preference for each element of X, even if that statement
is one of indifference.
It is assumed that such a preference ordering is
consistent with the qualities of R. The power set of ≻ is X × X
Notation 3.2 (Policies and Voters). Some simple set definitions are needed.
I is the set of voters or individuals I = {1, . . . , i, . . . , n}.
X a non -- empty set is the universal set of social alternatives, or policies, at least
one of which must be chosen by the voters.
X is a subset of the power set P (X) of X; it is a non -- empty set of non -- empty
subsets of X and describes the potential feasible policy sets of X.
Y is an element of X . It is the set of policies that are presented to an electorate
for them to vote on: the proposal set.
~D is a preference profile of all voters, it will be called a vote, but will contain
more than just the voters' preferences on the elements of the proposal set.
It contains the preference orders of all the individuals in the society for all
alternatives in X. For the n individuals, if individual i is presumed to have
a preference order ≻i, ~D can be written as the n-tuple (≻1, ≻2, . . . , ≻n)
of preference orders on X.
P(cid:16) ~D(cid:17) is the power set of all votes, feasible and infeasible. For a given set of
policies and a given set of individuals only a subset of these votes will
occur.
(Y, ~D) is an ordered pair called the situation. It is the feasible set of policies,
Y , presented to the electorate, and a vote ~D.
The important word is feasible. Only some votes will be feasible given the
preferences held by voters; therefore only some policy sets will be feasible.
Definition 3.1 (Sincere and Sophisticated). The following function defini-
tions clarify how voters make up their minds and form their preference orders.
They are, therefore, purely notional and one or the other is performed by each
individual, i. How a preference order is formed is dependent on whether the
individual votes sincerely or is sophisticated.
If an individual, or population, is voting sincerely, then:
sincere : I × X 7→ X × X
sincere : X 7→ P(cid:16) ~D(cid:17)
e.g. ≻i = sincere(i, X)
~D = sincere(X)
13
If an individual is a sophisticated voter, then a new set of histories is needed:
~D -- and its power set P(cid:16) ~D(cid:17) -- which is the set of all votes that have taken place.
sophisticated: I × X × P(cid:16) ~D(cid:17) 7→ X × X
sophisticated: X 7→ P(cid:16) ~D(cid:17)
e.g. ≻i = sophisticated(i, X, ~D)
~D = sophisticated(X, ~D)
Definition 3.2 (Promotion and Demotion). When stating conditions it is
useful to construct votes from other votes. These may be elements of P(cid:16) ~D(cid:17)
that are infeasible.
These functions promote and demote a policy within a vote.
promote: P(cid:16) ~D(cid:17) × X 7→ P(cid:16) ~D(cid:17) e.g. ~D′ = promote(x, ~D)
And similarly,
demote : P(cid:16) ~D(cid:17) × X 7→ P(cid:16) ~D(cid:17) e.g. ~D′ = demote(x, ~D)
demote() would be implemented as follows:
1. Every preference not involving x is unchanged:
∀x′, y′ ∈ X[(x′ 6= x, y′ 6= x, x′ R y′ ∈≻i, x′ R′ y′ ∈ ≻′
i)
→ (x′ R y′ ↔ x′ R′ y′)]
2. Everything that involves x is unchanged if x if preferred over something
else; or is changed so that x is now preferred over the other policy
∀y′ ∈ X[x > y′ ∈≻i→ x > y′ ∈ ≻′
i]
or
∀y′ ∈ X[x = y′ ∈≻i→ x > y′ ∈ ≻′
i]
Along the same lines, two other variants of promote and demote can be
defined, which promote or demote for a particular voter on a particular policy
and resolve any conflicts.
~D′ = promote(i, x, ~D) and ~D′ = demote(i, x, ~D)
Definition 3.3 (Collective Choice Function). Is a function that maps each
situation to a subset of the feasible subsets for that situation. The collective
choice function F yields the choice set of the proposal set.
F : X × P(cid:16) ~D(cid:17) 7→ X F (Y, ~D) ⊆ Y
Typically the choice set will contain only one policy, the one that is preferred
over all others. The chosen policy can then be removed from Y and the next
policy found.
In the event of a tie -- if ties are tolerated -- the choice set will
contain the tied policies.
14
Remark 3.1 (Quorums). A quorum is usually taken to be the minimum num-
ber of voters that can demand that a policy be a legitimate choice. However, it
may be that case that there is a quorum, but all votes, bar one, are abstentions
and that the choice of that single individual becomes mandatory.
For the time being, this anomaly should be noted, and there will be references
in the text to the validity of a policy decision.
3.2 Conditions on a Collective Choice Function
These conditions prescribe the behaviour of a collective choice function 2. These
are derived from Arrow's work[Arr63] and have been the subject of consid-
erable debate. This rendering is original and, it is hoped, is more explicit,
self -- contained and rigorous than that given in Arrow's work. Each condition is
expressed as a deduction rule in second -- order logic: if the premises are fulfilled
then the conclusion is required i.e. it is expected behaviour. The conditions
therefore constitute tests for an implementation of a collective choice function.
A more succinct rendering can be found in a paper by Batteau et al. [PB81].
It requires some familiarity with Arrow's conditions and the theory of games3,
but has the advantage of relating Arrow's conditions to work in games theory
and also to requirements on the behaviour of collective choice functions. This
latter task has been carried out very successfully by Fishburn[Fis74], but only for
two issue collective choice functions. The paper by Batteau et al. only addresses
collective choice functions that use strong orderings.
There are five conditions in all. There is a brief description of the meaning
of each.
Condition 3.1 (Admissible Orderings). This is a specification that the func-
tion need only operate on what are called admissible orderings, an individual's
ordering is admissible if it alone satisfies the collective choice function, viz.
∀Y ∈ X
∀i∃ ~D[ ~D = (≻i), ∅ 6= F (Y, ~D) ⊂ Y ]
∃ ~D′[ ~D′ = (≻1, . . . , ≻i, . . . , ≻n), ∅ 6= F (Y, ~D) ⊂ Y ]
By specifying that individual voters must present orderings that are proper
subsets of Y , this eliminates orderings that are completely indifferent or are
cyclical, e.g. X = {x, y, z}, ≻i= ( x > y, y > z, z > x) -- so it is a condition on
the vote set as well as on the collective choice function. In practice, it would be
best to ensure that the orderings in ~D are well -- formed.
The choice set F () cannot be empty and it must be a proper subset of Y . If
it is neither of these then there is either global indifference or a policy cycle.
Condition 3.2 (Monotonicity). or "positive association of social and indi-
vidual values"[Arr63, p. 25]: put simply if the individuals want something and
choose it for their society; if, in a later vote, more individuals choose it, then,
2The more common term is social choice function, but this due to its origin in social welfare
economics.
3See for example, [BO82].
15
ceterus paribus, it will be chosen for society again, or, more formally:
∀Y ∈ X
∀ ~D∃S[F (Y, ~D) = S]
∀x ∈ S∃ ~D′[ ~D′ = promote(x, ~D)]
∃S′[S′ = F (Y, ~D′), x ∈ S′]
Condition 3.3 (Independence). or "independence of irrelevant alternatives"
[Arr63, p. 27] requires that the collective choice function return a choice set
regardless of any individual's preferences for policies that are not explicitly part
of the proposal set. This means that individuals may take on or discard values,
or they may change their values regarding other matters, but these changes
should not effect those values that have not changed. Formally, this can be
expressed thus:
∀Y ∈ X
∃X, X ′[Y ⊆ X, Y ⊆ X ′, X 6= X ′]
∀ ~D, ~D′[ ~D = sincere(X), ~D′ = sincere(X ′), F (Y, ~D) = F (Y, ~D′)]
This condition on the implementation of the collective choice function is proba-
bly unimportant in practice; normally, the input to the collective choice function
is ~D which only contains the preferences on the contents of Y , but as can be de-
duced from the discussion of the effect of an irrelevant alternative in the Borda
preferendum, see table 2, these can affect a preference order.
The following two conditions are more contentious. They are different from
the other conditions in that they need not be applicable to all issues and there
are two types of tests one can apply.
Condition 3.4 (Non -- imposition). There is no bias in the collective choice
function that causes it, on some issues, to yield a choice set that is insensitive
to voters' preferences.
The first test is a unilateral test, viz.
∃Y ∈ X
∀ ~D∃S[S = F (Y, ~D)]
∀x ∈ S∃ ~D′[ ~D′ = demote(x, ~D)]
∄S′[S′ = F (Y, ~D′), S′ ∩ S 6= ∅]
That is, it should be possible on a particular set of issues to construct a vote
that does not return a particular policy for all votes that select that policy.
The second test is used in the event of a tie between some policies to ensure
that the collective choice function does not prefer one policy over the other.
∃Y ∈ X [#Y ≥ 3]
∃S[#S ≥ 1, F (Y, ~D) = S]
∀y, z ∈ Y \ S[y 6= z]
∀w∃Y ′[¬(w ∈ Y ), Y ′ = Y ∪ {w}]
∃ ~D′[ ~D′ = sincere(Y ′)]
∄S′[S′ = F (Y ′, ~D′), (y ∈ S′ ∧ ¬(z ∈ S′)) ∨ (z ∈ S′ ∧ ¬(y ∈ S′))]
16
It might not be immediately clear from this formulation but this is an exact
statement of the irrelevant alternative anomaly observed in the Borda preferen-
dum, see table 2.
Condition 3.5 (Non -- dictatorial). There is no one individual whose choice
on some issues is always returned by the collective choice function, a dictator,
nor is there any one individual who can reject some policies, a vetoer4. (Unfor-
tunately, there is some contention about the use of the term "one individual",
see the discussion following.)
1. Unilateral Tests
These test whether it is possible to overcome a dictator's choice or a
vetoer's rejection. The dictator or vetoer is placed in position 1 for con-
venience.
(a) Dictator
∃Y ∈ X
∃S∃ ~D[ ~D = (≻1), S = F (Y, ~D)]
∀x ∈ S∃ ~D′[∀ ≻i [i 6= 1, ≻i= demote(x, ≻i)], ~D′ = (≻1, . . . , ≻n)]
∄S′[S′ = F (Y, ~D′), S′ ∩ S 6= ∅
(b) Vetoer
∃Y ∈ X
∃ ~D∃Sc[ ~D = (≻1), Sc = Y \ F (Y, ~D)]
∀x ∈ Sc∃ ~D′[∀ ≻i [i 6= 1, ≻i= promote(x, ≻i)], ~D′ = (≻1, . . . , ≻n)]
∄S′c[S′c = Y \ F (Y, ~D′), S′c ∩ Sc 6= ∅]
2. Tie -- Breaking Tests
Subtler tests are those that are applied in the event of a tie.
(a) Dictator
The dictator is preferred in some way. In that, if the dictator changes
allegiance, policy changes, but if anyone else it does not.
∃Y ∈ X [#Y ≥ 3]
∃S[S ⊂ Y, #S ≥ 1]
∀i∃ ~D′
∀y[y ∈ Y \ S]
i = promote(i, y, ~D)]
i[ ~D′
∀i∃Si[F (Y, ~D′
i) = Si]
∄j∀i[i 6= j, Si 6= Sj]
4Vetoer is a noun constructed solely for the purposes of this exposition.
17
(b) Vetoer
∃Y ∈ X [#Y ≥ 3]
∃S[S ⊂ Y, #S = 1]
∀i∃ ~D′
∃x[x ∈ S]
i = demote(i, x, ~D)]
i[ ~D′
∀i∃Si[F (Y, ~D′
i) = Si]
∄j∀i[i 6= j, Si 6= Sj]
This definition states that "no one individual" can dictate a vote, which
would seem to suggest that individuals can change their minds, but must do
so en masse. Arrow requires that a deciding set 5 of voters must change their
preferences. How many need to be in that set can only be determined by
analyzing a vote. It may appear that simple majority rule does not appear to
meet this criteria, since, in a close result, any voter can invert the result, but
under simple majority rule on two issues, this can be dismissed, because every
voter has exactly the same capability, therefore the simple majority becomes
the decisive set.
So, as it stands, this condition is still not accurately expressed, it should
state that no subset of voters that is not a deciding set can change the outcome.
The problem is that the deciding set cannot be known until the votes have been
cast and counted.
4 Possibility and Impossibility Theorems
Arrow, who originally addressed the problem of the distribution of social welfare,
developed these theorems as general statements about a class of functions which
seek to combine hierachies of preference relations. Such functions would be of
great use in any field where a joint policy must be formulated. Clearly, that
includes distributed computing and the "globalization" of local access rules to
databases to form lattices of information flow, §1.4 .
There are two results:
1. Possibility theorem for two -- policy elections
Such a collective choice function does exist for elections which have only
a choice between two policies.
2. Impossibility theorem for elections having more than two policies
There is, in general, no such collective choice function for elections having
more than two policies.
The idea behind the proof of the possibility theorem has already been given
in the discussion of deciding sets, condition 3.5, but the same idea is used in the
proof of the impossibility theorem.
5Also called a preventing set in [PB81] or a winning set as defined in [Isb60]. Some useful
rules on winning sets are defined in the latter paper.
18
4.1 Conflict Resolution Mechanisms
Arrow's proof for the impossibility theorem consists of analyzing how a collective
choice function can choose one preference over another. The notation is as used
in §3.1 .
Definition 4.1 (Unanimous Choice). If the voters unanimously agree that
one policy is preferred over all others then that policy is chosen.
∀Y ∈ X
∀i∃x ∈ Y [∀ ~D[ ~D = (≻i)], x ∈ F (Y, ~D)]
∄S∀ ~D′[ ~D′ = (≻1, . . . , ≻i, . . . , ≻n), S = F (Y, ~D′), x ∋ S]
Note that with this formulation it is possible to be indifferent to x, but one
cannot oppose it6. For example, if some voter has ≻= (x = y, y > z, x > z)
then {x, y} = F ({x, y, z}, (≻)).
Definition 4.2 (Biased Choice). If pairs of voters contradict one another
over a policy, one policy is chosen over the other.
∀Y ∈ X
∀x′, y′ ∈ Y ∃i, j ∈ I[i 6= j, x′ > y′ ∈≻i, y′ > x′ ∈≻j]
∀ ~D∄S[ ~D = (≻i, ≻j), S = F (Y, ~D), y′ ∈ S]
Note that this rule can only be applied pair -- wise, it cannot be applied to the
population as a whole.
Definition 4.3 (Unresolved Choice). If pairs of voters contradict one an-
other over a policy, neither policy is chosen.
∀Y ∈ X
∀x′, y′ ∈ Y ∃i, j ∈ I[i 6= j, x′ > y′ ∈≻i, y′ > x′ ∈≻j]
∀ ~D∄S[ ~D = (≻i, ≻j), S = F (Y, ~D), x′, y′ ∈ S]
This, too, may only be applied pair -- wise.
4.2
Inadequacy of Conflict Resolution Mechanisms
Unanimity, with abstentions, does not resolve any conflicts. If one attempts to
do so using one of the other two choice methods, one or more of the conditions
will be breached.
1. Biased Choice
With two voters there is no majority decision, so the collective choice
function must prefer:
(a) Either a policy
6This is, in essence the Pareto principle of social welfare.
It can be stated as: "social
welfare is increased by a change that makes at least one individual better off, without making
anybody else worse off[SN89]." Clearly, if one abstains then one feels one is not going to be
worse off if x is chosen. If one opposes x then one would be worse off if x were chosen.
19
(b) Or a particular voter's choice of policy
(c) Or randomly choose one policy
The first two lead to an imposed policy or indicate a dictatorship, respec-
tively; the latter has been suggested [Zec69], but it is rather arbitrary.
2. Unresolved Choice
This method allows anyone to act as a vetoer.
One might think that one can improve the biased choice method so that
it decides in favour of whichever policy has a simple majority over the other.
Unfortunately, if one admits a third voter to make the biased choice decisive,
then one also allows that third voter to present a third policy choice; in which
case, one is attempting to choose between three policies, which is the problem
one is trying to solve.
If, in an attempt to overcome this, one requires that there always be more
voters than issues, then on some issues at least one voter will be a dictator, or
vetoer7. The dictator, or vetoer, is acting as a "Kingmaker". It is possible to
stand this dilemma on its head (or feet) and use it as the basis for a collective
choice function as in [How89].
4.3 Are the Conditions Reasonable?
Hopefully, it should now be clear that it is not possible to construct a collective
choice function that satisfies all the conditions given above simultaneously. That
said, one can argue that the requirements on the collective choice function's
behaviour are too demanding.
1. Decisiveness
Implicit in the definition of the collective choice function is that it is
decisive.
2. The Non -- imposition and Non -- dictatorial Conditions
These conditions come in two forms. The first form is unilateral and
quite acceptable, although it may even be desirable that one particular
voter has an absolute veto over a policy. The other form of the condition
is only invoked in the event of a tie, where the requirement is that no
policy or voter be preferred over another. This, it has been seen, is the
pivotal difference between simple majority rule for a two policy vote and
a three (or more) policy vote. In the two policy vote in the event of there
only being one vote separating those for and those against, every voter is
equally decisive and will be supported by the majority. In a close three (or
more) policy vote, some voters can choose the minority position to force
a tie.
The issue is, again, the decisiveness of the collective choice function and
would so close a result be acceptable.
7This minority power is often said to be the cause of the instability of proportional repre-
sentation parliaments.
20
3. Independence from Irrelevant Alternatives
The preferendum, see table 2, and Condorcet pairings are both examples of
collective choice functions that are non -- dictatorial but are not free from
the effects of irrelevant alternatives. This is very unfortunate, since an
irrelevant alternative is any policy that is ranked lower than the collective
winner, but is ranked higher than the collective winner by some voters.
This would happen when the voters are assessing the policies with different
underlying values which are more abstract.
If one looks at the rankings made in table 2, it is clear that voter k agrees
with the others that y > z but disagrees with them regarding the merits
of both y and z over both of w and x.
4. Monotonicity
It can also be argued that monotonicity should be sacrificed to achieve
a consensus. There is an attractive suite of collective decision functions
called "Kingmaker Trees"[How89]. These do not demonstrate monotonic
behaviour, but can be used to obtain a decision in the presence of sophis-
ticated voting.
4.4 Are three (or more) policy collective choice functions
usable?
Whether the conditions Arrow imposed on collective choice functions are rea-
sonable only arises when the votes are close and that a sophisticated voter would
vote in such a way that a policy cycle arises. Three (or more) policy votes can
give decisive results and it would be very useful to have them resolve issues: can
one therefore quantify how reliable a collective choice is? How likely is it that
an election has been subverted by sophisticated voters, given the distribution
of votes. Referring again to table 2, if it were possible to compare the two sets
of votes, impartially, an administrator would be able to make a better choice
of final policy. In which case, it would be better to introduce more irrelevant
alternatives to make a more certain choice.
Technologically, this is feasible. There are cryptographic algorithms which
would allow vote sets to be cast secretly [Sch96] and these could then be assessed
by an impartial arbitrator to make the most appropriate choice. The basis for
that choice would be probabilistic and an example of a probabilistic criterion
that could be employed is given next.
5 Max -- Min Probabilities in Condorcet Pairings
There is an interesting paper by Usiskin [Usi64], which quantifies the probabil-
ities for Condorcet pairings. The paper addresses the "Voting Paradox", but
this is slightly misleading, it addresses the organization of votes within commit-
tees. It covers the same ground as the seminal works of Black and Farquharson
[Bla58, Far69], which describe how committee procedures can be abused, if a
policy cycle exists.
In a committee procedure, if there is a policy cycle then for all those voting:
• A > B > C has exactly 1
3 of vote
21
• as does C > A > B
• and B > C > A
If the policies are voted on in pairs, then the order in which they are intro-
duced will determine which is chosen, viz.
• (A vs B) vs C = C ∵ A vs B = A
• (C vs A) vs B = C ∵ C vs A = C
The question that Usiskin resolves is how much more popular than one
another can they be. In the example above, they all have probability of beating,
or of being beaten, of 2
3 .
Denote by Xi a real -- valued random variable that represents the proportion
of a simple majority vote received for policy i. The probability of a simple
majority vote having the outcome that Xi > Xj will be: P (Xi > Xj). A policy
cycle will be revealed if the probabilities for all pairs, P (X1 > X2), P (X2 > X3)
and so on, for n policies is non -- zero and, finally, P (Xn > X1) is also non -- zero.
The maximum minimum value will represent how much more popular one policy
can be over another so that a policy cycle might still result.
If one then has at least one policy that beats another by an amount that is
greater than this, then there can be no policy cycle.
Theorem 5.1 (Arbitrary Random Variables). The maximum minimum value
for the joint probability distribution of a set of n arbitrary random variables is
given by:
max {min [P (X1 > X2), . . . , P (Xn−1 > Xn), P (Xn > X1)]} =
n − 1
n
(2)
This is as one would expect, looking at the example given above, where
each of A, B and C had a probability of 2
3 of beating the other, if one of these
had a greater probability than that, there could be no policy cycle.
In the
example given above, were there were only three voters, this would mean the
voters unanimously agreed on one policy being preferred over at least one other.
This may not be the winning policy, it would allow at least one policy to be
eliminated, then a two -- policy vote can be taken.
This result is rather depressing but one can appreciate its intuitive correct-
ness, because it tells us that the more policies there are, the more difficult it is
to have one policy beating all others.
This case of arbitrary random variables does not help in understanding the
behaviour of sophisticated voters. (The probabilistic events would all be con-
ditioned by at least the previous result, viz. P (Xi > Xi+1Xi > Xi−1) and
probably would need to be conditioned by all events.
However, Usiskin does present a result which could be used to interpret
sincere voting results. The election results P (Xi > Xj) would be based on
Xi, Xj being independent random variables
Theorem 5.2 (Independent Random Variables). The maximum minimum
value for the joint probability distribution of a set of n independent random vari-
ables is given by:
max {min [P (X1 > X2), . . . , P (Xn−1 > Xn), P (Xn > X1)]} = b(n)
where b(n + 1) > b(n) and lim
n→∞
b(n) =
3
4
22
This latter result is quite encouraging, because if there is an election where
4 then no policy cycle can
at least one votes has a probability of greater than 3
exist.
Usiskin also demonstrates a method for formulating the function b(n) and
presents some upper and lower bounds.
6 Some Collective Choice Functions
It was mentioned above, in the discussion of quorums, that a simple majority
rule collective choice would be valid even if only one voter expressed a choice
and all the others abstained. Simple majority rule is just one of a number
of collective choice functions that could be employed.
It is worthwhile just
listing the collective choice functions. These are only for two policy votes and,
because of Arrow's conditions, cannot be extended to three (or more) policy
votes, but they are insightful to the acceptability of collective choices. This
summary follows Fishburn, [Fis74] and the details are contained in appendix B,
but suffice to say that when only two policies, x and y, are under consideration
ternary logic can be used with x > y being 1, y < x being −1 and x = y 0.
There are some diagrams that illustrate the different types of voting rule and
outcomes, figure 10
From the discussion above, §4.1 , a Pareto -- optimal collective choice function
can also be specified, which is not in the appendix, but is discussed in [ARS98]
where it is called "unanimous with abstentions".
Rank Rule
1
2
3
4
Specified Majority
Simple Majority
Specified Majority
Simple Majority
Paretian
Yes
Yes
No
No
Table 4: Ranking of Binary Voting Rules
Simple Majority if s( ~D) > 1 then x > y is the collective choice, see (3).
Non -- minority if 1( ~D) > n/2 then x > y is the collective choice, see (4). This
is a special case of the next type of rule.
Specified Majority if 1( ~D) > αn then x > y, where α is some pre -- defined
constant in range (0, 1), see (5).
Absolute Majority if 1( ~D) > αn then x > y and y > x otherwise, where α
is some pre -- defined constant in range (0, 1), see (5).
Absolute Special Majority if 1( ~D) ≤ αn then y > x, see (6).
Pareto Majority: For if -1( ~D) = 0 and 1( ~D) > 0 then x > y.
Pareto Majority: Indifference if 0( ~D) > 0 and -1( ~D) = 1 = 0 then x = y.
23
Absolute special majority is a variant of absolute majority (it is an absolute
majority of votes against) and absolute majority is just a variant of specified
majority where the complement of the policy is not installed if the required vote
count is not reached and, as noted, non -- minority rule is a variant of Specified
Majority Rule. So all of these can be replaced by that rule.
It is possible to define a Paretian quality which can be added to any voting
rule and requires there is no dissenting vote.
(Paretian indifference can be
thought of as a vote for z , (x = y) as opposed to w =, (x > y) ∨ (y > x), so
it is actually a Paretian vote on a different pair of issues: z and w.) The rules
can be ranked in a qualitative order of difficulty of attaining them, see table 4.
7 Summary
As information systems become more sophisticated they will be used to support
human decision -- making. The prospect of constructing virtual organizations
based on how people interact is attractive: they could potentially be more
responsive -- Miller describes an evolving information processing organization,
[Mil95], which develops its internal structure using a genetic algorithm. There
are already some prototypes, [HK96, HH94], which share information based on
past usage.
This paper acts as a warning that collective decision -- making is not some-
thing to be taken lightly. Even the most sophisticated voting systems can give
rise to erroneous results,
§2.2 . Sophisticated voters making policy choices
could give rise to systems falling into stasis or being subverted to execute the
wrong policies. If a discretionary access control mechanism used to control the
release of information from databases were put under the control of collective
choice functions and determined access rights based on the criteria proposed by
Bertino et al. , §1.4 , it would almost certainly prove to be a vulnerable system.
There is clearly a need for a more sophisticated architecture to deal with
access requests which can only be expressed in an enterprise modelling language
which would have components similar to that described by ISO in their ODP,
[ISO95]. This type of information processing system would need software agents
acting on behalf of individuals to ensure that their information is protected.
This would necessarily be a probabilistic analysis, based on how trustworthy
potential information users are and it may prove expedient to develop systems
that are insured against loss or provide degrees of surety, like those proposed
by Neumann et al. [LMN97].
These information systems and their users would constitute an economy very
much like the everyday commercial world occupied by institutions, corporations
and people -- only it would be faster, less resource wasteful and, if information
system designers integrate the safeguards before they are used, safer.
A ODP Enterprise Entities
These are Booch [Boo94] class diagrams of the relationships that exist between
entities in a distributed processing system, or, indeed, any organization as de-
scribed in [ISO95, Enterprise Modelling Language].
1. Communities, see figure 5
24
Community(cid:13)
1(cid:13)
1(cid:13)
Resource(cid:13)
x(cid:13)
Resource(cid:13)
x(cid:13)
Policy(cid:13)
Figure 5: Communities, policies and resources
Communities comprise of collections of resources and policies. The com-
munity is itself a resource.
2. Enterprise Agents, see figure 6
Performative agents are also resources. There are three kinds of these:
• Administrators
• Arbitrators
• Policy -- Makers
Administrators and arbitrators have policies they follow, but policy -- makers
create policy.
3. Resource Users
These are also agents, but are not performative. They will have their own
policies, but they are not explicitly open to arbitration. Resource users
may contact administrators prior to using a resource or they may not, it
depends on the nature of the resource.
Resource users usually control the policy -- makers within communities and
this is the case with most societies, since legislative assemblies are usually
elected, but this need not be so. Companies are owned by its sharehold-
ers who appoint the board of directors, the administrators, but may not
use the resources the company makes available. Suffice to say, that, in
practice, in most business processes, the resource users have very little
influence on the policy -- makers.
4. Administrators, see figure 7
Administrators control sets of resources. Each consumable resource has a
behaviour and may use other consumable resources. Since there may be
25
Administrator(cid:13)
executive(cid:13)
1(cid:13)
Agent(cid:13)
performative()(cid:13)
x(cid:13)
Policy(cid:13)
x(cid:13)
1(cid:13)
Arbitrator(cid:13)
ju dicial(cid:13)
legislative(cid:13)
Resource(cid:13)
PolicyMaker(cid:13)
make()(cid:13)
Figure 6: Agents (Policy -- maker, Administrator, Arbitrator) and Policy
(cid:0)(cid:0)
Administrator(cid:13)
Resource(cid:13)
consumable(cid:13)
controller(cid:13)
Resource(cid:13)
1(cid:13)
1(cid:13)
u
s
e
r
(cid:13)
x(cid:13)
1(cid:13)
Resource(cid:13)
consumable(cid:13)
Set of Resources(cid:13)
x(cid:13)
x(cid:13)
Behaviour(cid:13)
Figure 7: Administrators, Resources and Behaviour
26
different administrators vieing for the same consumable resources conflicts
may arise.
Resource users must obey the policy for using a resource. If there is no
suitable policy, then the prospective user must have policy made. (This
might seem different from what is observed in most organizations, where
one can ask an administrator to apply policy differently in some way:
usually by asking the administrator's superior to become involved. The
administrator's superior is then acting as a policy -- maker.)
5. Arbitrators, see figure 8
Arbitrator(cid:13)
controller(cid:13)
Administratpr(cid:13)
Set of(cid:13)
Administrators(cid:13)
Figure 8: Arbitrators and Administrators
Arbitrators control adminstrators in that they resolve any conflicts that
arise between them. They control neither administrators nor resources
directly.
6. Policy, see figure 9
Policy is a set of prescriptive statements about the behaviour of a set of
resources: usually one sub -- set of those resources vis `a vis another sub -- set.
There are three kinds of statement:
• Permission what one sub -- set may do with (or for) the other sub -- set.
• Requirement (or Obligation) what one sub -- set must do with (or for)
the other sub -- set.
• Prohibition what one sub -- set must not do with (or for) the other
sub -- set.
B Collective Choice Functions
What follows is drawn mostly from [Fis74] (who refers to proofs from his own
text[Fis73]: and gives a precise definition of the conditions that a collective
27
Statement(cid:13)
Set of statements(cid:13)
1(cid:13)
Requirement(cid:13)
Permission(cid:13)
x(cid:13)
x(cid:13)
Prohibition(cid:13)
Policy(cid:13)
x(cid:13)
1(cid:13)
1(cid:13)
1(cid:13)
Behaviour(cid:13)
x(cid:13)
Statement(cid:13)
1(cid:13)
x(cid:13)
Resource(cid:13)
Set of Resources(cid:13)
Figure 9: Policy, Resources and Behaviour
28
choice function must fulfill for a particular voting procedure for two policy
alternative systems.
Definition B.1 (Principle of Choice). The basic materials for collective choice
functions are social alternatives (candidates, policies, etc.) and individuals (vot-
ers, members, etc.) who have preferences among the alternatives. The idea of
a collective choice function is to map a non -- empty subset of the potential feasi-
ble subset of alternatives to each ordered pair consisting of a potential feasible
subset of alternatives and a schedule of the voters' preferences. The assigned
set is often referred to as the choice set.
How that mapping is achieved is based on the properties of the collective
choice function, which decides whether the choice is:
• Egalitarian
• Weighted
• Representative
• Unbiased (or neutral)
• Decisive
• Unanimous
B.1 Two Policies
If each individual has only two policies to choose from, then the policy chosen
by the population as a whole will always be one of them, so two policy systems
cannot select a set of policies that an individual has not specified.
Definition B.2 (Sets for Two Policies). The sets can be enumerated quite
easily for two policies x, y. If a voter prefers x to y, then 1 else if y to x then
−1 else 0 signifies indifference -- ternary logic.
Policies
X = {x, y}
X = {X}
so write
F ( ~D) for F (X, ~D)
n voters
~D = (D1, D2, . . . , Dn)
where Di ∈ {1, 0, −1}
so {1, 0, −1}n , P(cid:16) ~D(cid:17) and ~D ⊆ {1, 0, −1}n
For the collective choice function over n individuals:
F : {1, 0, −1}n 7→ {1, 0, −1}
Note that the power set of the preferences is written {1, 0, −1}n as shorthand
and is the set of all permutations of vectors of length n where each component
can take one of three values -- #{1, 0, −1}n = 3n. When a condition is applied
to a preference profile, it is either applied with reference to the power set or the
vote set: the power set, although large, is denumerable a priori, the vote set is
not.
29
B.2 Egalitarian
Egalitarian8 collective choice functions treat each voter's vote as identical in
effect to every other's.
Condition B.1 (Strongly Neutral). A collective choice function F is Strongly
Neutral, if, for all ~D ∈ {1, 0, −1}n:
F (− ~D) = −F ( ~D)
(Strongly Neutral)
Condition B.2 (Strongly Monotonic). A collective choice function F is Strongly
Monotonic, if, for any ~D, ~D′ ∈ {1, 0, −1}n:
~D ≥ ~D′ ⇒ F ( ~D) ≥ F ( ~D′)
~D > ~D′, F ( ~D′) = 0 ⇒ F ( ~D) = 1
(Strongly Monotonic)
Condition B.3 (Egalitarian). A collective choice function F is Egalitarian,
if for all ~D ∈ {1, 0, −1}n:
F (D1, . . . , Dn) = F (Dσ(1), . . . , Dσ(n))
if σ is a permutation on {1, . . . , n}
(Egalitarian)
Theorem B.1 (Conditions for Simple Majority Rule[May52]). A collec-
tive choice function F implements simple majority rule over two policies and has
the following qualities: Strongly Neutral, Strongly Monotonic and Egalitarian.
Definition B.3 (Simple Majority Rule). If F applies ternary logic, it can
be implemented with:
F (D) , s( ~D)
(3)
Condition B.4 (Monotonic). A collective choice function is just Monotonic,
rather than Strongly Monotonic if, for any ~D, ~D′ ∈ {1, 0, −1}n:
~D ≥ ~D′ ⇒ F ( ~D) ≥ F ( ~D′)
(Monotonic)
Definition B.4 (Non -- minority Rule). If a collective choice function is Strongly
Neutral, Monotonic and Egalitarian and is implemented thus:
F ( ~D) = 1 ⇔ 1( ~D) > n/2
F ( ~D) = −1 ⇔ -1( ~D) > n/2
(4)
then the voting system is known as non -- minority rule.
8Fishburn in [Fis74] uses the term "anonymous" for egalitarian and "dual" for neutral.
30
Non -- minority rule is just one of a class of neutral, monotonic and egalitarian
collective choice functions; they differ in effect from the strongly monotonic sim-
ple majority rule by having a "dead -- band". A geometric insight into the speci-
fication of a collective choice function can be given using a unit simplex[Saa94].
There is only one dimension.
~q = (
1( ~D)
n
,
-1( ~D)
n
)
The election vector ~q emanates from the origin and will always be within [−1, 1].
Under simple majority rule, whichever point, −1 or 1, the vector is closest to
wins. Under non -- minority rule the vector has to be over half 1
2 way towards
the point, see figure 10. The indecisive region in the centre is symmetric. For
non -- minority rule, any boundary can be chosen, so long as it is symmetric about
the origin.
-1(cid:13)
-1/2(cid:13)
0(cid:13)
1/2(cid:13)
1(cid:13)
Simple Majority(cid:13)
Non-minority(cid:13)
(cid:0)(cid:0)(cid:0)
alpha(cid:13)
Absolute majority(cid:13)
Figure 10: Two policy voting systems
It should be clear from figure 10 that a simple majority voting system can
determine policy if only one voter is not indifferent to either of the policies, the
others abstaining. In this respect, non -- minority rule seems to impose a natural
quorum, since it requires that at least half of the voters have chosen one or
the other policy.
In this respect, non -- minority rule is less questionable as a
decision -- making device than simple majority rule. Most electoral systems do in
fact operate a non -- minority rule system.
Condition B.5 (Neutral). A collective choice function is Neutral, as opposed
to Strongly Neutral, if, for all ~D ∈ {1, 0, −1}n:
1( ~D) 6= -1( ~D) ⇒ F (− ~D) = −F ( ~D)
(Neutral)
31
Condition B.6 (Strongly Decisive). A collective choice function is Strongly
Decisive if, for all ~D ∈ {1, 0, −1}n:
F ( ~D) 6= 0
(Strongly Decisive)
Condition B.7 (Unanimity unambiguous). A collective choice function is
Unanimity unambiguous if:
F (~1) = 1 and F ( ~−1) = −1
(Unanimity unambiguous)
Condition B.8 (Pro -- biased). A collective choice function is Pro -- biased if,
for all ~D ∈ {1, 0, −1}n:
if ~D = ~D′ except that (Di, D′
i) = (0, 1)
for some i then F ( ~D) = F ( ~D′)
(Pro -- biased)
The number of electoral ties can be reduced by downgrading Strongly Neutral
to Neutral and adding Strongly Decisive. If Neutral is dropped then an electoral
preference is given to one policy over the other. The policy that is preferred is
usually already in force and is therefore called the incumbent, the other policy is
the challenging policy. See figure 10 for the asymmetry of the indecisive region
under an absolute majority rule.
Definition B.5 (Absolute Majority Rule). If the collective choice function
is no longer Neutral and is made Strongly Decisive, and the function is imple-
mented thus:
F ( ~D) = 1 ⇔ 1( ~D) > αn
F ( ~D) = −1 ⇔ 1( ~D) ≤ αn
for α ∈ (0, 1)
(5)
then these are the absolute majority rule functions. A special case is unaminous
rule which requires either all votes are −1 or 1.
Definition B.6 (Absolute Special Majority Rule). As for absolute major-
ity rule, but the function is also (Unanimity unambiguous) and (Pro -- biased).
The collective choice function can be implemented by:
F ( ~D) = −1 ⇔ 1( ~D) ≤ αn
for α ∈ (0, 1)
(6)
B.3 Non -- Egalitarian: Weighted
The effect of the Egalitarian property is that one voter's preference can be
exchanged for another within the decision profile and it will have no effect on
the evaluation of the collective choice function. An alternative system is to use a
weighted system, which is often used by some committees, where the chairman
is given both a deliberative and a casting vote. Weighted systems have a number
of attractions because they can be designed so that:
• No one person can dictate policy to the group.
• Ties can be readily resolved without a further vote.
The simplest way to define a weighted voting function is to use a weighting
vector.
32
Weighted Voting
Definition B.7 (Weighting Vector and Vote). A weighting vector can be
defined thus:
~ρhni > ~0hni,~ρ = (ρ1, . . . , ρn)
where ρi ≥ 0, ρi ∈ Z+
0 is the weight assigned to each voter
Then
Weighted Vote ~ρ · ~D
Weight Function W (~ρ) ,
n
X
i=1
ai
Weight of c ∈ {1, 0, −1}Wc( ~D) , W ((~ρ · ~D) : Di = c) =
n
X
i=1
ρiDi for those Di = c
It does not follow that n(~ρ) ≥ n because it is possible to set any number of
ρi equal to zero, but at least one must be non -- zero.
Theorem B.2 (Weighted Majority Function). A function F , F : {1, 0, −1}n 7→
{1, 0, −1}, is a weighted majority function if and only if it satisfies Monotonic,
Unanimity unambiguous and Neutral, the weighted majority function can be de-
fined as F ( ~rho · ~D) = s(~ρ · ~D).
The conditions can be readily tested by substituting F (~ρ · ~D) for F ( ~D).
The other systems can be added thus:
1. Non -- minority rule
F (~ρ · ~D) =
2. Absolute majority
1 W1(~ρ · ~D) > W (~ρ)
−1 W−1(~ρ · ~D) > W (~ρ)
0
Otherwise
2
2
F (~ρ · ~D) =
1 W1(~ρ · ~D) > αW (~ρ)
−1 W−1(~ρ · ~D) ≤ αW (~ρ)
0
Otherwise
Dictators and Vetoers The final choice may be wholly determined by only
one of the voters, in which case that voter is either a dictator or a vetoer.
1. Dictator
Definition B.8 (Dictator). A voter, j, is a Dictator with regard to a
collective choice function F and a weighting vector ~ρ if:
For all ~D ∈ {1, 0, −1}n such that
~D = (D1, . . . , Dj, . . . , Dn)
when Dj 6= 0, F (~ρ · ~D) = Dj
33
(Dictator)
Whether a dictator can effect all decision profiles is a condition on the
behaviour of the collective choice function and the weighting vector, not
on the voters, which is as follows.
Definition B.9 (Undominated by dictator). A collective choice func-
tion is Undominated by dictator if there is no Dictator. This can be stated
thus:
For all ~D ∈ ~D
There is no i such that
~D = (D1, . . . , Di, . . . , Dn),
Di 6= 0 and F ( ~D) = Di
(Undominated by dictator)
This is not particularly useful, since there may legitimately be a voter
whose vote is always in line with the choice of the group as a whole.
2. Vetoer
Definition B.10 (Vetoer). The first voter, 1, is a Vetoer with regard
to a collective choice function F and a weighting vector ~ρ if:
For
~D1 = (0, 1, . . . , 1)
~D−1 = (0, −1, . . . , −1)
(Vetoer)
F (~ρ · ~D1) = 0
F (~ρ · ~D−1) = 0
(The vetoer is in first position for convenience).
Whether a vetoer can effect those two very specific decision profiles is
a condition on the behaviour of the collective choice function and the
weighting vector, not on the voters, which is as follows.
Definition B.11 (Undominated by vetoer). A group of voters are Un-
dominated by vetoer if there is no voter j, such that:
If
Dj = 1, F (~ρ · ~D1) = 1, and
Dj = −1, F (~ρ · ~D−1) = −1
but
Dj 6= 1, F (~ρ · ~D1) = 0
Dj 6= −1, F (~ρ · ~D−1) = 0
For all ~D ∈ ~D.
(Undominated by vetoer)
This, again, is not particularly useful, since there may legitimately be a
voter whose always votes against the group.
34
Sensitivity
Definition B.12 (Essential). A voter i is said to be Essential with regard to
a collective choice function, and weighting vector, if at least one of the following
conditions holds:
Either F (~ρ · ~D1) 6= F (~ρ · ~D0)
Or F (~ρ · ~D−1) 6= F (~ρ · ~D0)
Or F (~ρ · ~D1) 6= F (~ρ · ~D−1)
(Essential)
For at least one of the vectors ~D1, ~D0, ~D−1 constructed as follows:
~D1 = (D1, . . . , Di−1, 1, Di+1, . . . , Dn)
~D0 = (D1, . . . , Di−1, 0, Di+1, . . . , Dn)
~D−1 = (D1, . . . , Di−1, −1, Di+1, . . . , Dn)
where the contents of those vectors can be taken from any of the vectors con-
structed thus:
~Dhn−1i = (D1, . . . , Di−1, Di+1, . . . , Dn) ∈ {1, 0, −1}n−1
This condition requires that a voter can be decisive in at least one decision
profile. It prevents a voter from being given so ineffectual a vote that it is never
decisive in any election.
B.3.1 Vetoer, Dictator and Essential
By this it is meant safe from dictators and vetoers and sensitive to voters; it
is desirable if a collective choice function and weighting vector could be chosen
so that for all decision profiles in {1, 0, −1}n there is no voter who is either a
Dictator or a Vetoer. It would also be desirable there is at least one voter who
is Essential.
The collective choice functions simple majority, non -- minority and absolute
majority are demonstrably safe from dictators and vetoers when used under an
egalitarian regime so only the weighting vector needs to be checked.
Dictators, Vetoers and Weighting Vectors
1. Weighted majority and weighted non -- minority rule
If the collective choice function is either of the above, for ~ρ = (ρ1, . . . , ρn)
and the weighting vector has been scaled so that ρ1 has weight 1.
In
which case, the worst case for the dictator is that all vote against, so for
the dictator to succeed ρdictator > W (~ρ)
Because both of these collective choice functions are dual; it should be
clear ρdictator = ρvetoer.
Consequently, ρmax < W (~ρ)
, is sufficient for both of these.
/ 2.
2
2. Absolute majority
To succeed, ρdictator > α · W (~ρ), so the converse is required.
A vetoer has it easier ρvetoer > (1 − α) · W (~ρ), so the converse.
35
Weighting Vectors that are Essential to Voters Under egalitarian rule,
all voting functions are sensitive to all voters, because if any voter is sensitive,
a permutation can put another voter in his place, so again the collective choice
function will not be at fault should a system of rule prove insensitive, it will be
the weighting vector.
There are two possibilities:
• The voter has a weighting of zero
• The voter has a weighting which can never be decisive
Whether a voter has a non -- zero vote can only be tested for; probably by
using the weighted majority rule function with all other voters not voting.
Having a vote that is never decisive is more subtle. The weighting vector is
as usual, with the lowest rated voter in first position having value 1. Construct
a dictator to each voter in the following manner.
ρ1 = 1
ρ2 = ρ1 + 1
ρ3 = ρ2 + ρ1 + 1
. . .
ρn−1 = ρn−2 + · · · + ρ2 + ρ1 + 1
ρn = ρn−1 + ρn−2 + · · · + ρ2 + ρ1 + 1
ρn = ρn−1 + ρn−1
ρn = 2ρn−1
ρn = 2 · 2 · · n − 1 times · 1
ρn = 2n−1
A necessary conditions on weighting vectors can be set, each ρi < 2n−1 if
W (~ρ) ≥ 2n − 1 -- geometric progression. There are effectively two choices:
• Set n ≥ 3, max ρi = 2n−1 − 1, in which case W (~ρ) ≥ 2n − 2 and all other
voters will tie with the largest voter.
• Set n ≥ 4, max ρi = 2n−1 − 2, in which case W (~ρ) ≥ 2n − 3 and all other
voters will defeat the largest voter.
For ~ρ = (ρ1, . . . , ρi, . . . , ρn)
∄ρi >
1
2
n
X
j=1
ρj
B.4 Representative Systems
A representative system can be thought of as a heirachy of voting councils in
which the outcomes of votes in lower councils become votes in higher councils.
A voter in one of the higher councils may be a voter or a voting council. Each
voting council can use weighted majority rule between its members. Voters (or
councils) can vote more than once in different councils.
36
Definition B.13 (Representation). Let a heirachy of voting councils be de-
fined as a representation R. Let each level of representation be denoted by a
suffix, the lowest level being R0.
To make the lowest level similar in mathematical structure to higher levels,
we shall introduce a selection function Si which selects from a preference profile,
D, the preference of voter i.
Si : {1, 0, −1}n 7→ {1, 0, −1}
Si(D) = Di
R0 can then be written as:
R0 = {S1(D), . . . , Sn(D)}
So R0 is simply the decision profile as a set.
Thereafter, there is a level m ∈ N which is such that:
Rm = {s(F1, . . . , FK)(Rm−1)}
This is effectively voting using a tree structure and is probably the most used
organizational control system. Unfortunately, it is proving to be very difficult
to analyze. Hopefully, more results will arise.
C Funding and Author Details
Research was funded by the Engineering and Physical Sciences Research Council
of the United Kingdom. Thanks to Malcolm Clarke, Russell -- Wynn Jones and
Robert Thurlby.
Walter Eaves
Department of Electrical Engineering,
Brunel University
Uxbridge,
Middlesex UB8 3PH,
United Kingdom
[email protected]
[email protected]
http://www.bigfoot.com/∼Walter.Eaves
http://www.brunel.ac.uk/∼eepgwde
References
[Arr63]
[ARS98]
Kenneth Arrow. Social Choice and Individual Values. Yale Univer-
sity Press, second edition, 1963.
Hajnal Andr´eka, Mark Ryan, and Pierre-Yves Schobbens. Opera-
tors and laws for combining preference relations. World -- Wide Web,
1998. http://www.dmi.ens.fr/∼vaudenay/spw97/.
37
[Bla58]
[BO82]
[Boo94]
[BW94]
[Cas97]
[Cor84]
[Cur90]
[Den76]
[Den82]
[Den89]
Duncan Black. The theory of committees and elections. Cambridge
University Press, 1958.
Tamer Baser and Geert Jan Olsder. Dynamic Noncooperative Game
Theory, volume 160 of Mathematics in Science and Engineering.
Academic Press, London, 1982.
Grady Booch. Object-oriented analysis and design with applications.
Benjamin/Cummings, 1994.
E. Bertino and H. Weigand. An approach to authorization modeling
in object-oriented database systems. Data & Knowledge Engineer-
ing, 12:1 -- 29, 1994.
Silvana Castano. An approach to deriving global authorizations
in federated database systems.
In P Samarati and R S Sandhu,
editors, Database Security X: Status and Prospects, International
Federation for Information Processing. Chapmand and Hall, 1997.
Digital Equipment Corp. VAX/VMS DCL commands and lexi-
cal functions. Technical Report AA-Z003A-TE, Digital Equipment
Corp. (DEC), Maynard, Mass., September 1984.
David A. Curry.
Improving the security of your unix system.
Technical Report ITSTD-721-FR-90-21, SRI International, 1990.
http://julmara.ce.chalmers.se/Security/security-doc.tar.gz.
Dorothy E. Denning. A lattice model of secure information flow.
Communications of the ACM, 19(5):236 -- 243, May 1976. Papers
from the Fifth ACM Symposium on Operating Systems Principles
(Univ. Texas, Austin, Tex., 1975).
Dorothy Elizabeth Robling Denning. Cryptography and data secu-
rity. Addison -- Welsey, 1982.
D E Denning. Secure databases and safety: some unexpected con-
flicts. In T Anderson, editor, Safe and Secure Computing Systems,
chapter 6. Blackwell, Oxford, 1989.
[Dor79]
Gideon Doron.
Is the hare voting scheme representative? The
Journal of Politics, 41(3):918 -- 922, August 1979. In Research Notes.
[Eav99a] Walter Eaves. mypack.m. Mathematica Package: Available by FTP,
January 1999. ftp://ftp.brunel.ac.uk/eepgwde/myPack.m.
[Eav99b] Walter D Eaves. ODP channel objects that provide services trans-
parently for distributing processing systems. Technical report, e --
print archives, April 1999, http://xxx.lanl.gov/abs/cs.DC/9904020.
[EFL+97] R J Ellison, D A Fisher, R C Linger, H F Lipson, T Longstaff,
and N R Mead. Survivable network systems: An emerging dis-
cipline. Technical Report CMU/SEI-97-TR-013, Carnegie Mel-
lon University, Software Engineering Institute, November 1997.
http://www.cert.org/research/97tr013.pdf.
38
[Far69]
Robin Farquharson. Theory of voting. Blackwell, Oxford, 1969.
[Fis73]
[Fis74]
[HH94]
[HK96]
Peter C Fishburn. The Theory of Social Choice. Princeton Univer-
sity Press, 1973.
Peter C Fishburn. Social choice functions. SIAM Review, 16(1):63 --
90, January 1974.
B A Hubermann and T Hogg. Communities of practice: Perfor-
mance and evolution. World-Wide Web, 1994. ftp://parcftp.xerox.
com/pub/dynamics/multiagent.html.
B A Hubermann and M Kaminsky. Beehive: A system for filtering
and sharing information. World-Wide Web, 1996.
ftp://parcftp.
xerox.com/pub/dynamics/multiagent.html.
[How89]
J V Howard. Implementing alternative voting in kingmaker trees.
Technical report, STCRED, London School of Economics, 1989.
[Ins99]
[Isb60]
[ISO95]
[Jum98]
[Lew91]
[LMN97]
[May52]
[Mil95]
Wolfram Research Institute. Mathematica. World Wide Web,
February 1999. http://www.wri.com.
J R Isbell. Homogeneous games - II. Proceedings of the American
Mathematical Society, 11(2):159 -- 161, April 1960.
ISO. Reference model: Architecture (10).
In A Herbert, edi-
tor, Information technology - Open distributed processing, number
903 in Series X Recommendations X.900 to X.1000. International
Telecommunication Union, International Telecommunication Union
(ITU), Place des Nations, CH-1211 Geneva 20, Switzerland, 1995.
http://info.itu.ch/itudoc/itu-t/rec/x/x500up.html.
James G. Jumes, editor. Microsoft Windows NT 4.0 Security, Au-
dit, and Control. Microsoft Technical Reference. Microsoft Press,
December 1998. ISBN 157231818X.
Donald A. Lewine. POSIX programmer's guide: writing portable
UNIX programs with the POSIX.1 standard. O'Reilly & Associates,
Inc., 981 Chestnut Street, Newton, MA 02164, USA, 1991. March
1994 printing with corrections, updates, and December 1991 Ap-
pendix G.
Charlie Lai, Gennady Medvinsky, and B. Clifford Neuman. En-
dorsements, licensing, and insurance for distributed services.
In
Lee W Knight and Joseph P Bailey, editors, Internet Economics,
chapter Internet Commerce. MIT Press, 1997.
Kenneth May. A set of independent necessary and sufficient con-
ditions for simple majority decisions. Econometrica, 20:680 -- 684,
1952.
John H Miller. Evolving information processing organizations. Re-
search note, Carnegie Mellon University, Social and Decision Sci-
ences, Carnegie Mellon University, Pittsburgh, PA 15213, 1995.
http://zia.hss.cmu.edu/miller/.
39
[Min89]
[Min95]
N. H. Minsky. Law-governed software processes. In D. E. Perry, ed-
itor, Proceedings of the 5th International Software Process Work-
shop, pages 98 -- 100, October 1989.
Naftaly Minsky. Coordination and trust in open distributed sys-
tems. Technical report, Rutgers University, NJ, March 1995.
http://www.cs.rutgers.edu/∼minsky/public-papers/trust-paper.ps.
[MSK99a] Event msg: The browser has forced an election... Microsoft Per-
sonal Supprt Center, January 1999. http://support.microsoft.com/
support/kb/articles/q103/0/42.asp, Article ID: Q103042.
[MSK99b]
Information on browser operation. Microsoft Personal Supprt
Center, January 1999.
http://support.microsoft.com/support/kb/
articles/q102/8/78.asp, Article ID: Q102878.
[PB81]
Bernard Monjardet Pierre Batteau, Jean-Marie Blin. Stability of
aggregation procedures, ultrafilters and simple games. Economet-
rica, 49(2):527 -- 534, March 1981.
[RBKW91] Fausto Rabitti, Elisa Bertino, Won Kim, and Darrell Woelk. A
model of authorization for next-generation database systems. ACM
Transactions on Database Systems, 16(1):88 -- 131, March 1991. http:
//www.acm.org/pubs/toc/Abstracts/tods/103144.html.
[Ricly]
Craig Richmond.
Setting up a basic DNS server for a do-
main. World -- Wide Web, 1993 July. http://web.syr.edu/∼jmwobus/
comfaqs/faq-dns.
[Saa94]
Donald Saari. Geometry of Voting. Springer, 1994. BLPES JF1001.
[Sch96]
[Ski90]
Bruce Schneier. Applied Cryptography: Protocols, Algorithms, and
Source Code in C. John Wiley & Sons, Inc, 1996.
Steven S. Skiena.
Implementing Discrete Mathematics: Combi-
natorics and Graph Theory with Mathematica, with programs by
Steven Skiena and Anil Bhansali. Addison-Wesley, Reading, MA,
USA, 1990.
[SN89]
P A Samuelson and W D Nordhaus. Economics. McGraw -- Hil, 13
edition, 1989.
[SUN98]
Java security. World -- Wide Web, 1998.
products/jdk/1.2/docs/guide/security/.
http://java.sun.com/
[Usi64]
Zalman Usiskin. Max-min probabilities in the voting paradox. An-
nals of Mathematical Statistics, 35(2):857 -- 862, June 1964.
[VOT98]
[Wel99]
Alternative systems for calculating winners of multi-option votes.
World -- Wide Web Document Queens University Belfast, Jan 1998.
http://www.qub.ac.uk/mgt/papers/prefer/prefcalc.html.
Brian Wellington. An introduction to Domain Name System se-
curity. World -- Wide Web Document from TISLabs at Network
Associates, January 1999. http://www.nai.com/products/security/
tis research/netsec/dnssec-paper.html.
40
[Zec69]
Richard Zeckhauser. Majority rule with lotteries on alternatives.
The Quarterly Journal of Economics, 83(4):696 -- 703, 1969. Nov.
41
|
1306.3134 | 6 | 1306 | 2016-10-31T21:06:12 | Opinion dynamics and wisdom under out-group discrimination | [
"cs.MA",
"cs.SI",
"nlin.AO"
] | We study a DeGroot-like opinion dynamics model in which agents may oppose other agents. As an underlying motivation, in our setup, agents want to adjust their opinions to match those of the agents of their 'in-group' and, in addition, they want to adjust their opinions to match the 'inverse' of those of the agents of their 'out-group'. Our paradigm can account for persistent disagreement in connected societies as well as bi- and multi-polarization. Outcomes depend upon network structure and the choice of deviation function modeling the mode of opposition between agents. For a particular choice of deviation function, which we call soft opposition, we derive necessary and sufficient conditions for long-run polarization. We also consider social influence (who are the opinion leaders in the network?) as well as the question of wisdom in our naive learning paradigm, finding that wisdom is difficult to attain when there exist sufficiently strong negative relations between agents. | cs.MA | cs |
Opinion dynamics and wisdom under out-group discrimination
Steffen Eger
Computer Science Department
Technische Universitat Darmstadt
Abstract
We study a DeGroot-like opinion dynamics model in which agents may oppose other agents. As an
underlying motivation, in our setup, agents want to adjust their opinions to match those of the agents
of their 'in-group' and, in addition, they want to adjust their opinions to match the 'inverse' of those of
the agents of their 'out-group'. Our paradigm can account for persistent disagreement in connected
societies as well as bi- and multi-polarization. Outcomes depend upon network structure and the
choice of deviation function modeling the mode of opposition between agents. For a particular choice
of deviation function, which we call soft opposition, we derive necessary and sufficient conditions for
long-run polarization. We also consider social influence (who are the opinion leaders in the network?)
as well as the question of wisdom in our naıve learning paradigm, finding that wisdom is difficult to
attain when there exist sufficiently strong negative relations between agents.1
1 Introduction
On many economic, political, social, and religious agendas, disagreement among individuals is pervasive.
For example, the following are or have been highly debated: whether abortion, gay marriage, or death
penalty should be legalized or not; whether Iraq had weapons of mass destructions; the scientific standing
of evolution; whether taxes/social subsidies/unemployment benefits/(lower bounds on) wages should be
increased or decreased; the right course of government in general; the effectiveness of alternative (or
'standard') medicine such as homeopathy.2
In fact, in certain contexts such as the political arena,
disagreement is 'built into' and essential part of the system of opinion exchange (Jones, 1995; Cohen,
2003). Yet, contradicting this factual basis, it has been observed that the phenomenon of disagreement
is not among the predictions of renown and widely used theoretical models of opinion dynamics in the
social and economic context.3 Namely, in these models, a standard prediction is that agents tend toward
a consensus opinion, that is, that all agents eventually hold the same opinion (or belief)4 about any
specific issue. Typically, this applies to both (fully rational) Bayesian frameworks -- which is the reason
why Acemoglu and Ozdaglar (2011) call them "[no] natural framework[s] for understanding persisent
disagreement" (p. 6) -- and non-Bayesian (boundedly rational) setups such as the famous DeGroot
model of opinion dynamics (DeGroot, 1974), where consensus obtains as long as the social network
wherein agents communicate with each other is strongly connected (and aperiodic).
Concerning the non-Bayesian DeGroot model, as we consider in this work, a few amendments have
more recently been suggested which are capable of producing disagreement among agents. In one strand
of literature, models including a homophily mechanism, whereby agents limit their communication to
individuals whose opinions are not too different from their own, can reproduce patterns of opinion diver-
sity and disagreement (Deffuant et al., 2000; Hegselmann and Krause, 2002). In another strand, Daron
Acemoglu and colleagues (cf. Acemoglu and Ozdaglar, 2011) introduce two types of agents, regular and
stubborn, whereby the latter never update their opinions but 'stubbornly' retain their old beliefs, which
1Earlier and more verbose working paper versions of this article can be found at http://arxiv.org/pdf/1306.3134 and
the author's personal website.
2Our examples are, i.a., taken from Abramowitz and Saunders (2005), Acemoglu and Ozdaglar (2011), and Golub and
Jackson (2012).
3See, e.g., the discussions Acemoglu and Ozdaglar (2011); Acemoglu et al. (2013). See also Abelson (1964).
4Typically, in the literature, the term belief is used when there exists a truth for an agenda, and the term opinion is
used when truth is not explicitly modeled. Like our related work, we more generally subsume under the term opinions also
beliefs, judgements, estimations, or even norms and values, depending on the application scenario.
1
may be considered an autarky condition. Multiple stubborn agents with distinct opinions on a certain
agenda may then draw society toward distinct opinion clusters. Such stubborn agents, it is argued, may
appear in the form of opinion leaders, (propaganda) media, or political parties that wish to influence
others without receiving any feedback from them. As a solution to the disagreement problem, however,
both of these model types rely on a problematic 'disconnectedness condition', insofar as disagreement
only obtains when there is no (uni- or bilateral) information flow between certain subclasses of agents.
In this work, we investigate an alternative explanation of disagreement, which can also explain dis-
agreement in connected societies. We consider a non-Bayesian DeGroot-like opinion dynamics model in
which agents are related to each other via two types of links. One link type represents the degree or
intensity of relationship between agents and is given by nonnegative real numbers. The other link type
represents whether agents follow or oppose (deviate from) each other, that is, it represents the kind of
relationship between agents. We assume that group identity causes agents to follow their in-group mem-
bers and to deviate from their out-group members. In-group favoritism and out-group discrimination are
important and well-established notions in social psychology (see, for instance, Tajfel et al., 1971; Brewer,
1979; Castano et al., 2002). They have also more recently been included in economists' models (e.g., in an
experimental context, Charness, Rigotti, and Rustichini, 2007; Ben-Ner et al., 2009; Chen and Li, 2009;
Hargreaves Heap and Zizzo, 2009; Fehrler and Kosfeld, 2013; Tsutsui and Zizzo, 2014). Experimentally,
it has been shown that even minimal group identities, induced by a random labeling of groups, may lead
to intergroup discrimination. When group membership is more salient, Charness, Rigotti, and Rustichini
(2007) show that there is much more cooperation between in-group members than between out-group
members in a prisoner's dilemma game, and Fehrler and Kosfeld (2013) show that individuals associating
with particular NGOs (non-governmental organizations) strongly discriminate against out-group mem-
bers (those that do not associate with an NGO) in a trust game.5 Analogously, we assume that agents
want to coordinate with their in-group members (have negative utility from holding different opinions
than in-group members) and want to anti-coordinate with their out-group members (have negative utility
from not deviating from the opinions that their out-group members hold).6 A special case of our model is
when an agent opposes everyone but himself, i.e., his in-group is himself and his out-group is all 'the rest'.
In some works, such agents have been referred to as rebels or anti-conformists in contrast to conformists
(Jackson, 2009; Cao et al., 2011; Javarone, 2014; Jarman et al., 2015).
Our model closely follows the literature on learning through communication in a given social network
(cf. DeGroot (1974); DeMarzo, Vayanos, and Zwiebel (2003); Golub and Jackson (2010); Acemoglu,
Ozdaglar, and ParandehGheibi (2010); Buechel, Hellmann, and Klossner (2015)). There, the standard
assumption is that agents learn from others in a naıve manner, not properly accounting for the repetition
of opinion signals, which DeMarzo, Vayanos, and Zwiebel (2003) call 'persuasion bias'. A now classical
argument is that if all agents' initial beliefs/opinions were independent and unbiased estimates of the true
value (of a discussion topic), then taking a weighted average of the agents' beliefs in one's social network
(where the weights are proportional to the inverses of the beliefs' variances) is an optimal aggregation
strategy. Then, continuing to average -- in order to incorporate more remote information, e.g., from
friends of friends -- in the same manner is a boundedly rational heuristic that treats the evolving
information signals as novel, not accounting for their cross-contamination. Such a heuristic aggregation
of opinion signals appears quite plausible given the processing costs involved in exact inference in this
setup (cf. Golub and Jackson, 2010). Also, recent experimental evaluations find that the naıve DeGroot
model is a much better approximation of information aggregation in network interactions than 'fully
rational' Bayesian approaches and that individuals are indeed affected by persuasion bias (Corazzini et
al., 2012). In our model, we posit that agents are subject to the same biases involving processing of cross-
contaminated information and are, in addition, susceptible to in-group bias, attempting to coordinate
5In a 'field' setting, the in-group/out-group distinction may prominently be seen as arising, e.g., in a (main stream)
culture/counterculture (e.g., hippies, punks, etc.) dichotomy (Yinger, 1977) or in classical party divisions (e.g., Republicans
vs. Democrats) in the field of politics, etc.
6Out-group discrimination (opposition) is also closely related to what has been termed rejection of beliefs, actions,
and values of dissimilar/disliked others. According to this concept, agents change their normative systems to become
more dissimilar to interaction partners they dislike (cf. Abelson, 1964; Kitts, 2006; Tsuji, 2002; cf. also Groeber, Lorenz,
and Schweitzer, 2013) insofar as disliked others may serve as 'negative referents' who inspire contrary behavior. While in
controlled experiments Tak´acs, Flache, and Mas (2014) do not find strong evidence for the tenet that individuals disassociate
from the opinions of a disliked source, their study explicitly excludes a group identity structure. Moreover, as the authors
argue, their laboratory experiment may have "suppressed the emotional processes that in field settings induce disliking and
rejection of others' opinions."
2
with in-group members and to anti-coordinate with out-group members.7
This work is structured as follows. Section 2 presents the model and gives two introductory exam-
ples. Sections 3 and 4 present our main results, on persistent disagreement (Theorem 3.1) and bi- and
multi-polarization (Proposition 3.1). For a special case of our model, we derive necessary and sufficient
conditions for long-run polarization (Theorem 4.1) as well as opinion leadership (Theorem 4.2) as further
main results. In Section 5, we summarize and conclude with a discussion on wisdom. To make this work
more or less self-contained, we provide concepts, e.g., from graph and matrix theory in the appendix, to
which we also relegate all our proofs.
2 Model
Let S be a finite set ('discrete model') or a subset of the real numbers ('continuous model'), which we
refer to as opinion spectrum.8 Let n ≥ 1 and let [n] = {1, . . . , n} be a set of n agents. Consider the
normal form game ([n], S1 × · · · × Sn, U ), where
• [n] is the set of players,
• S = S1 = · · · = Sn is the action set of each player,
• and U = (u1, . . . , un), where ui : Sn → R is the payoff/utility function of player i ∈ [n].
Let F be the identity function on S -- that is, F(x) = x for all x ∈ S -- and let D be a function D : S → S
that is not the identity function, and which we term deviation function (in the most general form of our
model, we let deviation functions D depend on particular agents i and j involved, that is, we subscript
D as in Di or Dij ). We assume that agents are connected via a (social) network W, where Wij ≥ 0
denotes the strength of relationship between agents i and j.9 More precisely, Wij signals the degree of
importance of agent j for agent i, and we do not require W to be symmetric, that is, Wij and Wji may
j=1 Wij = 1, for all i ∈ [n].10 Assume further that player i has payoff
differ. Assume that Wii = 0 and Pn
for action profile (b1, . . . , bn)
ui(b1, . . . , bn) = − Xj∈In(i)
Wij (bi − F(bj))2 − Xj∈Out(i)
Wij (bi − D(bj))2
(2.1)
for the continuous model. Here, In(i) ⊆ [n] is the in-group (set of friends) of player i and Out(i) ⊆ [n] is
the out-group (set of enemies) of player i. For the discrete model, assume that the analogous payoff is
ui(b1, . . . , bn) = − Xj:j∈In(i),bi6=F(bj )
Wij −
Xj:j∈Out(i),bi6=D(bj )
Wij .
(2.2)
Utility functions ui in Eq. (2.1) and Eq. (2.2) say that player i has disutility from choosing a different
action than his friends and has disutility from choosing a different action than the 'opposite' action of
his enemies, where D specifies what the opposite of an action is.
When each agent repeatedly plays a best response to the actions -- which in our setup are opinions
-- of the other players, i.e., i ∈ [n] chooses action bi that maximizes ui(·), then, in the continuous model,
opinions evolve over time according to the following weighted average of (possibly, via D, 'inverted') past
opinions:
bi(t + 1) = Xj∈In(i)
Wij bj(t) + Xj∈Out(i)
Wij D(bj(t)) =
n
Xj=1
Wij Fij (bj(t)),
(2.3)
7If one wanted to construct an argument that closely follows that of DeMarzo, Vayanos, and Zwiebel (2003), one might
posit that, in our model, agents 'correct' their out-group members' opinion signals -- possibly because of distrust -- before
averaging.
8For the continous DeGroot model as we discuss, S is typically modeled as a convex subset of the real numbers, that is,
Pj αj xj ∈ S for all finite numbers of elements xj ∈ S and all weights αj ∈ [0, 1] such that Pj αj = 1. For convenience, we
think of S as the whole of R or of some (closed and bounded) interval [α, β] for α ≤ β.
9Throughout, we denote the entries of a vector u as ui or [u]i and analogously for matrices.
10The assumption Wii = 0 can be relaxed, see Groeber, Lorenz, and Schweitzer (2013). In subsequent sections, we do not
always assume that Wii = 0.
3
for t = 0, 1, 2, . . ., starting from some particular initial actions bi(0). Here, Fij ∈ {F, D}, depending
on whether j is in i's in-group or out-group, respectively. For the discrete model, the analogous best
response action is weighted majority voting of agents' (possibly inverted) past opinions:
bi(t + 1) = arg max
,
(2.4)
n
Xj=1
s∈S
Wij 1(cid:0)Fij (bj(t)), s(cid:1)
where 1(r, t) = 1 if r = t and zero otherwise. Casting the updating processes (2.3) and (2.4) in a more
compact 'operator' notation, we write (F being the n × n 'matrix' with entries Fij )
b(t + 1) = (W ⊙ F)(b(t)).
(2.5)
Here, we let the 'operator' W ⊙ F act on a vector b ∈ Sn in the manner prescribed in (2.3) and (2.4),
j=1 Wij · Fij (bj) and analogously for the discrete model.11 Equation (2.5) may
i.e., (cid:2)(W ⊙ F)(b)(cid:3)i
again be rewritten as
def= Pn
b(t) = (W ⊙ F)t(b(0)),
(2.6)
by which we denote the t-fold application of operator W ⊙ F on b(0), that is, f t(b) = f (· · · f (f (b))),
where f = W ⊙ F. In the sequel, we refer to W ⊙ F as 'operator' or 'social network'.
Remark 2.1. In case F is the n × n matrix of identity functions, updating process (2.6) collapses to the
standard DeGroot learning model where (W ⊙ F)t is simply the t-th matrix power of matrix W. From
an alternative (equivalent) viewpoint, our model generalizes the standard DeGroot model insofar as the
latter posits that Out(i) = ∅ for all i ∈ [n].
We note that since the operator W ⊙ F in opinion updating process (2.5) retrieves best responses
of agents to an opinion profile b(t), under utility functions ui(·) as in (2.1) or (2.2), the fixed-points of
W ⊙ F -- that is, the points b such that (W ⊙ F)(b) = b -- are the Nash equilibria of the normal form
games ([n], Sn, U (·)). Namely, for each such a fixed-point, all players in [n] play best responses to the
other players' actions (opinions).
We now illustrate our model with two examples, outlining its relationship to other models discussed
in the literature and hinting at its potential for long-run disagreement.
Example 2.1. Let S = {0, 1} be a binary opinion space. Assume that
D(x) =(1
0
if x = 0,
if x = 1.
Let Wij = 1
N (i) , where N (i) denotes the set of neighbors of agent i in the social networks, i.e., the set
of agents j for which Wij > 0. When Fij = F for all j ∈ [n], and agent i updates opinions according
to (2.4), then, at each time step, agent i chooses the majority opinion among his neighbors' opinions.
Such individuals have also been called 'conformists' in some contexts; e.g., Jackson (2009); Cao et al.
(2011); Javarone (2014); Jarman et al. (2015). Conversely, when Fij = D for all j ∈ [n], and agent i
updates opinions according to (2.4), then, at each time step, agent i chooses the minority opinion (=
majority of inverted opinions) among his neighbors' opinions. Such individuals have also been called
'anti-conformists' or 'rebels'. When weights are non-uniform and/or agents follow some of their peers
while deviating from others, then the current setup may yield interesting generalizations of the basic
conformist/non-conformist model.12
11For short, we will usually write (W ⊙ F)b instead of (W ⊙ F)(b).
12 It is also worthy to note that binary opinion spaces (in a conformist/anti-conformist setup) are closely related to games
on networks (cf. Jackson, 2009 and references therein) with binary action spaces. In a society with only anti-conformists
on simple graphs (undirected graphs with no self-loops), maximally independent sets S ⊆ [n] of agents for which it holds
that Wi,S = Pj∈S Wij > 1/2 for all i ∈ [n]\S form pure strategy Nash equilibria of the underlying games in the sense that
assigning one action/opinion to all agents in S and the complementary action/opinion to all agents in [n]\S constitutes a
setting where each agent plays a best response to the actions/opinions of the others. In a society with both conformists
and anti-conformists, (pure strategy) Nash equilibria exist on networks in which each agent assigns weight mass > 1/2 to
conformists: an equilibrium is where all the conformists take one action (hold one opinion), and all the anti-conformists the
other (cf. Jackson, 2009, p.272).
4
Example 2.2. In this example, we let S = [−1, 1] and think of the opinion x = −1 as extreme left-wing
opinion, x = +1 as extreme right-wing opinion and of opinions −1 < x < 1 as more moderate opinions
(x = 0 as 'center' opinion). Assume there are six individuals, organized in four groups A, B, C, D,
members of each of which follow members of their own group and deviate from the members of the other
groups. Hence, let [n] = {1, 2, 3, 4, 5, 6}, A = {1, 2}, B = {3}, C = {4, 5}, D = {6}, and
Out(1) = Out(2) = {3, 4, 5, 6}, Out(3) = {1, 2, 4, 5, 6},
Out(4) = Out(5) = {1, 2, 3, 6}, Out(6) = {1, 2, 3, 4, 5}.
Assume the following specification of deviation functions for members in each group of agents:
DA(x) = 1, DB(x) = −1, DC (x) =
x
2
,
(2.7)
DD,AB(x) = −x, DD,C (x) = sgn(x)px,
for all x ∈ [−1, 1]. Put differently, agents in group A ignore the actual opinion signals of members of their
out-groups, simply interpreting any uttered opinion of an out-group individual as evidence of the opinion
1. Similarly, agents in group B interpret any opinion signal uttered by an out-group agent as evidence
for opinion −1. Agents in group C are moderate in that they 'discount' (extreme) viewpoints that their
out-group members hold. Finally, agents in group D more literally invert the opinions of members of
(their out-)groups A and B -- possibly knowing of these agents' predispositions for extreme opinions of
particular kinds. Moreover, they map the opinions of members of (their out-)group C to more extreme
opinions for any value of x between −1 and 1 -- possibly knowing of these agents' biases toward moderate
opinions. According to this specification, matrix F looks as follows:
F =
F
F
DB
DC
DC
F
F
DB
DC
DC
DA
DA
F
DC
DC
DA
DA
DB
F
F
DD,AB DD,AB DD,AB DD,C DD,C
DA DA
DA DA
DB DB
DC
F
F
DC
F
.
In Figure 1, we plot sample opinion dynamics when agents start with the initial consensus b(0) =
(µ, . . . , µ)⊺, where µ = 1/4, and for an arbitrarily selected positive row-stochastic matrix W ∈ R6×6.
Note how, in this case, agents' opinions polarize into extreme and moderate viewpoints and note how the
opinion dynamics process (apparently) converges and stabilizes as time progresses. In the same figure,
we also sketch the deviation functions defined in Equation (2.7).
3 Persistent disagreement, bi- and multipolarization
In this section, we consider our extended DeGroot model in the abstract situation of arbitrary deviation
functions D. Our results in this context concern the possible consensus opinions that agents may hold
in our model in the long-run (Theorem 3.1). We will find that the long-run consensus vectors can
be determined quite simply: under appropriate weight conditions, a certain consensus opinion vector
(c, . . . , c) is an equilibrium if and only if c is a (D-)neutral opinion with respect to each agent's deviation
function D, i.e., D maps the opinion c to itself. This is our persistent disagreement result: as long as there
exist (sufficiently strong) out-group relations between agents, society will disagree forever if there is no
opinion which is neutral for each agent, no matter the agents' initial opinions. After investigating long-run
persistent disagreement, we consider a particular form of disagreement, namely, bi- and multi-polarization
(for abstract and arbitrary deviation functions D).
To begin with our formal analysis, we define a few concepts. Let C be the set of consensus opinion
vectors in Sn, i.e.,
C = {(a1, . . . , an) ∈ Sn a1 = · · · = an}.
Let Y be an arbitrary set and let Q be an arbitrary function Q : Y → Y . By Fix(Q), we denote the set
of fixed-points of Q, that is, the set of all x ∈ Y such that Q(x) = x. For a deviation function D, we call
5
1
0.5
0
-0.5
-1
-1
1
x/2
−1
−x
sgn(x)x0.5
x
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
-0.5
0
0.5
1
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
Agent 6
0
2
4
6
8 10 12 14 16 18 20
t = 0, 1, 2, . . .
Figure 1: Left: Deviation functions DA, DB, DC , and DD,AB, DD,C specified in the Equation (2.7), as
well as identity function F, on S = [−1, 1]. Right: Opinion dynamics for the society described in the
example, starting from the initial consensus (1/4, . . . , 1/4) at time 0. Agents 1 and 2 belong to group A,
agent 3 makes up group B, agents 4 and 5 belong to group C, and agent 6 makes up group D. Weights
Wij set to arbitrary positive values.
an opinion x ∈ S for which D(x) = x (D-)neutral. Let A ⊆ [n] be an arbitrary subset of the set of agents
and let i ∈ [n] be a particular agent. We denote by Wi,A :=Pj∈A Wij the total weight mass i assigns to
group A.
Theorem 3.1. Let W ⊙ F be an arbitrary social network such that Fij ∈ {F, Di}, for all i, j ∈ [n].
Assume that either W ⊙ F refers to the discrete model or that each function Fij in F is continuous. Let
A = {i ∈ [n] Wi,Out(i) > C0} denote the set of agents whose weight mass assigned to out-group members
exceeds a particular threshold C0; in the discrete case, C0 = 1
2 , and in the continuous case, C0 = 0. Then:
P1[Lim(W ⊙ F) ∩ C] = \i∈A
Fix(Di),
where Lim(W ⊙ F) = {b ∈ Sn b = limt→∞(W ⊙ F)tb(0), for some b(0) ∈ Sn} denotes the set of
opinion profiles b ∈ Sn that may result in the limit of opinion updating process (2.6). Moreover, P1
projects consensus vectors (c, . . . , c) ∈ Sn on their first coordinate c ∈ S. If A is the empty set, we let
Remark 3.1. One of the implications of Theorem 3.1 is persistent disagreement, under the assump-
tions of the theorem and further rather mild conditions, as a prediction of our generalized DeGroot
Ti∈A Fix(Di) = S.
updating process (2.6). Namely, in particular, the relation Ti∈A Fix(Di) = ∅ follows when, for instance:
• One agent's deviation function is radical : Fix(Di) = ∅.
• Two agents' assessment of what constitutes a neutral opinion differs: Fix(Di) ∩ Fix(Dj) = ∅.
More generally, persistent disagreement follows whenever the agents in society have no common
interpretation of neutrality: there exists no opinion c ∈ S such that c is Di-neutral for all agents i.
Moreover, concerning deviation functions D, our only assumption was that there are points which they do
not fix. Modeling deviation, however, a plausible (stronger) restriction on D is that D(x) 6= x for many,
most, or all x ∈ S. Thus, potential long-run agreement, that is, Ti∈A Fix(Di) 6= ∅, would be particularly
difficult to obtain, because this consists of a condition that is unlikely on the level of individual agents,
and in addition contains a cross-agent constraint that deviation functions would have to fix the same
point(s) x ∈ S for all agents.
6
Example 3.1. We apply Theorem 3.1 to the examples discussed previously. In Example 2.1, Fix(D) = ∅,
so in a conformist/anti-conformist society, consensus cannot ensue if at least one agent opposes more than
half of his social network.
In Example 2.2, Fix(DA) = {1}, Fix(DB) = {−1}, Fix(DC ) = Fix(DD,AB) = {0}, and Fix(DD,C ) =
{0, 1, −1}. While Theorem 3.1 does not directly apply to Example 2.2, since agent 6's deviation functions
vary across out-group agents, we can nonetheless apply the theorem to the society consisting of agents
1 through 5 and conclude that reaching a consensus is not possible for this subsociety since, e.g., ∅ =
{1} ∩ {−1} (and assuming that respective weights satisfy the positivity assumption outlined in the
theorem). Hence, since agents 1 to 5 cannot reach a consensus, then also agents 1 to 6 -- the overall
society in the example -- cannot reach a consensus.
Polarization
We now investigate (bi-)polarization as an outcome of our opinion updating dynamics. We call an
opinion vector p ∈ Sn a (bi-)polarization if p consists of two elements a, b ∈ S exclusively, that is, if
[p]i ∈ {a, b} for all i = 1, . . . , n. Note that according to our definition, a consensus vector is a special case
of a polarization. We first define network structures that are sufficient for inducing polarization opinion
vectors.
Definition 3.1 (Opposition bipartite operator W ⊙ F). We call the operator W ⊙ F opposition bipartite
if there exists a partition (N1, N2) of the set of agents [n] into two disjoint subsets -- [n] = N1 ∪ N2,
with N1 ∩ N2 = ∅ -- such that agents in Nk follow each other, for k = 1, 2, while for all agents i ∈ Nk,
j ∈ N−k, for k = 1, 2, it holds that i deviates from j. More precisely, we require
∀ i, i′ ∈ Nk(cid:16)Wii′ > 0 =⇒ i′ ∈ In(i)(cid:17),
∀ i ∈ Nk, j ∈ N−k(cid:16)Wij > 0 =⇒ j ∈ Out(i)(cid:17),
for k = 1, 2,
for k = 1, 2.
Remark 3.2. Note that our above definition of opposition bipartiteness is equivalent to the condition
that (1) no two agents within each subsociety are enemies of each other and that (2) no two agents across
the two subsocieties are friends of each other.
Also note that we do not necessarily require N1 and N2 to be non-empty. Therefore, the standard
DeGroot model (with exclusively friendship relationships) constitutes a special case of an opposition
bipartite network in which N1 = [n] and N2 = ∅.
Remark 3.3. What we call 'opposition bipartite' operator -- or at least a special case of our concept
-- has also been called 'balanced signed network' in the literature (cf. Beasley and Kleinberg, 2010).
Definition 3.2 (Reverse opposition bipartite operator W ⊙ F). We call the operator W ⊙ F reverse
opposition bipartite if there exists a partition (N1, N2) of the set of agents [n] into two disjoint subsets
such that agents in Nk deviate from each other, for k = 1, 2, while for all agents i ∈ Nk, j ∈ N−k, for
k = 1, 2, it holds that i follows j. More precisely, we require
∀ i, i′ ∈ Nk(cid:16)Wii′ > 0 =⇒ i′ ∈ Out(i)(cid:17),
∀ i ∈ Nk, j ∈ N−k(cid:16)Wij > 0 =⇒ j ∈ In(i)(cid:17),
for k = 1, 2,
for k = 1, 2.
An example of an opposition bipartite operator is given in Example 3.2 below. An example of a
reverse opposition bipartite operator is given in Example 3.3 below. A schematic illustration of both
concepts is given in Figure 2.
Clearly, opposition bipartite networks have polarization opinion vectors as equilibria, when agents in
the same group hold one opinion and agents in the alternative group hold the 'opposite' opinion, as we
illustrate below.
Example 3.2. Let W be arbitrary row-stochastic. Consider
F =
F F D D
F F D D
D D F F
D D F F
.
7
2
1
3
4
6
5
7
2
1
3
4
6
5
7
Figure 2: Schematic illustration of the concepts of opposition bipartite (left) and reverse opposition
bipartite operators (right). We omit network links referring to weights W for clarity and we also draw
links as undirected for the same reason. We omit links Fij where Wij = 0. Red links denote opposition
(D), green links following (F).
Clearly, W ⊙ F is opposition bipartite; for example, take N1 = {1, 2} and N2 = {3, 4}. Moreover,
let S = {"impossible","unlikely","possible","likely","certain"} and let, e.g., "unlikely" and "likely" be
D-opposing viewpoints, i.e., D(a) = b and D(b) = a for a ="unlikely" and b ="likely". Then
p =
is a polarization fixed-point of W ⊙ F.
"unlikely"
"unlikely"
"likely"
"likely"
Reverse opposition bipartite networks may induce oscillating, or fluctuating, opinion updates (cf.
Kramer, 1971), very similar to ordinary periodic networks.
Example 3.3. Let W be arbitrary row-stochastic. Consider
F =
D D F F
D D F F
F F D D
F F D D
.
Clearly, W ⊙ F is reverse opposition bipartite; for example, take N1 = {1, 2} and N2 = {3, 4}. For p and
D as in Example 3.2, we have
"unlikely"
"unlikely"
"likely"
"likely"
7→W⊙F
"likely"
"likely"
"unlikely"
"unlikely"
7→W⊙F
"unlikely"
"unlikely"
"likely"
"likely"
7→W⊙F . . .
Also, note that in this example self-weights Wii may be zero for all agents i, so that agents do not
necessarily have to deviate from their own opinions in order for reverse opposition bipartiteness to be
satisfied.13
Next, we turn from bi-polarization to multi-polarization in which an opinion vector p ∈ Sn consists
of K distinct opinions s1, . . . , sK. We generalize our notion of opposition bipartite networks.
Definition 3.3 (Opposition multi-partite operator W ⊙ F). We call the operator W ⊙ F opposition
K-partite if there exists a partition (N1, . . . , NK), for K ≥ 2, of the set of agents [n] into disjoint subsets
-- [n] = N1 ∪ · · · ∪ NK, with Nk ∩ Nℓ = ∅ for all k 6= ℓ -- such that agents in Nk follow each other,
for k = 1, . . . , K, while for all agents i ∈ Nk, j ∈ Nℓ with k 6= ℓ, it holds that i deviates from j. More
13In fact, in reverse opposition bipartite networks we have either Wii = 0 or we have Wii > 0 and Fii = D, while in
opposition bipartite networks we have either Wii = 0 or we have Wii > 0 and Fii = F.
8
precisely, we require
∀ i, i′ ∈ Nk(cid:16)Wii′ > 0 =⇒ i′ ∈ In(i)(cid:17),
∀ i ∈ Nk, j ∈ Nℓ(cid:16)Wij > 0 =⇒ j ∈ Out(i)(cid:17),
for k = 1, . . . , K,
for k 6= ℓ.
We also call an opposition K-partite operator opposition multi-partite.
Opposition multi-partite networks admit multi-polarizations as outcomes, as a simple generalization
of Example 3.2.
Proposition 3.1. Let W ⊙ F be an opposition K-partite, for K ≥ 2, social network. Moreover, let F
be such that Fij ∈ {F, D1, . . . , DK} for deviation functions D1, . . . , DK that precisely correspond to the
K groups society is made up of (that is, agents in group k, for 1 ≤ k ≤ K, apply deviation function Dk).
Let s1, . . . , sK ∈ S be such that
Dk(sℓ) = sk,
∀ k 6= ℓ.
Then, there exists a multi-polarization opinion vector p consisting of opinions s1, . . . , sK, that is, [p]i ∈
{s1, . . . , sK}, such that (W ⊙ F)p = p.
Example 3.4. For a concrete example, let S = {L, M, R} and consider three different groups with
distinct deviation functions D1(x) = L, D2(x) = M , D3(x) = R for all x ∈ {L, M, R}. Group 1 may
be thought of as always deviating to a left wing opinion, provided that its members deviate from certain
agents; group 2 to a moderate position in the opinion spectrum; and group 3 to a right wing position.
Let, e.g., n = 6, W be arbitrary row-stochastic, and let
F =
F
F
F
F
F
F
D2 D2 D2
D3 D3 D3 D3
D3 D3 D3 D3
F D1 D1 D1
F D1 D1 D1
F D1 D1 D1
F D2 D2
F
F
F
F
.
Clearly, W⊙F is opposition 3-partite (see also Figure 3). It is easy to check that, e.g., p = (L, L, L, M, R, R)
is a fixed-point of W ⊙ F, in accordance with the proposition.
1
2
3
4
5
6
Figure 3: Graphical illustration of Example 3.4. Groups have individualized deviation functions, in
different colors. We omit many links for clarity.
We omit the introduction of reverse opposition K-partite networks, for K ≥ 2, as a straightforward
generalization of reverse opposition bipartite networks. The generalization is along the lines of the
generalization of opposition K-partite networks over opposition bipartite ones.
9
4 Continuous linear deviation functions
We now consider the situation when deviation functions take the affine-linear form
D(x) = a · x + b,
(4.1)
for a, b ∈ R. When deviation functions are of this form and when agents are in addition homogeneous
with respect to their deviation functions, i.e., D does not vary across agents, then the operator W ⊙ F
admits a particularly simple form, namely, that of an affine-linear operator; see Proposition A.2 in the
appendix. Moreover, we consider here the special case when a = −1 and b = 0.14 We consider such D
on opinion spectra S0 that are either of the form [−β, β] for some β > 0 or of the form S = R. We call
this D defined on such opinion spectra soft opposition.15 In this case, social networks W ⊙ F may be
represented by a matrix A that has entries Wij iff j ∈ In(i) and entries −Wij iff j ∈ Out(i).
4.1 A graph theoretical description
In the remainder, we consider the situation when graphs (i) contain no self-loops and are undirected
(simple) in the sense that Wij = Wji and Fij = Fji,16 (ii) contain at most one type of deviation function
D across all agents, and (iii) when D is linear on the opinion spectrum S0; in particular, D is soft
opposition. When networks W ⊙ F are so specified, then, as before, W ⊙ F admits a linear matrix
representation A where A is in addition symmetric: A⊺ = A. We denote this class of networks by
SLS(S0) (for S imple, Linear, S oft opposition, and where the argument refers to the opinion spectrum),
that is,
SLS(S0) = {W ⊙ F ∀i, j ∈ [n](cid:0)Wii = 0, Fij ∈ {F, D}, D soft opposition on S0,
Wij = Wji, Fij = Fji(cid:1)}
denotes the class of social networks on agent sets [n] that satisfy simplicity, symmetricity, etc., as de-
scribed.
2
1
3
4
6
5
7
2
1
3
4
6
5
7
Figure 4: Balanced and imbalanced networks. The left network is opposition bipartite (balanced) while
the right is neither opposition bipartite nor reverse opposition bipartite. In particular, agents 3 and 4
have mutual 'friends' (agents 1, 2) while agent 6 is in 3's out-group and 4's in-group. Alternatively:
agents 4 and 6, from different positively linked factions, have befriended each other.
We make an additional technical assumption here and in the remainder, namely, we generally assume
that the social networks wherein agents interact are aperiodic, that is, the greatest common divisor of the
lengths of their simple cycles is 1. Our main theorem in this context exhaustively categorizes long-run
opinions in terms of three different social network structures as outlined in the theorem. In the theorem,
recall that a graph G is said to be (strongly) connected if there exists a path in G from any node to any
other node. Moreover, in the theorem, we say that W ⊙ F is convergent if W ⊙ F is convergent for all
14The case a > 1 usually implies that opinions 'explode' over time because agents 'overscale' their out-group members'
opinions while a < 1 usually implies that society reaches a 'zero consensus' because agents iteratively discount their
out-group members' opinions.
15We call such D soft opposition because opinions are inverted the less strongly the closer they are to the neutral consensus
zero. In our working paper version, we also define a (discontinuous) deviation function called hard opposition that maps
opinions to extreme inverted values of the opinion spectrum, depending on whether they are positive or negative (depending
on which side of the opinion spectrum they lie, with 0 as reference point).
16This captures reciprocity: amity and enmity are mutual.
10
initial opinion vectors b(0) ∈ Sn, that is, limt→∞(W ⊙ F)tb(0) exists for all b(0) ∈ Sn. We say that
W ⊙ F induces a consensus if W ⊙ F induces a consensus for all initial opinion vectors b(0) ∈ Sn, that
is, W ⊙ F is convergent and limt→∞(W ⊙ F)tb(0) is always a consensus. We say that W ⊙ F is divergent
if it is not convergent. In other words, W ⊙ F is divergent when there exists a vector b(0) such that
limt→∞(W ⊙ F)tb(0) does not exist. We say that W ⊙ F induces a polarization if limt→∞(W ⊙ F)tb(0)
is a polarization, for all initial opinion vectors b(0). We say that W ⊙ F induces a non-zero polarization
when W ⊙ F induces a polarization and limt→∞(W ⊙ F)tb(0) 6= 0 for some initial opinion vectors b(0).
Theorem 4.1. Let W ⊙ F ∈ SLS(S0). Assume that W ⊙ F is strongly connected (since A is symmetric,
we might also simply say 'connected') and aperiodic. Then:
(i) W ⊙ F induces a polarization (that is not always zero) if and only if W ⊙ F is opposition bipartite.
(ii) W ⊙ F is divergent if and only if W ⊙ F is reverse opposition bipartite.
(iii) W ⊙ F induces a neutral17 consensus if and only if W ⊙ F is neither opposition bipartite nor reverse
opposition bipartite.
For better understanding, we give alternative characterizations of conditions (i) and (ii) of Theorem
4.1, which follow from the theorem and its proof. Namely, we find that:
(i) If and only if W ⊙ F is opposition bipartite, the following holds: b(∞) = limt→∞(W ⊙ F)b(0)
exists for all b(0) ∈ Sn and b(∞) is then always either the zero vector (e.g., when b(0) is the zero
vector; but it is not the zero vector for all b(0) ∈ Sn) or is some polarization vector where each
entry is a > 0 or −a. Moreover, when Wi,Out(i) > 0 for some i ∈ [n], then, if b(∞) is a non-zero
polarization vector, both a and −a are components of b(0); otherwise b(∞) is always a consensus,
as in the standard DeGroot model.
(ii) If and only if W ⊙ F is reverse opposition bipartite, the following holds: limt→∞(W ⊙ F)b(0) does
not exist for all b(0) in Sn. Moreover, when it exists it is the zero vector.
The fact that polarization requires 'exact' balance (opposition bipartiteness) and admits not a 'grain
of imbalancedness', as stated in Theorem 4.1 and exemplified in Figure 4, may appear odd since one
might expect, in reality, small perturbations to balance -- e.g., small-scale intra-group antagonisms or
individual friendships among enemies -- to be the rule rather than the exception, particularly in large
enough systems. We note that this result is closely connected to the continuous opinion spectrum and
the corresponding averaging updating that we have considered in this section. If one thinks that reality
is better perceived of as discrete, with weighted majority voting as a more plausible opinion updating
mechanism, then it is apparent that the discrete model is clearly robust against small such perturbations,
so that polarizing viewpoints can be Nash equilibria in this case even when the underlying networks exhibit
(marginal) imbalancedness.
We also note that our results may be generalized to periodic graphs (those that are not aperiodic)
and to graphs that are not connected. We leave this to future work.
4.2 Social influence and opinion leadership
In DeGroot learning, one of the important questions has been that of opinion leadership: whose agents'
initial opinions have most impact upon resulting limiting (long-run) opinions and how does this depend
on the network structure in which the agents are embedded.
In the context of connected SLS(S0)
networks as we have defined above, this question admits an elegant solution in our extended DeGroot
model with in-group/out-group relationships. Namely, if the network is reverse opposition bipartite,
the updating operator diverges (for at least some initial opinion vectors) and opinion leadership is thus
not well-defined. If, in contrast, the network is opposition bipartite and aperiodic, opinion leadership is
determined by eigenvector centrality (Bonacich, 1972) exactly in the same way as in the original DeGroot
model, except that a plus or minus sign indicates the group membership of the agents. Finally, if none of
those two conditions hold, then no agent is influential, since agents will converge to a fixed-point of the
deviation function no matter their initial opinions. The opposition bipartite case is the focus of the next
theorem.
17Here, we call a consensus (c, . . . , c) neutral if D(c) = c. For soft opposition on S0, c = 0.
11
Theorem 4.2. Let W ⊙ F ∈ SLS(S0). Assume that W ⊙ F is strongly connected and aperiodic. Then
the following are equivalent:
(i) There exists a unique (nonnegative) left unit eigenvector s of A, the matrix with entries Aij ,
whose entries sum to 1 such that each agent holds one of two long-run opinion values a and b
(b = −a), given by
g(j)sjbj(0),
(−g(j))sjbj(0),
a = Xj∈[n]
b = Xj∈[n]
where g(j) ∈ {±1}.
(ii) W ⊙ F is opposition bipartite.
Since in case of opposition bipartite and aperiodic strongly connected networks there are two long-run
opinion values and an agent j's 'influence' on each of them is g(j)sj and −g(j)sj, respectively, where
g(j) ∈ {±1}, we may speak of agent j's absolute power vj = g(j)sj = −g(j)sj = sj. Thus, in
summary, the last theorem and our previous discussion lead to the following characterization for strongly
connected networks W ⊙ F ∈ SLS(S0):
(i) Each agent's absolute power vi is given by vi = si where s is the unique left unit eigenvector of
A with normalization Pi∈[n] si = 1 if and only if W ⊙ F is opposition bipartite.
(ii) Opinion leadership is not (well-)defined if and only if W ⊙ F is reverse opposition bipartite.
(iii) Each agent's power is given by vi = 0 if and only if W ⊙ F is neither opposition bipartite nor
reverse opposition bipartite.
It is noteworthy that in case (i), absolute power is independent of the kinds of relationships between
agents and only depends on their intensities.
In other words, an agent may also be prominent when
she attracts strong negative links. When society partitions into several subsocities, these results may be
applied independently to each of them.
5 Concluding remarks
Opinions are important in an economic context (and other contexts) since they shape the demand for
products, set the political course, and guide, in general, socio-economic behavior. Models of opinion
dynamics describe how individuals form opinions or beliefs about an underlying state or a discussion
topic. Typically, in the social networks literature, subjects may communicate with other individuals,
their peers, in this context, enabling them to aggregate dispersed information. Bayesian models of
opinion formation assume that agents form their opinions in a fully rational manner and have an accurate
'model of the world' at their disposal, both of which are questionable and unrealistic assumptions for
actual social learning processes of human individuals. Non-Bayesian models, and most prominently the
classical DeGroot model of opinion formation, while also not unproblematic (cf. Acemoglu and Ozdaglar,
2011), posit that agents employ simple 'rule-of-thumb' heuristics to integrate the opinions of others.
Unfortunately, both the non-Bayesian and Bayesian paradigms typically lead individuals to a consensus,
which apparently contradicts the facts as people disagree with others on many issues of (everyday) life.
In the context of DeGroot learning models, some approaches can address this, either by assuming a
homophily principle whereby agents limit their communication to those who hold similar opinions as
themselves or by introducing stubborn agents, modeling, e.g., opinion leaders, who never update their
opinions. Both approaches are, again, debatable since they assume a complete lack of flow of information
between some classes of agents (from some time onwards). In addition, models based on homophily and
stubborn agents both ignore negative relationships between individuals as potential sources for conflict
and disagreement.
12
In the current work, we have investigated opinion dynamics under out-group discrimination (in-group
bias) as such a potentially alternative explanation for disagreement.
In our setup, agents are driven
by two forces: they want to adjust their opinions to match those of the agents of their in-group and, in
addition, they want to adjust their opinions to match the 'inverse' of those of the agents of their out-group.
Best responses in this setting lead us to a DeGroot-like opinion updating process in which agents form
their next period opinions via weighted averages of their neighbors' (possibly inverted) opinion signals.
Unlike in the standard DeGroot model where opinions may converge to arbitrary consensus opinion
profiles, in our model only neutral consensus opinion profiles may be attained in the long-run, that is,
consensus vectors where the consensus opinion is a fixed-point of each agent's deviation function (modeling
the mode of opposition between agents). Thus, if there exists no opinion that is 'globally' neutral in
this sense, our model predicts persistent disagreement provided that negative relations between agents
are sufficiently strong. When we specialize our model to undirected networks containing no self-loops
and where the only allowable deviation function is 'soft opposition', we derive necessary and sufficient
conditions for bi-polarization in connected societies. These say that long-run opinions bi-polarize if and
only if the underlying network wherein agents communicate satisfies 'opposition bipartiteness': it consists
of two groups of agents exhibiting positive within-group links and negative between-group links. We also
investigate social influence in this special case, that is, the question of whose initial opinions matter most
for resulting long-run opinions. We find that in opposition bipartite networks (satisfying aperiodicity),
opinion leadership, in terms of absolute power, is determined by eigenvector centrality exactly in the same
way as in the standard DeGroot model. This means that an agent is prominent to the degree that she is
interlinked with prominent agents; it is noteworthy that even strong negative ties increase prominence.
Finally, considering the question of wisdom (Golub and Jackson, 2010; Jadbabaie et al., 2012; Buechel,
Hellmann, and Klossner, 2015), that is, whether agents can (jointly) learn the true state of nature of their
discussion topic, provided that such a truth exists, we observe the following. The case for wisdom is a
weak one in our model since negative ties typically prevent consensus formation, so clearly not everybody
can be wise in the long-run. This holds even when agents are initially perfectly informed in the sense that
their initial beliefs coincide with truth. In particular, if agents (multi-)polarize, then at most one group
of agents may be wise in the long-run, but due to the mutual dependence of long-run beliefs, we might
suspect none to be.18 Ultimately, this result must be interpreted by reference to the rationality of the
agents involved. The standard interpretation of DeGroot learners is that of naıve individuals susceptible
to persuasion bias. Golub and Jackson (2010) show that such agents can still learn the true state of
nature under not too demanding conditions. Jadbabaie et al. (2012) show that slightly more rationality
increases the case for wisdom. In contrast, we show that an additional bias such as in-group bias may
significantly worsen this case.
Concerning future research directions within our context, both weight links and opposition links
between agents, W and F, have been assumed exogenous in the current work. Prospectively, it might
be worthwhile to consider endogenous link formation processes. In particular, the origin and evolution of
out-group relations, and their interdependence with agents' opinions and external factors, such as, most
importantly, external truth, might be of interest.
Acknowledgments
I am indebted to the associate editor Juan D. Moreno-Ternero and two anonymous reviewers for various
comments that have greatly improved the layout of the current work. I also thank Matthias Blonski for
valuable suggestions that led to improvements in the paper. All remaining errors are my own.
Appendix A Definitions, theorems, and proofs
Sections A.1 and A.2 review notation and concepts for graphs and matrices, respectively. Section A.3
states basic results on signed social networks from Altafini (2013). Finally, Section A.4 provides the proofs
18In particular, observe that an agent's wisdom is expected to decrease in her amount of in-group bias (i.e., how strongly
she discriminates against out-group individuals) provided that out-group individuals' opinions are close to truth, because
this controls how strongly she desires to match the 'opposite' of truth. From a reverse perspective, her wisdom is expected
to increase in her out-groups' biases.
13
of our own results in this work.
A.1 Graphs
Definition A.1 ((Weighted) Network). A network, or graph, is a tuple G = (V, E) where V is a finite set
and E ⊆ V × V = {(u, v) u, v ∈ V }. We call V the vertices or nodes of graph G and E the edges or links
of G. Moreover, we call the network G weighted if there exist weights wuv for each edge (u, v) ∈ E.19
In a multigraph, instead of having only one link type between nodes, there may exist multiple link
types. The networks investigated in this work may be considered multigraphs with exactly two types of
links, one denoting intensity of connection and one denoting kind of connection.
Definition A.2. A walk in a network G = (V, E) is a sequence of nodes i1, i2, . . . , iK, not necessarily
distinct, such that (ik, ik+1) ∈ E for all k ∈ {1, . . . , K − 1}. A path is a walk consisting of distinct nodes.
A cycle is a walk i1, . . . , iK such that i1 = iK. The length of cylce i1, . . . , iK is defined to be K − 1. A
cycle is called simple if the only node appearing twice is i1 = iK.
Remark A.1. We use the same terminology -- 'strongly connected', 'aperiodic', etc. -- whether we
speak of (our) multigraphs, in which there exist exactly two types of relationships between agents, or
ordinary graphs. In the case of multigraphs, we refer to their underlying ordinary graphs. We also use
the same terminology for n × n matrices A and their underlying graphs ([n], {(i, j) Aij 6= 0}).
A.2 Matrix and Markov chain theory
We first state one of the main theorems for the DeGroot updates (2.6) in the non-opposition case (cf.
Golub and Jackson (2010)). We assume that W is row-stochastic.
Theorem A.1. Consider the opinion updating process (2.6) with Fij = F for all i, j ∈ [n], where F is
the identity function. Let the multigraph corresponding to the operator W ⊙ F = W -- an ordinary
graph -- be strongly connected and aperiodic. Then W ⊙ F is convergent and induces a consensus.
In case W ⊙ F is an affine-linear map, whether or not W ⊙ F converges can be fully determined by
reference the notion of eigenvalues, which we introduce now.
Definition A.3. Let A ∈ Rn×n be an n × n matrix. An eigenvalue of A is any value λ ∈ C such that
Ax = λx for some non-zero vector x ∈ Rn. The set of distinct eigenvalues of matrix A is called its
spectrum and denoted by σ(A). By ρ(A), we denote the spectral radius of A, the largest absolute value
of all the eigenvalues of A, that is, ρ(A) = max{(cid:12)(cid:12)λ(cid:12)(cid:12) λ ∈ σ(A)}.
Theorem A.2 (Meyer, 2000, p.630). For A ∈ Rn×n, limt→∞ At exists if and only if
ρ(A) < 1,
ρ(A) = 1 and λ = 1 is the only eigenvalue on the unit circle, and λ = 1 is semisimple,
or else,
where an eigenvalue is called semisimple if its algebraic multiplicity equals its geometric multiplicity. The
algebraic multiplicity of an eigenvalue λ is the number of times it is repeated as a root of the characteristic
polynomial χ(λ) = det (A − λIn), where In is the n × n identity matrix. The geometric multiplicity is
the number of linearly independent eigenvectors associated with λ.
A.3 Signed networks
Here, we assume social networks W ⊙ F such that Fij ∈ {F, D} where D is soft opposition on S0, i.e.,
D(x) = −x. Such operators admit a matrix representation A in which each entry has a positive or
negative sign (or is zero), see Proposition A.2. We assume that A is connected and aperiodic.
Lemma A.1. Let W ⊙ F be such that Aii = 0 and Aij = Aji. Then, W ⊙ F is opposition bipartite if
and only if there exists a diagonal matrix ∆ such that ∆A∆ = A, where A denotes the matrix with
entries Aij.
19Weights may typically be real numbers but we more generally allow them to be arbitrary mathematical objects.
14
Lemma A.2. Let A = ∆A∆ as in Lemma A.1. Then A and A have the same eigenvalues with the
same multiplicities.
Lemma A.3. Let W ⊙ F be such that Aii = 0 and Aij = Aji. Then, W ⊙ F is opposition bipartite if
and only if λ = 1 is an eigenvalue of A.
Proof. Altafini (2013), Lemma 1, shows that 0 ∈ σ(L) if and only if A is opposition bipartite where
L = I − A. Clearly, 1 ∈ σ(A) ⇐⇒ 0 ∈ σ(L).
A.4 Proofs of main results
Proposition A.1. Let W ⊙ F be an arbitrary social network such that Fij ∈ {F, Di}, for all i, j ∈ [n].
Then, for all c ∈ S,
c ∈ \(i,j)∈[n]×[n]
Fix(Fij ) =⇒ (c, . . . , c) ∈ Fix(W ⊙ F).
Moreover, let A = {i ∈ [n] Wi,Out(i) > C0} denote the set of agents whose weight mass assigned to
out-group members exceeds a particular threshold C0; in the discrete case, C0 = 1
2 , and in the continuous
case, C0 = 0. Then, for any i ∈ A, it holds that
c /∈ Fix(Di) =⇒ (c, . . . , c) /∈ Fix(W ⊙ F).
Combining both implications yields that
P1[Fix(W ⊙ F) ∩ C] = \i∈A
Fix(Di).
Remark A.2. If D is allowed to vary across both i and j, then c /∈ Fix(Dij) does not necessarily imply
that (c, . . . , c) /∈ Fix(W ⊙ F). To see this, assume, for example, that in a three-player society {1, 2, 3}
agent 1 has Out(1) = {2, 3} with W12 = W13 = 1
4 . For a c ∈ S = R, let D12(c) = c + ǫ and let
D13(c) = c − ǫ, for some ǫ > 0. Assuming that Out(2) = Out(3) = ∅, we have (W ⊙ F)c = c, since, in
particular, for agent 1,
W11c + W12D12(c) + W13D13(c) =
1
2
c +
1
4
(c + ǫ) +
1
4
(c − ǫ) = c.
Proof of Proposition A.1. We only provide the proof for the continuous model. The discrete model proof
is similar.
If c = Fij (c) for some c ∈ S and all (i, j) ∈ [n] × [n], then clearly -- letting c = (c, . . . , c) --
(W ⊙ F)c = c by the definition of W ⊙ F since for each agent i ∈ [n],
n
(cid:2)(W ⊙ F)c(cid:3)i =
Xj=1
Wij Fij(c) = c Xj∈[n]
Wij = c = [c]i.
Conversely, let c 6= Di(c) for some c ∈ S and some i ∈ A. If c = (c, . . . , c) were a fixed-point of
W ⊙ F, then
c = Xj∈Out(i)
Wij Di(c) + Xj∈In(i)
Wij c = Di(c)Wi,Out(i) + c(1 − Wi,Out(i)),
which implies that
This is a contradiction since Wi,Out(i) > 0 by assumption.
Wi,Out(i)c = Wi,Out(i)Di(c).
Lemma A.4. Let W ⊙ F be an arbitrary social network. Assume that either W ⊙ F refers to the discrete
model or that each function Fij in F is continuous. Then:
Lim(W ⊙ F) = Fix(W ⊙ F).
15
Proof of Lemma A.4. The relation Fix(W ⊙ F) ⊆ Lim(W ⊙ F) is obvious. Conversely, if each Fij is
continuous, then W ⊙ F is a continuous operator and thus, each limit vector b(∞) ∈ Lim(W ⊙ F) is a
fixed-point of W ⊙ F:
(W ⊙ F)b(∞) = (W ⊙ F) lim
t→∞
(W ⊙ F)tb(0) = lim
t→∞
(W ⊙ F)t+1b(0) = b(∞).
If S is finite and W ⊙ F is convergent (for b(0)), then b(∞) is a fixed-point of W ⊙ F no matter the
specification of F.
Proof of Theorem 3.1. This is an application of Proposition A.1 and Lemma A.4.
Proposition A.2. Let D be of the form ax + b for some constants a and b, and let Fij ∈ {F, D}. Then,
W ⊙ F is an affine-linear operator of the form Ax + d, that is, (W ⊙ F)(x) = Ax + d for all x ∈ Sn.
Proof of Proposition A.2. For each agent i ∈ [n], we have
(cid:2)(W ⊙ F)x(cid:3)i = Xj∈In(i)
= Xj∈In(i)
= Xj∈In(i)
Wij xj + Xj∈Out(i)
Wij xj + Xj∈Out(i)
Wij xj + Xj∈Out(i)
Wij D(xj ) = Xj∈In(i)
aWij xj + b Xj∈Out(i)
(aWij )xj + bWi,Out(i).
Wij (axj + b)
Wij xj + Xj∈Out(i)
Wij
Thus, we can set A ∈ Rn×n, d ∈ Rn with
Aij =(Wij
aWij
if Fij = F,
if Fij = D,
di = bWi,Out(i).
(A.1)
Lemma A.5. Let W ⊙ F be an arbitrary social network with Fij ∈ {F, D} for an arbitrary deviation
function D. Then, W ⊙ F is opposition bipartite if and only if W ⊙ ¯F is reverse opposition bipartite,
where ¯F is the matrix with entries ¯Fij = ¬Fij , whereby we define ¬D = F and ¬F = D.
Proof. See Figure 2, in Section 3, for a graphical proof.
Remark A.3. If D is soft opposition on S0, let (A, 0) be the representation of W ⊙ F. Then, the lemma
specializes to the statement that, in this situation, (A, 0) is opposition bipartite if and only if (−A, 0) is
reverse opposition bipartite.
Proof of Theorem 4.1. (i) If W ⊙ F induces a (non-zero) polarization, then, necessarily, 1 ∈ σ(A). But,
1 ∈ σ(A) ⇐⇒ W ⊙ F is opposition bipartite by Lemma A.3. Conversely, let W ⊙ F be opposition
bipartite. Then, A and A are isospectral, that is, they have the same eigenvalues and with the same
associated multiplicities by Lemmas A.1 and A.2. Now, (⋆) a strongly connected and aperiodic row-
stochastic matrix A has exactly one eigenvalue on the unit circle, λ = 1, with algebraic and geometric
multiplicity of 1. Therefore, A has exactly one eigenvalue on the unit circle, λ = 1, with algebraic
and geometric multiplicity of 1 and, consequently, converges by Theorem A.2. Moreover, since each
polarization vector x with xi = 1 if i ∈ N1 and xi = −1 if i ∈ N2 satisfies Ax = (W ⊙ F)x = x when
W ⊙ F is opposition bipartite with partition (N1, N2), W ⊙ F induces a polarization that is not always
zero (note that the geometric multiplicity of λ = 1 of A is 1).
Part (ii) "⇐" follows from the fact that 1 ∈ σ(A) ⇐⇒ W ⊙ F is opposition bipartite and the
fact that W ⊙ F with representation A is opposition bipartite if and only if −A is reverse opposition
bipartite by Lemma A.5. Thus, −1 ∈ σ(A) ⇐⇒ W ⊙ F is reverse opposition bipartite, whence A
diverges by Theorem A.2. Conversely, when W ⊙ F diverges, then ρ(A) = 1 (we have ρ(A) ≤ 1 for all
such matrices A as we consider since A = W is row-stochastic and therefore, ρ(A) ≤ ρ(A) = 1). If 1
were in σ(A), then W ⊙ F were opposition bipartite and W ⊙ F would converge by (i). Hence, 1 /∈ σ(A)
and consequently, −1 ∈ σ(A) -- since a symmetric matrix A has no complex eigenvalues. Consequently,
W ⊙ F is reverse opposition bipartite.
16
Finally, for part (iii), if W ⊙ F is neither opposition bipartite nor reverse opposition bipartite, then,
by our above reasonings, ±1 /∈ σ(A), and since A is symmetric, A has no complex eigenvalues, whence
ρ(A) < 1. Thus, W ⊙ F induces the unique neutral consensus (0, . . . , 0). Conversely, if W ⊙ F induces
the neutral consensus (0, . . . , 0) for each initial belief vector b(0), then necessarily ρ(A) < 1. Hence,
W ⊙ F is neither opposition bipartite nor reverse opposition bipartite.
Now, fact (⋆) is a classical theorem for row-stochastic matrices, which is, e.g., based on the famous
Perron-Frobenius theorem; in our context, it is given by combining Theorems A.1 and A.2, for example.
Proof of Theorem 4.2. (i) =⇒ (ii): For an appropriate initial opinion vector, let each agent hold limiting
opinions a 6= 0 or −a as indicated. Place agents in a group N1 resp. N2 depending on whether they hold
limit opinions a or −a, respectively. We show that (N1, N2) forms an opposition bipartite partition of
[n]. Take i, i′ ∈ N1 and assume that Aii′ < 0 (i.e., i and i′ are enemies). Then, since the limit opinion
vector is a fixed-point of A, we have:
a(±Ai1 + · · · + Aii′ + · · · + ±Ain) = a.
But this cannot be, since Aii′ + C < Aii′ + C′ = 1 where C = Pj6=i′ ±Aij and C′ = Pj6=i′ Aij .
Similarly, we can show that no two agents i ∈ N1, j ∈ N2 are friends of each other. Hence, W ⊙ F is
opposition bipartite.
(ii) =⇒ (i): Conversely, let W ⊙ F be opposition bipartite and aperiodic. By Lemma A.1, W ⊙ F
is opposition bipartite if and only if there exists a diagonal matrix ∆ (with entries ±1) such that A =
∆A∆. We know that
At p = s⊺p1
lim
t→∞
for all p ∈ Sn (see, e.g., Golub and Jackson, 2010, Proposition 1) and that ∆−1 = ∆. Therefore,
lim
t→∞
Atp = ∆ lim
t→∞
At (∆p) = ∆s⊺(∆p)1.
This proves the theorem.
References
[1] Robert P. Abelson. "Mathematical models of the distribution of attitudes under controversy". In:
Contributions to mathematical psychology. Ed. by N. Frederiksen and H. Gulliksen. New York:
Rinehart Winston, 1964, pp. 142 -- 160.
[2] Alan Abramowitz and Kyle Saunders. "Why can't we all just get along? The reality of a polarized
America". In: The Forum 3 (2005).
[3] Daron Acemoglu and Asuman Ozdaglar. "Opinion Dynamics and Learning in Social Networks".
In: Dynamic Games and Applications 1 (1 2011), pp. 3 -- 49.
[4] Daron Acemoglu, Asuman Ozdaglar, and Ali ParandehGheibi. "Spread of (Mis)Information in
Social Networks". In: Games and Economic Behavior 70 (2 2010), pp. 194 -- 227.
[5] Daron Acemoglu, Giacomo Como, Fabio Fagnani, and Asuman Ozdaglar. "Opinion Fluctuations
and Disagreement in Social Networks". In: Mathematics of Operations Research 38.1 (Feb. 2013),
pp. 1 -- 27. issn: 0364-765X. doi: 10.1287/moor.1120.0570. url: http://dx.doi.org/10.1287/moor.1120.0570.
[6] Claudio Altafini. "Consensus Problems on Networks With Antagonistic Interactions". In: IEEE
Transactions on Automatic Control 58 (4 2013).
[7] David Beasley and Dan Kleinberg. Networks, Crowds, and Markets: Reasoning about a Highly
Connected World. Cambridge, UK: Cambridge University Press, 2010.
[8] Avner Ben-Ner, Brian McCall, Massoud Stephane, and Hua Wang. "Identity and in-group/out-
group differentiation in work and giving behaviors: Eperimental evidence". In: Journal of Economic
Behavior and Organization 72 (2009), pp. 153 -- 170.
17
[9] Philip Bonacich. "Factoring and weighting approaches to status scores and clique identification".
In: The Journal of Mathematical Sociology 2 (1 1972), pp. 113 -- 120.
[10] Marylinn B. Brewer. "In-Group Bias in the minimal intergroup situation: A cognitive-motivational
analysis". In: Psychological Bulletin 86 (2 1979), pp. 307 -- 324.
[11] Berno Buechel, Tim Hellmann, and Stefan Klossner. "Opinion dynamics and wisdom under con-
formity". In: Journal of Economic Dynamics and Control 52.0 (2015), pp. 240 -- 257. issn: 0165-1889.
doi: http://dx.doi.org/10.1016/j.jedc.2014.12.006. url: http://www.sciencedirect.com/science/article/pii/S0165188914003315.
[12] Zhigang Cao, Mingmin Yang, Xinglong Qu, and Xiaoguang Yang. "Rebels Lead to the Doctrine of
the Mean: Opinion Dynamic in a Heterogeneous DeGroot Model". In: The 6th International Con-
ference on Knowledge, Information and Creativity Support Systems. Beijing, China, 2011, pp. 29 --
35.
[13] Emanuele Castano, Vincent Yzerbyt, David Bourguignon, and El´eonore Seron. "Who may Enter?
The Impact of In-Group Identification on In-Group/Out-Group Categorization". In: Journal of
Experimental Social Psychology 38 (2002), pp. 315 -- 322.
[14] Gary Charness, Luca Rigotti, and Aldo Rustichini. "Individual Behavior and group membership".
In: American Economic Review 97 (4 2007), pp. 1340 -- 1352.
[15] Yan Chen and Sherry Xin Li. "Group Identity and Social Preferences". In: American Economic Re-
view 99.1 (2009), pp. 431 -- 57. doi: 10.1257/aer.99.1.431. url: http://www.aeaweb.org/articles.php?doi=10.1257/aer.99.1.431.
[16] Geoffrey L. Cohen. "Party Over Policy: The Dominating Impact of Group Influence on Political
Beliefs". In: Journal of Personality and Social Psychology 85 (5 2003), pp. 808 -- 822.
[17] Luca Corazzini, Filippo Pavesi, Beatrice Petrovich, and Luca Stanca. "Influential listeners: An
experiment on persuasion bias in social networks". In: European Economic Review 56 (6 2012),
pp. 1276 -- 1288.
[18] Guillaume Deffuant, David Neau, Frederic Amblard, and Gerard Weisbuch. "Mixing beliefs among
interacting agents". In: Advances in Complex Systems 3 (2000), pp. 87 -- 98.
[19] Morris H. DeGroot. "Reaching a Consensus". English. In: Journal of the American Statistical Asso-
ciation 69.345 (1974), pp. 118 -- 121. issn: 01621459. url: http://www.jstor.org/stable/2285509.
[20] Peter M. DeMarzo, Dimitri Vayanos, and Jeffrey Zwiebel. "Persuasion Bias, Social Influence, And
Unidimensional Opinions". In: The Quarterly Journal of Economics 118.3 (2003), pp. 909 -- 968.
url: http://ideas.repec.org/a/tpr/qjecon/v118y2003i3p909-968.html.
[21] Sebastian Fehrler and Michael Kosfeld. "Can you trust the good guys? Trust within and between
groups with different missions". In: Economics Letters 121.3 (2013), pp. 400 -- 404.
[22] Benjamin Golub and Matthew O. Jackson. "How Homophily Affects the Speed of Learning and
Best-Response Dynamics". In: The Quarterly Journal of Economics 127 (3 2012), pp. 1287 -- 1338.
[23] Benjamin Golub and Matthew O. Jackson. "Naıve Learning in Social Networks and the Wisdom
of Crowds". In: American Economic Journal: Microeconomics 2 (1 2010), pp. 112 -- 149.
[24] Patrick Groeber, Jan Lorenz, and Frank Schweitzer. "Dissonance minimization as a microfoundation
of social influence in models of opinion formation". In: Journal of Mathematical Sociology (2013).
[25] Shaun P. Hargreaves Heap and Daniel John Zizzo. "The Value of Groups". In: American Economic
Review 99.1 (2009), pp. 295 -- 323. doi: 10.1257/aer.99.1.295. url: http://www.aeaweb.org/articles.php?doi=10.1257/aer.99.1.295.
[26] Rainer Hegselmann and Ulrich Krause. "Opinion dynamics and bounded confidence: models, anal-
ysis and simulation". In: J. Artificial Societies and Social Simulation 5.3 (2002).
[27] Matthew O. Jackson. Social and Economic Networks. Princeton: Princeton University Press, 2009.
[28] Ali Jadbabaie, Pooya Molavi, Alvaro Sandroni, and Alireza Tahbaz-Salehi. "Non-Bayesian social
learning." In: Games and Economic Behavior 76.1 (2012), pp. 210 -- 225. url: http://dblp.uni-trier.de/db/journals/geb/geb76.html#JadbabaieMST12.
[29] Matthew Jarman, Andrzej Nowak, Wojciech Borkowski, David Serfass, Alexander Wong, and Robin
Vallacher. "The Critical Few: Anticonformists at the Crossroads of Minority Opinion Survival and
Collapse". In: Journal of Artificial Societies and Social Simulation 18.1 (2015).
18
[30] Marco Alberto Javarone. "Social influences in opinion dynamics: The role of conformity". In: Phys-
ica A: Statistical Mechanics and its Applications 414.0 (2014), pp. 19 -- 30. issn: 0378-4371. doi:
http://dx.doi.org/10.1016/j.physa.2014.07.018. url: http://www.sciencedirect.com/science/article/pii/S0378437114005871.
[31] Mark P. Jones. Electoral Laws and the Survival of Presidential Democracies. Notre Dame: University
of Notre Dame Press, 1995.
[32] James A. Kitts. "Social influence and the emergence of norms amid ties of amity and enmity". In:
Simulation Modelling Practice and Theory 14 (2006), pp. 407 -- 422.
[33] Gerald H. Kramer. "Short-term fluctuations in U.S. voting behavior, 1896 -- 1964". In: American
Political Science Review 65 (1 1971), pp. 131 -- 143.
[34] Carl D. Meyer. Matrix analysis and applied linear algebra. Philadelphia: SIAM, 2000.
[35] Henri Tajfel, M. G. Billig, R. P. Bundy, and Claude Flament. "Social categorization and intergroup
behaviour". In: European Journal of Social Psychology 1.2 (1971), pp. 149 -- 178. issn: 1099-0992.
doi: 10.1002/ejsp.2420010202. url: http://dx.doi.org/10.1002/ejsp.2420010202.
[36] K´aroly Tak´acs, Andreas Flache, and Michael Mas. "Is there negative social influence? Disentan-
gling effects of dissimilarity and disliking on opinion shifts". In: CoRR abs/1406.0900 (2014). url:
http://arxiv.org/abs/1406.0900.
[37] Ryuhei Tsuji. "Interpersonal influence and attitude change toward conformity in small groups: a
social psychological model". In: Journal of Mathematical Sociology 26 (2002), pp. 17 -- 34.
[38] Kei Tsutsui and Daniel Zizzo. "Group status, minorities and trust". In: Experimental Economics
17.2 (2014), pp. 215 -- 244. url: http://EconPapers.repec.org/RePEc:kap:expeco:v:17:y:2014:i:2:p:215-244.
[39] Ercan Yildiz, Daron Acemoglu, Asuman Ozdaglar, Amin Saberi, and Anna Scaglione. Discrete
Opinion Dynamics with Stubborn Agents. LIDS report 2870. to appear in ACM Transactions on Eco-
nomics and Computation. 2012. url: http://web.mit.edu/asuman/www/documents/voter-submit.pdf.
[40] John M. Yinger. "Countercultures and social change". In: American Sociological Review 42 (6 1977),
pp. 833 -- 853.
19
|
1703.06416 | 1 | 1703 | 2017-03-19T10:35:17 | A Passivity-Based Distributed Reference Governor for Constrained Robotic Networks | [
"cs.MA",
"cs.DC",
"cs.RO",
"eess.SY"
] | This paper focuses on a passivity-based distributed reference governor (RG) applied to a pre-stabilized mobile robotic network. The novelty of this paper lies in the method used to solve the RG problem, where a passivity-based distributed optimization scheme is proposed. In particular, the gradient descent method minimizes the global objective function while the dual ascent method maximizes the Hamiltonian. To make the agents converge to the agreed optimal solution, a proportional-integral consensus estimator is used. This paper proves the convergence of the state estimates of the RG to the optimal solution through passivity arguments, considering the physical system static. Then, the effectiveness of the scheme considering the dynamics of the physical system is demonstrated through simulations and experiments. | cs.MA | cs |
A Passivity-Based Distributed Reference
Governor for Constrained Robotic
Networks (cid:63)
Tam Nguyen ∗ Takeshi Hatanaka ∗∗ Mamoru Doi ∗∗
Emanuele Garone ∗ Masayuki Fujita ∗∗
∗ Service d'Automatique et d'Analyse des Syst`emes, Universit´e Libre de
Bruxelles, Brussels, Belgium (e-mails: [email protected],
∗∗ Department of Systems and Control Engineering, Tokyo Institute of
[email protected])
Technology, Tokyo, Japan (e-mails: [email protected],
[email protected], [email protected])
Abstract: This paper focuses on a passivity-based distributed reference governor (RG) applied
to a pre-stabilized mobile robotic network. The novelty of this paper lies in the method used to
solve the RG problem, where a passivity-based distributed optimization scheme is proposed. In
particular, the gradient descent method minimizes the global objective function while the dual
ascent method maximizes the Hamiltonian. To make the agents converge to the agreed optimal
solution, a proportional-integral consensus estimator is used. This paper proves the convergence
of the state estimates of the RG to the optimal solution through passivity arguments, considering
the physical system static. Then, the effectiveness of the scheme considering the dynamics of
the physical system is demonstrated through simulations and experiments.
Keywords: Distributed control and estimation, Control of networks, Mobile robots, Convex
optimization, Control under communication constraints
1. INTRODUCTION
In recent years, teleoperative robotic networks have at-
tracted the interest of several researchers around the
world. In bilateral teleoperation, Lee and Spong (2005)
and Shokri-Ghaleh and Alfi (2014) propose control frame-
works for the synchronization of bilateral teleoperation
systems with communication delays. A number of basic
algorithms running on synchronous robotic networks to
achieve rendez-vous are analyzed by Mart´ınez et al. (2007).
The study of motion coordination algorithms for robotic
networks is rich in the literature and Bullo et al. (2009)
summarize the basic tools for coordination algorithms.
However, the main problems of huge telecommunication
networks are hardware limitations and communication
signal strength, where the latter has been thoroughly
analyzed and modeled in Mostofi et al. (2010). Reich and
Sklar (2006) have shown that the performance of simple
and inexpensive onboard hardwares can be similar to the
performance of sophisticated and expensive systems that
are applied to search and rescue robotic networks. This
is why, as a rule of good practice, we usually limit the
communication capabilities of each agent and select a
small group of leaders that are able to communicate with
the teleoperation station.
(cid:63) This work is supported by EM-EASED, FRIA, and JSPS KAK-
ENHI Grant Number JP15H04019. The stay of Tam Nguyen in
Tokyo Institute of Technology has been supported by the Eras-
mus Mundus EASED programme (Grant 2012-5538/004-001) coor-
dinated by CentraleSup´elec.
An important feature to take into account for the con-
trol of mobile robotic networks is the ability to manage
constraints present in the environment (e.g. walls, holes,
obstacles, etc.) and the limitations of the actuators. For
geometric constraints, numerous research in the literature
for single robot path planning have been carried out us-
ing potential fields (Ge and Cui (2000)) and grid search
(Thorpe and Matthies (1984)). However, in the case of
robotic networks, the importance to manage constraints
in an efficient and distributed fashion is brought to light
since in general each agent does not share its information
to all the agents of the network and the teleoperator is only
able to send requests to the leaders. A possible way to deal
with the constraints in a distributed way is presented in
Soleymani et al. (2015), where a solution based on the
reference governor (RG) is proposed.
The RG (Gilbert et al. (1994)) is an add-on control
scheme, which manages the constraints to a pre-stabilized
system by suitably changing the applied reference. The
first idea to introduce an RG scheme using passivity argu-
ments applied to constrained mobile robotic networks has
been proposed in Nguyen et al. (2016). The proposed set
invariance-based RG uses the same passivity arguments for
the pre-stabilization of the robotic network, introducing
a paradigm for the control of constrained mobile robotic
networks. However, the proposed RG solves the problem
in a global way but not in a distributed fashion, which will
be one of the contribution of this paper.
In theory, passivity has been shown to have close relations
with distributed convex optimization. Firstly, Fan et al.
(2006) developed new decentralized control algorithms
that globally stabilize the desired Nash equilibrium by
exploiting the passivity property of the feedback loop.
Then, the results in Burger et al. (2014) established a
strong and explicit connection between passivity-based co-
operative control theory and network optimization theory.
The second contribution of this paper is to extend those
results to both inequality and equality constraints. The
proposed solution is then applied to robotic networks to
solve the specific RG problem.
This paper is organized as follows. First, the problem of
the mobile robotic network is described. The system is pre-
stabilized and then the RG optimization problem is for-
mulated. Then, we focus on a passivity-based distributed
optimization scheme to solve the RG problem. We prove
that the states of the proposed scheme converge to the
optimal solution using passivity arguments for a static
robotic network. To demonstrate the effectiveness of the
passivity-based distributed RG in real-time, the proposed
algorithm is implemented to the pre-stabilized system,
where simulations and experiments are carried out.
2. SYSTEM DESCRIPTION AND LOCAL
FEEDBACK
Consider a system of n mobile robots V = {1, . . . , n}
operating on a plane. The model of the ith robot proposed
in De La Croix and Egerstedt (2012) is
qi = ui, i ∈ V,
(1)
where qi ∈ R2 and ui ∈ R2 are the position and the
velocity input of the ith robot, respectively. The robots are
able to communicate and the inter-agent communication
is modeled by a graph G = (V,E), E ⊆ V ×V. Accordingly,
robot i has access to the information of the robots that are
belonging to the set of neighbors Ni = {j ∈ V(i, j) ∈ E}.
It is assumed that Vh is the subset of V corresponding to
the subset of robots that are able to communicate with
the teleoperation station, while all the robots in V \Vh are
able to communicate only with their neighbors. We also
introduce the notation δi such that δi = 1 if i ∈ Vh, and
δi = 0 otherwise. The system is guided by an operator that
sets a reference r ∈ R2 to the leaders.
Assumption 1. The graph G is fixed, undirected and con-
nected.
is(cid:20) q
(cid:21)
= −
(cid:20) ¯L − ¯L
(cid:21)(cid:20) q
(cid:21)
+
(cid:20) D ⊗ I2
(cid:21)
The system dynamics is pre-stabilized through a proportional-
integral (PI) consensus estimator (Freeman et al. (2006))
and a proportional feedback loop. The aggregate system
αr((1n⊗I2)r−q) (2)
0
ξ
¯L 0
ξ
n ]T is the coordinate vector, the
1 , . . . , qT
where q := [qT
symbol ⊗ is the Kronecker product, ξ the additional
integral state, D ∈ Rn×n a diagonal matrix whose (i, i)-
element is equal to δi, 1n the n-unit vector, I2 the 2-
dimensional identity matrix, αr ∈ R+
0 the gain of the
system, and ¯L := L ⊗ I2, where L ∈ Rn×n is the Graph
Laplacian associated to the graph G.
To ensure the formation objective, a pre-defined bias is
applied to the robots with respect to the leader. We can
then prove the following theorem.
Theorem 1. The states q of system (2) converge to the
pre-defined formation reference in absence of constraints.
Proof. The proof can be found in Nguyen et al. (2016).
The next section formulates the RG optimization problem
to deal with the constraints.
3. REFERENCE GOVERNOR OPTIMIZATION
PROBLEM FORMULATION
Similar to the paradigm proposed in Nguyen et al. (2016),
the constraints of the pre-stabilized system are managed
by a set invariance-based RG (Gilbert and Kolmanovsky
(2001)). The control scheme is depicted in Fig. 1.
Fig. 1. Proposed control scheme. The system is first pre-
stabilized then augmented with RG for constraints
enforcement.
We assume that each robot has sensors detecting obstacles
around the robot. Moreover, we assume that the sensors
provide a line specifying the boundary of the obstacles
detected. The collision-free space is given by the half-
space where the robot lies in and this space is denoted by
Hq
i ⊆ R2. As a consequence, the condition for all robots
to avoid collisions with the obstacle detected by robot i is
(3)
[qT ξT ]T ∈ Qi :=(cid:8)[qT ξT ]T qj ∈ Hq
Because the set Qi is a polyhedron, it is formulated as
i ∀j ∈ V(cid:9) .
(cid:9)
Qi :=(cid:8)[qT ξT ]T Aq
i [qT ξT ]T ≤ bq
(4)
i ∈ Rn. If robot i does not
i ∈ Rn×4n and bq
using some Aq
detect any obstacle, then Qi = R4n
Let us assume that the input ui is constrained within a
convex polytope Hu
i . Following from (2), ui is given by
ui = −(ei ⊗ I2)( ¯Lq − ¯Lξ) + δi(r − qi),
(5)
where ei is the i-th standard basis of Rn. Therefore, the
condition can be formulated as
[qT ξT ]T ∈ Ui(r)
(6)
Ui(r) := {[qT ξT ]T (ei ⊗ I2)( ¯Lq − ¯Lξ) + δi(r − qi) ∈ Hu
i }.
i
(cid:1)Passivity-based distributed RG !"#$(cid:1)%&−&&0)(cid:1)PI consensus estimator *+(cid:1),!-#0.(cid:1)[!-#0](cid:1)(cid:2)(cid:1)(cid:2)(cid:1) (cid:1) (cid:1)1(cid:1)2(cid:1)(cid:4)(cid:1),23.(cid:1)Robotic network system i
Since Hu
i + Bu
(7)
i ∈ Rγ×2, where
using some Au
γ corresponds to the number of conditions specifying Hu
i .
Define
i [qT ξT ]T ≤ bu
i ∈ Rγ and Bu
(cid:21)
i ∈ Rγ×4n, bu
is a convex polytope, it is formulated as
Ui(r) :=(cid:8)[qT ξT ]T Au
(cid:20)bq
i r(cid:9)
(cid:21)
Ci(r) :=(cid:8)[qT ξT ]T Ai[qT ξT ]T ≤ bi + Bir(cid:9) .
Bu
i
Then, the set Ci(r) := Qi ∩ Ui(r) is given as
(8)
In the sequel, we assume that the set Ci, namely Ai, bi and
Bi, is a local information of robot i and the other robots
do not have access to these information.
(cid:20)Aq
(cid:20) 0
, Bi :=
, bi :=
Ai :=
i
Au
i
(cid:21)
i
bu
i
.
Let us now assume that the robotic network modifies the
reference r in order to ensure constraints (3) and (6), where
the modified reference is denoted by m. Suppose that a
constant signal m is added to (2) instead of r and define
qm = q − (1n ⊗ I2)m. System (2) becomes
(cid:20) ¯L + αr(D ⊗ I2) − ¯L
(cid:21)(cid:20) qm
(cid:20) qm
(cid:21)
(cid:21)
(9)
¯L
0
ξ
= −
ξ
Under Assumption 1, the matrix ¯L is symmetric and
positive semi-definite. Thus, defining
(cid:107)qm(cid:107)2 +
V (qm, ξ) :=
(cid:107)ξ(cid:107)2,
V ≤ 0 holds along the
it is immediately proved that
trajectories of (9) and hence any level set of the function is
positively invariant for system (9). More specifically, define
the set
1
2
1
2
Lm(c) := {[qT ξT ] V (qm, ξ) ≤ c}
Then, at a time t, the state trajectories never get out of
the set Lm(V (qm(t), ξ(t))) as long as the constant m will
be applied to the system in the future. Thus, if
Lm(V (qm(t), ξ(t))) ⊂ Ci(m),
constraints (3) and (6) are never violated. It is worth
to note that this set inclusion is a sufficient condition
for constraint fulfilment. It is well-known for linear-time
invariant systems like (2) that, once the constraint sets
Ci(r) (r = 1, 2, . . . , n) are given a priori, a necessary and
sufficient condition is provided through offline computa-
tion (Gilbert and Tan (1991)). However, this approach
cannot be taken in the present case, since these sets are
provided online according to the sensing information and
also no robot can collect all of the sets. This is why we
take the sufficient condition at the cost of conservatism.
In the next subsections, the global optimization problem
is formulated and then the local problem is derived.
3.1 Global Optimization Problem
(cid:88)
Since the constraints have to be satisfied for all i, the global
problem to be solved by the robotic network at time t is
i∈Vh
(cid:107)r − m(cid:107)2, subject to
min
m∈R2
Lm(V (qm(t), ξ(t))) ⊂ Ci(m) ∀i
Denoting the l-th row of Ai and Bi by A(l)
, re-
i
spectively, and the l-th element of bi by b(l)
, the constraint
Lm(V (qm(t), ξ(t))) ⊂ Ci in (11) becomes
i
and B(l)
(11)
(10)
i
(cid:18)
[((1n ⊗ I2)m)T 0]T − b(l)
A(l)
i
1
(cid:107)A(l)
i (cid:107)
(cid:20)(1n ⊗ I2)m
(cid:21)
(cid:19)
i − B(l)
i m ≤ 0 ∀l
(cid:21)(cid:13)(cid:13)(cid:13)(cid:13) ∀l (13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:20)(1n ⊗ I2)m
(cid:21)
+ b(l)
i + B(l)
i m
(cid:20)q(t)
−A(l)
(12)
ξ(t)
≥
−
0
0
i
At this point, define a function gi : R2×R2n×R2n such that
gi(m, q(t), ξ(t)) coincides with (12) and (13). Remark that
the constraint (12) is linear in m and hence convex. The
left-hand side of (13) is also a convex function in m and
the left-hand side is also proved to be convex by showing
positive semi-definiteness of its Hessian. Thus, the function
gi(m, q(t), ξ(t)) is convex in m for any given q(t) and ξ(t)
and the problem (10) and (11) is a convex optimization.
In summary, the problem which the robotic network solves
at time t can be compactly formulated as
(cid:88)
(cid:107)r − m(cid:107)2 subject to:
min
m
gi(m, q(t), ξ(t)) ≤ 0 ∀i
i∈Vh
(14)
(15)
Following the strategy of the standard reference governors,
the optimal solution to the problem is applied to the
system and then, at the next time t + 1, the network
again solves the problem by replacing q(t) and ξ(t) by
new measurements q(t + 1) and ξ(t + 1). Remark now that
since the cost function is strictly convex, if there exists
an optimal solution for given q(t) and ξ(t), it must be
unique. However, feasibility of the problem is not always
ensured depending on the set Qi, namely the locations
of the obstacles and of the robots. Although the issue is
basically left to future works, an approach is to expect
the human decision to be flexible enough to overcome the
problem.
Moreover, another problem is faced and this problem is
the main focus of this paper. Problem (14),(15) depends
on Ai, bi and Bi for all i but they are local information
as mentioned above. To ensure the information restriction,
we need to solve the problem in a distributed fashion. To
this end, we equivalently transform (14) and (15) into
min
z=(m,zq,zξ)
(cid:107)r − m(cid:107)2 subject to:
gi(m, zq, zξ) ≤ 0 ∀i
(ei ⊗ I2)zq = qi(t) and (ei ⊗ I2)zξ = ξi(t) ∀i
(16)
(17)
(18)
From the equivalence to (14) and (15), if (16) -- (18) is
feasible, the optimal solution to (16) -- (18) is also unique.
Following the same procedure as above, we can also
confirm that each element of gi is convex in z and hence
it is also a convex optimization.
In the next subsection, the local optimization problem is
derived from the global problem.
3.2 Local Optimization Problem
Let us now decompose the problem (16) -- (18) into n local
problems as follows.
(cid:88)
i∈Vh
fi(z) subject to:
(19)
(20)
z=(m,zq,zξ)
min
gi(z) ≤ 0
hi(z) = 0
(21)
where fi(z) = (cid:107)r − m(cid:107)2 if i ∈ Vh and fi(z) = 0 otherwise,
and hi
is defined so that hi(z) = 0 is equivalent to
(ei ⊗ I2)zq = qi(t) and (ei ⊗ I2)zξ = ξi(t). Then, the
local problem consists only of the local information Ai,
bi and Bi. Moreover, it does not depend on the states of
other robots. Thus, if each robot i with the local problem
(19) -- (21) were able to compute the optimal solution to the
global problem (16) -- (18), they would not need to share the
state information among robots. For future developments,
denote the k-th element of gi by gik and the l-th element
of hi by hil. Moreover, define z∗ as the optimal solution of
(19) -- (21).
In the next section, we present a solution to the above
problem. Although several solutions have already been
presented in the literature, we propose a novel passivity-
based solution since it allows one to integrate other passive
components like communication delays while ensuring
the entire system stability as exemplified in Hatanaka
et al. (2016). A passivity-based distributed optimization
algorithm is already presented in Hatanaka et al. (2016),
but the equality constraint like (21) is not taken into
account therein. We thus extend Hatanaka et al. (2016)
to problems with equality constraints in order to apply
the algorithm to the above reference governor problem.
4. DISTRIBUTED OPTIMIZATION ALGORITHM
AND PROOF OF CONVERGENCE FOR A STATIC
PROBLEM
In this section, a scheme to solve (19) -- (21) is proposed.
The proof of the states convergence of the scheme to the
optimal solution is provided using passivity arguments
when the physical system is static,
q = 0. Then,
in the next section, the effectiveness of the proposed
solution considering the dynamics of the robotic system
is demonstrated through simulations and experiments.
i.e.
4.1 Passivity-Based Distributed Optimization Scheme
This subsection gives details about the algorithm to solve
(19) -- (21). The RG block in Fig. 1 is detailed in Fig. 2,
using the state estimate vector z = [zT
n ]T , where
zi is the estimate of the optimal solution z∗ to (19) -- (21)
and zi ∈ R(4n+2). For the sake of completeness, consider
mi ∈ N0 inequality constraints and pi ∈ N0 equality
constraints for agent i.
1 , . . . , zT
The dual problem associated to (19) -- (21) is
n(cid:88)
i=1
Hi(λi, νi)
λi ≥ 0,
maximize
subject to
where Hi(λi, νi) := minz∈RN Li(z, λi, νi) is the Hamilto-
nian with Li(z, λi, νi) := fi(z) + λigi(z) + νihi(z), and
λi ∈ Rmi, νi ∈ Rpi (i = 1, . . . , n) are the Lagrange multi-
i=1 Hi(λi, νi)
i=1 Li(z, λi, νi), where
pliers. Define the dual function H(λ, ν) :=(cid:80)n
and the Lagrangian L(z, λ, ν) :=(cid:80)n
λ := [λT
1 , . . . , λT
n ]T and ν := [νT
1 , . . . , νT
n ]T .
Fig. 2. Proposed passivity-based distributed optimization
architecture, dealing with inequality and equality con-
straints. I4n+2, Im, Ip, In are the 4n + 2, m, p, n-
dimensional
identity matrices, respectively, where
i=1 mi and p =:=(cid:80)n
m :=(cid:80)n
i=1 pi.
Since strong duality holds, the optimal solution of (22),
denoted as d∗, satisfies
d∗ = z∗.
i=1
i=1
i=1
(23)
Moreover, z∗ satisfies the Karush-Kuhn-Tucker (KKT)
conditions (Boyd and Vandenberghe (2004))
n(cid:88)
n(cid:88)
∇fi(z∗) +
i ≥ 0, gi(z∗) ≤ 0, ∀i = 1, . . . , n
λ∗
ikgik(z∗) = 0, ∀k = 1, . . . , mi,∀i = 1, . . . , n,
λ∗
i ∇gi(z∗) +
λ∗
n(cid:88)
i ∇hi(z∗) = 0
ν∗
i ∈ Rmi is the optimal Lagrange multiplier and
where λ∗
λ∗
ik is the k-th element of λ∗
i .
The main idea to solve (19) -- (21) is to minimize the
global objective function f (z) through the gradient descent
method (Nedic and Ozdaglar (2009)) while maximizing
the dual function H(λ, ν) through the dual ascent method
(Boyd et al. (2011)). Then, using the PI consensus estima-
tor, the agents will cooperate to converge to the optimal
solution.
(24)
(22)
The next subsection provides the algorithm of the passivity-
based distributed optimization scheme.
Algorithm The algorithm dynamics depicted in Fig. 2 is
(cid:20) z
(cid:21)
ζ
(cid:21)
(cid:20) L −L
(cid:21)(cid:20) z
(cid:21)
(cid:20) ΓT (z)λ
L 0
ζ
− α
0
(cid:20) φ(z)
(cid:21)
(cid:21)
(cid:20) ΨT (z)ν
− α
0
0
= −
− α
(25)
(cid:1)1"#$%&'(cid:1)(cid:1)(cid:1)[#'%&'0](cid:1)+,−,,0.(cid:1)/#'%&'00(cid:1)1(cid:1)Agent i's optimization dynamics PI consensus estimator 2(cid:1)3(cid:1)#4"(cid:1)Γ(cid:1)6(cid:1)#7"(cid:1)8(cid:1)9(cid:1)Gradient descent method :(cid:1)−(cid:1)−(cid:1)−(cid:1)+(cid:1)+(cid:1)=:>?(cid:1)Dual ascent method @(cid:1)/::0(cid:1)/AB0(cid:1)C(cid:1)E(cid:1)C(cid:1)E(cid:1)−(cid:1)+(cid:1) where ζ ∈ RN is the additional
integral variable, L
the Graph Laplacian (Ren and Beard (2008)) associated
to the graph G, α ∈ R the passivity-based distributed
optimization gain, φ(z) := [φ1(z1)T , . . . , φn(zn)T ]T , where
φi := ∇fi is the gradient of the global objective function,
and the functions Γ(z) and Ψ(z) are
∇g1(z1)T
∇h1(z1)T
Γ(z) :=
Ψ(z) :=
,
.
. . .
. . .
O
∇gn(zn)T
O
∇hn(zn)T
(26)
(27)
The Lagrange multipliers λ, ν are updated as follows. First,
define the constraint matrices G := [gT
n ]T and
H := [hT
1 , . . . , gT
n ]T . The update algorithm for λ is
1 , . . . , hT
λ = η(λ, G),
η1,1 . . . η1,m
...
...
ηn,1 . . . ηn,m
. . .
(28)
is
where the switch function η(λ, G) :=
(cid:26)0,
ηij(λij, gij) :=
gij, otherwise,
if λij = 0 and gij < 0
(29)
where λij (i = 1, . . . , n, j = 1, . . . , mi) denotes the j-th
element of λi. This switch block η ensures λi ≥ 0 (see
KKT conditions (24)), where the initial condition must be
λi(0) ≥ 0.
The update algorithm of ν is
ν = h(z).
(30)
4.2 Convergence to the Optimal Solution
Define z∗ := 1n ⊗ z∗ as the goal state. This subsection
will prove the convergence to the optimal solution in three
steps:
(1) z∗ is a point of equilibrium of the system;
(2) the optimization scheme is passive;
(3) proof of asymptotic convergence to the optimal solu-
tion z∗.
Point of Equilibrium We will prove that z∗ is a point of
equilibrium of the system.
1 , . . . , λ∗T
Define the optimal Lagrange multipliers λ∗ = [λ∗T
n ]T
and ν∗ = [ν∗T
n ]T . The following lemma holds true.
Lemma 2. There exists ζ∗ such that (z∗, ζ∗) is an equilib-
rium of (25) for the equilibrium input Γ(z∗)λ∗ + Ψ(z∗)ν∗.
In addition, (λ∗, ν∗) is an equilibrium of (25) for the
equilibrium input z∗.
1 , . . . , ν∗T
Proof. See the proof of Lemma 7 in Hatanaka et al.
(2016).
Passivity Property We will study the passivity property
of the passivity-based distributed optimization scheme.
Define u ∈ R4n+2 as the input of the PI consensus
estimator/gradient descent method subsystem as depicted
in Fig. 2, zc = z − z∗ as the shifted state estimate,
ζ := ζ − ζ∗ as the shifted integral additional state, and
∆ := [µT , ωT ]T as the output of the dual ascent method
subsystem, where [µT , ωT ]T := [Γ(z)λ, Ψ(z)ν]T (see Fig.
2). For the sake of completeness, define the shifted input
∆ := [µT , ωT ]T , where [µT , ωT ]T := [(µ − Γ(z∗)λ∗)T , (ω −
Ψ(z∗)ν∗)T ]T .
First, the passivity of the PI consensus estimator/gradient
descent method subsystem is proved (see blue block in Fig.
2).
Lemma 3. The PI consensus estimator/gradient descent
method subsystem is passive from u := u − u∗ to zc with
2ζ2, where
respect to the storage function S := 1
u∗ := −αφ(z∗).
2zc2+ 1
Proof. The proof can be found in Hatanaka et al. (2015).
At this point, the passivity of the dual ascent method
subsystem is proved (see gray block in Fig. 2).
Lemma 4. The dual ascent method subsystem is passive
from Z := [zT
c ]T to ∆ with respect to the storage
function
c , zT
Ui, Ui :=
1
2
(λi − λ∗
i + νi − ν∗
i ) ,
where the initial condition for λi is λi(0) ≥ 0.
Proof. The proof consists in proving that (cid:80)n
i=1
∆T Z.
(31)
Ui ≤
n(cid:88)
i=1
The first step is to compute the time derivative of Ui in
(31), which is
m(cid:88)
j=1
p(cid:88)
ij and νij := νij − ν∗
ij.
j=1
Ui =
λijηij(λij, gij(zi)) +
νijhij(zi),
(32)
where λij := λij − λ∗
The first sum of (32) is analyzed. The KKT conditions (24)
and the switch block (29) are used to create an inequality.
Following from the switch block (29), the terms in the first
sum of (32) becomes
(cid:26)λijgij(zi) + λ∗
λijgij(zi),
λijηij =
ijgij(zi),
if λij = 0 and gij(zi) < 0
otherwise.
(33)
In the first case of (33), gij(zi) < 0 and following from
KKT conditions (24),
ijgij ≤ 0.
λ∗
Following from (33),(34), we deduce
λijηij(λij, gij(zi)) ≤ λijgij(zi).
As a consequence,
Ui ≤ (λi − λ∗
i )T gi(zi) + (νi − ν∗
i )T hi(zi).
(34)
(35)
(36)
At this point, it is needed to make the gradients appear
in the inequality to prove passivity. Using artifices to
introduce gi(z∗) and hi(z∗) in the inequality, Eq. (36) can
be rewritten as
Ui ≤(λi − λ∗
i )T{gi(zi) − gi(z∗)} + (λi − λ∗
i )T gi(z∗)
+ (νi − ν∗
i )T{hi(zi) − hi(z∗)} + (νi − ν∗
i )T hi(z∗).
(37)
Note that (λ∗
i )T gi(z∗) = 0 from KKT condition (24).
i gi(z∗) ≤ 0 holds, and since the optimal solution
Since λT
z∗ satisfies the equality constraint hi(z∗) = 0, Eq. (37)
becomes
Ui ≤ (λi−λ∗
i )T{hi(zi)−hi(z∗)},
(38)
i )T{gi(zi)−gi(z∗)}+(νi−ν∗
which can be rewritten as
[λij{gij(zi) − gij(z∗)} − λ∗
ij{gij(zi) − gij(z∗)}]
Ui ≤ m(cid:88)
p(cid:88)
j=1
+
[νij{hij(zi) − hij(z∗)} − ν∗
ij{hij(zi) − hij(z∗)}].
j=1
(39)
5. COMPLETE PROPOSED PASSIVITY-BASED
SCHEME
In this section, the schemes of Fig. 1 and 2 are combined as
shown in Fig. 3. The proof of the convergence of the entire
system is not provided and is let for future works. Instead,
the effectiveness of the proposed method in real-time is
demonstrated through simulations and experiments.
Consider the terms in the first sum of (39). Because of the
convexity of gij, Boyd and Vandenberghe (2004) proves
that (cid:26)gij(zi) − gij(z∗) ≥ (∇gij(z∗))T (zi − z∗),
gij(zi) − gij(z∗) ≤ (∇gij(zi))T (zi − z∗).
(40)
Since hij is affine, hij is also convex and ∇hij is constant,
i.e.
∇hij(z∗) = ∇hij(zi).
As a consequence, we can deduce that
(cid:26)hij(zi) − hij(z∗) = (∇hij(z∗))T (zi − z∗),
hij(zi) − hij(z∗) = (∇hij(zi))T (zi − z∗).
(41)
(42)
Therefore, since λi ≥ 0 and λ∗
(39), the gradients appear as
i ≥ 0, using (40) and (42) in
Ui ≤{(∇gi(zi))T λi − (∇gi(z∗)T )λ∗
i }T (zi − z∗)
+ {(∇hi(zi))T νi − (∇hi(z∗)T )ν∗
i }T (zi − z∗).
(43)
n(cid:88)
Therefore, using (43) in (31), we deduce
Ui ≤ [(Γ(z)λ − Γ(z∗)λ∗)T + (Ψ(z)ν − Ψ(z∗)ν∗)T ]zc
i=1
= µzc + ν zc = ∆T Z,
which concludes the proof.
Fig. 3. Complete scheme considering the dynamics of the
(44)
robotic network system.
5.1 Simulations
In the next subsection, using hybrid Lassale's principle, we
can prove asymptotic convergence to the optimal solution
z∗.
Convergence At this point, we proved that (z∗, ζ∗) is a
point of equilibrium of the system (Lemma 2) and that the
two subsystems (blue and gray blocks in Fig. 2) are passive
(Lemmas 3 and 4). Using hybrid Lassale's principle, the
following theorem proves that the states converge to the
optimal solution z∗.
Theorem 5. Consider the system (25),(28),(30). The state
estimate of the i-th agent zi asymptotically converges to
the optimal solution z∗ for all i ∈ V.
Proof. See the proof of Theorem 3 in Hatanaka et al.
(2016).
In the next section, simulations and experiments are car-
ried out to demonstrate the effectiveness of the passivity-
based distributed optimization scheme to solve (19) -- (21)
in real-time.
The system is composed of 5 agents, which communicate
in a circle according to the adjacency matrix
0 1 0 0 1
1 0 1 0 0
0 1 0 1 0
0 0 1 0 1
1 0 0 1 0
.
A =
(45)
The RG (25) is implemented to the pre-stabilized system
(2).
The leader of the system is agent 5. The formation is
added to the system through biases with respect to the
leader. The formation is a triangle formation, which is
[−1, 1]T , [−1,−1]T , [0,−1]T , [1,−1]T for agents 1, 2, 3 and
4, respectively, biased from agent 5. The initial conditions
are q = [−2, 0.7,−1.4,−1, 0.2,−1.4, 1.2,−3]T , and ξ and
all the state estimate variables zi are initialized at 0. The
parameters of the system are α = 2, αr = 1. For simplicity,
input constraints are neglected. The line detected by the
leader is x + y = 3(m).
(cid:1)!"#$(cid:1)%&−&&0)(cid:1)PI consensus estimator *+(cid:1),!-#0.(cid:1)[!-#0](cid:1)(cid:2)(cid:1)(cid:2)(cid:1) (cid:1) (cid:1)1(cid:1)2(cid:1)(cid:4)(cid:1),23.(cid:1)Robotic network system (cid:1)1"#$%&'(cid:1)(cid:1)(cid:1)[#'%&'0](cid:1)+,−,,0.(cid:1)/#'%&'00(cid:1)1(cid:1)Agent i's optimization dynamics PI consensus estimator 2(cid:1)3(cid:1)#4"(cid:1)Γ(cid:1)6(cid:1)#7"(cid:1)8(cid:1)9(cid:1)Gradient descent method :(cid:1)−(cid:1)−(cid:1)−(cid:1)+(cid:1)+(cid:1)=:>?(cid:1)Dual ascent method @(cid:1)/::0(cid:1)/AB0(cid:1)C(cid:1)E(cid:1)C(cid:1)E(cid:1)−(cid:1)+(cid:1) Consider first a reference chosen close to the constraint
r = [1, 2]T . The passivity-based distributed RG modifies
the reference r so that constraints are enforced during the
transients as shown in Fig. 4. Then, consider an inadmis-
the built Simulink program is D-Space 1.1.0.4. Fig. 6 shows
the experimental environment.
Fig. 4. Simulations for a reference close to the constraint.
Trajectories and modified references of the agents for
r = [1, 2]T .
sible reference r = [2, 3]T . The passivity-based distributed
RG modifies the reference r so that the reference remains
admissible and ensures constraints satisfaction as shown
in Fig. 5.
Fig. 6. Experimental environment captured by cameras.
The agents communicate in a circle according to the
adjacency matrix (45). The RG (25) is implemented to the
pre-stabilized system (2). The leader of the system is agent
5. The formation is added to the system through biases
with respect to the leader. The formation is a triangle
formation similar to subsection 5.1. The initial conditions
are q = q(0) and r = q(0), where q(0) is the initial
real position of the robots, and ξ is initialized at 0. The
parameters of the system are α = 7, αr = 2.5. The line
detected by the leader is x + y = 3(m).
The first experiment consists in moving the robots to a
reference close to the constraint solving (19) -- (21) at each
time instant. As seen in Fig. 7, the trajectories followed
by each robot do not violate the constraints at each time
instant. Note the small static error due to the difference
of static friction that depends on the working area.
Fig. 5. Simulations for an inadmissible reference. Tra-
jectories and modified references of the agents for
r = [2, 3]T .
Fig. 7. Experimental results with a reference close to the
constraint. Trajectories and modified references of the
agents for r = [1.4, 1.18]T .
5.2 Experimental Results
The experimental environment consists of five TDO-48
robots, whose positions are captured by a 30-fps Fire
Fly MV camera. The input signal of the robots' motors
is a PWM signal, which is sent through a Bluetooth
communication using OpenCV 2.4.11. The platform to run
The second experiment consists in moving the robots to
a reference chosen outside the admissible region. As seen
in Fig. 8, the system behaves as expected and the tra-
jectories followed by each robot do not violate the con-
straints at each time instant. The video of the experiments
can be found on https://www.youtube.com/watch?v=
TeEE9gXt3qQ.
-2-101234x position (m)-4-2024y position (m)ConstraintInitial positionFinal positionReference02040time (s)00.20.40.60.811.2x reference (m)02040time (s)00.511.522.5y reference (m)-2-101234x position (m)-4-2024y position (m)ConstraintInitial positionFinal positionReference02040time (s)00.511.52x reference (m)02040time (s)00.511.522.53y reference (m)0.511.52x position (m)00.511.522.5y position (m)ConstraintInitial positionFinal positionReference0204060time (s)0.811.21.4x reference (m)0204060time (s)0.40.60.811.2y reference (m) Gilbert, E. and Kolmanovsky, I. (2001). A generalized ref-
erence governor for nonlinear systems. In Decision and
Control, 2001. Proceedings of the 40th IEEE Conference
on, volume 5, 4222 -- 4227. IEEE.
Gilbert, E.G., Kolmanovsky, I., and Tan, K.T. (1994).
Nonlinear control of discrete-time linear systems with
state and control constraints: A reference governor with
global convergence properties. In Decision and Control,
1994., Proceedings of the 33rd IEEE Conference on,
volume 1, 144 -- 149. IEEE.
Gilbert, E.G. and Tan, K.T. (1991). Linear systems
with state and control constraints: The theory and
application of maximal output admissible sets.
IEEE
Transactions on Automatic control, 36(9), 1008 -- 1020.
Hatanaka, T., Chopra, N., Ishizaki, T., and Li, N. (2016).
Passivity-Based Distributed Optimization with Com-
munication Delays Using PI Consensus Algorithm.
ArXiv e-prints.
Hatanaka, T., Chopra, N., and Fujita, M.
(2015).
Passivity-based bilateral human-swarm-interactions for
cooperative robotic networks and human passivity anal-
ysis.
In 2015 54th IEEE Conference on Decision and
Control (CDC), 1033 -- 1039. IEEE.
Lee, D. and Spong, M.W. (2005). Bilateral teleoperation of
multiple cooperative robots over delayed communication
networks: theory.
In Proceedings of the 2005 IEEE
International Conference on Robotics and Automation,
360 -- 365. IEEE.
Mart´ınez, S., Bullo, F., Cort´es, J., and Frazzoli, E. (2007).
On synchronous robotic networkspart ii: Time complex-
ity of rendezvous and deployment algorithms.
IEEE
Transactions on Automatic Control, 52(12), 2214 -- 2226.
Mostofi, Y., Malmirchegini, M., and Ghaffarkhah, A.
(2010). Estimation of communication signal strength in
robotic networks. In Robotics and Automation (ICRA),
2010 IEEE International Conference on, 1946 -- 1951.
IEEE.
Nedic, A. and Ozdaglar, A. (2009). Distributed sub-
gradient methods for multi-agent optimization. IEEE
Transactions on Automatic Control, 54(1), 48 -- 61.
Nguyen, T., Mamoru, D., Hatanaka, T., Garone, E., and
Fujita, M. (2016). A passivity-based approach for
constrained mobile robotic networks. In 2016 55th IEEE
Conference on Decision and Control (CDC). IEEE.
Reich, J. and Sklar, E. (2006). Robot-sensor networks for
search and rescue. In IEEE International Workshop on
Safety, Security and Rescue Robotics, volume 22.
Ren, W. and Beard, R.W. (2008). Distributed consensus
in multi-vehicle cooperative control. Springer.
Shokri-Ghaleh, H. and Alfi, A. (2014). A comparison be-
tween optimization algorithms applied to synchroniza-
tion of bilateral teleoperation systems against time delay
and modeling uncertainties. Applied Soft Computing, 24,
447 -- 456.
Soleymani, T., Garone, E., and Dorigo, M. (2015). Dis-
tributed constrained connectivity control for proxim-
ity networks based on a receding horizon scheme.
In
2015 American Control Conference (ACC), 1369 -- 1374.
IEEE.
Thorpe, C. and Matthies, L. (1984). Path relaxation: Path
planning for a mobile robot. In OCEANS 1984, 576 -- 581.
IEEE.
Fig. 8. Experimental results with an inadmissible ref-
erence. Trajectories and modified references of the
agents for r = [1.8, 2]T .
6. CONCLUSIONS
A passivity-based distributed RG has been proposed for
the control of constrained mobile robotic networks. The
distributed RG problem is solved through a passivity-
based optimization scheme. This scheme solves a convex
optimization problem that is subject to inequality and
equality constraints. Passivity arguments are used to prove
the convergence of the states to the optimal solution. The
main limitation of the passivity-based distributed RG lies
in the fact that the problem may be infeasible. In the
present paper, the convergence of the interconnected sys-
tem is demonstrated through simulations and experiments.
Future works will aim at proving the convergence of the
states of the complete system.
REFERENCES
Boyd, S., Parikh, N., Chu, E., Peleato, B., and Eckstein, J.
(2011). Distributed optimization and statistical learning
via the alternating direction method of multipliers.
Foundations and Trends R(cid:13) in Machine Learning, 3(1),
1 -- 122.
Boyd, S. and Vandenberghe, L. (2004). Convex optimiza-
tion. Cambridge university press.
Bullo, F., Cortes, J., and Martinez, S. (2009). Distributed
control of robotic networks: a mathematical approach to
motion coordination algorithms. Princeton University
Press.
Burger, M., Zelazo, D., and Allgower, F. (2014). Dual-
ity and network theory in passivity-based cooperative
control. Automatica, 50(8), 2051 -- 2061.
De La Croix, J.P. and Egerstedt, M.B. (2012). Control-
lability characterizations of leader-based swarm interac-
tions.
Fan, X., Alpcan, T., Arcak, M., Wen, T., and Ba¸sar, T.
(2006). A passivity approach to game-theoretic cdma
power control. Automatica, 42(11), 1837 -- 1847.
Freeman, R.A., Yang, P., and Lynch, K.M. (2006). Sta-
bility and convergence properties of dynamic average
consensus estimators. In Proceedings of the 45th IEEE
Conference on Decision and Control.
Ge, S.S. and Cui, Y.J. (2000). New potential functions
for mobile robot path planning. IEEE Transactions on
robotics and automation, 16(5), 615 -- 620.
0.40.60.811.21.41.61.82x position (m)00.511.522.5y position (m)ConstraintInitial positionFinal positionReference0102030time (s)0.60.811.21.41.6x reference (m)0102030time (s)0.511.52y reference (m) |
1612.08552 | 1 | 1612 | 2016-12-27T09:37:56 | A Hybrid Network/Grid Model of Urban Morphogenesis and Optimization | [
"cs.MA",
"physics.soc-ph"
] | We describe a hybrid agent-based model and simulation of urban morphogenesis. It consists of a cellular automata grid coupled to a dynamic network topology. The inherently heterogeneous properties of urban structure and function are taken into account in the dynamics of the system. We propose various layout and performance measures to categorize and explore the generated configurations. An economic evaluation metric was also designed using the sensitivity of segregation models to spatial configuration. Our model is applied to a real-world case, offering a means to optimize the distribution of activities in a zoning context. | cs.MA | cs | Raimbault, J., Banos, A. & Doursat, R. (2014) A hybrid network/grid model of urban morphogenesis and optimization.
Proceedings of the 4th International Conference on Complex Systems and Applications (ICCSA 2014), June 23-26, 2014,
Universite de Normandie, Le Havre, France; M. A. Aziz-Alaoui, C. Bertelle, X. Z. Liu, D. Olivier, eds.: pp. 51-60.
Proceedings of ICCSA 2014
Normandie University, Le Havre, France - June 23-26, 2014
A Hybrid Network/Grid Model of
Urban Morphogenesis and Optimization
Juste Raimbault ∗, Arnaud Banos †, and Ren´e Doursat ‡
6
1
0
2
c
e
D
7
2
]
A
M
.
s
c
[
1
v
2
5
5
8
0
.
2
1
6
1
:
v
i
X
r
a
Abstract. We describe a hybrid agent-based model and sim-
ulation of urban morphogenesis. It consists of a cellular au-
tomata grid coupled to a dynamic network topology. The in-
herently heterogeneous properties of urban structure and func-
tion are taken into account in the dynamics of the system. We
propose various layout and performance measures to catego-
rize and explore the generated configurations. An economic
evaluation metric was also designed using the sensitivity of
segregation models to spatial configuration. Our model is ap-
plied to a real-world case, offering a means to optimize the
distribution of activities in a zoning context.
Keywords.
agent-based modeling, cellular automata, bi-
objective pareto optimization, evidence-based urbanism, ur-
ban morphogenesis.
1
Introduction
Recent progress in many disciplines related to urban plan-
ning can be interpreted as the rise of a "new urban sci-
ence" according to Batty [7]. From agent-based mod-
els in quantitative geography [17], in particular the suc-
cessful Simpop series by Pumain et al. [30], to other ap-
proaches termed "complexity theories of cities" by Portu-
gali [28], involving physicists of information theory such
as Haken [15] or architects of "space syntax theory" such
as Hillier [18], the field is very broad and diverse. Yet,
all these works share the view that urban systems are
quintessentially complex systems,
i.e. large sets of ele-
ments interacting locally with one another and the en-
vironment, and collectively creating a emergent struc-
ture and behavior. Taking into account the intrinsic het-
erogeneity of geographical and urban systems, this view
lends itself naturally to an agent-based modeling and sim-
ulation (ABMS) approach.
∗Graduate School, ´Ecole Polytechnique, Palaiseau, France; and
LVMT, Ecole Nationale des Ponts et Chauss´ees, Paris, France.
E-mail: [email protected]
†G´eographie-cit´es, CNRS UMR8504, Paris, France. E-mail:
‡Complex Systems Institute, Paris Ile-de-France (ISC-PIF),
[email protected]
CNRS UPS3611. E-mail: [email protected]
Among the most popular ABMS methods are cellular
automata (CA), in which agents are cells that have fixed
locations on a grid and evolve according to the state of
their neighbors. CA models of urban planning, in particu-
lar the reproduction of existing urban forms and land-use
patterns, have been widely studied, notably by White and
Engelen [39], then analyzed [5, 9] and synthesized [6] by
Batty. A recent review by Iltanen [19] of CA in urban
spatial modeling shows a great variety of possible system
types and applications. They include, for example, "mi-
croeconomic" CA for the simulation of urban sprawl [11],
"linguistic" CA (including real-time rule update via feed-
back from the population) for the measure of sustain-
able development in a fast growing region of China [41],
and one-dimensional CA [26] showing discontinuities and
strong path-dependence in settlement patterns.
In this context, we propose a hybrid model of urban
growth that combines a CA approach with a graph topol-
ogy containing long-range edges. It is inspired by Moreno
et al.'s work [24, 25], which integrates a network dynam-
ics in a CA model of urban morphology. Its goal was to
test the effects of physical proximity on urban develop-
ment by introducing urban mobility in a network whose
evolution was coupled with the evolution of urban shape.
We generalize this type of model to take into account het-
erogeneous urban activities and the functional properties
that they create in the urban environment. This idea
was introduced by White [38] and explored by van Vliet
et al. [36] but, to our knowledge, never considered from
the perspective of physical accessibility and its impact on
sprawl patterns.
The rest of the paper is organized as follows. The
model and indicator functions used to quantify the gener-
ated patterns are explained in Section 2. Next, Section 3
presents the results of internal and external validations
of the model by sensitivity analysis and reproduction of
typical urban patterns. It is followed by an application
to a concrete case, proposing a bi-objective optimization
heuristic of functional configuration based on the rele-
vant objective functions from the validation study. We
end with a discussion and conclusion in Sections 4 and 5.
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
51
J. Raimbault, A. Banos & R. Doursat
2 Model description
2.1 Agents and rules
The world is represented by a square lattice (Li,j)1≤i,j≤N
composed of cells that are empty or occupied (Fig. 1).
This is denoted by a function δ(i, j, t) ∈ {0, 1}, where
time t follows an iterative sequence t ∈ T = τN =
{0, τ, 2τ, ...} [14] with a regular time step τ . Another
evolving structure is laid out on top of the lattice: a
Euclidean network G(t) = (V (t), E(t)) whose vertices V
are a finite subset of the world and edges E (its agents)
represent roads. In the beginning, the lattice is empty:
δ(i, j, 0) = 0, and the network is either initialized ran-
domly (e.g. uniformly) or set to a user-specified configu-
ration G(0) = (V0, E0). In order to translate functional
mechanisms into the growth of a city, we assume that
the initial vertices include a subset formed by city cen-
ters, C0 ⊂ V0, which have integer activities, denoted by
a : C0 → {1, . . . , amax}.
To characterize the urban structures emerging in this
world, we define in general a set of k functions of the lat-
tice, (dk(i, j, t))1≤k≤K, called explicative variables. These
variables are here: d1, the density,
i.e. the average δ
around a cell (i, j) in a circular neighborhood of radius ρ;
d2, the Euclidean distance of a cell to the nearest road; d3
the network-distance of a cell to the nearest city center,
i.e. the sum of d2 and edge lengths; and d4, the accessi-
bility of activities (or rather difficulty thereof), written
(cid:32)
(cid:33)1/p4
amax(cid:88)
a=1
1
amax
d4(i, j, t) =
d3(i, j, t; a)p4
(1)
where d3(i, j, t; a) is the network-distance of the cell to the
nearest center with an activity a, and p4 ≥ 1 (typically 3)
defines a p-norm.
A set of weights (αk)1≤k≤K ∈ [0, 1]K is assigned to
these variables to tune their respective influence on what
we define as the net land value of a cell, as follows:
dk,max(t) − dk(i, j, t)
dk,max(t) − dk,min(t)
1(cid:80)
v(i, j, t) =
αk
.
(2)
K(cid:88)
k=1
k αk
Figure 1: Example of urban morphology generated by the
model. Houses (blue squares) were built in some cells of a
56 × 56 lattice. City centers and roads (red edges) compose
the added network. Cell shades (yellow) represent distances
to the built cells (the brighter, the closer).
Figure 2: The hybrid network/grid model. Blue arrows: feed-
back interactions. Red arrows: output evaluation functions.
of randomness. In any case, growth is halted after a con-
stant amount of time T , evaluated from experiments, so
that the final structure is neither "unfinished" nor filling
out the world (see 3.1). Fig. 2 displays the core ABMS
flowchart with feedback interactions between agents.
Houses are preferentially built where v is high, i.e. dk's
are low. Thus the evolution of the system proceeds in
three phases at each time step: (a) all values v(i, j) are
updated, (b) among the cells that have the best values, n
new cells are randomly chosen and "built" (set to δ = 1);
(c) for each built cell, if d2 is greater than a threshold θ2
(maximum isolation distance), then that cell is directly
connected to the network by creating a new road branch-
ing out orthogonally from the nearest edge.
Network initialization is random (see details in 3.1),
and the selection of new cells is also random among iden-
tical values of v. A sensitivity analysis and model explo-
ration is conducted in the next section to determine the
relative effect of parameters with respect to these sources
2.2 Evaluation functions
Once a structure is generated, its properties need to be
quantified so that it can be categorized or compared
to other structures for optimization purposes. To this
goal, we define various evaluation functions, both objec-
tive quantification measures and structural fitness values.
The measures described in this section take into account
all the explicative variables, whose distributions over the
grid are emergent properties that cannot be known in ad-
vance and are therefore essential to monitor.
Morphology To assess the morphological structure of
an urban configuration, we map it onto a 2D metric space
52
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
A Hybrid Network/Grid Model of Urban Morphogenesis and Optimization
defined by a pair of global indicators (D, I) called the
integrated density and the Moran index (Fig. 4). The
density D ∈ [0, 1] is calculated by taking the p-norm (with
exponent pD ≥ 1, typically 3) of the local densities d1:
1/pD
A(t) =
(3)
1(cid:80)
i,j δ(i, j, t)
N(cid:88)
i,j=1
δ(i,j,t)(cid:54)=0
D(t) =
d1(i, j, t)pD
Moran's I, an index of spatial autocorrelation, is widely
used in quantitative geography [35, 20] to evaluate the
"polycentric" character of a distribution of populated
cells. It is defined by
(cid:80)
(cid:80)M 2
µ(cid:54)=ν(Pµ − P )(Pν − P )/dµν
µ=1(Pµ − P )2
(4)
M 2(cid:80)
µ(cid:54)=ν 1/dµν
I(t) =
where the lattice is partitioned into M × M square areas,
at an intermediate scale between cell size and world size
(1 (cid:28) M (cid:28) N ), dµν is the distance between the centroids
of areas µ and ν, (Pµ)1≤µ≤M 2 denotes the number of oc-
cupied cells in each area, and P is their global average.
We can recognize in this formula the normalized ratio of
a modified covariance (pairwise correlations divided by
distances) and the variance of the distribution. Moran's
I belongs by construction to the interval [−1, 1], where
values near 1 correspond to a strong monocentric distri-
bution, values around 0 to a random distribution, and
values near −1 to a checkered pattern (every other cell
occupied). Usually, polycentric distributions have rela-
tively small positive I values, depending on the size and
distance between centers.
Network performance Due to the branching nature
of the growth algorithm, the network of roads G cannot
contain any other loops than the ones initially present in
G0. Therefore, notions of "clustering coefficient" or "ro-
bustness" (with respect to node removal) are not relevant
here. On the other hand, since G is intended to simulate
a mobility network, we can evaluate its performance by
defining a relative speed [4] S, representing the "detours"
imposed by G with respect to direct, straight travels:
S(t) =
1(cid:80)
i,j δ(i, j, t)
N(cid:88)
(cid:18) d3(i, j, t)
(cid:19)pS
e3(i, j, t)
i,j=1
δ(i,j,t)(cid:54)=0
1/pS
(5)
on the relative local accessibility from each cell, which is
d4 over its maximum:
1(cid:80)
i,j δ(i, j, t)
N(cid:88)
(cid:18) d4(i, j, t)
(cid:19)pA
d4,max(t)
i,j=1
δ(i,j,t)(cid:54)=0
1/pA
(6)
This normalization puts A in [0, 1] and allows comparing
configurations of different sizes. Like S, "better" urban
configurations are characterized by a lower A.
Economic performance It was shown by Banos [2]
that the Schelling segregation model, a standard ABM of
socio-economic dynamics [33], was highly sensitive to the
spatial structures in which it could be embedded, since
segregation rules depended on proximity. This justifies
the use of this model as an evaluation of economic perfor-
mance of our urban configurations, measuring how much
structure influences segregation. To this aim, we imple-
mented a model of residential dynamics based on the work
of Benenson et al. [10]. The output function is a segre-
gation index H(t) calculated on the residential patterns
that emerge inside a distribution of built patches. For ur-
ban structures produced in a practical case (Section 3.3),
we obtained densities of mobile agents between 0.1 and
0.2. Following Gauvin et al. [13], the phase diagram of
the Schelling model indicates that in such a density range,
tolerance thresholds of 0.4 to 0.8 lead to clustered frozen
states, where the calculation of a spatial segregation in-
dex is indeed relevant. The detailed description of this
economic model is out of the scope of this paper.
3 Results
Our hybrid network/grid model was implemented in Net-
Logo [40]. Plots and charts were created in R [32] from ex-
ported data. Processing of GIS Data (for vectorization by
hand of simple raster data) was done in QGIS [31]. Explo-
ration of the 4D space of explicative variables' weights αk
was conducted inside the [0, 1]4 hypercube with a linear
increment of 0.2. This created 64 − 1 = 1295 points, from
(0, 0, 0, 0) excluded to (1, 1, 1, 1) included, via (0.2, 0, 0, 0),
etc. Unless otherwise noted, the output values of the eval-
uation functions were averaged over 5 simulations for each
combination of the αk's.
where pS ≥ 1 (also 3), and e3(i, j, t) is the direct Eu-
clidean distance between cell (i, j) and the nearest city
center over the network, i.e. the one that realizes the value
of d3(i, j, t). Note that S ≥ 1 and is actually higher for
more convoluted networks (thus it is a measure of "slow-
ness", but we still employ "speed").
Functional accessibility The global functional acces-
sibility A to city centers is another p-norm (also 3), based
3.1 Generation
of
urban
patterns:
external validation of the model
Typical patterns We ran the model on different ini-
tial configurations, in which a few city centers C0 (typ-
ically 4) were randomly positioned on a 56 × 56 lattice,
and their activity values drawn in [1, amax] (both uni-
formly). The initial network G0 was built progressively
and quasi deterministically over increasing distances, by
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
53
J. Raimbault, A. Banos & R. Doursat
Figure 4: Morphological classification of urban patterns.
Scatterplot in the (D, I) evaluation plane with four typical
structures highlighted. Three of them, despite visual differ-
ences, are relatively close to each other in this space, indicat-
ing that two metrics are not sufficient for a full classification.
The area near the origin corresponds to unfinished patterns,
i.e. which occupy only a small part of the world and cannot be
compared with larger ones. Almost all I values were positive.
Figure 3: Typical patterns obtained from our model, repro-
ducing Le Corbusier's analysis of "human settlements". In his
1945 attempt to theorize urban planning, Le Corbusier ana-
lyzed the form of cities by hand and outlined three types of
settlements: radial-concentric cities, linear cities along com-
munication roads, and rural communities. We were able to
reproduce this typology by setting the weights of the ex-
plicative variables of our model to corner values: Top-right:
(αk) = (1, 0, 0, 0), i.e. density-based only. Middle: (0, 1, 0, 0),
i.e. distance-to-road only. Bottom: (0.2, 0, 1, 0), i.e. network-
distance combined with a little density. Left: source [22].
creating isolated clusters and linking them until they per-
colated into one component. The initial grid was empty
(δ = 0 everywhere). Simulations were cut off at 30 iter-
ations (T = 30τ ), before the sprawl of urban settlements
reached the square boundaries of the world and started
"reverberating". Since this artifact occurred the fastest
in a density-driven model, αk = (1, 0, 0, 0), we empirically
assessed T in that case and applied it everywhere.
Different parameter settings generated very diverse
structures. In particular, we observed striking similarities
between the patterns obtained for binary values of αk's
in some "corners" of the hypercube (one or two measures
dk with weight 1, the others 0), and the fundamental ur-
ban configurations that Le Corbusier had identified in his
1945 analysis of human settlements [22] (Fig. 3).
Classification of structures Using the pair of mor-
phological indicators (D, I) defined above, and by vary-
ing the αk's, we constructed a 2D map of the dynamical
regimes of our system (Figs. 4-5), in which qualitatively
different morphological "classes" could be distinguished.
The projected locations of urban configurations in this
plane allowed a better understanding and comparison of
their features and growth process. Again, for certain cor-
ner parameter values (all of them 0 except one or two at
1), the results ended up in distinct locations on the map,
Figure 5: Influence of each explicative variable dk on urban
morphogenesis. Color darkness corresponds to the relative
value of weight αk used during the growth of mapped struc-
tures. Whereas Figs. 4-5 showed distinct classes at expected
locations, this plot displays a rather uniform and chaotic dis-
tribution of high weights for d2, d3, and d4, revealing a perva-
sive role of roads, city centers, and accessibility. Only density
d1 correlates better with its own evaluation function D (a high
influence of density results in low-density patterns), except for
the low-I cluster on the right.
54
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
0.00.10.20.30.40.50.60.00.20.40.60.8Morphological classificationDensity (D)Moran's Index (I)centers only
D = 0.48,
I = 0.07
densities onlyD = 0.9,
I = 0.49
roads only
D = 0.45,
I = 0.08
roads and centersD = 0.58,
I = 0.13
"unachieved"
sprawls
0.00.10.20.30.40.50.60.00.20.40.60.8Density influenceDensityMoran0.00.10.20.30.40.50.60.00.20.40.60.8Accessibility influenceDensityMoran0.00.10.20.30.40.50.60.00.20.40.60.8Centers' influenceDensityMoran0.00.10.20.30.40.50.60.00.20.40.60.8Roads' influenceDensityMoran A Hybrid Network/Grid Model of Urban Morphogenesis and Optimization
Figure 6: Statistical distribution of the output evaluations.
For each of the 15 corner points of the 4D hypercube of αk's
(excluding the origin), we ran 500 simulations from random
initializations of 4 city centers C0. Three resulting distribu-
tions out of these 15 are displayed, each in the form of a his-
togram of evaluation function values, D, S, I, and A, fit-
ted with a Gaussian curve. Green: (αk) = (1, 0, 0, 0), i.e. a
simulation taking into account only the density d1. Yellow :
(0, 1, 0, 0), i.e. Euclidean distance d2 only. Red : (0, 0, 0, 1),
i.e. accessibility d4 only. These three histograms were chosen
for their minimum overlap and clarity of display; the other 17
are similar. The narrow peaks (except one), spread about the
mean by ±10%, attest to the low sensitivity of the model with
respect to the spatial initialization, and validates its internal
consistency. This also allowed us to rely on a smaller number
of runs in our experiments.
Figure 7: Assessing the influence of the update scheme on
the morphologies. In the (D, I) classification plane, each point
corresponds to 3 runs of a given combination of αk parameters,
repeated under a sequential (n = 1) and under a parallel (n =
20) update scheme. For each run, the symmetric difference ∆
between the two patterns is computed and its average over the
3 runs is projected on the map. The color of a point highlights
its "significance", defined as the product of its local density
(clustered points represent more frequent configurations) and
its pattern size, ∆ (large patterns are more significant). The
scattered points indicate that the model is sensitive to the
update scheme for certain parameters. On the other hand, the
concentration of significant points near the origin and D = 0.5
means that corner cases, such as (1, 0, 0, 0), are more robust.
which could be relatively well explained.
Intermediate
combinations of parameters, however, seemed to project
the structures quite literally "all over the map", which
might be interpreted as the emergence of chaos in the
system.
3.2 Sensitivity analysis and parameter
space exploration: internal validation
Sensitivity to initial conditions To ensure the va-
lidity of the results, we investigated the sensitivity of the
model to the spatial conditions, the initial set of nodes
C0, estimating in particular the number of repetitions
necessary to obtain statistically meaningful values for the
evaluation functions. If conclusions drawn from one case
were highly susceptible to small changes in the initial lay-
out, then the model would obviously have less significance
than if there was some invariance with respect to abstract
topological features (in particular the distribution of cen-
ters' activities). The optimization heuristics would have
to be designed very differently in these two cases.
Toward this assessment, we ran a large number of simu-
lations under the same parameter values but starting from
different initial C0 configurations, and collected statistics
on the output. For each of the 15 binary combinations
of αk's (excluding all zero), standard deviations were cal-
culated over 500 runs. We obtained narrow peak distri-
butions in most cases, with Gaussian widths typically at
10% of the mean function value (Fig. 6).
In order to
ensure that these were the typical widths on all the pa-
rameter space and not only on extreme binary points, we
also explored the grid {0; 0.5; 1}4 with 100 runs per point,
assessing in a more representative subspace the relative
spread of distributions. This confirmed that the eval-
uation functions were significantly less sensitive to the
exact spatial locations than the parameters and over-
all topology, and justified our use of a smaller number
of trials in subsequent experiments. Typically, assum-
ing a normal distribution of width σ = 0.1, we needed
n = (2σ·1.96/0.05)2 (cid:39) 60 trials to reach a 95% confidence
interval of length 0.05, and 5 trials for a length 0.17. For
practical reasons of computing speed, we chose the latter.
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
55
DensityFrequency0.00.20.40.60.81.001020304050MoranFrequency0.00.20.40.60.81.0051015202530SpeedFrequency1.01.52.02.53.0010203040AccessibilityFrequency0.00.20.40.60.81.00510152025300.10.20.30.40.50.20.40.60.8Morphologic imprint of differencesDensityMoran J. Raimbault, A. Banos & R. Doursat
Sensitivity to update scheme On the other hand,
the emergent urban patterns depended on the number n
of cells filled at every iteration, before land values were re-
calculated at the next iteration, i.e. whether the update
scheme was a sequential (n = 1) or parallel (n > 1).
Building several houses "simultaneously" between two
market reevaluations is consistent with the view that real-
world functions have a response delay, here of the order of
τ . There must be a limit, however, and an intermediate
n must be found to obtain reasonable simulations.
To this aim, we explored the 4D parameter space of the
αk's as in Figs. 4-5 and ran one sequential update scheme
and one parallel update scheme with n = 20 in each case.
At the end of the simulation, t = T , the two correspond-
ing output patterns δseq and δpar were compared by calcu-
lating their symmetric difference, i.e. the subset of lattice
cells that were built either in one or the other but not
in both: ∆ = {(i, j); δseq(i, j, T ) (cid:54)= δpar(i, j, T )}. Then,
these difference patterns ∆ were projected on the same
classification map (D, I) used previously (Fig. 7). The
results showed that for many combinations of parame-
ters, the model's behavior could be noticeably influenced
by the update scheme, as many difference patterns ex-
hibited a nontrivial structure with high density or high
Moran's index or both. On the other hand, it exhibited a
stronger invariance for the corner quadruplets of αk's: in
these cases the ∆'s clustered near the origin and D = 0.5.
Based on this study, we decided to adopt a parallel up-
date scheme with n = 15 built cells per time step in the
remainder of the experiments.
Figure 8: Sample surface plots of the evaluation functions.
For each 4D field of evaluation values in the hypercube, we
select two out of four parameters and display the 2D slice cor-
responding to the other two parameters set to (0, 0). Horizon-
tal axes are reoriented in each case to minimize visual clutter.
This exhaustive exploration has an intrinsic explanatory value
(see text), and allows us to predict with some level of confi-
dence how the model responds to certain input parameters.
Exploration of parameter space The above two pre-
liminary studies validated the robustness of the model
with respect to the initialization and update scheme, and
helped us choose a reasonable number of runs (about 5)
for each parameter combination, and a adequate degree
of parallelism in the simulations (n = 15). Next, we revis-
ited the αk hypercube (same 1295 points in the partition
of step 0.2), this time calculating the complete charts of
all evaluation functions. Other parameters with a direct
correspondence to the real-world, depending on the scale
adopted, were set to fixed values. For example, the neigh-
borhood radius ρ or the road-triggering distance θ2 were
both equal to 5 cells: this number could represent 50m,
characteristic of a block at the scale of a district, or 500m
for a district in a city, or 5km between cities in a region.
Examples of evaluation surfaces in 2D projection spaces
are shown in Fig. 8. Each function, D, I, S, and A, was
plotted against two parameters out of four, chosen for
their higher "influence" (variations in amplitude) on the
function. The economic index H was not calculated here
(see 3.3). This exhaustive exploration of parameter space
was necessary to gain deeper insight into the behavior of
the model. It also represents a crucial step toward making
computational simulations more rigorous [3].
Altogether, we observed that outputs varied for the
most part smoothly, except Moran's index which ap-
peared more chaotic. Variations were greater in cases
where one parameter was dominant. For example, the
measures of density D, speed S and (global) accessibil-
ity A all exhibited a significant jump when including
the effect of (local) accessibility d4 in the simulations,
i.e. when transitioning from α4 = 0 to α4 > 0. In par-
ticular, the more activities were influent, the denser the
city became-a nonintuitive emergent effect, compared
to top-down planning alternatives that would try to op-
timize accessibility while keeping density low. Speed,
or rather "sluggishness", exploded when density was the
only influence on urban sprawl: this confirmed that pure
density-driven dynamics creates anarchic growth, without
concern for network performance.
As for global accessibility, or rather the difficulty
thereof, it was minimal for α4 = 0 : an interesting para-
doxical effect suggesting that when individual agents took
into account local accessibility (α4 > 0), a few of them
might have occupied the "best spots" too quickly, signif-
icantly diminishing the others' prospects. Therefore, at
the collective level, it would be better for everyone to ig-
nore that dimension-an example where competition at
the individual level does not produce the most efficient
system for all. Finally, no meaningful conclusion could
be formulated about the chaotic variations of Moran's in-
dex, except for its extreme sensitivity to spatial structure.
56
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
alpha_40.00.20.40.60.81.0alpha_20.00.20.40.60.81.00.200.250.300.350.400.45Densityalpha_30.00.20.40.60.81.0alpha_10.00.20.40.60.81.00.10.20.30.4Moran Indexalpha_40.00.20.40.60.81.0alpha_10.00.20.40.60.81.01.41.61.82.0Speedalpha_40.00.20.40.60.81.0alpha_10.00.20.40.60.81.00.50.60.70.8Accessibility0.00.00.00.50.5 A Hybrid Network/Grid Model of Urban Morphogenesis and Optimization
Figure 9: Practical application: optimizing the distribution of activities over urban centers. Left: existing masterplan of the
Massy Atlantis district: 9 building areas are identified by their centers (black squares). Buildings are color-coded by type of
activity (orange: residential, blue: tertiary). Railroad tracks traverse the upper left corner (light blue strip) and three train
transportation hubs are also represented (green locomotive icons): these are part of the fixed environment. Right: example
of housing configuration obtained after a simulation in which the 9 centers' activities were initialized to the same values as
their real-world counterpart (red house icons: residential, blue computer icons: tertiary; one of 510 possible distributions).
The evaluation functions of the outcome are (H, A) = (0.067, 0.76), a point close to the Pareto front of Fig. 10 (blue circle).
3.3 Practical example
In this section, we apply our model to the optimization of
activities on top of a real-world urban structure obtained
from a geographic file, as opposed to an randomly gener-
ated, artificial configuration. This type of scenario occurs
in a planning problem where one must decide about the
possible land use of predefined zones.
The practical example under study here concerns the
It is based on a real-
planning of a new district.
world neighborhood, Massy Atlantis (Paris metropolitan
area), built in 2012 (Fig. 9). We would like to investi-
gate whether a more efficient planning could have been
achieved. The goal of this exercise is to find an optimal
assignment of two types of activities, "residential" (apart-
ments) or "tertiary" (offices), to the centers of 9 areas
located on a map. The transportation infrastructure is
already in place and the train station is also considered a
center with a fixed, third type of activity. A network of
avenues is laid out and passes through the 9 centers. The
district is initially empty (unbuilt). The particular spa-
tial configuration was automatically imported from a GIS
shapefile, so the computation could be readily applied to
other cases.
Parameters of the model were set as follows: high in-
fluence of activities, α4 = 1, reflecting the fact that ac-
cessibility to home, workplace, and train station are of
special importance to the agents of this district; medium
influence of density, α1 = 0.7, because, not far from Paris,
housing must reasonably fill the available areas; no influ-
ence of road proximity, α2 = 0, since the initial network
is already built and the scale is relatively small; and no
influence of network-distance, α3 = 0, because centers in
this problem are abstract entities representing areas.
For every possible distribution of binary activities over
the 9 areas, excluding the two uniform cases (all resi-
dential or all tertiary), the model was simulated 5 times,
producing a total of (29 − 2)× 5 = 2550 runs. The result-
ing configurations were examined here via a morphologi-
cal projection in the (H, A) plane, instead of (D, I) used
in the previous sections, as we judged it to be a more
meaningful measure of fitness in this application. The
calculation of the economic segregation index H involved
a secondary agent-based simulation on top of the main
urban development model (details not provided here).
Results are shown in Fig. 10. We obtained a Pareto
front of "optimal solutions" trying to minimize both H
and A, while observing that the actual configuration is
not far from being optimal itself, and appears to be a
compromise between accessibility and economic perfor-
mance. After closer examination of the Pareto front and
its vicinity, we found that the distribution of activities
was highly mixed in these points. More precisely, we de-
fined a spatial heterogeneity index of center activities by
(cid:80)
(cid:80)
d(c, c(cid:48))−1
a(c)(cid:54)=a(c(cid:48))
c(cid:54)=c(cid:48)
c(cid:54)=c(cid:48) d(c, c(cid:48))−1
λ = amax
(7)
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
57
J. Raimbault, A. Banos & R. Doursat
4 Discussion
The reproduction of typical urban morphologies and the
possible application to a real-world problem shown in the
previous sections indicate that a model like ours can be
useful for evidence-based decision-making in urban plan-
ning. Several questions remain open, however, and would
need further investigation.
Scale of the model One ambiguity of the model is
that it can be applied at different scales, therefore there
is no unique correspondence between its agents and the
real world. As the above results illustrate, the simulated
urban configurations may represent a system of cities at
the macroscopic scale, the neighborhoods of one city at
the mesoscopic scale, or the buildings of one district at
the microscopic scale. Without engaging in an ontolog-
ical debate over levels of abstraction, this could still be
pointed out as a potential issue.
We wish to argue, however, that the multiscale appli-
cability of our model is legitimate as a great number of
urban systems and associated dynamics have been shown
to be "scale-free", in particular by Pumain [29], and even
to possess fractal properties, by Batty [8]. It means that
scaling laws may also operate in our model, therefore
qualitative results should remained unchanged while the
quantitative evolution of variables and relations should
only depend on the underlying power law's exponent.
Barth´elemy [21] warns that most multi-agent urban
models fail because they do not focus on the "dominating"
physical process but, instead, integrate too many aspects
that bear no relevance to the emergent properties of the
system. Following up on this advice, we believe that we
have successfully identified "good" proxies for the domi-
nating processes of urban morphogenesis, namely: urban
density, accessibility to road network, and accessibility to
main functionalities.
Local scope When the model is considered at a meso-
scopic or microscopic scale, another objection could be
that it seems to limit itself to an artificially "closed" ur-
ban system, neglecting important contextual phenomena
such as economic exchanges. Yet, although input and out-
put flows are greatly simplified here, they are still present
in implicit form. Our simulated world is not truly closed,
since newly built houses are associated with a net influx
of resources. Moreover, despite the absence of a direct
economic force in the growth dynamics (the H index is
only a post-hoc metric), the attractivity of centers con-
stitutes a proxy for underlying activity, and a form of
interdependence among urban processes. Finally, other
models that have taken into account the global complex
network of cities [1] have reproduced well-known patterns
of urban systems much like ours.
Therefore, here too, local or global approaches appear
to be equivalent and the modeling decisions and compro-
mises made in each case must be compared. This ques-
Figure 10: Scatterplot of all configurations in the (H, A) mor-
phological plane. A Pareto front (red circles) is apparent in the
bottom left part of the plot: it corresponds to "optimal" con-
figurations trying to minimize both H and A objectives. The
real situation (blue circle in H = 0.067, A = 0.76) corresponds
to Fig. 9 and is not far from this front. Points are colored ac-
cording to their level of heterogeneity λ, from low (black) to
high (yellow). More homogeneous configurations are concen-
trated in the central cluster, whereas Pareto points and their
neighbors have higher heterogeneity levels. This lends sup-
port to the principle of "functional diversity", which is often
adopted by planners and urbanists today but has never been
backed up by computational simulations.
where c = (i, j) and c(cid:48) = (i(cid:48), j(cid:48)) are two centers, d(c, c(cid:48))
their Euclidean distance, and a(c), a(c(cid:48)) their activities.
Points in the scatterplot were colored according to their
level of λ. Highly heterogeneous configurations appeared
in regions of the plot distinct from homogeneous configu-
rations, which were for the most part located in the cen-
tral cluster. Optimal solutions and their neighbors all
corresponded to high heterogeneity. This interesting re-
sult is a step toward evidence-based justification of mixed
land use in planning-a principle often invoked by urban-
ists but never quantitatively demonstrated.
In conclusion, this case study is encouraging as it pro-
poses a concrete methodology of optimal planning with
respect to criteria that are relevant to a particular situa-
tion. It could be used by generically planners in similar
situations, while remaining cautious on the conditions of
its applicability. We discuss this point next.
58
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
0.0600.0650.0700.0750.0800.0850.0900.600.650.700.750.80Pareto plot of configurationsEconomic segregation (H)Distance to activities (A) A Hybrid Network/Grid Model of Urban Morphogenesis and Optimization
tion also ties in with the fundamental issue, contained in
the previous point, of the existence of a "minimal dimen-
sion" for a generalized representation of urban systems.
The challenge is to understand how universal the depen-
dence between a system and its dimension may be, and if
a generalized minimalist formulation can be constructed.
Speculations toward that ambitious goal have been for-
mulated by Haken [15] through a notion of "semantic in-
formation" linked to properties of attractors in dynamical
systems. This theory, however, has not been quantified,
i.e. neither confirmed nor falsified.
Quantitative calibration The question of the validity
of the model is also linked to the need for a finer quan-
titative calibration based on real patterns, which creates
a dilemma: on the one hand, calibration on the errors of
output function proxies does not influence the formation
of spatial patterns; on the other hand, calibration on the
spatial patterns themselves is too constraining and may
preclude the emergence of other, similar patterns. Pre-
vious works addressing the issue of calibration [23] have
not been conclusive so far.
To revisit this question, we would need to apply our
model at a finer grain of spatial resolution, i.e. a very
large world in terms of data size.
In this scenario, it
would be particularly important to keep processing time
under control by reducing computational complexity, for
example through a cache of the network's shortest paths.
The potential increase in size can also create methodolog-
ical hurdles, not just computational, as a huge amount
of details in the resulting patterns might contribute to
more noise than signal and significantly bias the indica-
tors. One solution would be to create a new operator
extracting the morphological envelope of the generated
pattern, along the lines of an original method proposed
by Frankhauser et al. [12, 34]. Other ways to deal with
noise may involve Gaussian smoothing.
Complex coupling with economic model Our
method of economic evaluation consists of "simple cou-
pling", i.e. running a secondary agent-based model (the
basis of H's calculation, not described here) after the pri-
mary urban growth simulation has finished. Another im-
portant direction of research would implement a "complex
coupling" between the two models in the sense proposed
by Varenne [37]: the study of urban sprawl on other time
scales would require the simultaneous and mutually in-
teracting evolutions of the population, the building rents,
and the terrain values. Obviously, this would lead to a
more sophisticated model oriented toward a whole new
set questions, such as the evaluation of long-term rent
policies to foster social diversity.
5 Conclusion
We have proposed a hybrid network/grid model of ur-
ban growth structures, and studied their morphological
and functional properties by simulation. Results showed
that it could reproduce the characteristic urban facts of a
classical typology of "human settlements", and was also
applicable to a concrete scenario by calculating "optimal"
solutions (in the Pareto sense) to a planning challenge in
an existing zoning context. Our work provide evidence
in favor of the "mixed-use city", a topic on which lit-
erature is still scarce and future work is needed. This
paradigm is now commonly advocated by urbanists, such
as Mangin [22] through his concept of "ville passante" (a
pun on "evolving/flowing/pedestrian city"), and would
require more validation through quantitative results.
Finally, beyond its technical achievements and poten-
tial usefulness as a decision-making tool, our work also
fuels a contemporary debate on the state-of-the-art in
"quantitative urbanism". Siding with Portugali [27], we
certainly agree that the conception and application of
computational models is a delicate matter, which can lead
to more confusion than explanation if not properly han-
dled and validated. Depending on the scale, a careless
choice of parameter values can produce dubious results.
Yet, we support the idea that quantitative insights are
paramount for a better understanding of urban and so-
cial systems. With the recent explosion in data size and
computing power, evidence-based analysis and solutions
are becoming a real alternative to older attitudes, such as
Lefebvre's [16], which doubted that scientific approaches
could ever translate or predict the mechanisms of a city.
References
[1] C. Andersson, A. Hellervik, K. Lindgren, A. Hagson, and
J. Tornberg. Urban economy as a scale-free network. Physical
Review E, 68(3):036124, 2003.
[2] A. Banos. Network effects in schelling's model of segregation:
new evidences from agent-based simulation. Environment and
Planning B: Planning and Design, 39(2):393–405, 2012.
[3] A. Banos. Pour des pratiques de mod´elisation et de simula-
tion lib´er´ees en G´eographie et SHS. PhD thesis, UMR CNRS
G´eographie-Cit´es, ISCPIF, D´ecembre 2013.
[4] A. Banos and C. Genre-Grandpierre. Towards new metrics for
urban road networks: Some preliminary evidence from agent-
based simulations. In Agent-based models of geographical sys-
tems, pages 627–641. Springer, 2012.
[5] M. Batty. Cellular automata and urban form: a primer.
Journal of the American Planning Association, 63(2):266–274,
1997.
[6] M. Batty. Cities and Complexity: Understanding Cities with
Cellular Automata, Agent-based Models, and Fractals. MIT
Press, 2007.
[7] M. Batty. The New Science of Cities. MIT Press, 2013.
[8] M. Batty and P. Longley. Fractal Cities: A Geometry of Form
and Function. Academic Press, London, 1994.
[9] M. Batty and Y. Xie. Possible urban automata. Environment
and Planning B, 24:175–192, 1997.
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
59
J. Raimbault, A. Banos & R. Doursat
[10] I. Benenson. Multi-agent simulations of residential dynam-
ics in the city. Computers, Environment and Urban Systems,
22(1):25–42, 1998.
[30] D. Pumain. Multi-agent system modelling for urban systems:
The series of simpop models. In Agent-based models of geo-
graphical systems, pages 721–738. Springer, 2012.
[31] QGIS Development Team. QGIS Geographic Information Sys-
tem. Open Source Geospatial Foundation, 2009.
[32] R Core Team. R: A Language and Environment for Statistical
Computing. R Foundation for Statistical Computing, Vienna,
Austria, 2013.
[33] T. C. Schelling. Models of segregation. The American Eco-
nomic Review, 59(2):488–493, 1969.
[34] C. Tannier, G. Vuidel, and P. Frankhauser. D´elimitation
d'ensembles morphologiques par une approche multi-´echelle. In
J.-C. Foltete, editor, Actes des huiti`emes Rencontres de Th´eo
Quant, page 14, Besan¸con, France, 2008. http://thema.univ-
fcomte.fr/theoq/.
[35] Y.-H. Tsai. Quantifying urban form: compactness versus'
sprawl'. Urban Studies, 42(1):141–161, 2005.
[36] J. van Vliet, J. Hurkens, R. White, and H. van Delden. An
activity-based cellular automaton model to simulate land-use
dynamics. Environment and Planning-Part B, 39(2):198, 2012.
[37] F. Varenne, M. Silberstein, et al. Mod´eliser & simuler.
Epist´emologies et pratiques de la mod´elisation et de la sim-
ulation, tome 1. 2013.
[38] R. White. Modeling multi-scale processes in a cellular au-
tomata framework. In Complex artificial environments, pages
165–177. Springer, 2006.
[39] R. White and G. Engelen. Cellular automata and fractal urban
form: a cellular modelling approach to the evolution of urban
land-use patterns. Environment and planning A, 25(8):1175–
1199, 1993.
[40] U. Wilensky.
Center for Connected Learn-
ing and Computer-Based Modeling, Northwestern University,
Evanston, IL., 1999.
Netlogo.
[41] F. Wu. A linguistic cellular automata simulation approach for
sustainable land development in a fast growing region. Com-
puters, Environment and Urban Systems, 20:367–87, 1996.
[11] G. Caruso, G. Vuidel, J. Cavailhes, P. Frankhauser, D. Peeters,
and I. Thomas. Morphological similarities between dbm and
a microeconomic model of sprawl. Journal of Geographical
Systems, 13:31–48, 2011.
[12] P. Frankhauser and C. Tannier. A multi-scale morpholog-
ical approach for delimiting urban areas.
In 9th Comput-
ers in Urban Planning and Urban Management conference
(CUPUM'05), University College London, 2005.
[13] L. Gauvin, J. Vannimenus, and J.-P. Nadal. Phase diagram of
a schelling segregation model. The European Physical Journal
B, 70(2):293–304, 2009.
[14] B. Golden, M. Aiguier, and D. Krob. Modeling of complex sys-
tems ii: A minimalist and unified semantics for heterogeneous
integrated systems. Applied Mathematics and Computation,
218(16):8039–8055, 2012.
[15] H. Haken and J. Portugali. The face of the city is its infor-
mation. Journal of Environmental Psychology, 23(4):385–408,
2003.
[16] L. Henri. Le droit `a la ville. Anthropos, 1968.
[17] A. J. Heppenstall, A. T. Crooks, and L. M. See. Agent-based
models of geographical systems. Springer, 2012.
[18] B. Hillier, A. Leaman, P. Stansall, and M. Bedford. Space
syntax. Environment and Planning B: Planning and Design,
3(2):147–185, 1976.
[19] S. Iltanen. Cellular automata in urban spatial modelling.
In Agent-based models of geographical systems, pages 69–84.
Springer, 2012.
[20] F. Le N´echet and A. Aguilera. D´eterminants spatiaux et
sociaux de la mobilit´e domicile-travail dans 13 aires urbains
fran¸caises : une approche par la forme urbaine, `a deux ´echelles
g´eographiques.
In ASRDLF 2011, SCHOELCHER, Mar-
tinique, July 2011. http://asrdlf2011.com/.
[21] R. Louf and M. Barthelemy. Modeling the polycentric transi-
tion of cities. ArXiv e-prints, Sept. 2013.
[22] D. Mangin. La ville franchis´ee: formes et structures de la ville
contemporaine. ´Editions de la Villette Paris, 2004.
[23] C. Maria de Almeida, M. Batty, A. M. Vieira Monteiro,
G. Camara, B. S. Soares-Filho, G. C. Cerqueira, and C. L. Pen-
nachin. Stochastic cellular automata modeling of urban land
use dynamics: empirical development and estimation. Com-
puters, Environment and Urban Systems, 27(5):481–509, 2003.
[24] D. Moreno, D. Badariotti, and A. Banos. Un automate cellu-
laire pour exp´erimenter les effets de la proximit´e dans le pro-
cessus d'´etalement urbain :
le mod`ele raumulus. Cybergeo :
European Journal of Geography, 2009.
[25] D. Moreno, A. Banos, and D. Badariotti. Conception d'un
automate cellulaire non stationnaire `a base de graphe pour
mod´eliser la structure spatiale urbaine: le mod`ele remus. Cy-
bergeo: European Journal of Geography, 2007.
[26] D. Peeters and M. Rounsevell. Space time patterns of urban
sprawl, a 1d cellular automata and microeconomic approach.
Environment and Planning B: Planning and Design, 36:968–
988, 2009.
[27] J. Portugali. Complexity theories of cities: Achievements, crit-
icism and potentials. In Complexity Theories of Cities Have
Come of Age, pages 47–62. Springer, 2012.
[28] J. Portugali, H. Meyer, E. Stolk, and E. Tan. Complexity theo-
ries of cities have come of age: an overview with implications
to urban planning and design. Springer, 2012.
[29] D. Pumain. Scaling laws and urban systems. Santa Fe Insti-
tute, Working Paper n 04-02, 2:26, 2004.
60
ICCSA 2014, Normandie University, Le Havre, France June 23-26, 2014
|
1910.09508 | 1 | 1910 | 2019-10-21T16:54:49 | Multi-agent Hierarchical Reinforcement Learning with Dynamic Termination | [
"cs.MA",
"cs.AI",
"cs.LG"
] | In a multi-agent system, an agent's optimal policy will typically depend on the policies chosen by others. Therefore, a key issue in multi-agent systems research is that of predicting the behaviours of others, and responding promptly to changes in such behaviours. One obvious possibility is for each agent to broadcast their current intention, for example, the currently executed option in a hierarchical reinforcement learning framework. However, this approach results in inflexibility of agents if options have an extended duration and are dynamic. While adjusting the executed option at each step improves flexibility from a single-agent perspective, frequent changes in options can induce inconsistency between an agent's actual behaviour and its broadcast intention. In order to balance flexibility and predictability, we propose a dynamic termination Bellman equation that allows the agents to flexibly terminate their options. We evaluate our model empirically on a set of multi-agent pursuit and taxi tasks, and show that our agents learn to adapt flexibly across scenarios that require different termination behaviours. | cs.MA | cs |
Multi-agent Hierarchical Reinforcement
Learning with Dynamic Termination
Dongge Han, Wendelin Bohmer, Michael Wooldridge, and Alex Rogers
Dept. Computer Science, University of Oxford, Oxford, UK
{dongge.han, wendelin.boehmer, michael.wooldridge,
alex.rogers}@cs.ox.ac.uk
Abstract. In a multi-agent system, an agent's optimal policy will typ-
ically depend on the policies chosen by others. Therefore, a key issue
in multi-agent systems research is that of predicting the behaviours of
others, and responding promptly to changes in such behaviours. One ob-
vious possibility is for each agent to broadcast their current intention,
for example, the currently executed option in a hierarchical reinforce-
ment learning framework. However, this approach results in inflexibility
of agents if options have an extended duration and are dynamic. While
adjusting the executed option at each step improves flexibility from a
single-agent perspective, frequent changes in options can induce incon-
sistency between an agent's actual behaviour and its broadcast intention.
In order to balance flexibility and predictability, we propose a dynamic
termination Bellman equation that allows the agents to flexibly terminate
their options. We evaluate our model empirically on a set of multi-agent
pursuit and taxi tasks, and show that our agents learn to adapt flexibly
across scenarios that require different termination behaviours.
Keywords: Multi-agent Learning · Hierarchcial Reinforcement Learning
1
Introduction
Many important real-world tasks are multi-agent by nature, such as taxi coor-
dination [10], supply chain management [6], and distributed sensing [9]. Despite
the success of single-agent reinforcement learning (RL) [17,13], multi-agent RL
has remained as an open problem. A challenge unique to multi-agent RL is that
an agent's optimal policy typically depends on the policies chosen by others [16].
Therefore, it is essential that an agent takes into account the behaviours of oth-
ers when choosing its own actions. One possible solution is to let each agent
model and broadcast its intention, in order to indicate the agent's subsequent
behaviours [3]. As an example, Figure 1(a) shows a taxi pickup scenario where
taxi A is choosing its next direction. Given the information that taxi B is cur-
rently heading towards Q, taxi A can determine passenger P as its preferred
option over Q.
* Preprint presented at PRICAI 2019. The final authenticated version is available online at
https://doi.org/10.1007/978-3-030-29911-8_7.
2
D. Han et al.
(a) Taxi A choosing a target
(b) Taxi B switching target
Fig. 1: Taxi Scenario Examples
Fig. 2: The effect of terminating options early, i.e., after T steps
Fortunately, hierarchical RL provides a simple solution for modeling agents'
intentions by allowing them to use options, which are subgoals that an agent aims
to achieve in a finite horizon. Makar et al. [11] proposed multi-agent hierarchical
RL, where hierarchical agents broadcast their current options to the others.
However, despite the advantage brought by using options, there can be a delay
in an agent's responses towards changes in the environment or others' behaviours,
due to the temporally-extended nature of options, which forbids the agents from
switching to another option before the current one is terminated. In the scenario
depicted in Figure 1(b), while taxi A is going for passenger R, taxi B finished
picking up passenger Q and also switched towards R. In this case, taxi A will
miss the target R, but it cannot immediately switch its target.
A potential solution to the delayed response challenge is to terminate options
prematurely. Figure 2 shows the performance of a multi-agent taxi experiment
where the agents' current options are interrupted after T timesteps. By reduc-
ing T , the agents gain higher flexibility for option switching, which also leads to
increasing rewards. This has been studied previously to address the problem of
imperfect options in single-agent settings, where an agent can improve its per-
formance by terminating and switching to an optimal option at each step [18].
However, this approach may no longer prove advantageous in a multi-agent sce-
nario. When an agent frequently switches options, the broadcast option will be
inconsistent with its subsequent behaviour. Consequently, the agent's behaviour
becomes less predictable and the advantage of broadcasting options is dimin-
ished.
010000200003000040000episodes101234567average rewards per episode45673000040000T = 1T = 2T = 3T = 4T = 5 Multi-agent Hierarchical Reinforcement Learning with Dynamic Termination
3
This poses a dilemma that is specific to multi-agent systems: excessive termi-
nations makes an agent's behaviour unpredictable, while insufficient termination
of options results in agents' inflexibility towards changes[8]. We will refer to an
agent's flexibility as the ability to switch options in response to changes in oth-
ers or the environment. Furthermore, we will use predictability to measure how
far an agent will commit to its broadcast option. In this paper, we propose an
approach called dynamic termination, which allows an agent to choose whether
to terminate its current option according to the state and others' options. This
approach balances flexibility and predictability, combining the advantages of
both.
An obvious approach to modelling dynamic termination is to use an addi-
tional controller, which decides whether to terminate or to continue with the
current option at each step. In this paper, we incorporate termination as an
additional option for the high-level controller. In this way, the Q-value of the
newly introduced option is associated consistently with the Q-values of the orig-
inal options, and our approach introduces negligible additional complexity to the
original model. We evaluate our model on the standard multi-agent pursuit and
taxi coordination tasks across a range of parameters. The results demonstrate
that our dynamic termination model can significantly improve hierarchical multi-
agent coordination and that it outperforms relevant state-of-the-art algorithms.
The contributions of our work are as follows:
1. Based on the decentralized multi-agent options framework, we propose a
novel dynamic termination scheme which allows an agent to flexibly termi-
nate its current option. We show empirically that our model can greatly
improve multi-agent coordination
2. We propose a delayed communication method for an agent to approximate
the joint Q-value. This method allows us to use intra-option learning, and
reduces potentially costly communication
3. We incorporate dynamic termination as an option to the high level controller
network. This design introduces little additional model complexity, and al-
lows us to represent the termination of all options in a consistent manner
In addition, we adopted several methods that benefits the model architecture
and training: deep Q networks and parameter sharing reduce state space and
model complexity; adapting intra-option learning [18] to multiple agents yields
better sample efficiency; and an off-policy training scheme [7] for exploration.
2 Related Work
Makar et al.[11] appear to have been the first to combine multi-agent and hi-
erarchical RL, through the MaxQ framework [2]. We build on their work, with
the following changes: First, the use of tabular Q-learning is insufficient for large
state spaces. Therefore, we adopt deep Q networks for parameterizing state and
action spaces. Second, we adapt intra-option learning to multi-agent systems [18],
4
D. Han et al.
which greatly improves the sample efficiency. Third, we adopt a delayed com-
munication channel to prevent costly communication, and joint optimization.
And finally, as options cannot be terminated before their predefined termination
condition, tasks are limited to the use of perfect options and agents experience
the delayed response problem.
Our solution to the delayed response problem is related to works on inter-
rupting imperfect options, i.e., when the set of available options are not perfectly
suited to the task, an agent can choose to terminate its options dynamically in
order to improve its performance. Sutton et al.[18] introduced a mechanism for
interrupting options whenever a better option appears, and Harutyunyan et al.[7]
proposed a termination framework which improves upon this idea with better
exploration. This is achieved by off-policy learning, which uses an additional
behaviour policy for longer options.
Bacon et al.[1] proposed a dynamically terminating model for their Option-
critic framework. In comparison, we use the Q-learning framework instead of
policy gradient; and we focus on addressing the coordination problems in a
multi-agent system. Moreover, our Q-value for dynamic termination does not
depend on the currently executing option, which significantly reduces the model
complexity and also improve sample efficiency due to off-policy training, i.e., the
value of terminating can be learned with any executed option.
In the multi-agent learning literature, Riedmiller et al.[15] proposed the
multi-option framework. This is a centralized model in which multiple agents are
considered as a single meta-agent that chooses a joint option o = (o1, . . . , on).
In contrast, our model uses a decentralized scheme where each agent i chooses
and executes its own option oi. This reduces the action space of the high level
controller from On to O, where O is the set of all options, and n is the number
of agents.
Our model also draws upon the independent Q-learning framework proposed
by Tan[20], where each agent independently learns its own policy on primitive
actions, while treating other agents as part of the environment. Additionally
in our model, each agent conditions on the others' broadcast options as part
of its observation when choosing the next option. We will discuss the detailed
formulation in section 4.
3 Basic Definitions
We first introduce the essential concepts in reinforcement learning (RL), fol-
lowed by multi-agent RL, hierarchical RL, intra-option learning and off-policy
termination.
A Markov Decision Process [17] is given by a tuple (cid:104)S,A, R, P, γ(cid:105), where S
denotes a set of states, A a set of actions, P the stationary transition probability
P (st+1st, at) from state st to state st+1 after executing action at, R is the aver-
age reward function rt := R(st, at), and γ ∈ [0, 1) is the discount factor. A policy
π(atst) is a distribution over actions at given the state st. The objective of a
RL agent is to learn an optimal policy π∗, which maximizes the expected cumu-
Multi-agent Hierarchical Reinforcement Learning with Dynamic Termination
5
lative discounted future rewards. The Q-value of the optimal policy conditions
this return on an action at that has been selected in a state st:
(st, at) = E(cid:104) ∞(cid:80)
∗
Q
(cid:105)
γτ rt+τ
τ =0
= rt + γ max
a(cid:48)
E(cid:2)Q
∗
(cid:48)
(st+1, a
)(cid:3).
(1)
Q-learning learns the Q-value of the optimal policy by interacting with a discrete
environment [22]. Continuous and high-dimensional states require function ap-
proximation [17], for example deep convolutional neural networks (DQN) [12,13].
To improve the stability of gradient decent, DQN introduces an experience replay
buffer to store transitions that have already been seen. Each update step samples
a batch of past transitions and minimizes the mean-squared error between the
left and right side of Equation 1.
In multi-agent reinforcement learning, n agents interact with the same envi-
ronment. The major difference to the single agent case is that the joint action
space A = A1 × ··· × An of all agents grows exponential in n. Independent
Q-learning addresses this by decentralizing decisions [20]: each agent learns a Q-
value function that is independent of the actions of all other agents. This treats
others as part of the environment and can lead to unstable DQN learning [5].
Other approaches combine decentralized functions with a learned centralized
network [14] or train decentralized Actor-Critic architectures with centralized
baselines [4].
We now describe some important concepts related to hierarchical reinforce-
ment learning (HRL). The Options Framework [18] is one of the most common
HRL frameworks, which defines a two-level hierarchy, and introduces options as
temporally extended actions. Options o are defined as triples (cid:104)I o, βo, πo(cid:105), where
I o ⊆ S is the initiation set and βo : S → [0, 1] is the option termination condi-
tion. πo : S → A is a deterministic option policy that selects primitive actions
to achieve the target of the option. On reaching the termination condition in
state s(cid:48), an agent can select a new option from the set O(s(cid:48)) := {o s(cid:48) ∈ I o}. A
Semi-Markov Decision Process (SMDP) [18] defines the optimal Q-value:
Q(st, ot) = E(cid:104)k−1(cid:80)
γτ rt+τ + γk max
o(cid:48)∈O(st+k)
τ =0
Q(st+k, o(cid:48))
,
(2)
(cid:105)
where k refers to the number of steps until the termination condition βot(st+k) =
1 is fulfilled.
To improve sample efficiency, Intra-option Learning [18,19] was proposed as
an off-policy learning method which at each time step t updates all options that
are in agreement with the executed action, i.e. ∀o ∈ {o πo(st) = at} holds:
Q(st, o) = rt + γ E(cid:2)U (st+1, o)(cid:3)
U (s, o) =(cid:0)1 − βo(s)(cid:1) Q(s, o) + βo(s) max
(3)
Q(s, o(cid:48)).
o(cid:48)∈O(s)
Here U (st+1, o) is the TD-target [21]: if o is terminating in the next state, the
TD-target will be the value of choosing the next optimal option. If not, the
target will be the value of continuing with option o. Updating multiple options
vastly improves the efficiency of training. Consider a grid-world navigation case
6
D. Han et al.
Fig. 3: Dynamic termination Q-value network architecture.
where an agent is going for some goal location, and each coordinate corresponds
to the sub-goal of an option. When the agent takes a primitive action aj
t and
reaches the next position, all options oj that would have chosen that action
will be updated. Each transition updates therefore a significant fraction of the
options, which massively improves sample efficiency.
As introduced by Sutton et al. [18], when the set of available options are
not suited to the task, an agent can improve its performance by terminating at
each step and switching to an optimal option. Harutyunyan et al. [7] have shown
that this approach improves the agent's performance significantly, but has an
adverse effect on exploration: temporally extended options can explore the state
space more consistently, which is lost by early termination. The authors therefore
advocate the use of intra-option learning to update the Q-value off-policy, while
executing a different exploration policy that follows a selected option for multiple
steps before terminating.
4 Method
In this section we will present our framework for deep decentralized hierarchical
multi-agent Q-learning. Our model uses delayed communication to approximate
the decisions of a centralized joint policy, which avoids many problems usually
associated with joint optimization. This induces new challenges such as a delayed
response of agents, and requires us to define a novel dynamic termination update
equation.
Delayed Communication: A straightforward application of decentralized multi-
agent approaches like independent Q-learning (IQL) [20] to the Options frame-
work [18] would yield agents that make decisions independent of each other.
Agent j would estimate the Q-value (see Equation 2)
t ) := E(cid:104)k−1(cid:80)
τ =0
Qj
iql(st, oj
γτ rt+τ +γk max
o(cid:48) j∈Oj
(cid:105)
iql(st+k, o(cid:48)j)
Qj
,
(4)
and select the option oj
t that maximizes it. Here other agents are treated as
stationary parts of the environment, which can lead to unstable training when
ConvolutionFully-connectedO1O2O3…Oi…OnTOther agents' options (0,1)(0,1)Gridworld ImageAgentOther AgentsPassengersGridworld Image Multi-agent Hierarchical Reinforcement Learning with Dynamic Termination
7
those agents change their policy. The best way to avoid this instability would
be learn the Q-value w.r.t. the joint option of all agents ot := (o1
t ),
i.e. Qjoint(st, ot) [15]. While these joint Q-values allow training in a stationary
environment, decisions require to maximize over Q-values of all possible joint
options. As the number of joint options grows exponentially in the number of
agents n, and joint optimization would require a vast communication overhead,
this approach is not feasible in decentralized scenarios.
t , . . . , on
Instead we propose to use a delayed communication channel over which agents
signal the new option they switched to after each termination. This reduces po-
tentially costly communication and allows each agent j access to all other agents'
t−1 , . . . , on
t−1). Agents
options of the previous time step o
can approximate the joint Q-value by conditioning on this information, that is,
−j
t−1, oj
by choosing options oj
t ). Note
that the approximation is exact if no other agent terminates at time t. The op-
timality of the agents' decisions depends therefore on the frequency with which
agents terminate their options.
t that maximize the delayed Q-value Qj(st, o
t−1, . . . , oj−1
−j
t−1 := (o1
t−1 , oj+1
Multi-agent Intra-Option Learning: As introduced in the previous section,
the intra-option learning method (Equation 3) efficiently associates options with
primitive actions. In our decentralized multi-agent options model, agent j selects
an option according to Qj(st, o
−j
t−1, oj), which is defined as
t−1, oj) := E(cid:2)rt + γU j(st+1, o
, oj) :=(cid:0)1 − βoj
, oj)(cid:3)
(st+1)(cid:1) Qj(st+1, o
−j
−j
t
−j
t
Qj(st, o
U j(st+1, o
−j
t
−j
Qj(st+1, o
t
, oj)
, o(cid:48)j).
+ βoj
(st+1) max
o(cid:48)j∈Oj
(5)
t as the actually executed option oj
We can learn Qj by, for example, minimizing the mean-squared TD error [17]
between the left and right side of Equation 5. In line with intra-option learning,
we update the Q-values of all options oj that would have executed the same
action aj
t . Note that due to our delayed com-
munication channel, the executed options of all other agents are known after the
transition to st+1 and can thus be used to compute the target U j(st+1, o
, oj),
that is, the Q-value of either following the option oj if βoj
(st+1) = 0, or termi-
nating and choosing another option greedily if βoj
Dynamic Option Termination: As mentioned above, the delayed Q-value
defined in Equation 5 only approximates the joint Q-value function. This ap-
proximation will deteriorate when other agents terminate, but sometimes agents
can also benefit from early termination, as shown in Figure 1(b). Additionally,
options are usually pre-trained and have to cover a large range of tasks, without
being able to solve any one task perfectly. Being able to prematurely terminate
options can increase the expressiveness of the learned policy dramatically.
(st+1) = 1.
−j
t
The easiest way to use partial options is to modify the termination conditions
βoj
(s). In particular, we denote choosing the option with the largest Q-value
(Eq. 5) at each time step as greedy termination. Following [7] we combined this
approach with an exploration policy that terminates executed options with a
8
D. Han et al.
fixed probability ρ = 0.5 to allow for temporally extended exploration. During
testing the agent is nonetheless allowed to terminate greedily at every step if the
Q-value of another option is larger.
Although greedy termination has been shown to improve the performance of
individual agents with imperfect options [7], the agent's behaviour will become
less predictable for others. In particular, agents that utilize the delayed Q-value
of Equation 5 will make sub-optimal decisions whenever another agent termi-
nates. To increase the predictability of agents, while allowing them to terminate
flexibly when the task demands it, we propose to put a price δ on the deci-
sion to terminate the current option. Option termination is therefore no longer
hard-coded, but becomes part of the agent's policy, which we call dynamic ter-
mination. This can be represented by an additional option oj = T for agent j
to terminate. Note that, unlike in the Options framework, we no longer need
a termination function βoj
(st) for each option oj. It is sufficient to compare
the value of the previous option Qj(st, o
t−1) with the value of termination
−j
Qj(st, o
t−1, T ). Evaluating oj = T is computationally similar to evaluating the
termination condition βo. Dynamic termination therefore has a similar cost to
traditional termination.
−j
t−1, oj
The optimal behaviour for a given punishment δ is the fix-point of the novel
dynamic termination Bellman equation:
o(cid:48)j∈{oj ,T} Qj(st+1, o
−j
t−1, o(cid:48)j) − δ .
Qj(st, o
(6)
t−1, oj (cid:54)= T ) := E(cid:2)rt + γ max
−j
Qj(st, o
Qj(st, o
−j
t−1, oj = T ) := max
o(cid:48)j∈Oj
t , o(cid:48)j)(cid:3),
−j
Similarly to Equation 5, Equation 6 allows intra-option learning and can be
applied to all options oj that would have selected the same action as the executed
option oj
t . Note that the termination option T can always be updated, as it does
not depend on the transition.
Deep Q-learning: A group of n agents can be trained using a deep Q-network
Qθ [12], parameterized by θ. The architecture is shown in Figure 3: each agent
j selects and executes the next option based on the current state (i.e. grid-word
−j
t−1 of all the other agents. A centralized
image) st and the last known options o
manager is not needed, and the options must only be broadcast after an agent
chose to select a new option. The Q-value of choosing an option is updated
by temporal difference learning with experience replay, which is the established
standard procedure in deep Q-learning [12]. To reduce the number of parameters,
we let all the agents share θ, and the model is thus updated using the experiences
collected by all the agents. To differentiate the behaviour of different agents, the
presented grid-world image contains a dedicated channel that encodes the current
agent's state. These design decisions follow previous work in deep multi-agent
learning [4,14] and drastically reduce training time with very little impact on
the performance in large domains.
At each transition t, the Q-values of all options oj, that would execute the
t , and the termination option T are updated
same action as the executed option oj
Multi-agent Hierarchical Reinforcement Learning with Dynamic Termination
9
predictability
flexibility
option
changes
option
changes
steps to
change
all
near
far
all
dynamic 24.1% 28.1% 16.8%
63%
greedy 59.9% 57.9% 52.8% 77%
option 10.9% 10.9% 9.5%
3%
all
1.61
1.15
6.86
Table 1: Flexibility and Predictability Results.
Near/far refers to whether an agent is within
distance = 4 to a passenger. Steps to change
denotes the number of steps from the new pas-
senger is placed to the agent's option change.
Fig. 4: Example 16 × 16 Grid-world
by gradient descent on the sum of their respective losses
j,t
Lo
LT
j,t
(cid:16)
(cid:16)
(cid:2)θ(cid:3) :=
(cid:2)θ(cid:3) :=
L(cid:2)θ(cid:3)
rt +γ max
o(cid:48)∈{o,T} Qθ(st+1, o
Qθ(st, o
max
o(cid:48)∈Oj
−j
t , o(cid:48)) − Qθ(st, o
−j
t−1, T )
−j
t−1, o(cid:48)) − δ − Qθ(st, o
(cid:88)
n(cid:88)
m−1(cid:88)
(cid:2)θ(cid:3) +
(cid:16)LT
j,t
t=0
j=1
Lo
πo(st)=aj
t
j,t
:=
1
mn
(cid:17)2
−j
t−1, o)
.
(cid:17)2
(cid:2)θ(cid:3)(cid:17)
.
,
(7)
The total loss for a batch of m transitions with n agents is
5 Experiments
We will first evaluate the flexibility and predictability of our dynamically ter-
minating agent, followed by the impact of dynamic termination on the agents'
performance.
Experimental Setup: Figure 4 shows a 16×16 grid-world of the taxi pickup as
observed by the green agent, which includes the passengers, the other agents and
their broadcast options. The landmarks of distance L = 3 show the destinations
of options that are currently visible to the agent. This raises our first challenge:
in order to reach a passenger that stands outside the landmarks, an agent needs
to correctly switch between options.
In the Taxi Pickup Task m passengers are randomly distributed in each
episode. An agent is rewarded r = 1 when occupying the same grid as a passen-
ger, and each step incurs a cost of −0.01. Apart from landmark switching, the
agents need to interpret others' behaviours to avoid choosing the same passen-
ger, as well as responding quickly to changes such as when a passenger is picked
up by another agent.
In the Pursuit Task agents try to catch randomly distributed prey by coop-
erating with others. We refer to the task as k-agent pursuit, where a successful
capture requires at least k agents occupying k positions adjacent to the prey,
10
D. Han et al.
which rewards each participating agent r = 1. This task relies heavily on agents
coordination. In particular, when close to a specific prey, agents need to observe
others and switch between options to surround the prey; whereas when faraway,
agents need to agree on and commit to go for the same prey.
Algorithms and Training: Having described the settings, we now introduce
the detailed training procedures of the SMDP and option policies, before com-
paring the four types of agents.
The Policy of Options adopts a local perspective, and navigates the agent
to the option's destination. Specifically, we use a DQN of 2 convolutional lay-
ers (kernel size 2) with max-pooling, followed by 4 fully-connected layers (size
300). The input is the destination coordinate with the grid-world image, and the
output is a primitive action in {N, S, E, W, Stay}.
The SMDP Policies are trained through intra-option learning for all agent
types. The inputs are 4-channel grid-world images as in Figure 3, which rep-
resents the agent, the preys (or passengers), the other agents, and lastly, the
options broadcast by other agents (except for IQL). The DQN contains 2 con-
volutional layers (kernel size 3), max pooling, and 4 fully-connected layers (size
512). We use experience replay with a replay buffer of size 100,000.
Self-Play is used in the experiments, and our decentralized agents share the
same DQN parameters (not states) [14]. This allows us to scale up the number
of agents without additional parameters; and the trained model can directly
transfer to more agents during testing. Moreover, self-play creates an important
link between the predictability of an individual agent and of the society.
The four types of agents are as follows:
1. Option Termination Agent executes its option until the natural termination
condition is met.
2. Greedy Termination Agent terminates every step and switches to the optimal
option. For better exploration during training, an additional behaviour policy
is used for experience collection. For fairness of comparison, this exploration
policy which terminates with probability ρ = 0.5 (tuned for the greedy agent)
is applied across all agent types.
3. Dynamic Termination Agent is our proposed algorithm that chooses whether
to terminate the current option at each step. δ is the termination penalty.
4. IQL is independent Q-learning, where agents option broadcasts are disabled.
IQL (greedy) and IQL (δ) refers to IQL agents using greedy or dynamic
termination.
Results: The delayed response problem reveals that agents need to be flexible
enough to change their options when the situation changes, but also predictable
enough not to interfere in other agents' plans too frequently. Table 1 shows
experimental measurements to showcase these conflicting goals for the investi-
gated termination methods. We measure the agents' flexibility in the single-agent
taxi domain. 100 episodes are initialized with 5 random passengers. During each
episode, one additional passenger is placed near the agent at step T and we
Multi-agent Hierarchical Reinforcement Learning with Dynamic Termination
11
(a) 19 × 19 taxi with 10 agents
(b) 19 × 19 pursuit with 10 agents
Fig. 5: Results from Taxi Pickup and Pursuit Tasks
Note: Every point per 500 episode is the testing result averaged over 100 random
episodes and 5 seeds. The shaded area shows standard deviation across seeds.
taxi
2 agent pursuit
3 agent pursuit
n=5, m=10 n=10, m=20 n=3, m=5 n=10, m=10
n=3, m=5
Agents
19x19 25x25
19x19
16x16
(r=1)
19x19
(r=1)
Dynamic
Greedy
Option
IQL
δ = 0.1 7.89 5.75
15.29
10.24 9.30
δ = 0
6.58
3.28
6.62
3.23
-0.32 -0.94
δ = 0.1 7.11
5.09
greedy
6.08
2.79
11.81
12.39
0.52
12.02
9.06
6.73
4.07
7.36
3.74
5.47 -1.82
-1.57 -2.29
-2.12 -2.49
19x19
(r=1)
12.50
5.38
4.65
-1.42
-0.84
-1.64
10x10
(r=1)
16x16
(r=2)
6.71
10.38
5.53
5.89
-3.77
-1.62
-2.13
6.54
6.40
5.20
-0.59
-0.42
Table 2: Average reward after training for Taxi and Pursuit tasks. n is the number
of agents and m is the number of passengers (preys). NxN denotes grid-world
size, k agent pursuit denotes the required number of agents for capture, and r is
the capture range.
observe how quickly the agent adjusts to the new situation. We report the per-
centage of option changes at step T + 1 and the average number of steps till the
agent changes options. Note that dynamic termination allows to react almost as
flexible to the changed situation as the greedy termination.
For predictability, we measure the average probability to change the option
in the multi-agent taxi task for two cases: when the agent is near (within dis-
tance 4) or far from its closest passenger. This is an imperfect measurement, as
we cannot distinguish the effect of termination on other agents. While options
need to change close to a passenger due to imperfect options, the behavior of
dynamic termination is much closer to standard option termination when far
away. Note that this effect is marginal for the other techniques, which indicates
that our method may purposefully refrain from changing to better options to
avoid interrupting other agent's plans.
0100002000030000400005000060000episodes−20246810121416averagerewardsperepisodedynamic(δ=0.15)dynamic(δ=0)greedyoptionIQL(δ=0.15)IQL(greedy)0500010000150002000025000300003500040000episodes−202468101214averagerewardsperepisodedynamic(δ=0.1)dynamic(δ=0)greedyoptionIQL(δ=0.1)IQL(greedy) 12
D. Han et al.
Performance: Figure 5(a) shows the results from the taxi pickup task. The
option termination agent fails due to its inflexibility to switch options. In con-
trast, our dynamic (δ = 0.15) agent is highly flexible. Moreover compared with
greedy and IQL, its high predictability indeed helps the agents to interpret oth-
ers' intentions and better distribute their target passengers. Figure 5(b) shows
the results on the pursuit task, where at least two agents need to surround a prey
within capture range = 1. Seen from the IQL agents' low performance, option
broadcasting and interpreting others' behaviours are crucial to this task. Our
dynamic termination agent (δ = 0.1) significantly outperforms all other agents.
Compared with the greedy agents, we can conclude that predictability signifi-
cantly helps our dynamic agents to stay committed and succeed in cooperation.
Finally, we present the performance of all agents across different tasks and
varying parameters in Table 2. Firstly, the option termination agent has diffi-
culty with tasks which require higher level of accuracy and quick responses, such
as the taxi tasks and pursuit with capture range = 1. However, it works well
with tasks which require coordination but less flexibility, such as the 16 × 16 3
agent pursuit with capture range 2, which shows the advantage of predictability
on cooperation. The performance of greedy termination agents decreases signif-
icantly in larger grid-world sizes, and when commitment is essential, such as
the 16 × 16 3 agent pursuit with capture range 2. Our dynamically terminating
agent performs well across all tasks, as it balances well between flexibility and
predictability. The IQL agents performs well in the taxi task. However, they
fail to learn the pursuit tasks where foreseeing others' behaviours is essential to
coordination.
6 Conclusions and Future Work
In this paper, we identified the delayed response problem, that occurs when hi-
erarchical RL is combined with multi-agent learning. To address this challenge,
we investigated existing approaches of greedy option termination in single agent
learning. However, this method introduces a new dilemma specific to multi-agent
systems: as an agent broadcasts its current options to indicate its subsequent
behaviours, frequent changes in options will result in its behaviour being less
predictable by others. Therefore, to balance flexibility with predictability, we in-
troduced dynamic termination, which enables agents to terminate their options
flexibly according to the current state. We compared our model with current
state of the art algorithms on multi-agent pursuit and taxi tasks with varying
task parameters, and demonstrated that our approach outperformed the base-
lines through flexibly adapting to the task requirements. For future work, we are
interested in applying the dynamic termination framework to traffic simulations,
such as junction and highway management.
References
1. Bacon, P.L., Harb, J., Precup, D.: The option-critic architecture. In: AAAI. pp.
1726 -- 1734 (2017)
Multi-agent Hierarchical Reinforcement Learning with Dynamic Termination
13
2. Dietterich, T.G.: Hierarchical reinforcement learning with the maxq value function
decomposition. Journal of Artificial Intelligence Research 13, 227 -- 303 (2000)
3. Foerster, J., Assael, I.A., de Freitas, N., Whiteson, S.: Learning to communicate
with deep multi-agent reinforcement learning. In: Advances in Neural Information
Processing Systems. pp. 2137 -- 2145 (2016)
4. Foerster, J., Farquhar, G., Afouras, T., Nardelli, N., Whiteson, S.: Counterfactual
multi-agent policy gradients. arXiv preprint arXiv:1705.08926 (2017)
5. Foerster, J., Nardelli, N., Farquhar, G., Afouras, T., Torr, P.H.S., Kohli, P., White-
son, S.: Stabilising experience replay for deep multi-agent reinforcement learn-
ing. In: Proceedings of the 34th International Conference on Machine Learn-
ing. Proceedings of Machine Learning Research, vol. 70, pp. 1146 -- 1155 (2017),
http://proceedings.mlr.press/v70/foerster17b.html
6. Giannakis, M., Louis, M.: A multi-agent based system with big data processing
for enhanced supply chain agility. Journal of Enterprise Information Management
29(5), 706 -- 727 (2016)
7. Harutyunyan, A., Vrancx, P., Bacon, P.L., Precup, D., Nowe, A.: Learning with
options that terminate off-policy. arXiv preprint arXiv:1711.03817 (2017)
8. Jennings, N.R.: Commitments and conventions: The foundation of coordination in
multi-agent systems. The knowledge engineering review 8(3), 223 -- 250 (1993)
9. Lesser, V., Ortiz Jr, C.L., Tambe, M.: Distributed sensor networks: A multiagent
perspective, vol. 9. Springer Science & Business Media (2012)
10. Lin, K., Zhao, R., Xu, Z., Zhou, J.: Efficient large-scale fleet management via
multi-agent deep reinforcement learning. arXiv preprint arXiv:1802.06444 (2018)
11. Makar, R., Mahadevan, S., Ghavamzadeh, M.: Hierarchical multi-agent reinforce-
ment learning. In: Proceedings of the fifth international conference on Autonomous
agents. pp. 246 -- 253. ACM (2001)
12. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D.,
Riedmiller, M.: Playing Atari with deep reinforcement learning. In: NIPS Deep
Learning Workshop (2013)
13. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G.,
Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., et al.: Human-level
control through deep reinforcement learning. Nature 518(7540), 529 (2015)
14. Rashid, T., Samvelyan, M., de Witt, C.S., Farquhar, G., Foerster, J., Whiteson, S.:
Qmix: Monotonic value function factorisation for deep multi-agent reinforcement
learning. arXiv preprint arXiv:1803.11485 (2018)
15. Riedmiller, M., Withopf, D.: Effective methods for reinforcement learning in large
multi-agent domains (leistungsfahige verfahren fur das reinforcement lernen in
komplexen multi-agenten-umgebungen). it-Information Technology 47(5), 241 -- 249
(2005)
16. Stone, P., Veloso, M.: Multiagent systems: A survey from a machine learning per-
spective. Autonomous Robots 8(3), 345 -- 383 (2000)
17. Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press
(1998)
18. Sutton, R.S., Precup, D., Singh, S.: Between mdps and semi-mdps: A framework
for temporal abstraction in reinforcement learning. Artificial intelligence 112(1-2),
181 -- 211 (1999)
19. Sutton, R.S., Precup, D., Singh, S.P.: Intra-option learning about temporally ab-
stract actions. In: ICML. vol. 98, pp. 556 -- 564 (1998)
20. Tan, M.: Readings in agents. chap. Multi-agent Reinforcement Learning: Indepen-
dent vs. Cooperative Agents, pp. 487 -- 494 (1998)
14
D. Han et al.
21. Tesauro, G.: Temporal difference learning and td-gammon. Communications of the
ACM 38(3), 58 -- 68 (1995)
22. Watkins, C., Dayan, P.: Q-learning. Machine Learning 8, 279 -- 292 (1992)
|
1010.0155 | 1 | 1010 | 2010-10-01T13:25:31 | An Investigation of the Advantages of Organization-Centered Multi-Agent Systems | [
"cs.MA"
] | Whereas classical multi-agent systems have the agent in center, there have recently been a development towards focusing more on the organization of the system. This allows the designer to focus on what the system goals are, without considering how the goals should be fulfilled. This paper investigates whether taking this approach has any clear advantages to the classical way of implementing multi-agent systems. The investigation is done by implementing each type of system in the same environment in order to realize what advantages and disadvantages each approach has. | cs.MA | cs |
An Investigation of the Advantages of
Organization-Centered Multi-Agent Systems
Andreas Schmidt Jensen
Department of Informatics and Mathematical Modelling
Technical University of Denmark
Richard Petersens Plads, Building 321, DK-2800 Kongens Lyngby, Denmark
Abstract. Whereas classical multi-agent systems have the agent in cen-
ter, there have recently been a development towards focusing more on the
organization of the system. This allows the designer to focus on what the
system goals are, without considering how the goals should be fulfilled.
This paper investigates whether taking this approach has any clear ad-
vantages to the classical way of implementing multi-agent systems. The
investigation is done by implementing each type of system in the same
environment in order to realize what advantages and disadvantages each
approach has.
1
Introduction
Within the area of multi-agent systems there has recently been a development
towards making the organization of such systems explicit [3,4,5]. However, while
[3] lists some drawbacks of classical (agent-centered) multi-agent systems, the
actual advantages of making the organization explicit has not been thoroughly
investigated.
This paper summarizes our work with such investigation of the organization
of multi-agent systems. The investigation was conducted by implementing two
systems: a classical (agent-centered) (ACMAS) and an organization-centered
(OCMAS).
In classical multi-agent systems the agent is in focus. The programmer de-
veloping the agents is able to decide what the agents can do and how the choose
to do it. In an OCMAS we are more concerned with the organization; i.e. the
structure of the multi-agent system. Naturally all multi-agent systems have a
structure, but it is most often implicitly defined by the agents and their rela-
tions.
By explicitly defining the organization it is possible to focus on what the
agents should do without at the same time deciding how they should do so. In
other words, the organization makes it possible to create the structure of the
system without specifying details about the implementation.
1
2 The Setup
The two types of systems have been compared using a team-based version of
the well-known game Bomberman. However, the nature of an implementation of
intelligent agents does not guarantee a certain quality and a comparison based
on the overall performance of a team of agents may not be adequate; the results
may merely be caused by better or worse strategies.
Since the two approaches are quite different in many ways, it seems more
natural to employ other measures of comparison. The comparison of ACMAS
and OCMAS is therefore based on the following measures:
-- Structure of the source code
-- Development speed
-- Performance
-- Error handling
-- Debugging
-- Complexity of the scenario
-- Number of intelligent agents
The original Bomberman game consists of five key elements: bombs, boxes,
solid obstacles, exitways and power-up panels. Whereas boxes are destructible,
solid obstacles are not. This means that Bomberman will always be able to take
cover behind solid obstacles. Most boxes must be destroyed since they hide both
power-ups and exitways. Exitways are what Bomberman must find to be able
to complete a level. A power-up can be used to enhance bomberman's abilities
and bombs. In the beginning, his bombs are weak, but by using power-ups he
will be able to drop several stronger bombs at a time.
In a multi-agent context, the enemies could be considered a team of agents
(i.e. a multi-agent system) with the general purpose of stopping Bomberman
from escaping and by implementing the enemies using each of the approaches
we could perform the comparison. However, to be able to experiment with the
cooperative aspects of intelligent agents we instead propose an altered version
of Bomberman in which two teams attempt to eliminate each other:
Definition 1 (Team-based Bomberman). The multi-agent system is similar
to Bomberman. It consists of two teams fighting against each other. Each team
consists of at least two "bombermen" (or agents). The teams are situated in
a maze-like environment consisting of solid obstacles and boxes. An agent can
place bombs which at some point will explode. An agent dies when he is hit by
an explosion. Explosions will also destroy boxes. A team wins when all players
from the other team have been eliminated.
This version of Bomberman consists of some of the same key elements as
the original game: a maze, destructible and indestructible obstacles, and bombs.
This should allow the agents to employ the strategies intended for the game,
while at the same time competing in teams. This fact creates a new aspect of
the game, since a group of agents potentially is able to trap enemies by placing
bombs strategically.
2
The concepts of exitways and power-ups have been excluded in this version.
Exitways have been removed, since the overall goal of "getting to the surface"
is no longer relevant (as the goal instead is to eliminate the other team). Power-
ups are not included to avoid making the overall system too complex since the
intention is not to make a perfect implementation of Bomberman; rather is it
to compare and discuss two different approaches to implementing multi-agent
systems. In this case we believe a simple, yet strategically challenging system
will be adequate.
2.1 Jason
The implementation of the ACMAS is done using Jason, "a Java-based inter-
preter for an extended version of AgentSpeak" 1. We provide an overview of the
interpreter by introducing how to program multi-agent system using it, however
we will not go into details with all parts of the system. The overview should give
a foundation for building simple systems using Jason. A thorough description
of Jason is found in [2].
The language of Jason, AgentSpeak, is a Prolog-like logic programming
language. AgentSpeak allows the developer to create a plan library for the agent.
A plan in AgentSpeak is basically of the form
+triggering event : context <- body.
Roughly speaking, if an event matches a trigger, the context is matched with
the current state of the agent. If the context matches the current state, the body
is executed; otherwise the engine continues to match contexts of plans with the
same trigger. If no plan is applicable, the event fails.
The fact that AgentSpeak is a logic programming language allows one to
easily transfer specifications written in logic formulas of a multi-agent system to
an implementation written in Jason. For instance, part of a plan for a vacuum
cleaner agent [9] is shown below:
+!cleaning : in(X,Y) & dirt(X,Y) <- do(suck).
The plan is triggered by the goal !cleaning, so if the vacuum cleaner is in a
"cleaning state", this triggering event would be applicable. The context specifies
that this plan is relevant if the agent currently is somewhere in the environment
which is dirty. If the context can be unified with data from the database of the
agent, it will perform the body, which in this case means that it will perform
the action do(suck). However, as mentioned it is possible to have several plans
for the same triggering event if those plans have different contexts:
+!cleaning : in(X,Y) & dirt(X+1,Y) <- do(right).
This plan will then be applicable if the agent has perceived dirt in an area to
the right of its current area. In that case, it will perform the action do(right).
1 http://jason.sourceforge.net/
3
2.2 The Moise+ organizational model
The implementation of the OCMAS is based on the Moise+ organizational
model, in which it is possible to create a structural, functional and deontic
specification of an organization. The organizational model has been combined
with Jason in the middleware called J -Moise+.
Moise+ is an organizational model for multi-agent systems which makes
it possible to specify the organization in a MAS structurally, functionally and
deontically. The model takes an organizations-centered approach, meaning that
an organization will exist a priori (created at design-time) and the agents ought
to follow it [5].
Structural Specification: Moise+ uses the concepts of roles, role relations and
groups in the structural specification of an organization. Each agent plays one
or more roles. The roles are related by links, which specify how agents are ac-
quainted and can communicate. In order to further structure the organization,
the agents can join different groups depending on the roles they play.
Functional Specification: The functional specification consists of a goal decom-
position tree, known as a Social Scheme (SCH), where the root is the goal of the
SCH and each node is a sub-goal that can be delegated to different agents. In
[5] three operators are defined for decomposing a goal into sub-goals: sequence,
choice and parallelism. These operators allow us to create complex schemes in
which the agents can commit to advanced missions.
Deontic Specification: The relation between the structural and functional speci-
fication is made explicit by the deontic specification. Using it, we can constrain
the agents further by specifying what missions an agent ought to follow and what
missions an agent is allowed to follow when playing certain roles. We write
obl(ρ, m, tc),
when agents playing role ρ are obliged to complete mission m under the time
constraint tc. Analogously we write per for permissions.
The Moise+ organizational model gives a foundation for defining and using
an organizational model for multi-agent systems -- in other words to create an
OCMAS. However, the model itself is not directly associated with any multi-
agent framework and the intention is that it should be usable for all kinds of
frameworks for multi-agent systems.
The software implementation called J -Moise+ is an implementation of
Moise+ which should enable multi-agent systems implemented in Jason to
follow an organizational structure [6].
2.3 Implementational details
The team-based Bomberman introduced above has been implemented in both an
ACMAS and OCMAS version. The following describes some of the implementa-
tional details of these systems. Both systems are built for the same environment,
4
and a lot of their abilities will be similar (if not identical). The differences are
mostly how the systems make use of their knowledge and actions, and not as
much what abilities they have. This ensures a somewhat identical setup for both
systems.
It is the intention that some cooperation
is implicitly present. When two agents pur-
sue the goal of killing the same enemy, they
should at least be able to avoid putting bombs
at the same spots, and instead attempt to
trap the enemy. This will generally be pos-
sible because of the autonomy of the agents;
they should choose paths and bomb locations
which seems reasonable, i.e. not place bombs
which will potentially hit allies or go through
a path in which a bomb may explode soon.
Fig. 1. The agent is stuck between
the boxes.
Some cooperation can, however, not be
done implicitly. Consider the situation in fig-
ure 1. Agent "26" is stuck between a number
of boxes without the possibility of placing a bomb to destroy them; it would kill
the agent as well.
In such situation the agent has two options: (1) wait for another agent to
autonomously choose to help the agent or (2) ask for help. Option (1) may be
possible but it seems irrational to wait for another agent to detect the situation
by himself. The agent being stuck should therefore always ask for help. This is
done using the contract net protocol [10].
The agents uses the well-known A*-algorithm [9] for pathfinding. In addition
to the heuristics used by the algorithm, we use a punishment value for every
location in the environment. The main difference of this algorithm compared to
A* is that included in the tentative value of a neighbor of the current location
is a punishment value depending on the objects on the location of that neigh-
bor. This will make the algorithm consider other, perhaps longer, paths, which
however may prove to be safer.
For instance, by specifying a punishment of 5 on a field containing a box, the
algorithm will consider paths, that avoids going through that box, which are up
to 5 steps longer. This may not seem as a big improvement, but it means that
if a single box is blocking a path, the agent will consider a path which is a little
longer. Compared to a situation where he blows up the box and continues, this
is usually more efficient, since he will have to wait for the bomb to explode and
the explosion to disappear. However, consider the situation depicted in figure
2(a). The agent wants to move from the current position, A, to the target, B, so
it will be highly inefficient to compute a path avoiding the boxes. A much more
efficient path would be to compute a path, in which a box must be destroyed.
To do this we introduce the notion of an intermediate target. An intermediate
target is a target in which a bomb should be placed in order to clear the way to
5
(a) A path avoiding boxes.
(b) A path with an intermediate target,
C.
Fig. 2. Different ways of computing a path from a location A to a target B.
the "real" target. In figure 2(b) we have an intermediate target, C. This path is
clearly preferred over the other.
The fact that AgentSpeak is Prolog-like makes it possible to specify a model
for the plans in a way that is easily transferable to Jason. The agent needs the
following predicates from its knowledge base:
-- pos(X, Y ): The current position of the agent.
-- target(X, Y ): A final target.
-- intermediate(X, Y ): A possible intermediate target.
-- clear(X, Y ): An intermediate target is clear, meaning that the targeted box
has been removed, creating a passage.
-- bombs(N ): The number of bombs currently available. As a shorthand we
write bombs(N>0) for (bombs(N ) ∧ N > 0).
We now present a model for agent a describing the situation above. Note
that in a scenario of Bomberman, there will generally be an enormous amount
of epistemic states, since there will be a state for each possible position, and
an agent will have different knowledge every time he is in a state, yielding even
more states. Therefore, we consider a more abstract and general model with
three possible locations: (XA, YA), (XB, YB) and (XC, YC), corresponding to
the agent's location, his target and intermediate target, respectively. These will
be referred to as A, B and C. We write pred( ) to match any value of that
predicate (i.e. only in the fact that the predicate exists in the knowledge base is
of interest).
For each of the three possible locations, a number of possible states exist.
In epistemic logic indistinguishable states are typically states the agent cannot
distinguish because of lack of knowledge. In the following we also refer to states
as indistinguishable if they, even though they are somewhat different in terms of
knowledge, will result in the agent performing the same action. This simplifies
the model greatly and can be considered as a model created from another agent's
point of view.
Figure 3 shows the model for the path-finding problem. The predicates in a
state are the predicates which the agent knows (or believes) to be true at that
6
Fig. 3. Model of the path-finding problem depicted in figure 2.
state. That is, if the agent is at the final target, B, and the intermediate target
is clear, he will be in the lower right state of the figure. We use this model to
create a plan for how to decide which path to choose.
Notice the two possible outcomes of moving towards C when the agent is
at A and has no bombs (with the intermediate target being blocked). This is
because of the fact that during the move towards C, a bomb may become avail-
able, meaning that the agent will be in a state where bombs(N>0) rather than
bombs(0).
Committing to a mission in J -Moise+ When the agents commit to a mission
in a scheme the J -Moise+ engine will generate goal achievement events for the
goals that are currently available. For instance, when an explorer commits itself
to the mission of exploration, it will automatically generate the goal achievement
event of finding an unexplored area. Whenever a goal is completed, an event for
the next goal of the plan is generated. In this case the next available goal will
be to move to the unexplored area. In this way, it is very easy to follow the plan
of a mission, since the goals are automatically generated when they have been
specified in the organization.
Basically, the explorer has the following plans:
+!exploreMap[scheme(Sch)]
<- jmoise.set_goal_state(Sch, exploreMap, satisfied).
7
pos(A)target(B)∧intermediate(C)∧¬clear(C)∧bombs(N>0)pos(A)target(B)∧intermediate(C)∧clear(C)pos(A)target(B)∧intermediate()pos(A)target(B)∧intermediate(C)∧¬clear(C)∧bombs(0)pos(B)target(B)∧clear(C)pos(B)target(B)∧¬clear(C)pos(C)target(B)∧intermediate(C)∧¬clear(C)∧bombs(N>0)pos(C)target(B)∧intermediate(C)∧clear(C)pos(C)target(B)∧intermediate()pos(C)target(B)∧intermediate(C)∧¬clear(C)∧bombs(0)movetowards(B)movetowards(B)movetowards(B)do(bomb)movetowards(B)movetowards(C)movetowards(C)movetowards(C)wait +!findUnexploredArea[scheme(Sch)]
: <context>
<- <plan to find unexplored area>;
jmoise.set_goal_state(Sch, findUnexploredArea, satisfied).
+!moveToUnexploredArea
: <context>
<- <plan to move to unexplored area>.
+near(_,_)
<- ?scheme(exploration, Sch);
jmoise.set_goal_state(Sch, moveToUnexploredArea, satisfied).
Notice that since the organizational specification shows exactly how to ex-
plore the map, it is only necessary to create plans for each goal event. When
the plan is successfully executed, the agent informs J -Moise+ that the goal
has been satisfied. It is then the responsibility of J -Moise+ to generate the
next goal event. Note that moving to an unexplored area is a bit different since
it uses the path-finding algorithm described in the previous chapter. Therefore,
the goal is satisfied only when the agent is near the unexplored area.
3 Results
We now present the main results gained during the work with Jason and J -
Moise+.
3.1 Agent-Centered Multi-Agent Systems
The ACMAS, along with its structure, is built form the ground. Having to build
everything from the ground gives a lot of freedom with regards to the structure of
the implementation; there are no constraints as to where specific details must be
implemented. This has lead to a solution where plans for achieving sub-goals and
reacting to percepts can be implemented concisely, while still doing as intended.
The resulting agents are therefore reacting quite fast to changes in the en-
vironment; with short code and only few precisely defined responsibilities, the
agents are easily able to prioritize during a game, if it, for instance, is necessary
to take cover from a bomb.
But the freedom one has with regards to structure has also been the biggest
issue during the implementation. Ensuring successful transition between goals
has caused some trouble during the implementation. The debugging functionality
can be quite tricky to master, and the only way to test the transition from one
goal to another is by executing the system. This can render the process quite
slow.
Overall though, we are quite satisfied with the resulting system; it satisfies
the proposed strategy and even though the agents may be quite simple, they are
8
able to cooperate to complete their tasks and use their knowledge to decide how
to move through the environment.
3.2 Organization-Centered Multi-Agent Systems
Building an OCMAS is a more well-structured process than that of building
an ACMAS, since it consists of two parts, (1) specifying organization and (2)
implementing the details of the organization where the latter depends on the
completion of the first. While this does not automatically result in a more struc-
tured program, it does force the user to think more about what, why and how.
When specifying the organization the focus is on what the overall goals are.
This leads to considering why these are the goals and in that way it allows us to
justify the choices made, even before they are implemented. Finally, when the
plans are implemented the focus is on how the agents are supposed to complete
their goals.
Generally, since only sub-goals, and not their relations, need to be imple-
mented, the code tends to be quite clear. However, without being able to study
the specification of the organization, it is not possible to see the relation be-
tween the goals (as it is handled automatically). Furthermore, as required by
J -Moise+, the code is often quite verbose, because of the statements required
to setup and manage the organizational structure (jmoise.create group(...)
etc.) and the extensive use of annotations (+!goal[scheme(Sch)] etc.). While
this in general makes the code quite clear, it also can result in situations where
one need to include a plan for a goal event in which the goal is simply set to be
satisfied. Consider the example below:
+!goal // available when subgoal is satisfied
<- jmoise.set_goal_state(goal, satisfied).
+!subgoal
<- <complete subgoal>;
jmoise.set_goal_state(subgoal,satisfied).
In this case, even though the primary goal is completed, when the subgoal
is completed, one has to explicitly state that the goal is satisfied even though
nothing else happens. This is not a serious problem, but in large scenarios with
complicated schemes, it may result in many "empty" plans.
Since the agents are part of an organization, they are required to do what-
ever their obligations tells them to do. To be able to do so, they need access
to the specification of the organization. As described, this access is provided
through a special agent. This means that when an agent has satisfied a goal,
this information is sent to the J -Moise+ agent which then determines what
the next goals are and informs the agent. This can decrease performance in very
active environments if the agent has no goals to pursue while waiting for new
information.
9
This can be seen quite clearly in the implementation of the explorer of the
OCMAS team. When the exploreMap goal is available, it is immediately satis-
fied. When this happens, the scheme for exploring is finished and a new scheme
must be created in order to continue the exploration. Compared to the explorer
of the ACMAS, the performance differences are quite clear; in the ACMAS,
the agent immediately chooses a new spot to move to, while in the OCMAS,
the agent waits for the generation of an appropriate goal events. The resulting
implementation fulfills the proposed strategy, however the road towards this re-
sult has been more bumpy than when working with the ACMAS. This will be
discussed below.
3.3 Using Jason and Moise+
Jason uses AgentSpeak which is an agent-oriented programming language. Such
a programming language is perfect for implementing goal-directed and reactive
behavior since one builds a set of plans for how to react to such events. AgentS-
peak is very similar to the logic programming language Prolog, and both the
language of AgentSpeak and the general features of Jason have been quite ex-
tensively documented in [2]. This makes it possible to quite easily understand
and exploit the features of the interpreter.
The debugging feature of Jason has lead to a few issues during the imple-
mentation. Often when attempting to debug, the entire system pauses and then
when attempting to perform a stepwise operation through the system, noth-
ing happens. This has lead to much trial and error and has overall slowed the
development process. Generally though, the system provides descriptive error
messages and the more acquainted one gets with the system, the easier errors
are spotted.
The J -Moise+ extension is built on Jason and uses the Moise+ model,
so in general the same things apply to this. However, since it is an extension, it
allows for more actions and there are a few more things one needs to be aware
of.
As mentioned, having an organization often leads to a very well structured
result since the user is required to really think about what the agents are sup-
posed to do. This is even more the case in J -Moise+ since goal events are
automatically generated, meaning that the user need not consider the transition
between the goals. Furthermore, the schemes that can be specified in the func-
tional specification of Moise+ makes coordination of tasks very easy. Simply
by specifying the cardinality of a goal in a scheme, the user specifies how many
agents must complete this goal before it is completed within the scheme. For
instance, a goal event can be synchronized by having a sub-goal that all agents
must satisfy before the actual goal event is created.
The Moise+ organizational model has been quite extensively described in
[5,6,7] along with a tutorial of the details of how to use it in [8]. This makes it
very easy to understand how the different concepts are related and should be
used.
10
Organizational knowledge When a group or scheme is created, the agents will
perceive certain events so that they are able to react accordingly. In order to be
able to distinguish between similar events, annotations are added that among
other things include which agent created the organizational object.
This can be used to let an agent decide not to join a group if a specific agent
has created it, or only committing to a mission it is permitted to commit to, if
it is related to a specific group. This is a great use of the Jason annotations,
as it is perfectly clear how to use them. Furthermore, because it is annotations
they will not be shown if the programmer chooses not to use them.
What Moise+ is lacking in term of organizational knowledge is the ability
for an agent to know whether it is allowed to join a group before it attempts to
join it. The reason for this is that if it is not permitted to join a group or play
a role, an error event is created. This should be okay, but it is not possible for
the agent to reason about the error in details so it will not know why it could
not join the group.
Overall, both tools are quite pleasant to work with once acquainted with
them.
4 OCMAS vs ACMAS: When to Use What?
The work with the two approaches has lead to a discussion of the implementation
of each team as well as the two tools used for building the implementations.
Generally speaking, one approach is not better than the other but given the
results above it is clear that there are situations more suited for one approach
than the other.
Figure 4 gives an overview of the main results of the comparison. The fig-
ure uses two parameters as basis: the number of agents in the system and the
structural complexity. When the system has a high structural complexity, there
are greater advantages of dividing the implementation into two distinct parts:
what and how. This means that while an organization can be applied to simple
systems, they will in most cases not benefit much from this.
Notice that the two approaches overlap. The reason is that there will always
be situations where it is not clear whether one system is an advantage over the
other. The Team-Based Bomberman is an example of such system. While the
results show a bias towards the ACMAS in this case, it is partly due to the
communication overhead in J -Moise+. If this problem is solved, the OCMAS
could be performing just as well as the ACMAS.
Personal software assistant: A personal software assistant is a simple agent
resposible for assisting an end-user with certain tasks that can be automated.
For such purpose it seems an organization will be inappropriate. The primary
reason for this is that the system will generally consist of very few agents (in
many cases only one) and the complexity will be low. In such cases there is not
much sense in creating an organization, since the system will not benefit from
the OS. It will probably consist of a single group with few roles that the agents
11
Fig. 4. An overview of the main results.
can play. Building an entire organization for very few agents can in most cases
not be justified.
Distributed calculations: Consider a system of intelligent agents, which have one
or more sensors. This could for instance be the "distributed sensing" scenario
described in [10]. Here an agent has a clear responsibility of sensing the environ-
ment and using its calculations in some well-defined way. The role of an agent
is defined by the sensors it can use, i.e. it has a static role. At all time the
information it is computing will be used for the same purpose.
In such cases there is no need to build an organization since the missions are
very simple and there is no explicit need of coordination. Furthermore, being in
a group would not change the behavior of an agent in the system since its role
and responsibilities remains.
In other words: Even though the system can contain many agents, the struc-
tural complexity remains low. Therefore the agents will not benefit much from
making the organization explicit.
12
manyagentsfewagentshighstructuralcomplexitylowstructuralcomplexityOCMASACMASPersonal SoftwareAssistantDistributedCalculationsPaper ReviewTeam-BasedBombermanMulti-AgentProgrammingContest 2009Multi-AgentProgrammingContest 2006 Paper review process: In [3] an example of the "reviewing process" of papers
in a conference is considered. In this example we have a group for submission
of papers and one for evaluation. Being in a certain group then gives an agent
certain responsibilities, such as evaluating the papers that are being submitted.
This is a very specific example of the use of an organization but it can eas-
ily be generalized to situations where certain agents are depending on results
from other agents. By grouping such agents and creating schemes they can com-
mit to, the general structure of the dependence-relation is immediate and the
implementation is easily constructed using it.
While the number of agents may vary it is clear that the structural complexity
is higher than in the previous examples. An explicit organization will definitely
make the implementation much easier since responsibilities and acquaintances
are well-structured.
Games: Games can be quite different and naturally there is no definite answer to
whether using ACMAS or OCMAS would be better. In such situations it is impor-
tant to realize how complex the game is. If the game consists of one well-defined
type of controllable character, an organization is probably not a good choice.
However, if the game consists of several different characters, all with different
possibilities, it may be reasonable at least to consider whether an organization
could be useful.
This seems to indicate that using an organization for a Bomberman game
may not be the best choice. In this specific case the ACMAS solution is better
than the OCMAS; the agents react faster, can more easily adapt to changes
and is in general more robust. An organization can be justified in a game such
as Bomberman if we include features from the original game that would make
the game more complex (e.g. power-ups). An OCMAS will benefit in this case,
since it is possible to specify advanced roles and missions in the OS that would
otherwise be difficult to implement.
Multi-Agent Programming Contest: The area of multi-agent systems is quite
active, which for instance can be seen by the annual multi-agent programming
contest. The primary aim of the competition is to "stimulate research in the area
of multi-agent system development and programming" [1]. This is achieved by
developing a scenario of a dynamic environment in which cooperation is the key
to success. Different multi-agent systems are competing in the scenario in a set of
games to determine their performance. As illustrated in figure 4 the complexity
and number of agents in the contest has increased over the years. Therefore,
while the implementations would not benefit much from an explicit organization
in the first scenarios (where only a few agents where required to solve simple
tasks), the increased complexity has made this approach a reasonable choice
(in scenarios where more than 10 agents are required to cooperate in order to
succeed).
The overall complexity of the scenarios in the competition has increased,
meaning that it may be easier to implement a better strategy using OCMAS
13
simply because the complexity is easier handled when the structure of the solu-
tion has been made explicit.
5 Future Work & Conclusions
By taking both the agent-centered and organization-centered approach for im-
plementing the same strategy we have gained insights about both the advantages
and disadvantages of each approach. The focus has been on a single scenario,
which means that not all corners of the approaches have been investigated. Even
so, the results have made several differences of the approaches clear, differences
that in some situations makes one approach highly advantageous compared to
the other.
The focus has been on differences between different types of multi-agent
systems and in particular the use of two specific tools. However, it was shown
that both types of systems are useful in different situations and that there is
no definitive answer to when one system is a better choice than another. The
main reason is the many factors the programmer must consider when choosing
between the agent- and organization-oriented approaches. These factors include
the quality of the implementation, the actual tools used and the strategy to be
implemented.
While we have been able to discuss the differences and make suggestions
for which system is most suitable in different situations, it would be interesting
to be able to create systems of both types which exhibit the same behavior in
most situations. This would make it possible to compare the actual performance
difference between the systems. However, since this requires specialized systems,
there is a chance that the results would not apply to real-world applications.
The area of multi-agent systems is still somewhat new and is continuously
growing. With the addition of the organizational aspects it has been made pos-
sible to create even more sophisticated and advanced systems. This comparison
has shown that both approaches have advantages and disadvantages and are
well-suited for different situations.
In the end, it is hard to say which approach is better and a decision should
be justified by doing some research on the application at hand and the possible
tools for creating the system.
There is still much work to be done in the area of organizational multi-agent
systems, specifically in the Moise+ organizational model, but also the principles
of OCMAS in general.
The tools available makes it possible to implement advanced systems (both
ACMAS and OCMAS) which are very useful in both research and practical
applications and there is no doubt that the area will continue to develop even
more efficient and intelligent solutions to research problems and real-world ap-
plications.
14
References
1. Tristan M. Behrens, Jurgen Dix, Jomi Hubner, and Michael Koster. Multi Agent
Contest. http://www.multiagentcontest.org/, October 2010.
2. Rafael H. Bordini, Jomi Fred Hubner, and Michael Wooldridge. Programming
Multi-Agent Systems in AgentSpeak using Jason. John Wiley & Sons Ltd, 2007.
3. Jacques Ferber, Olivier Gutknecht, and Fabien Michel. From Agents to Organiza-
tions: an Organizational View of Multi-Agent Systems. Agent-Oriented Software
Engineering (AOSE) IV, pages 214 -- 230, 2004.
4. Mahdi Hannoun, Olivier Boissier, Jaime Simao Sichman, and Claudette Sayettat.
MOISE: An Organizational Model for Multi-agent Systems. Proceedings of the In-
ternational Joint Conference, 7th Ibero-American Conference on AI, 15th Brazilian
Symposium on AI, 2000.
5. Jomi Fred Hubner, Jaime Simao Sichman, and Olivier Boissier. A Model for the
Structural, Functional, and Deontic Specification of Organizations in Multiagent
Systems. Proceedings of the 16th Brazilian Symposium on Artificial Intelligence,
2002.
6. Jomi Fred Hubner, Jaime Simao Sichman, and Olivier Boissier. S-Moise+: A
Middleware for developing Organised Multi-Agent Systems. Proceedings of the
International Workshop on Organizations in Multi-Agent Systems, from Organiza-
tions to Organization Oriented Programming in MAS, 2005.
7. Jomi Fred Hubner, Jaime Simao Sichman, and Olivier Boissier. Developing Or-
ganised Multi-Agent Systems Using the Moise+ Model: Programming Issues at
the System and Agent Levels. International Journal of Agent-Oriented Software
Engineering, 2007.
8. Jomi Fred Hubner, Jaime Simao Sichman, and Olivier Boissier. Moise+ tutorial.
http: // moise. sourceforge. net/ , 2008.
9. Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach.
Prentice-Hall, 2nd edition, 2003.
10. Michael Wooldridge. An Introduction to MultiAgent Systems. John Wiley & Sons
Ltd, 2nd edition, 2009.
Acknowledgement
I thank associate professor Jørgen Villadsen, DTU Informatics, Denmark.
15
|
1904.12771 | 1 | 1904 | 2019-04-29T15:28:36 | Consensus Control for Leader-follower Multi-agent Systems under Prescribed Performance Guarantees | [
"cs.MA",
"eess.SY"
] | This paper addresses the problem of distributed control for leader-follower multi-agent systems under prescribed performance guarantees. Leader-follower is meant in the sense that a group of agents with external inputs are selected as leaders in order to drive the group of followers in a way that the entire system can achieve consensus within certain prescribed performance transient bounds. Under the assumption of tree graphs, a distributed control law is proposed when the decay rate of the performance functions is within a sufficient bound. Then, two classes of tree graphs that can have additional followers are investigated. Finally, several simulation examples are given to illustrate the results. | cs.MA | cs |
Consensus Control for Leader-follower Multi-agent Systems under
Prescribed Performance Guarantees
Fei Chen and Dimos V. Dimarogonas
Abstract -- This paper addresses the problem of distributed
control
for leader-follower multi-agent systems under pre-
scribed performance guarantees. Leader-follower is meant in
the sense that a group of agents with external
inputs are
selected as leaders in order to drive the group of followers
in a way that the entire system can achieve consensus within
certain prescribed performance transient bounds. Under the
assumption of tree graphs, a distributed control law is proposed
when the decay rate of the performance functions is within a
sufficient bound. Then, two classes of tree graphs that can have
additional followers are investigated. Finally, several simulation
examples are given to illustrate the results.
I. INTRODUCTION
The consensus problem has attracted great interest due
to its wide applications in robotics, cooperative control [7],
formation [1] and flocking [20]. Consensus or agreement is
achieved when a group of agents converge to a common
value. The first order consensus protocol was first introduced
in [15], where the authors discussed the consensus problem
of directed and undirected graphs with fixed or switching
topologies and time delays. Second order consensus protocol
has been investigated in [18], where the states of the agents
converge to a constant or a linear function.
In this work, we study the consensus problem in a
leader-follower framework, that is, one or more agents are
selected as leaders with external inputs in addition to the
first order consensus protocol. The remaining agents are
followers only obeying the first order consensus protocol.
Recent research that has been done in the leader-follower
framework can be divided into two parts. The first part
deals with the controllability of leader-follower multi-agent
systems. For instance, controllability of networked systems
was first investigated in [19] by deriving conditions on the
network topology, which ensures that the network can be
controlled by a particular member which acts as a leader.
In [6], [17], the authors identify necessary conditions for the
controllability of the corresponding leader-follower networks
using equitable partitions of graphs. Controllability condi-
tions for leader-follower multi-agent systems with double
integrator dynamics and their connection with graph topol-
ogy properties are addressed in [9]. The second part targets
leader selection problems [21], [16], [8]. These involve the
problem of how to choose the leaders among the agents such
This work was supported by the EU H2020 Co4Robots Project,
the
Swedish Research Council (VR) and the Knut och Alice Wallenberg
Foundation (KAW).
Fei Chen and Dimos V. Dimarogonas are with the Division of Decision
and Control Systems, KTH Royal Institute of Technology, SE-100 44
Stockholm, Sweden {fchen,dimos}@kth.se
that the leader-follower system satisfies the requirements
such as controllability, optimal performance or formation
maintenance.
Prescribed performance control (PPC) was originally pro-
posed in [3], with the aim to prescribe the evolution of
system output or the tracking error within some predefined
region. For example, an agreement protocol that can ad-
ditionally achieve prescribed performance for a combined
error of positions and velocities is designed in [13] for
multi-agent systems with double integrator dynamics, while
PPC for multi-agent average consensus with single integrator
dynamics is presented in [11]. In [2], the authors consider the
formation control problem for nonlinear multi-agent systems
with prescribed performance guarantees and connectivity
constraints. Funnel control, which uses a similar idea as PPC
was first introduced in [10] for reference tracking. In [4], the
authors utilize funnel control for uncertain nonlinear systems
that have arbitrary strict relative degree and input-to-state
stable internal dynamics.
In this work, we are interested in how to design control
strategies for the leaders such that the leader-follower multi-
agent system achieves consensus within certain performance
bounds. Compared with existing work of PPC for multi-agent
systems [13], we apply a PPC law only to the leaders while
most of the related work, including [13], applies PPC to all
the agents to achieve consensus. The benefit of this work
is to lower the cost and control effort since the followers
will follow the leaders by obeying first order consensus
protocols without any additional control. Unlike other leader-
follower consensus approaches using PPC [12], in which
the multi-agent system only has one leader and the leader
is treated as a reference for the followers, we focus on
a more general framework in the sense that we can have
more than one leader and the leaders are designed in order
to steer the entire system achieving consensus within the
prescribed performance bounds. The difficulties in this work
are due to the combination of uncertain topologies, leader
amount and leader positions. In addition, the leader can only
communicate with its neighbouring agents. The contributions
of the paper can be summarized as: i) within this general
leader-follower framework, under the assumption of tree
graphs, a distributed control law is proposed when the decay
rate of the performance functions is within a sufficient bound;
ii) the specific classes of chain and star graphs that can have
additional followers are investigated.
The rest of the paper is organized as follows. In Section
II, preliminary knowledge is introduced and the problem is
formulated, while Section III presents the main results, which
are further verified by simulation examples in Section IV.
Section V closes with concluding remarks and future work.
that Lx = 0. By stacking (1) and (2), the dynamics of the
leader-follower multi-agent system is rewritten as:
II. PRELIMINARIES AND PROBLEM STATEMENT
Σ : x = −Lx + Bu,
(3)
A. Graph Theory
i is defined as di = Pj∈Ni
An undirected graph [14] G = (V, E) comprises of the
vertices set V = {1, 2, . . . , n} and the edges set E =
{(i, j) ∈ V × V j ∈ Ni} indexed by e1, e2, . . . , em.
Here, m = E is the number of edges and Ni denotes the
agents in the neighbourhood of agent i that can commu-
nicate with i. The adjacency matrix A of G is the n × n
symmetric matrix whose elements aij are given by aij = 1,
if (i, j) ∈ E, and aij = 0, otherwise. The degree of vertex
aij . Then the degree matrix
is ∆ = diag(d1, d2, . . . , dn). The graph Laplacian of G is
L = ∆ − A. A path is a sequence of edges connecting two
distinct vertices. A graph is connected if there exists a path
between any pair of vertices. By assigning an orientation
to each edge of G we can define the incidence matrix
D = D(G) = [dij] ∈ Rn×m. The rows of D are indexed
by the vertices and the columns are indexed by the edges
with dij = 1 if the vertex i is the head of the edge (i, j),
dij = −1 if the vertex i is the tail of the edge (i, j) and
dij = 0 otherwise. Based on the incidence matrix, the graph
Laplacian of G can be described as L = DDT . In addition,
Le = DT D is the so called edge Laplacian [14] and cij
denotes the elemnts of Le.
B. System Description
In this work, we consider a multi-agent system with
vertices V = {1, 2, . . . , n}. Without loss of generality, we
suppose that the first nf agents are selected as followers
while the last nl agents are selected as leaders with respective
vertices set VF = {1, 2, . . . , nf }, VL = {nf + 1, nf +
2, . . . , nf + nl} and n = nf + nl.
Let xi ∈ R be the position of agent i, where we only
consider the one dimensional case, without loss of generality.
Specifically, the results can be extended to higher dimensions
with appropriate use of the Kronecker product. The state
evolution of each follower i ∈ VF is governed by the first
order agreement protocol:
xi = Xj∈Ni
(xj − xi),
(1)
while the state evolution of each leader i ∈ VL is governed by
the first order agreement protocol with an assigned external
input ui ∈ R:
where L is the graph Laplacian and B = h 0nf ×nl
Inl
C. Prescribed Performance Control
i .
The aim of PPC is to prescribe the evolution of the system
output or the tracking error within some predefined region
described as follows:
− Mijρij(t) < xij (t) < ρij(t)
if xij (0) > 0
− ρij(t) < xij (t) < Mijρij(t)
if xij (0) < 0
(4)
(5)
ρij(t) : R+ → R+ \ {0} are positive, smooth and stirctly de-
creasing performance functions that introduce the predefined
bounds for the target system outputs or the tracking errors.
One example choice is ρij(t) = (ρij0 − ρij∞)e−lij t + ρij∞
with ρij0, ρij∞ and lij positive parameters and ρij∞ =
limt→∞ρij (t) represents the maximum allowable tracking
error at
represents the maximum
allowed overshot.
the steady state; Mij
Normalizing xij (t) with respect to the performance func-
tion ρij (t), we define the modulated error as xij (t) and the
corresponding prescribed performance region Dij :
xij (t) =
xij (t)
ρij (t)
Dij , {xij : xij ∈ (−Mij, 1)}
if xij (0) > 0
Dij , {xij : xij ∈ (−1, Mij)}
if xij (0) < 0
(6)
(7)
(8)
Then the modulated error is transformed through the trans-
formed function Tij that defines the smooth and strictly
increasing mapping Tij : Dij → R and Tij(0) = 0. One
example choice is
Tij(xij ) = ln(cid:18)−
xij + 1
xij − Mij(cid:19) .
Hence the transformed error is defined as
εij(xij ) = Tij(xij )
(9)
(10)
if the transformed error εij (xij )
It can be verified that
is bounded, then the modulated error xij
is constrained
within the regions (7), (8). This also implies the error xij
evolves within the predefined performance bounds (4) and
(5), respectively. Differentiating (10) with respect to time,
we derive
xi = Xj∈Ni
(xj − xi) + ui.
(2)
where
Let x = [x1, . . . , xnf , . . . , xn]T ∈ Rn be the stack
the agents and u =
vector of absolute positions of all
[u1, . . . , unl]T ∈ Rnl be the control input vector . Denote
¯x = [¯x1, . . . , ¯xm]T as the stack vector of relative positions
between the pair of communicating agents (i, j) ∈ E, where
¯xk , xij = xi −xj, k = 1, 2, . . . , m. It can be easily verified
that Lx = D ¯x and ¯x = DT x. In addition, if ¯x = 0, we have
εij(xij ) = JTij (xij , t)[ xij + αij (t)xij ]
(11)
JTij (xij , t) ,
∂Tij(xij )
1
∂ xij
ρij (t)
> 0
αij (t) , −
ρij(t)
ρij(t)
> 0
(12)
(13)
are the normalized Jacobian of the transformation function
Tij and the normalized derivative of the performance func-
tion, respectively.
D. Problem Statement
In this work, we are interested in how to design a control
strategy for the leader-follower multi-agent system given by
(3) such that the controlled system can achieve consensus
within the prescribed performance requirements. The control
strategy is only applied to the leaders and these drive the
followers to guarantee the entire multi-agent system meet
the requirements. Formally,
Problem 1. Let the leader-follower multi-agent system Σ
defined by (3) with the communication graph G = (V, E) and
the prescribed performance functions ρij, (i, j) ∈ E. Derive
a control strategy such that the controlled leader-follower
multi-agent system achieves consensus within ρij .
III. MAIN RESULTS
In this section, we design the control for the leader-
follower multi-agent system (3) such that the system can
achieve consensus within the prescribed performance func-
tions
ρij(t) = (ρij0 − ρij∞)e−lij t + ρij∞.
(14)
Here the performance functions are chosen as (14) with-
out loss of generality and the communication agents share
information about their performance functions and trans-
formation functions, that is, ρij (t) = ρji(t), Mij = Mji
and Tij(xij ) = −Tji(xji). This means the communication
between the neighbouring agents are bidirectional and the
graph G is assumed undirected.
Consensus is achieved in the sense that the stack vector
¯x of relative positions converges to zero as t → ∞. We
then rewrite the dynamics of the leader-follower multi-agent
system (3) into the edge space in order to characterise
the dynamics of the relative positions. We first rewrite (3)
into the dynamics corresponding to followers and leaders,
respectively. The corresponding incidence matrix is denoted
with Df , Di denoting the incidence
matrices that characterise how followers and leaders are
connected with other agents. Then (3) is reorganised as
as D = (cid:2)DT
i (cid:3)T
f DT
load (less edges) for their implementation. Note however
that further results for a general graph could be built based
on the results of tree graphs, for example, through graph
decompositions [22]. For the leader-follower multi-agent
system (16), the proposed controller applied to the leader
agents is the composition of the term based on prescribed
performance of the positions of the neighbouring agents:
ui = − Xj∈Ni
gijJTij (xij , t)εij(xij ),
i ∈ VL,
(17)
where gij = gji is a positive scalar gain to be appropriately
tuned. Then the stack input vector is
u = −DiJT (¯x, t)Gε(¯x),
(18)
where ¯x is the stack vector of transformed errors xij , G ∈
Rm×m is a positive definite diagonal gain matrix with entries
gij . JT (¯x, t) ∈ Rm×m is a time varying diagonal matrix
with diagonal entries JTij (xij , t), ε(¯x) ∈ Rm is a stack
vector with entries εij (xij ). Then the edge dynamics (16)
with input (18) can be written as
¯x = −Le ¯x − DT
i DiJT (¯x, t)Gε(¯x),
(19)
In the sequel, we develop the following result and will use
Lyapunov-like methods to prove that the prescribed perfor-
mance can be guaranteed and consensus can be achieved.
Theorem 1. Consider the leader-follower multi-agent system
Σ under Assumption 1 with dynamics (3), the predefined
performance functions ρij as in (14) and the transformation
function s.t. Tij(0) = 0, ∀(i, j) ∈ E, and assume that the
initial conditions xij (0) are within the performance bounds
(4) or (5). If the following condition holds:
¯γ ≥ l = max
(i,j)∈E
(lij),
(20)
where l is the largest decay rate of ρij(t) and ¯γ is the
maximum value of γ that ensures:
Σ : (cid:20) xf
xl(cid:21) = (cid:20)Af Bf
BT
f Ai(cid:21)(cid:20)xf
(cid:2)x1 x2
xl(cid:21) +(cid:20)0nf ×nl
Inl (cid:21) u,
xnf(cid:3)T
=
, xl
f , Bf =
i . Multiplying with DT on both sides of
and Af = Df DT
xnf +nl(cid:3)T
· · ·
=
i , Ai = DiDT
Df DT
(15), we obtain the dynamics on the edge space as
xf
· · ·
where
(cid:2)xnf +1
(15)
Γ = (cid:20)
DT
i Di
1
2 (Le−γ(Im−DT
i Di))
1
2 (Le−γ(Im−DT
i Di))
γLe
(cid:21) ≥ 0. (21)
Then, the controlled system achieves consensus within the
prescribed performance bounds ρij (t) when applying the
control (18).
Proof. Consider the Lyapunov-like function
Σe : ¯x = −Le ¯x + DT
i u,
(16)
with the edge Laplacian Le. We know that Le is positive
definite if the graph is a tree [5]. We thus here assume the
following
Assumption 1. The leader-follower multi-agent system (3)
described by the graph G = (V, E) is a connected tree.
V (ε¯x, ¯x) =
1
2
εT
¯x Gε¯x +
γ
2
¯xT ¯x,
(22)
with ε¯x denoting ε(¯x) and JT ¯x denoting JT (¯x, t). Then,
V = εT
¯x G ε¯x + γ ¯xT ¯x. Replacing ε¯x according to (11), we
obtain
V = εT
¯x GJT ¯x ( ¯x + α(t)¯x) + γ ¯xT ¯x,
(23)
We consider tree graphs as a starting point since we
need the positive definiteness of Le in the analysis, and
motivated by the fact that they require less communication
where α(t) is the diagonal matrix with diagonal entries
αij (t). According to (13) and (14), we know that αij(t) <
lij, ∀t. Substituting (19), we can further derive that
V =εT
¯x GJT ¯x (−Le ¯x − DT
+ γ ¯xT (−Le ¯x − DT
¯x GJT ¯x Le ¯x + εT
¯x GJT ¯x DT
= − εT
− εT
− γ ¯xT DT
i DiJT ¯x Gε¯x
i DiJT ¯x Gε¯x + α(t)¯x)
i DiJT ¯x Gε¯x)
¯x GJT ¯x α(t)¯x
(24)
i DiJT ¯x Gε¯x − γ ¯xT Le ¯x
Adding and subtracting γεT
of (24), we obtain
¯x GJT ¯x ¯x on the right hand side
V = − εT
− εT
= − εT
¯x GJT ¯x (γIm − α(t))¯x − εT
¯x GJT ¯x Le ¯x − γ ¯xT Le ¯x + γεT
¯x GJT ¯x (γIm − α(t))¯x
¯x GJT ¯x DT
i DiJT ¯x Gε¯x
¯x GJT ¯x (Im − DT
i Di)¯x
− yT (cid:20)
DT
i Di
1
2 (Le−γ(Im−DT
i Di))
1
2 (Le−γ(Im−DT
i Di))
γLe
(cid:21) y
= − εT
¯x GJT ¯x (γIm − α(t))¯x − yT Γy
with
y = (cid:20)JT ¯x Gε¯x
¯x
(cid:21) .
t → ∞ by applying Barbalat's Lemma. This implies ¯x → 0
as t → ∞ and consensus will be achieved.
lim
t→∞
n Pn
n Pn
Remark 1. We are always interested in specifying the
state of the multi-agent system at the equilibrium. Denote
xc = 1
i=1 xi as the centroid of the network. In most of
xc(t) = xc(0) =
the work regarding PPC like [13],
1
i=1 xi(0). This is because a PPC input for every agent
exists. In our work, if we have an external input for every
agent, i.e. B = In in (3), we can also obtain lim
xc(t) =
t→∞
1
i=1 xi(0). This can be verified by multiplying 1T on
both sides of (3), where 1 ∈ Rn with all entries 1. Then, we
can conclude xc(t) = 0. The main difference is that when we
choose some leaders, we can achieve a varying equilibrium
state of each agent by tuning the gain matrix, which is quite
useful in practical design as we can decide where all the
agents should gather.
n Pn
(25)
(26)
In the sequel, we will discuss the results for two specific
classes of tree graphs: chain and star graph. First we con-
sider the chain graph, which is wildly used for instance in
autonomous vehicle platooning.
Since G, JT ¯x are both diagonal and positive definite matri-
ces, we have that GJT ¯x is also a diagonal positive definite
matrix. (γIm − α(t)) is a diagonal positive definite matrix
if γ ≥ l = max(lij) > ¯α = sup αij (t). Due to Tij(0) = 0,
we have εij(xij )xij ≥ 0. Then, by by setting γ := θ + ¯α,
with θ being a positive constant we get:
− εT
¯x GJT ¯x(γIm − α(t))¯x ≤ −θεT
¯x GJT ¯x ¯x
Then, according to (6), (12), we further obtain
− θεT
¯x GJT ¯x ¯x = −θεT
¯x G
∂ε¯x
∂ ¯x
¯x ≤ 0.
(27)
(28)
(28) holds because the transformed function is smooth and
strictly increasing and εij(xij )xij ≥ 0. Therefore, in order
V ≤ 0 to hold, it suffices that γ ≥ l = max(lij) >
for
sup αij(t) and in addition, Γ should be semi-positive def-
inite. Here, in order for Γ ≥ 0 to be feasible, we need
the assumption that the communication graph is a tree. This
further means that Le is positive definite and (21) is then
equivalent to:
DT
i Di≥ 1
4γ (Le−γ(Im−DT
i Di))L−1
e (Le−γ(Im−DT
i Di)).
(29)
Then, based on condition (20), and choosing γ = ¯γ, we
obtain −εT
¯x GJT ¯x(¯γIm − α(t))¯x ≤ 0 and Γ ≥ 0. Finally, we
V ≤ 0 when γ = ¯γ. This also implies
can conclude that
V (ε¯x, ¯x) ≤ V (ε¯x(0), ¯x(0)). Hence if ¯x(0)) is chosen within
the region (7) or (8) then V (ε¯x(0), ¯x(0)) is finite, which
implies that V (ε¯x, ¯x) is bounded ∀t. Therefore ε¯x, ¯x are
bounded and the boundedness of the transformed error ε¯x
implies that the relative position ¯x(t) evolves within the
prescribed performance bounds ∀t. Then we can prove the
boundedness of V (ε¯x, ¯x) based on the boundedness of ε¯x, ε¯x.
The boundedness of V (ε¯x, ¯x) implies the uniform continuity
V (ε¯x, ¯x) → 0 as
of
V (ε¯x, ¯x), which in turn implies that
Definition 1. A chain Gc = (V c, E c) is a tree graph with
vertices set V c = {1, 2, . . . , n}, n ≥ 2 and edges set E c =
{(i, i + 1) ∈ V c × V c i ∈ V c \ {n}} indexed by ei =
(i, i + 1), i = 1, 2, . . . , n − 1.
Note that (20) in Theorem 1 is a sufficient but not
necessary condition. For a chain graph, the matrix inequality
(21) may be actually infeasible when the graph has 2 or more
followers. The following result for Gc is derived.
Proposition 1. Consider the leader-follower multi-agent sys-
tem Σ described by (3) with the communication chain graph
Gc = (V c, E c) and the followers set V c
F = {1, 2, . . . , nf },
the predefined performance functions ρij as in (14) and
the transformation function s.t. Tij(0) = 0, ∀(i, j) ∈ E,
and assume that the initial conditions xij (0) are within the
performance bounds (4) or (5). Then, the chain can only
have at most 3 followers (nf ≤ 3) in order to achieve
consensus within the prescribed performance bounds ρij(t)
when applying (18). Specifically, when the chain has 2 and
3 followers,
max
(i,j)∈E
max
(i,j)∈E
(lij ) = l ≤ 2,
nf = 2;
(lij ) = l ≤ 1,
nf = 3
(30)
are the respective sufficient conditions under which the
chain achieves consensus within the prescribed performance
bounds ρij(t) when applying (18).
Proof. When the chain graph has only one follower, that
is nf = 1, the result can be proved by using Theorem
1. Let ¯γ be the maximum value of γ that ensures (21)
holds. By further choosing the decay rate of the performance
functions (14) to satisfy (20), we can conclude that the
controlled system achieves consensus within the prescribed
performance bounds by applying (18) based on Theorem
1. When the chain has additional followers, the condition
in Theorem 1 may be infeasible since it is a sufficient but
not necessary condition. But for this kind of special chain
structure, we can resort to checking the edge dynamics (16)
directly. It can be shown that −Le has elements given by
cij = −2 when i = j, cij = 1 when i − j = 1 and cij = 0
otherwise when the graph is a chain. We then rewrite (16)
as
(cid:20) ¯xf
¯xl (cid:21) = (cid:20) A B
BT C (cid:21)(cid:20) ¯xf
¯xl (cid:21) +(cid:20) 0
D (cid:21) u,
(31)
where ¯xf ∈ R(nf −1) represents the edges between followers,
while ¯xl ∈ Rnl represents the edge that connects the leader
node {nf +1} and the follower node {nf }, and the edges be-
tween leaders. Both A ∈ R(nf −1)×(nf −1), C ∈ Rnl×nl have
the same structure as −Le but with different dimensions, B
has an element 1 at row (nf − 1), column 1 (bottom left
corner) that represents the connection between the follower
node {nf } and the leader node {nf +1}. 0 is a (nf −1)×nl
zero matrix. D ∈ Rnl×nl has elements given by dij = 1
when i = j, dij = −1 when i − j = 1 and dij = 0
otherwise. Then we can analyse the leader part ¯xl and the
follower part ¯xf separately. For ¯xl, it can be proved that
¯xl achieves consensus within the performance bounds based
on the positive definiteness of DDT when applying control
(18). We further rewrite the follower part as
¯xf = A¯xf + b¯x⋆,
(32)
where b ∈ R(nf −1) is the first column of B, i.e., with the
last element equals to 1 and all other elements equal to 0.
¯x⋆ represents the edge between the follower node {nf } and
the leader node {nf + 1}. We can furture solve the state
evolution of (32) as follows:
¯xf (t) = eAt ¯xf (0) +Z t
0
eA(t−τ )b¯x⋆(τ )dτ
= M T eΛtM ¯xf (0) +Z t
0
eA(t−τ )b¯x⋆(τ )dτ,
(33)
eA(t−τ )b¯x⋆(τ )dτ,
= ¯x0
f (t) +Z t
f (t) = (cid:2)¯x0
0
¯x0
. . .
1(t)
¯x0
2(t)
nf −1(t)(cid:3)T
where ¯x0
is zero
input trajectories, that is when ¯x⋆(t) = 0, ∀t; A = M T ΛM ,
where Λ is a diagonal matrix with diagonal entries negative
and equal to the eigenvalues of A, which is due to A
having the same structure as −Le, and M is the matrix
composed with the corresponding eigenvectors. Without loss
of generality, suppose all performance functions are the same
and described by
Then, ¯x1(t) is within the performance bound ρ(t),
¯x1(t) < ρ(t), ∀t, when l ≤ 2 and in addition,
i.e.,
Z t
0
e−2(t−τ )¯x⋆(τ )dτ < (ρ0 − ¯x1(0))e−2t + ρ∞(1 − e−2t),
(36)
which can be ensured by tuning a large enough gain g32
to the leader indexed by node 3. From (36), we know that
when the relative position between the two followers is close
to the boundary, we need to tune a larger gain for the leader
that connects the followers. When nf = 3, we can derive a
similar result. In particular, we now have that
(cid:20)¯x0
2(t)(cid:21) = M T eΛtM (cid:20)¯x1(0)
¯x2(0)(cid:21) < k(cid:20)ρ0
1(t)
¯x0
ρ0(cid:21) e−t,
(37)
i (t) < ρ0e−t, i = {1, 2}.
with k = 1, which implies that ¯x0
Similarly, we can conclude that when l ≤ 1, and in addition
the tuning gain g43 for the leader indexed by node 4
is large enough, the controlled system achieves consensus
within the prescribed performance bounds. When nf ≥ 4,
i (t) < kρ0eλmax(A)t, i =
it can be proved similarly that ¯x0
{1, 2, . . . , nf − 1}, but with k > 1. This means that ¯x0
i (t)
cannot be bounded by ρ0eλmax(A)t for any initial conditions
within the performance bounds. Therefore, we can conclude
that in order to achieve consensus within the performance
bounds for all initial condition xij (0) within the performance
bounds (4) or (5), nf should be less or equal to 3.
Remark 2. Proposition 1 indicates that for a chain graph,
in order to achieve consensus within the prescribed perfor-
mance bounds, we can only have at most 3 consecutive
followers at the end of the graph. In addition, when the
initial relative position between 2 followers is close to the
prescribed performance boundary, we need to tune a large
enough gain for the leader that connects the followers.
Now we consider another specific class, in particular the
star graph Gs = (V s, E s) which is defined as follows.
Definition 2. A star Gs = (V s, E s) is a tree graph with
vertices set V s = {1, 2, . . . , n}, n ≥ 2 where vertice n is
called the centering node, and the edges set E s = {(i, n) ∈
V s × V s i ∈ V s \ {n}} indexed by ei = (i, n), i =
1, 2, . . . , n − 1.
Then, the following result can be derived.
Proposition 2. Consider the leader-follower multi-agent
system Σ described by (3) with the communication star graph
Gs = (V s, E s) and the leader set V s
L = {n}, the predefined
performance functions ρij as in (14) and the transformation
function s.t. Tij(0) = 0, ∀(i, j) ∈ E, and assume that the
initial conditions xij (0) are within the performance bounds
(4) or (5). If
ρ(t) = (ρ0 − ρ∞)e−lt + ρ∞.
(34)
max
(i,j)∈E
(lij) = l ≤ 1.
(38)
When nf = 2, ¯xf = ¯x1 and A = −2, we have that
1(t) = M T eΛtM ¯x1(0) = e−2t ¯x1(0) < ρ0e−2t.
¯x0
(35)
Then, the controlled system achieves consensus within the
prescribed performance bounds ρij (t) when applying the
control (18).
i Di, DT
f Df = Le − DT
f Df have special structures. DT
Proof. For a star graph defined as Definition 2 with the
centering node n as the only leader, the edge Laplacian Le
and matrices DT
i Di
has all elements equal to 1, while DT
i Di is
an identity matrix. Le has the elements given by cij = 2
when i = j, and cij = 1 otherwise. Under this special
structure of star graphs and according to Theorem 1, it can
be verified that (20) is always feasible with ¯γ = 1, and from
(38), we know the condition ¯γ ≥ l = max
(lij) holds.
(i,j)∈E
Finally, by applying Theorem 1, for a star graph, when the
performance functions (14) are chosen such that (38) holds,
then we can conclude that the controlled system achieves
consensus within the prescribed performance bounds when
applying (18).
We conclude this section with the following observations.
A sufficient condition for a general tree graph was derived
in Theorem 1, under which the leader-follower multi-agent
system (3) achieves consensus within the prescribed perfor-
mance bounds (14). It can be seen that (20) may be infeasible
when the decay rate of the performance functions is too large.
This means that we need to constrain the decay rate of the
performance functions in order to achieve consensus under
prescribed performance guarantees within the leader-follower
framework. This is reasonable since the followers only obey
the first-order consensus protocol without any additional ex-
ternal input. And the decay rate constraint differs for different
graph topologies, leader amount and leader positions. For the
specific class of star graphs, we have proven that when the
largest decay rate of the performance functions is less than or
equal to 1, the closed loop system achieves consensus within
the prescribed performance bounds by applying Theorem 1.
We have also shown that the condition in Theorem 1 is
a sufficient but not necessary condition by discussing the
specific class of chain graphs. That is, for a chain graph
with 2 or 3 followers, we can still achieve the result of
consensus within performance bounds although the condition
in Theorem 1 may be infeasible.
IV. SIMULATIONS
In this section three simulation examples are presented
in order to verify the results of the previous sections. The
communication graphs are shown as Fig. 1, where the
leaders and followers are represented by grey and white
nodes, respectively. Regarding the prescribed performance
functions, for all (i, j) ∈ E, we choose Mij = 1 and
Tij(xij ) = ln(cid:18)−
xij + 1
xij − 1(cid:19) .
3
1
4
5
6
10
2
(a)(a)
1
2
9
8
11
7
6
5
3
4
1
2
3
4
5
(b)(b)
(c)(c)
Fig. 1. Communication graphs with tree topologies.
(cid:2)4.6 4.9 4.5 4.7 4.5(cid:3)T
In Fig. 1.(a), We first consider a tree graph with leaders set
as VL = {4, 5, 6}, and the relative positions are initialised as
. According to Theorem 1, the
matrix inequality is feasible with ¯γ = 1, hence it suffices
that l ≤ ¯γ = 1. The simulation result when applying the
PPC law (18) with a gain matrix G whose diagonal entries
are all equal to 1 is shown on the right side of Fig. 2. As
a comparison, the simulation result without PPC is shown
on the left side of Fig. 2. We can see from Fig. 2 that the
trajectories intersect the performance bound without extra
control, which can be improved by applying the PPC law (18)
such that the controlled system achieves consensus within the
performance bound. Here the decay rate of the prescribed
performance function is 1.
6
4
2
j
i
x
0
j
i
x
-2
-4
-6
0
1
2
3
4
5
t
5
4
3
2
1
0
-1
-2
-3
-4
-5
0
1
2
3
4
5
t
Fig. 2. The left figure shows the trajectories of relative positions without
PPC, while the controlled system with PPC law is shown in the right figure
under the communication graph as in Fig. 1.(a).
The prescribed performance bounds are chosen as in (39)
with different decay rate l for different simulation examples.
For each graph, choosing the same ρij for all edges is
done without loss of generality. In addition, the prescribed
performance bounds are depicted in black color for the
following simulation graphs.
ρij(t) = 4.9e−lt + 0.1.
(39)
are initialised as (cid:2)4.8 3 −2 1(cid:3)T
In Fig. 1.(b), we consider a chain graph with followers set
as VF = {1, 2} and VF = {1, 2, 3}, the relative positions
. When the system has
2 followers, we know that the performance function can have
a higher decay rate of 2, while the maximum decay rate is
1 when the system has one more follower (agent 3). When
VF = {1, 2}, the simulation results are shown in Fig. 3,
where the left figure shows the simulation result without ad-
ditional control. Here the decay rate of the prescribed perfor-
mance function is 2. We can see that the trajectories intersect
the performance bound, which is improved as shown in the
middle figure by applying the PPC law (18) with gain matrix
G = diag(1, 10, 1, 1), where diag(a1, a2, . . . , an) represents
the diagonal matrix with diagonal entries a1, a2, . . . , an and
g32 = 10 is tuned for leader {3} that connects the followers.
However, it can be seen that the trajectories still intersect
the performance bound. We then increase g32 to 200, and
the simulation result is shown in the right figure. We can
see that the controlled system achieves consensus within the
performance bound. When VF = {1, 2, 3}, the simulation
results are shown as in Fig. 4, in which the decay rate of
the prescribed performance function is 1. Similarly, it can
be seen in the left figure that the trajectories intersect the
performance bound when there is no extra input, which is
improved as shown in the middle and right figure by applying
the PPC law (18) with gain matrix G = diag(1, 1, 10, 1) and
G = diag(1, 1, 100, 1), respectively. Here, the large gain is
tuned for agent 4 because it is the leader that connects the
followers. We can also conclude that the controlled system
achieves consensus within the performance bound.
j
i
x
5
4
3
2
1
0
-1
-2
-3
-4
-5
0
5
t
j
i
x
5
4
3
2
1
0
-1
-2
-3
-4
-5
0
5
t
j
i
x
5
4
3
2
1
0
-1
-2
-3
-4
-5
j
i
x
5
4
3
2
1
0
-1
-2
-3
-4
-5
0
5
t
j
i
x
5
4
3
2
1
0
-1
-2
-3
-4
-5
0
5
t
j
i
x
5
4
3
2
1
0
-1
-2
-3
-4
-5
0
5
t
Fig. 4. The left figure shows the trajectories of relative positions without
PPC, while the controlled system with PPC law but different gain matrix is
shown in the middle and right figure, respectively under the communication
graph as in Fig. 1.(b) with VF = {1
3}.
2
,
,
j
i
x
5
4
3
2
1
0
-1
-2
-3
-4
-5
0
1
2
3
4
5
t
j
i
x
5
4
3
2
1
0
-1
-2
-3
-4
-5
0
1
2
3
4
5
t
Fig. 5. The left figure shows the trajectories of relative positions without
PPC, while the controlled system with PPC law is shown in the right figure
under the communication graph as in Fig. 1.(c).
0
5
t
V. CONCLUSIONS
Fig. 3. The left figure shows the trajectories of relative positions without
PPC, while the controlled system with PPC law but different gain matrix is
shown in the middle and right figure, respectively under the communication
graph as in Fig. 1.(b) with VF = {1
2}.
,
4.7 −4 1 4.8(cid:3)T
(cid:2)4 3 −2 −3 4.9 1
In Fig. 1.(c), We consider a star graph with only one leader
as VL = {11}, and the relative positions are initialised as
. The sim-
ulation result when applying PPC law (18) with a gain matrix
G whose diagonal entries are all equal to 1 is shown on the
right side of Fig. 5. As a comparison, the simulation result
without PPC is shown on the left side of Fig. 5. It is shown
that the trajectories intersect the performance bound when
there is no extra input, which can be improved by applying
the PPC law (18) such that the controlled system achieves
consensus within the performance bound. Here the decay rate
of the prescribed performance function is 1.
In this paper, we have studied consensus problems of
leader-follower multi-agent systems with prescribed perfor-
mance bounds. Under the assumption of tree graphs, a
distributed prescribed performance control
law has been
proposed for a group of selected leaders in order to drive
the followers such that the entire system can achieve con-
sensus under the prescribed performance guarantees. We
have proved that when the decay rate of the performance
functions is within a sufficient bound, consensus together
with performance guarantees can be obtained. In addition,
two specific classes of chain and star graphs that can have
additional followers have been investigated.
Future research directions include considering more gen-
eral graphs with circles, applying other transient approaches
to this leader-follower framework and also investigating
leader selection problems.
REFERENCES
[1] T. Balch and R. C. Arkin. Behavior-based formation control for
IEEE transactions on robotics and automation,
multirobot teams.
14(6):926 -- 939, 1998.
[2] C. P. Bechlioulis and K. J. Kyriakopoulos. Robust model-free forma-
tion control with prescribed performance and connectivity maintenance
for nonlinear multi-agent systems.
In 53rd IEEE Conference on
Decision and Control, pages 4509 -- 4514. IEEE, 2014.
[3] C. P. Bechlioulis and G. A. Rovithakis. Robust adaptive control
of feedback linearizable mimo nonlinear systems with prescribed
performance. IEEE Transactions on Automatic Control, 53(9):2090 --
2099, 2008.
[4] T. Berger, H. H. Le, and T. Reis. Funnel control for nonlinear systems
with known strict relative degree. Automatica, 87:345 -- 357, 2018.
[5] D. V. Dimarogonas and K. H. Johansson. Stability analysis for multi-
agent systems using the incidence matrix: quantized communication
and formation control. Automatica, 46(4):695 -- 700, 2010.
[6] M. Egerstedt, S. Martini, M. Cao, K. Camlibel, and A. Bicchi. Inter-
acting with networks: How does structure relate to controllability in
single-leader, consensus networks? IEEE Control Systems, 32(4):66 --
73, 2012.
[7] J. A. Fax and R. M. Murray. Information flow and cooperative control
of vehicle formations. 2003.
[8] A. Franchi and P. R. Giordano. Online leader selection for improved
collective tracking and formation maintenance. IEEE transactions on
control of network systems, 5(1):3 -- 13, 2018.
[9] D. Goldin and J. Raisch. Controllability of second order leader-
follower systems. In 2nd IFAC Workshop on Distributed Estimation
and Control in Networked Systems 2010-NecSys 10, pages 233 -- 238,
2010.
[10] A. Ilchmann, E. P. Ryan, and S. Trenn. Tracking control: Performance
funnels and prescribed transient behaviour. Systems & Control Letters,
54(7):655 -- 670, 2005.
[11] Y. Karayiannidis, D. V. Dimarogonas, and D. Kragic. Multi-agent
average consensus control with prescribed performance guarantees. In
Decision and Control (CDC), 2012 IEEE 51st Annual Conference on,
pages 2219 -- 2225. IEEE, 2012.
[12] I. Katsoukis and G. A. Rovithakis. Output feedback leader-follower
with prescribed performance guarantees for a class of unknown non-
linear multi-agent systems. In 2016 24th Mediterranean Conference
on Control and Automation (MED), pages 1077 -- 1082. IEEE, 2016.
[13] L. Macellari, Y. Karayiannidis, and D. V. Dimarogonas. Multi-agent
second order average consensus with prescribed transient behavior.
IEEE Transactions on Automatic Control, 62(10):5282 -- 5288, 2017.
[14] M. Mesbahi and M. Egerstedt. Graph theoretic methods in multiagent
networks, volume 33. Princeton University Press, 2010.
[15] R. Olfati-Saber and R. M. Murray. Consensus problems in networks
of agents with switching topology and time-delays. IEEE Transactions
on automatic control, 49(9):1520 -- 1533, 2004.
[16] S. Patterson and B. Bamieh. Leader selection for optimal network
coherence. In 49th IEEE Conference on Decision and Control (CDC),
pages 2692 -- 2697. IEEE, 2010.
[17] A. Rahmani, M. Ji, M. Mesbahi, and M. Egerstedt. Controllability of
multi-agent systems from a graph-theoretic perspective. SIAM Journal
on Control and Optimization, 48(1):162 -- 186, 2009.
[18] W. Ren and E. Atkins. Distributed multi-vehicle coordinated control
via local information exchange. International Journal of Robust and
Nonlinear Control: IFAC-Affiliated Journal, 17(10-11):1002 -- 1033,
2007.
[19] H. G. Tanner. On the controllability of nearest neighbor interconnec-
tions. In Decision and Control, 2004. CDC. 43rd IEEE Conference
on, volume 3, pages 2467 -- 2472. IEEE, 2004.
[20] H. G. Tanner, A. Jadbabaie, and G. J. Pappas. Flocking in fixed
IEEE Transactions on Automatic control,
and switching networks.
52(5):863 -- 868, 2007.
[21] A. Y. Yazicioglu and M. Egerstedt. Leader selection and network
assembly for controllability of leader-follower networks. In American
Control Conference (ACC), 2013, pages 3802 -- 3807. IEEE, 2013.
[22] D. Zelazo and M. Mesbahi. Edge agreement: Graph-theoretic perfor-
mance bounds and passivity analysis. IEEE Transactions on Automatic
Control, 56(3):544 -- 555, 2011.
|
1502.03212 | 2 | 1502 | 2015-02-12T15:09:34 | Mathematical Modeling of Insurance Mechanisms for E-commerce Systems | [
"cs.MA"
] | Electronic commerce (a.k.a. E-commerce) systems such as eBay and Taobao of Alibaba are becoming increasingly popular. Having an effective reputation system is critical to this type of internet service because it can assist buyers to evaluate the trustworthiness of sellers, and it can also improve the revenue for reputable sellers and E-commerce operators. We formulate a stochastic model to analyze an eBay-like reputation system and propose four measures to quantify its effectiveness: (1) new seller ramp up time, (2) new seller drop out probability, (3) long term profit gains for sellers, and (4) average per seller transaction gains for the E-commerce operator. Through our analysis, we identify key factors which influence these four measures. We propose a new insurance mechanism which consists of an insurance protocol and a transaction mechanism to improve the above four measures. We show that our insurance mechanism can reduce the ramp up time by around 87.2%, and guarantee new sellers ramp up before the deadline $T_w$ with a high probability (close to 1.0). It also increases the long term profit gains and average per seller transaction gains by at least 95.3%. | cs.MA | cs | Mathematical Modeling of Insurance Mechanisms for E-commerce
Systems
Hong Xie
John C.S. Lui
Department of Computer Science & Engineering
The Chinese University of Hong Kong
{hxie,cslui}cse.cuhk.edu.hk
5
1
0
2
b
e
F
2
1
]
A
M
.
s
c
[
2
v
2
1
2
3
0
.
2
0
5
1
:
v
i
X
r
a
Abstract
Electronic commerce (a.k.a. E-commerce) systems such as eBay and Taobao of Alibaba are becoming
increasingly popular. Having an effective reputation system is critical to this type of internet service
because it can assist buyers to evaluate the trustworthiness of sellers, and it can also improve the revenue
for reputable sellers and E-commerce operators. We formulate a stochastic model to analyze an eBay-like
reputation system and propose four measures to quantify its effectiveness: (1) new seller ramp up time, (2)
new seller drop out probability, (3) long term profit gains for sellers, and (4) average per seller transaction
gains for the E-commerce operator. Through our analysis, we identify key factors which influence these
four measures. We propose a new insurance mechanism which consists of an insurance protocol and a
transaction mechanism to improve the above four measures. We show that our insurance mechanism can
reduce the ramp up time by around 87.2%, and guarantee new sellers ramp up before the deadline Tw
with a high probability (close to 1.0). It also increases the long term profit gains and average per seller
transaction gains by at least 95.3%.
1 Introduction
E-commerce systems are becoming increasingly popular and typical examples include eBay[9], Amazon[1],
and Taobao [22] of Alibaba (the largest E-commerce system in China), etc. Through an E-commerce system,
geographically distributed sellers and buyers can transact online. Sellers advertise products in their online
stores (which reside in the E-commerce's website), while buyers can purchase products from any online
stores. The E-commerce system charges a transaction fee from sellers for each completed transaction. In an
E-commerce system, it is possible to purchase products from a seller whom the buyer has never transacted
with, and this seller may not even be trustworthy[19]. This situation results in a high risk of buying low
quality products. To overcome such problems, E-commerce systems deploy reputation systems[19].
Usually, E-commerce operators maintain and operate a reputation mechanism to reflect the trustworthi-
ness of sellers[9, 22]. A high reputation seller can attract more transactions leading to higher revenue[19].
The eBay-like reputation system is the most widely deployed reputation policy, which is used in eBay and
Tabao, etc. This type of reputation system is a credit based system. More precisely, a seller needs to collect
enough credits from buyers in order to improve his reputation. These credits are obtained in form of feedback
ratings, which are expressed by buyers after each transaction. Feedback ratings in eBay and Taobao are of
three levels: positive (+1), neutral (0), and negative (−1). The cumulative sum of all the past feedback
ratings (i.e., reputation score) reflects the trustworthiness of a seller. The reputation score and feedback
ratings are public information and accessible by all buyers and sellers in such an E-commerce system.
Consider this eBay-like reputation system, a new seller may spend a long time to collect enough credits
(i.e., ramp up). This is because new sellers are initialized with a reputation score of zero, and buyers are
less willing to buy products from a seller with low reputation scores. The ramp up time is critical to the
effectiveness of a reputation system. A long ramp up time discourages new sellers to join an E-commerce
system. Furthermore, a new user starts an online store with certain budgets, and maintaining such online
stores involves cost. If a new seller uses up his entire budget and has not yet ramped up his reputation,
he may discontinue his online business (i.e, or drops out) due to low revenue. Therefore, a long ramp up
time increases the risk that a new seller drops out and discourages potential new sellers to join. Finally, a
long ramp up time also results in a low profit gain for a seller. Because before ramping up, a seller can only
attract few transactions due to his low reputation score. To an E-commerce operator, this also results in an
indirect loss on transaction gains. This paper aims to identify key factors that influence the ramp up time
and to design a mechanism to improve this measure.
1
Reducing ramp up time is challenging and to the best of our knowledge, this is the first work which
explores how to reduce the ramp up time for an eBay-like reputation system. This paper aims to explore
the following fundamental questions: (1) How to identify key factors which influence the ramp up time? (2)
How to take advantage of these factors to reduce the ramp up time? Our contributions are
• We propose four performance measures to quantify the effectiveness of eBay reputation systems: (1)
new seller ramp up time, (2) new seller drop out probability, (3) long term profit gains for sellers, and
(4) average per seller transaction gains for an E-commerce operator.
• We develop a stochastic model to identify key factors which many influence these four measures.
Through we gain important insights on how to design a new mechanism these performance measures.
• We propose and design an insurance mechanism which can reduce the ramp up time and the new seller
drop out probability. We show that our insurance mechanism can reduce the ramp up time by around
87.2%, and guarantee new sellers ramp up before the deadline Tw with a high probability (close to 1.0).
It also increases the long term profit gains and average per seller transaction gains by at least 95.3%.
This paper organizes as follows.
In §3
we formulate four measures to explore the ramp up time problem, i.e., ramp up time, new seller drop out
probability, long term profit gains and average per seller transaction gains.
In §4, we derive analytical
expressions for these four measures. In §5 present the design of our insurance mechanism. Related work is
given in §6 and we conclude in §7.
In §2, we present the system model for E-commerce systems.
2 E-commerce System Model
An E-commerce system consists of users, products and a reputation system. A user can be a seller or a
buyer or both. Sellers advertise products in their online stores and set a price for each product. Buyers,
on the other hand, purchase products through online stores and provide feedbacks to indicate whether a
buyer advertises products honestly or not. A reputation system is maintained by E-commerce operators to
reflect the trustworthiness of sellers. A high reputation seller can attract more transactions leading to a high
revenue. The reputation system aggregates all the feedbacks, and computes a reputation score for each seller.
The reputation score is public information which is accessible by all buyers and sellers.
Products are categorized into different types. For example, eBay categorizes products into "Fashion",
"Electronics", "Collectibles & Art", etc [9]. We consider L ≥ 1 types of product. Consider a type ℓ ∈
{1, . . . , L} product. A seller sets a price pℓ ∈ [0, 1] and the E-commerce operator charges a transaction fee of
T , αpℓ, where α ∈ (0, 1), after the product is sold1. It has a manufacturing cost of cℓ ∈ [0, 1]. A seller earns
a profit of uℓ by selling one product, we have
For the ease of presentation, our analysis focuses on one product type. It can be easily generalized to multiple
product types, and we omit the subscript unless we state otherwise.
uℓ = (1 − α)pℓ − cℓ.
(1)
2.1 Transaction Model
Sellers advertise the product quality in their online stores. Let Qa ∈ [0, 1] be the advertised quality. The larger
the value of Qa implies the higher the advertised quality. Buyers refer to the advertised quality Qa in their
product adoption. Each online store also has an intrinsic quality. Let Qi ∈ [0, 1] be the intrinsic quality (i.e.,
the ground truth of the product's quality). The larger the value of Qi implies the higher the intrinsic quality.
Since sellers aim to promote their products, so we have Qa ≥ Qi. We emphasize that the intrinsic quality Qi
is private information, e.g., it is only known to the seller. On the other hand, the advertised quality Qa is
public information which is accessible by all buyers and sellers.
Buyers estimate the product quality by referring to the advertised quality Qa (we will present the esti-
mating model later). Let Qe ∈ [0, 1] be the estimated quality. The larger the value of Qe implies the higher
the estimated quality. To purchase a product, a buyer must submit a payment p to the E-commerce system,
which will be given to the corresponding seller when he receives the product. There is usually a shipment
delay in any E-commerce systems. We denote the delay as d. Upon receiving a product, a buyer can evaluate
its quality and at that moment, he has the perceived quality, which we denote as Qp ∈ [0, 1]. The larger
the value of Qp implies the higher the perceived quality. We assume that buyers can perceive the intrinsic
1We can also consider a fixed transaction fee model and our analysis is still applicable. But for brevity, let us consider a
transaction fee which is proportional to the selling price.
2
quality, i.e., Qp = Qi. Buyers are satisfied (disappointed) if they find out that the product is at least as good
as (less than) it is advertised, or Qp ≥ Qa (Qp < Qa).
To attract buyers, an E-commerce system needs to incentivize sellers to advertise honestly, i.e., Qa = Qi.
Many E-commerce systems achieve this by deploying a reputation system. We next introduce a popular
reputation system used by many E-commerce systems such as eBay [9] or Taobao [22]. Table 1 summarizes
key notations in this paper.
p, c
T, u
Qa, Qi, Qe, Qp
d, CS
γ
F
r
rh, θ
β
P(Qe, p)
Pba, Pbr
λ1(λ2)
Tw
Tr, Pd
Gs, Ge
δ
λT (τ )
CI , DI, Td, Tc
bDI
λI
r , P I
T I
r
GI
s, GI
e
price and manufacturing cost of a product
transaction fee, unit profit of selling a product
advertised, intrinsic, estimated, perceived product quality
shipment delay, shipment cost
critical factor in expressing feedback ratings
reputation profile for a seller
reputation score for a seller
reputation threshold, consistency threshold
discounting factor in estimating product quality
probability that a buyer buys a product with
an estimated quality quality Qe and a price p
probability that a buyer buys a product from a seller
labelled as average (reputable)
buyer's arrival rate before (after) a seller ramps up
the maximum time that a seller
is willing to wait to get ramped up
ramp up time, new seller drop out probability
long term expected profit gains for a seller, average per
seller transaction gains for the E-commerce operator
discount factor in the long term expected profit gains Gs
transaction's arrival rate at time slot τ
insurance price, deposit, duration time and clearing time
insurance deposit threshold to revoke insurance certificate
transaction's arrival rate to an insured seller
ramp up time with insurance,
new seller drop out probability with insurance
long term expected profit gains with insurance,
average per seller transaction gains with insurance
Table 1: Notation list
2.2 Baseline Reputation System
The eBay-like system maintains a reputation system to reflect the trustworthiness of sellers. It consists of a
feedback rating system and a rating aggregation policy.
Buyers express feedback ratings to indicate whether a seller advertises honestly or not. The eBay-like
system adopts a feedback rating system consisting of three rating points2, i.e., {−1, 0, 1}. A positive rating
(rating 1) indicates that a product is at least as good as it is advertised, i.e., Qp ≥ Qa. A neutral rating
(rating 0) indicates that a buyer is indifferent about the product that he purchased. This happens when the
perceived quality is slightly lower than it is advertised, i.e., Qp ∈ [Qa − γ, Qa), where γ ∈ [0, 1] denotes the
critical factor. The smaller the value of γ implies that buyers are more critical in expressing ratings, e.g.,
γ = 0 means that buyers have zero tolerance on seller overstating the product quality. A negative rating
(rating −1) represents that the perceived quality is far smaller than the advertised quality, i.e., Qp < Qa − γ.
We have
All the historical ratings are known to all buyers and sellers.
2We can easily generalize the model to consider more rating points.
3
feedback rating =
1,
0,
− 1,
if Qp ≥ Qa,
if Qa − γ ≤ Qp < Qa,
if Qp < Qa − γ.
For the rating aggregation policy, each seller is associated with a reputation score, which is the summation
of all his feedback ratings. We denote it by r ∈ Z. A new seller who enters the E-commerce system is initialized
with zero reputation score, or r = 0. A positive feedback rating increases r by one, a negative feedback rating
decreases r by one, and a neutral feedback rating 0 does not change r. Figure 1 depicts the transition diagram
of a seller's reputation score. Note that r is a public information accessible by all buyers and sellers.
To assist buyers to evaluate the trustworthiness of a seller, E-commerce systems not only announce the
seller's reputation score r, but also his reputation profile. Let F , (r, n+, n0, n−) be the reputation profile,
where n+, n0, n− represent the cumulative number of feedback ratings equal to 1, 0, −1 respectively. Note that
this form of reputation is commonly deployed, say in eBay [9] and Taobao [22].
0
.!!.!!.
r - 1
0
r
1
-1
1
-1
0
r + 1
.!!.!!.
Figure 1: Transition diagram of a seller's reputation score r
Shipment delay in real-world E-commerce systems usually results in certain delay in the reputation up-
date. To characterize the dynamics of a reputation updating process, we consider a discrete time system
and divide the time into slots, i.e., [0, d), [d, 2d), . . ., where d is the shipment delay. We refer to a time
slot τ ∈ N as [τ d, (τ + 1)d). Let N (τ ) be the number of products sold in the time slot τ . Suppose
N +(τ ), N 0(τ ), N −(τ ) of these transactions result in positive, neutral and negative feedbacks respectively. Let
F (τ ) , (r(τ ), n+(τ ), n0(τ ), n−(τ )) be the reputation profile at time slot τ . Then we have F (0) = (0, 0, 0, 0).
We update the reputation profile F (τ ) as
n+(τ + 1) = n+(τ ) + N +(τ ),
n0(τ + 1) = n0(τ ) + N 0(τ ),
n−(τ + 1) = n−(τ ) + N −(τ ),
r(τ + 1) = r(τ ) + N +(τ ) − N −(τ ).
(2)
For simplicity, we drop the time stamp τ in the reputation profile, when no confusion involved.
We next present a model to characterize the impact of sellers' reputation profiles on buyers' product
adoption behavior. This model serves as an important building block to explore the effectiveness of this
baseline reputation system.
2.3 Model for Product Adoption Behavior
The reputation system forges trust among sellers and buyers. This trust plays a critical role in product
adoption. More precisely, buyers evaluate the trustworthiness of sellers from sellers' reputation profiles.
Buyers seek to minimize the risk in product purchase and they prefer to buy from reputable sellers.
Based on the reputation profile F , our model classifies sellers into two types: "reputable" and "average".
To be labeled as reputable, a seller's reputation profile must satisfy two conditions. The first one is that
a seller needs to collect enough credits, i.e., positive feedbacks from buyers. More precisely, his reputation
score must be at least greater than or equal to some positive reputation threshold rh, i.e., r ≥ rh. A new
seller is initialized with zero reputation score, i.e., r = 0. To accumulate a reputation score of at least rh,
he needs to accomplish enough number of honest transactions. The second condition is that a seller should
be consistently honest. More concretely, the fraction of positive feedbacks should be larger than or equal to
a consistency threshold θ ∈ (0, 1], i.e., n+/(n+ + n− + n0) ≥ θ. The larger the value of θ implies that the
E-Commerce operators are more critical about the honest consistency. We formally define a reputable seller
and an average seller as follows.
Definition 2.1. A seller is labeled as reputable if and only if the follwing two conditions are met
C1: r ≥ rh and,
C2: n+/(n+ + n− + n0) ≥ θ.
Otherwise, a seller is labeled as an average seller.
4
Hence, the reputation threshold rh and consistency threshold θ quantify how difficult is it to earn a reputable
label. The larger the rh and θ, the more difficult it is to earn a reputable label.
A buyer estimates the product quality referring to the advertised quality Qa and the reputation profile
of a seller. More concretely, if a seller's reputation profile indicates that this seller is reputable, then a
buyer believes that this seller advertises honestly. This buyer therefore estimates the product quality as the
advertised quality, i.e., Qe = Qa. On the contrary, if the reputation profile indicates that a seller is average, a
buyer believes that this seller is likely to overstate the product quality. Hence the estimated quality is lower
than the advertised quality, i.e., Qe = βQa, where β ∈ [0, 1] denotes the discounting factor. The smaller the
value of β implies that buyers are less willing to trust an average seller. We have
Qe =( Qa,
βQa, otherwise.
if r ≥ rh and n+/(n+ + n− + n0) ≥ θ,
A buyer makes the purchasing decision based on the estimated quality Qe and the product price p. More
concretely, the probability that a buyer buys a product increases in Qe and decreases in p. Formally, we have
Pr[adopts a product] , P (Qe, p) ,
(3)
where the function P increases in Qe and decreases in p.
In the following section, we evaluate the effectiveness of the baseline reputation system in E-commerce
applications. Our goal is to identify key factors that influence the effectiveness of this reputation system and
if possible, improve it.
3 Problems Formulation
We propose four performance measures to quantify the effectiveness of the baseline reputation system men-
tioned in Section 2. These measures are: (1) ramp up time Tr, (2) new seller drop out probability Pd, (3) long
term expected profit gains for a seller Gs, and (4) average per seller transaction gains for the E-commerce
system operator Ge. We also present our problem formulations and our objective is to identify key factors
which can influence these measures. Lastly, we raise an interesting question of whether there are other mech-
anisms which can reduce the ramp up time and the new seller drop out probability, and improve the long
term expected profit gains and average per seller transaction gains.
3.1 Ramp Up Time
Sellers and E-commerce system operators are interested in the minimum time that a new seller must spend
to collect enough credits, i.e., positive feedbacks from buyers, so that the seller can be classified as reputable.
For one thing, a reputable seller can attract more buyers which may result in more transactions, and higher
transaction volume implies higher transaction gains to the E-commerce operator. We next formally define
the ramp up process and the ramp up condition.
Definition 3.1. A new seller's reputation is r = 0. He needs to collect enough credits, i.e., positive feedbacks
from buyers, so that his reputation r can increase to at least rh. The process of increasing his reputation
to rh is called the ramp up process. Furthermore, when r ≥ rh, then we say that the ramp up condition is
satisfied.
Recall that r(τ ) denotes the reputation score of a seller at time slot τ . We formally define the ramp up
time as follows.
Definition 3.2. Ramp up time is the minimum time that a seller must spend to accumulate a reputation
score of rh. Let Tr denote the ramp up time, we have
Tr , d · arg min
τ
{r(τ ) ≥ rh}.
(4)
The ramp up time quantifies how long it will take to collect enough credits from buyers. It is critical to the
profit gains for a seller. To see this, we next quantify how the ramp up time can affect the transaction's
arrival rate.
A seller can attract more buyers when he satisfies the ramp up condition because his online store will
receive higher click rate by buyers, therefore increasing his profit gains. Let λ1 (λ2) be the buyer's arrival
rate before (after) a seller satisfies the ramp up condition. We assume that the buyer's arrival process, either
before or after a seller satisfies the ramp up condition, follows a Poisson counting process with parameter
5
λ1 (before ramping up) and λ2 (after ramping up) respectively, where λ1 < λ2 to signify that a ramped up
seller can attract more buyers. Recall that in Equation (3) we express the probability that a buyer adopts
a product as P(Qe, p). If a buyer adopts a product, we say a seller obtains a transaction. Based on the
Poisson property, it is easy to see that the transaction's arrival process is also a Poisson counting process.
Let λT (τ ) be the transaction's arrival rate at time slot τ . Let P(Qe(τ ), p) be the probability that a buyer
adopts a product at time slot τ , where Qe(τ ) denotes the estimated quality at time slot τ . We can express
the transaction's arrival rate as
λT (τ ) =( λ1P(Qe(τ ), p),
λ2P(Qe(τ ), p),
if r(τ ) < rh,
if r(τ ) ≥ rh,
.
(5)
Equation (5) serves as an important building block for us to explore the key factors which influence the ramp
up time Tr. Let us formulate our first problem.
Problem 1: Identify key factors which influence the ramp up time Tr, and design a mechanism which can
take advantage of these factors to reduce Tr.
3.2 New Seller Drop Out Probability
In real-world E-commerce systems, a new seller may drop out, or move to another E-commerce system,
if he does not collect enough credits (i.e., ramp up) within certain time because he cannot obtain enough
transactions. For example, a new seller in eBay may drop out if he does not ramp up in one year. This is
because a new seller starts an online store with certain budgets and there are costs associated with maintaining
this online business. Let Tw > 0 denote the maximum time that a new seller is willing to wait to get ramped
up.
In other words, if the ramp up time is longer than Tw, a new seller will quit or drop out from that
E-commerce system. We assume Tw/d ∈ N to accommodate the delay (d) in reputation update.
Definition 3.3. A new seller drops out, if and only if Tr > Tw.
Sellers and the E-commerce operator are interested in this new seller drop out probability. Let
Pd , Pr[Tr > Tw]
(6)
denote the probability that a new seller drops out. The smaller the value of Pd implies that sellers are more
likely to continue his online business in the E-commerce system. This is an important measure since a small
Pd can attract more new sellers to join the E-commerce system, which will result in higher transaction gains
for the E-commerce operator. On the other hand, a large Pd discourages new sellers to participate and can
reduce the transaction gains for the E-commerce operator. We therefore consider the second problem.
Problem 2: Identify key factors which influence the new seller drop out probability Pd, and design a mech-
anism which can take advantage of these factors to reduce Pd.
3.3 Long Term Profit Gains and Transaction Gains
The profit gain (transaction gain) is critical to sellers (E-commerce system operators). We focus on the
scenario that sellers are long lived and they aim to maximize their long term profit gains. Recall that u, the
unit profit of selling one product, is expressed in Equation (1). Also recall that N (τ ) denotes the number
of products sold in the time slot τ . We emphasize that N (τ ) is a random variable and follows a Poisson
distribution with parameter λT (τ )d, where λT (τ ) is derived in Equation (5). A seller earns a profit of uN (τ )
in the time slot τ . We consider a discounted long term profit gain with a discounting factor of δ ∈ (0, 1]. Let
Gs denote the long term expected profit gains for a seller. We express it as
Gs , E" ∞Xτ =0
δτ uN (τ )# .
(7)
Note that when a seller earns a profit u, he also contributes a transaction fee T = αp to the E-commerce
operator. Let Ge denote the average per seller transaction gains that a seller pays to the E-commerce operator.
We can express it as
αp
u
Gs.
(8)
Note that Gs is important to a seller while Ge is important to the E-commerce operator. We consider the
following problem.
Ge , E" ∞Xτ =0
δτ T N (τ )# = E" ∞Xτ =0
δτ αpN (τ )# =
6
Problem 3: Identify key factors which influence the profit gains Gs and average per seller transaction gains
Ge, and design a mechanism to use these factors to improve Ge and Gs.
We next derive E[Tr], Pd, Gs, and Ge. Through this analysis, we identify key factors which influence
the above mentioned performance measures. These insights will serve as important building blocks for us to
design a mechanism.
4 Analyzing the Baseline Reputation System
We derive analytical expressions for the expected ramp up time (E[Tr]), the new seller drop out probability
(Pd), the long term expected profit gains (Ge) and the average per seller transaction gains (Ge). Through
this we identify that the reputation threshold (rh), as well as the probability that a buyer buys a product
from an "average labeled" seller (Pab) are two critical factors which influence Tr, Pd, Gs and Ge. Our results
indicates that the baseline reputation mechanism described in Section 2, suffers from long ramp up time,
high new seller drop out probability, and small long term profit gains or transaction gains. These insights
show that one need to have a new mechanism to reduce Tr, Pd, and to improve Gs and Ge. We will present
this new mechanism in Section 5.
4.1 Deriving the Expected Ramp Up Time E[Tr]
Let us derive the analytical expression for the expected ramp up time E[Tr]. This measure quantifies on
average, how long it will take to ramp up a new seller under the baseline reputation mechanism mentioned
in Section 2. We consider the scenario that buyers advertise the product quality honestly, i.e., Qa = Qi. As
to how an eBay-like reputation mechanism can guarantee rational sellers to advertise honestly, one can refer
to [8]. We like to point out that new sellers can achieve the lowest ramp up time by advertising honestly
(Qa = Qi). Hence, the assumption that Qa = Qi can be viewed as deriving the best case of Tr for the baseline
reputation system. Let us first define the following notations.
Definition 4.1. Let Pba , P(βQi, p) and Pbr , P(Qi, p) denote the probability that a buyer buys a product
from an "average labeled" seller and a "reputable" seller respectively.
In the following theorem, we state the expected ramp up time.
Theorem 4.1. The expected ramp up time is
E[Tr] = d
∞Xτ =1 1 −
∞Xk=rh
e−λ1Pba(τ −1)d (λ1Pba(τ − 1)d)k
k!
! .
The E[Tr] increases in the reputation threshold rh, and decreases in the transaction's arrival rate λ1Pba.
Proof: Please refer to the appendix for derivation.
Remark: Theorem 4.1 states that a new seller is more difficult to get ramped up if the E-Commerce operator
sets a high reputation threshold rh, or the transaction's arrival rate to an "average labeled" seller (λ1Pba) is
low.
Table 2 presents numerical examples on the expected ramp up time E[Tr], where we fix Pba = 0.02, i.e.,
buyers buy products from an "average labeled" seller with probability 0.02, and fix d = 3, i.e., it takes three
days to ship a product to a buyer (or for the E-commerce operator to update sellers' reputation). We vary
the buyer's arrival rate λ1 from 5 to 25, i.e., on average each day an "average labeled" seller attracts 5 to 25
buyers to visit his online store. We vary the reputation threshold (rh) from 100 to 200. When rh = 200, as
λ1 increases from 5 to 25, the expected ramp up time (E[Tr]) drops from 2001.7 to 401.6 days, a deduction
ratio of 80%. When the buyer's arrival rate is low, say λ1 = 5, as the reputation threshold rh drops from 200
to 100, the expected ramp up time E[Tr] drops from 2001.7 to 1001.4 days, a reduction ratio of 50%. These
results indicates that the expected ramp up time (E[Tr]) is large in general. Namely, it is difficult for new
sellers to quickly get ramped up under the baseline reputation system. We next explore the new seller drop
out probability.
4.2 Deriving the New Seller Drop Out Probability Pd
We now derive the analytical expression for Pd. This probability quantifies how difficult it is for a new seller
to survive in the E-commerce system. Note that Pd is also crucial for new sellers to decide whether or not to
open an online store in an E-commerce system. Namely, a low drop out probability Pd is attractive to new
sellers, while a high Pd discourages new sellers to join.
7
λ1
E[Tr] (rh = 200)
E[Tr] (rh = 150)
E[Tr] (rh = 100)
5
2001.7
1501.4
1001.4
10
1001.4
751.5
501.5
15
668.2
501.5
334.8
20
501.5
376.5
251.5
25
401.6
301.5
201.5
Table 2: Expected ramp up time E[Tr] in days (Pba = 0.02, d = 3).
Theorem 4.2. The new seller drop out probability is
Pd =
rh−1Xk=0
e−λ1PbaTw (λ1PbaTw)k
k!
.
The Pd decreases in λ1Pba, Tw and increases in rh.
Proof: Please refer to the appendix for derivation.
Remark: Theorem 4.2 states that a new seller can reduces the drop out probability by extending his ramp
up deadline line (Tw), and a new seller is more likely to drop out if the reputation threshold (rh) increases
or the transaction's arrival rate to an "average labeled" seller (λ1Pba) decreases.
Table 3 presents numerical examples on the new seller drop out probability Pd, where we set λ1 = 20, i.e.,
on average, each day an "average labeled" seller attracts 20 buyers to visit his store, d = 3, and Tw = 180, i.e.,
sellers drops out if they do not ramp up in 180 days. We vary Pba, the probability that a buyer buys products
from an "average labeled" seller, from 0.01 to 0.05, and vary the reputation threshold rh from 100 to 200.
Consider rh = 200. As Pba increases from 0.01 to 0.05, the new seller drop out probability Pd decreases from
1 to 0.92514. This implies a very high drop out probability. Consider Pba = 0.03. As the reputation threshold
rh drops from 200 to 100, we see that Pd drops from 1 to 0.20819, a reduction ratio of around 80%. It is
interesting to observe that when the Pba is small, the new seller drop out probability is quite high. In fact
when Pba = 0.01, Pd is very close to 1. In other words, if buyers are less willing to buy from "average labeled"
sellers, new sellers will be more likely to drop out. We next explore key factors which influence long term
expected profit gains and average per seller transaction gains.
Pba
0.01
0.02
0.03
0.04
0.05
Pd (rh = 200)
Pd (rh = 150)
Pd (rh = 100)
1.00000
1.00000
1.00000
1.00000
1.00000
0.99897
1.00000
0.99992
0.20819
0.99999
0.68056
0.00005
0.92514
0.00991
0.00000
Table 3: New seller drop out probability Pd (λ1 = 20, Tw = 180, d = 3).
4.3 Deriving the Long Term Profit Gains Gs and Ge
Let us now derive analytical expressions for the long term expected profit gains Gs and the average per seller
transaction gains Ge respectively. They are important measures because a large Gs is attractive to new
sellers and a small Gs discourages new sellers to join the E-commerce system, while the average per seller
transaction gains Ge is crucial to the E-commerce system operator.
Theorem 4.3. The long term expected profit gains for a new seller can be expressed as
(λ1PbaTw)k
Gs =
rh−1Xk=0
δ⌈tj /d⌉dt1 . . . dtk +Z Tw
0
(λ1Pbatrh)rh−1
(rh − 1)!
Z
k!
T k
w
u
kXj=1
λ1Pbae−λ1Pbatrh
0<t1<...<tk<Tw
(k)!
e−λ1PbaTwZ . . .
Z
dtrhZ . . .
(trh )rh−1 u(cid:18) rhXj=1
(rh − 1)!
0<t1<...<trh−1<trh
δ⌈tj /d⌉ +λ2Pbrd
δ⌈trh /d⌉+1
1 − δ
+ δ⌈trh /d⌉λ1Pba(d⌈trh /d⌉ − trh )(cid:19)dt1 . . . dtrh−1.
Furthermore, Ge = αp
u Gs.
8
Proof: Please refer to the appendix for derivation.
Remark: Theorem 4.3 quantifies the impact of various factors on the long term expected profit gains (Gs)
and average per seller transaction gains (Ge), e.g., the reputation threshold rh, the buyer's arrival rate λ1, λ2,
etc. However the analytical expressions are in complicated forms. It is hard to draw some insights on how
critical are these factors in influencing Gs and Ge by directly examining these expressions. We next present
some numerical examples to illustrate their impact.
Table 4 presents numerical examples on the long term expected profit gains (Gs) and the average per
seller transaction gains (Ge), where we set λ1 = 20, T = αp = 0.1, λ2 = 50, u = 1, Tw = 180, d = 3, Pbr = 0.1
(i.e., buyers buy products from a reputable seller with probability 0.1). We vary Pba, the probability that a
buyer purchases products from an "average labeled" seller, and the ramp up threshold rh to examine their
impact on Gs and Ge. Consider rh = 200. As Pba increases from 0.01 to 0.05, Gs improves from 26.833 to
198.059, an improvement ratio of 6.38 time. Similarly, the average per seller transaction gains Ge is also
improved by 6.38 times. This implies that Pba is critical to both sellers' profit gains and the E-commerce
system operator's transaction gains. Consider Pba = 0.05. As rh drops from 200 to 100, Gs improves from
198.059 to 1142.670, an improvement ratio of 4.77. This improvement ratio also holds for the average per
seller transaction grains Ge. It is interesting to observe that when Pba is small, both Gs and Ge are quite
small. In fact when Pba = 0.01, the Gs is around 26.833 and Ge is around 2.6833. Namely, if buyers are less
willing to buy from "average labeled" sellers, sellers (E-commerce operators) will have low long term profit
gains (average per seller transaction gains).
Pba
Gs (rh = 200)
Gs (rh = 150)
Gs (rh = 100)
Ge (rh = 200)
Ge (rh = 150)
Ge (rh = 100)
0.01
26.833
26.980
26.941
2.6833
2.6980
2.6941
0.02
53.342
53.594
54.433
5.3342
5.3594
5.4433
0.03
80.705
80.812
760.511
8.0705
8.0812
76.0511
0.04
0.05
107.312
369.951
1054.507
10.7312
36.9951
105.4507
198.059
1006.017
1142.670
19.8059
100.6017
114.2670
Table 4: Long term expected profit gains Gs and average per seller transaction gains Ge (λ1 = 20, λ2 =
50, u = 1, T = αp = 0.1, δ = 0.99, Tw = 180, Pbr = 0.1, d = 3).
Summary: The reputation threshold rh and Pba are critical to the ramp up time, the new seller drop out
probability, the long term profit gains and the average per seller transaction gains. The baseline (or eBay-like)
reputation system presented in Section 2 suffers from long ramp up time, high seller drop out probability,
small long term profit gains and small average per seller transaction gains. Hence, it is important to ask
whether we can design a new mechanism that an E-commerce system can use to improve all the performance
measures E[Tr], Pd, Gs and Ge. We next explore this interesting question.
5 Insurance Mechanism
In the previous section, we showed that the baseline reputation system is not efficient. Here, we present a
new approach which we call the "insurance mechanism" to reduce both the expected ramp up time (E[Tr])
and new seller drop out probability (Pd), and to improve the long term expected profit gains (Gs) and
the average per seller transaction gains (Ge). We also quantify the impact of our insurance mechanism on
E[Tr], Pd, Gs and Ge. We show that our insurance mechanism can reduce the ramp up time by around
87.2%, and guarantee new sellers ramp up before the deadline Tw with a high probability (close to 1.0). It
also increases the long term profit gains and average per seller transaction gains by at least 95.3%.
5.1
Insurance Mechanism Design
The objective of our insurance mechanism is to help new sellers ramp up quickly. Reducing ramp up time
brings the benefit of reducing new seller drop out probability and improving long term expected profit gains
and average per seller transaction gains (Gs and Ge). Our insurance mechanism consists of an insurance
protocol and a transaction mechanism.
We first describe the insurance protocol. The E-commerce system operator provides an insurance service
to new sellers. Each insurance has a price of CI > 0, a duration time of Td > 0, and a clearing time of Tc > 0.
The insurance clearing time takes effect when an insurance expires. To buy an insurance, a seller must
provide the E-commerce operator an insurance deposit of DI . Hence, the total payment by the new seller
9
to the E-commerce system operator is CI + DI . We refer to this insurance as the (CI , Td, Tc, DI )-insurance.
Only new sellers can subscribe to this insurance. If a seller subscribes an insurance, the E-commerce system
operator issues an insurance certificate to him, and this certificate is known to the public (i.e., all buyers
and sellers). This certificate only takes effect within the insurance duration time Td. The E-commerce
system operator treats a seller with an insurance certificate as trustworthy. To guarantee that such sellers
will advertise their product quality honestly, the E-commerce system operator requires such sellers obey the
following transaction mechanism.
We now describe the transaction mechanism. Only sellers with an insurance certificate must obey this
transaction mechanism. Let us focus on a seller with an insurance certificate. When ordering a product from
this seller, a buyer sends his payment p to the E-commerce system operator. After receiving the product, if this
buyer express a positive feedback, then this transaction completes, i.e., the E-commerce operator forwards
the payment (1 − α)p to the seller and charges a transaction fee of αp. This transaction also completes
if this buyer expresses a neutral feedback. A neutral feedback means that a seller slightly overstated his
product quality, i.e. Qi < Qa < Qi + γ. To avoid such overstating, the E-commerce company revokes a
seller's insurance certificate once the fraction of positive feedbacks falls below the consistency factor (θ), i.e.,
n+/(n+ + n0 + n−) < θ. A negative feedback results in the transaction being revoked. More concretely, the
E-commerce operator gives the payment p back to the buyer and does not charge any transaction fee from
the seller (provided that it is within the duration time Td, or the clearing time Tc). The buyer needs to
ship the product back to the seller but the buyer does not need to pay for the shipment cost CS, for it will
be deducted from a seller's insurance deposit DI . If the insurance deposit is not enough to cover CS, the
E-commerce operator makes a supplemental payment. To avoid this undesirable outcome, the E-commerce
company revokes a seller's insurance certificate, once a seller's deposit reaches a threshold bDI < DI . The
insurance clearing time takes effect when an insurance is invoked. At the end of the clearing time, the
E-commerce company returns the remaining deposit of DI (if it is not deducted to zero) back to the seller.
Remark. Sellers may collude with buyers to inflate their reputation by fake transactions [10]. One way to
avoid such collusion is by increasing the transaction fee such as [2]. The shipment cost may exceed DI due to
a large number of products to be returned. This can be avoided with high probability by setting a large bDI
(Theorem 5.2). We also derive the minimum clearing time (Tc) to guarantee that a seller with an insurance
certificate needs to obey the transaction mechanism (Theorem 5.2).
5.2 Analyzing the Insurance Mechanism
We first show that buyers treat a seller having an insurance certificate as trustworthy. Through this, we
derive the transaction rate that a seller with an insurance certificate can attract. We then derive the improved
E[Tr], Pd, Gs and Ge.
Buyers treat sellers having an insurance certificate as trustworthy. This is an important property of
our insurance mechanism because it influences the probability that a buyer adopts a product from a seller.
Suppose in time slot τ , a seller has an insurance certificate. If this seller advertise honestly Qa = Qi, then
the buyer who buys a product from this seller will be satisfied (express positive feedback rating). In this
case, the payment from the buyer will be forwarded to the seller. Hence this seller earns a profit of u. If
this seller overstates his product quality beyond the lenient factor (γ), i.e., Qa > Qi + γ. Then according to
our insurance mechanism, the payment by the buyer will be returned back to the buyer. The seller needs
to pay a shipment cost of CS to ship back the product and CS will be deducted from his insurance deposit
DI . Hence, if a seller overstates the product quality beyond the lenient factor, he will lose a total shipment
cost of at least min{bDI, CSN (τ )} in time slot τ , where N (τ ) denotes the number of product selling. A seller
with an insurance certificate must obey the same consistency factor (θ) as reputable sellers in being honest,
i.e., n+/(n+ + n0 + n−) ≥ θ, because if not his insurance certificate will be revoked by the E-commerce
operator. Given these properties, buyers trust a seller with an insurance certificate. Recall that the E-
commerce operator also trusts a seller with an insurance certificate. Therefore, an insured seller can attract
transactions with an arrival rate being equivalent to those reputable sellers. Let λI
T denote the transaction's
arrival rate to a seller with an insurance certificate. We have
λI
T = λ2Pbr.
(9)
We now quantify the impact of our insurance mechanism on the four performance measures. Let
e denote the ramp up time, the new seller drop out probability, the long term profit gains
T I
r , P I
and the average per seller transaction gains respectively, when a new seller subscribes our insurance.
d , GI
s, GI
Theorem 5.1. Suppose a new seller subscribes to our proposed insurance mechanism. We express the
10
expected ramp up time and new seller drop out probability as
∞Xτ =1 1 −
e− PTw/d−1
ℓ=0
e− Pτ −2
∞Xk=rh
eλT (ℓ) (PTw /d−1
ℓ=0 eλT (ℓ) (Pτ −2
ℓ=0eλT (ℓ))k
eλT (ℓ))k
ℓ=0
k!
k!
,
! ,
E[Tr] = d
Pd =
rh−1Xk=0
whereeλT (ℓ) = λ2Pbrd for all ℓ = 0, 1, . . . , ⌊Td/d⌋−1, andeλT (⌊Td/d⌋) = λ2Pbr(Td −d⌊Td/d⌋)+λ1Pba(d⌊Td/d⌋+
d − Td), andeλT (ℓ) = λ1Pbad for all ℓ = ⌊Td/d⌋ + 1, . . . , ∞. The long term expected profit gains for an insured
seller is:
(λ2Pbr min{Td, Tw})k
e−λ2Pbr min{Td,Tw}
k!
0<t1<...<tk<min{Td,Tw}
rh−1Xk=0
GI
s =
Z . . .
Z
rh−1−kXi=0
Z
+ I{Td<Tw}
Td<tk+1<...<tk+i<Tw
Z . . .
+(cid:18)I{Td≥Tw}Z Tw
Z . . .
0<t1<...<trh −1<trh
(λ2PbrTd)k
0
k!dt1 . . . dtk
(min{Td, Tw})k(cid:18)I{Td≥Tw }
kXj=1
uδ⌈tj /d⌉
(λ1Pba(Tw − Td))i
i!
e−λ1Pba(Tw−Td)
i!
(Tw − Td)i dtk+1 . . . tk+i
uδ⌈tj /d⌉(cid:19)
k+iXj=1
λ2Pbre−λ2Pbr trh
(λ2Pbrtrh)rh−1
(rh − 1)!
dtrh
(rh − 1)!
(trh )rh−1 dt1 . . . dtrh−1 + I{Td<Tw}
rhXk=0
Z
k!
T k
d
dt1 . . . dtk
0<t1<...<tk<Td
Z
e−λ2Pbr TdZ . . .
k!
Td
Z Tw
Z . . .
λ1Pbae−λ1Pba(trn −Td) (λ1Pba(trn − Td))rh−k−1
dtrh
(rh − k − 1)!
Z
rhXj=1
+
(trh − Td)rh−k−1
(rh − k − 1)!dtk+1 . . . dtrh−1
(cid:19)u(cid:18)λ2dPbr
δ⌈tj /d⌉ + δ⌈trh /d⌉λ2Pbr(d⌈trh /d⌉ − trh)(cid:19)
Td<tk+1<...<trh−1<trh
δ⌈trh /d⌉+1
1 − δ
Furthermore, GI
e = αp
u GI
s.
Proof: Please refer to the appendix for derivation.
Remark: Theorem 5.1 quantifies the impact of our insurance mechanism on the four important performance
measures. Before we talk more about how to select the insurance price CI and deposit DI , let us illustrate
the effectiveness of our insurance mechanism using some numerical examples.
Table 5 presents numerical examples on E[T I
r ], P I
d , GI
s and GI
e. We use the following setting: λ1 = 20, λ2 =
50, u = 1, Tw = 180, d = 3, Pbr = 0.1, Pba = 0.03, CI = 100, DI = 100,bDI = 50, CS = 0.5, T = 0.1, Td = 100, Tc =
3, δ = 0.99. We also presents numerical examples on E[Tr], Pd, Gs and Ge for comparison studies. When
rh = 100, we have E[Tr] = 168.1 and E[T I
r ] = 21.5. In other words, our insurance mechanism reduces the
expected ramp up time from 168.1 days to only 21.5 days, or over 87.2% reduction.
It is interesting to
observe that our incentive mechanism reduces the new seller drop out probability from Pd = 0.20819 to
P I
d = 0. Namely, our insurance mechanism can guarantee that new sellers ramps up before the deadline line
Tw with a high probability (very close to 1.0). In addition, our insurance mechanism improves long term
expected profit gains from Gs = 760.51 to GI
s = 1485.04, a 95.3% improvement. This improvement ratio
also holds for average per seller transaction gains. As rh increases from 100 to 200, the improvement on the
E[Tr], Pd, Gs, Ge, becomes more significant. We next state the appropriate values for CI , DI, bDI and Tc in
the following theorem.
11
r ])
(E[Tr], E[T I
(Pd, P I
d )
(Gs, GI
s)
(Ge, GI
e)
rh = 100
(168.1, 21.5)
(0.20819, 0)
rh = 150
(251.6, 31.5)
(0.99992, 0)
rh = 200
(334.9, 41.5)
(1.0, 0)
(760.51, 1485.04)
(76.051, 148.504)
(80.81, 1485.03)
(8.081, 148.503)
(80.71, 1485.01)
(8.071, 148.501)
Table 5: Impact of our insurance on E[Tr], Pd, Gs and Ge.
Theorem 5.2. An upper bound for the insurance price CI is CI < GI
s − Gs. If DI and bDI satisfies
then Pr[shipment cost exceeds DI ] ≤ ǫ. If Tc ≥ d, then all products sold by a seller with an insurance certificate
can be guaranteed to obey the insurance mechanism.
DI > bDI ≥ CS max{ln ǫ−1 − λ2PbrTd, e2λ2PbrTd},
Proof: Please refer to the appendix for derivation.
Remark: The insurance price should be lower than GI
s − Gs. The clearing time should be larger or equal
to d. To guarantee that the insurance deposit covers the shipment cost for returning products with high
probability, DI and bDI need to be no less than CS max{ln ǫ−1 − λ2PbrTd, e2λ2PbrTd}.
6 Related Work
Research on reputation systems [19] for internet services has been quite active. Many aspects of reputation
systems have been studied, i.e., reputation metric formulation and calculation [11, 15, 20], attacks and defense
techniques for reputation systems [5, 10, 24, 23], and effectiveness of reputation systems [7]. A survey can be
found in [12].
Theoretical aspects of reputation system have been studied extensively. First, many works studied rep-
utation metric formulation and calculation. Two most representative reputation calculating models are the
eBay-like reputation model [11] and the transitive trust based model [5]. The eBay-like reputation system is
a typical example of reputation model which computes the reputation score by summarizing explicit human
feedbacks (or ratings) [3, 11, 21, 25]. The transitive trust based model [5, 6, 20, 15, 24] assumes that if user
A trusts user B and user B trusts user C, then user A trusts user C. More precisely, each user is represented
by a node in a graph, and the weighted directed link from A to B quantifies the degree that user A trusts
user B. For this model, many algorithms were developed to compute an overall reputation score for each
user [5, 6, 20, 15, 24]. These works provided theoretical foundations for reputation computing. Second,
many works explored attack and defense techniques for reputation systems. One type of potential attacks is
that users may not give honest feedbacks. Peer-prediction method based mechanisms were proposed to elicit
honest feedbacks [13, 14, 17]. Another type of potential attacks is reputation inflation, or self-promotion.
Many works have been done to address this issue [5, 10, 24, 23]. A survey on attack and defense techniques
for reputation systems can be found in [10]. The main difference between our work and theirs is that we
propose a new mechanism to improve eBay system.
The most closely related works are [2, 7, 8, 16], which studied the eBay reputation mechanism. Authors
in [2] derived the minimum transaction fee to avoid ballot stuffing (i.e., fake positive feedbacks). Authors
in [7] proposed an algorithm based on buyer friendship relationship to filter out unfair ratings. In [7], authors
explored the impact of buyers biases' (i.e., leniency or criticality) in express feedback ratings on sellers in
advertising product quality. The impact of negative feedbacks on buyers in expressing feedback ratings was
studied in [16]. The difference between our work and theirs is that we propose a new mechanism to improve
eBay system.
7 Conclusion
This paper presents an Insurance mechanism to improve eBay-like reputation mechanisms. We proposed
four performance measures to analyze eBay reputation system: (1) new seller ramp up time, (2) new seller
drop out probability, (3) long term profit gains for sellers and (4) average per seller transaction gains for an
E-commerce operator. We developed a stochastic model to identify key factors which influence the above
four measures. We proposed an insurance mechanism to improve the above four measures. We show that
our insurance mechanism can reduce the ramp up time by around 87.2%, and guarantee new sellers ramp up
12
before the deadline Tw with a high probability (close to 1.0). It also increases the long term profit gains and
average per seller transaction gains by at least 95.3%.
References
[1] Amazon. http://www.amazon.com/.
[2] R. Bhattacharjee and A. Goel. Avoiding ballot stuffing in ebay-like reputation systems.
In Proc. of
P2PECON, 2005.
[3] S. Buchegger and J.-Y. Le Boudec. A robust reputation system for peer-to-peer and mobile ad-hoc
networks. In Proc. of P2PECON, 2004.
[4] X. Chen and B. Miu. Stochastic processes. The University of Science and Technology of China Press,
2002.
[5] A. Cheng and E. Friedman. Sybilproof reputation mechanisms. In Proc. of P2PECON, 2005.
[6] R. Delaviz, N. Andrade, J. Pouwelse, and D. Epema. Sybilres: A sybil-resilient flow-based decentralized
reputation mechanism. In Proc. of IEEE ICDCS, 2012.
[7] C. Dellarocas. Immunizing online reputation reporting systems against unfair ratings and discriminatory
behavior. In Proc. ACM EC, 2000.
[8] C. Dellarocas. Analyzing the economic efficiency of ebay-like online reputation reporting mechanisms.
In Proc. of ACM EC, 2001.
[9] eBay. http://www.ebay.com/.
[10] K. Hoffman, D. Zage, and C. Nita-Rotaru. A survey of attack and defense techniques for reputation
systems. ACM Comput. Surv., 42(1):1:1 -- 1:31, Dec. 2009.
[11] D. Houser and J. Wooders. Reputation in auctions: Theory, and evidence from ebay. Journal of
Economics & Management Strategy, 15(2), 2006.
[12] A. Jøsang, R. Ismail, and C. Boyd. A survey of trust and reputation systems for online service provision.
Decis. Support Syst., 43(2), 2007.
[13] R. Jurca and B. Faltings. Minimum payments that reward honest reputation feedback. In Proc. of ACM
EC, 2006.
[14] R. Jurca and B. Faltings. Collusion-resistant, incentive-compatible feedback payments. In Proc. of ACM
EC, 2007.
[15] S. D. Kamvar, M. T. Schlosser, and H. Garcia-Molina. The eigentrust algorithm for reputation manage-
ment in p2p networks. In Proc. of WWW, 2003.
[16] T. Khopkar, X. Li, and P. Resnick. Self-selection, slipping, salvaging, slacking, and stoning: The impacts
of negative feedback at ebay. In Proc. of ACM EC, 2005.
[17] N. Miller, P. Resnick, and R. Zeckhauser. Eliciting informative feedback: The peer-prediction method.
Management Science, 51(9):1359 -- 1373, Sept. 2005.
[18] M. Mitzenmacher and E. Upfal. Probability and computing. Cambridge University Press, 2005.
[19] P. Resnick, K. Kuwabara, R. Zeckhauser, and E. Friedman. Reputation systems. Commun. ACM,
43(12):45 -- 48, Dec. 2000.
[20] P. Resnick and R. Sami. Sybilproof transitive trust protocols. In Proc. of ACM EC, 2009.
[21] A. Singh and L. Liu. Trustme: anonymous management of trust relationships in decentralized p2p
systems. In Proc. of P2P, 2003.
[22] Taobao. http://www.taobao.com/.
[23] B. Viswanath, M. Mondal, K. P. Gummadi, A. Mislove, and A. Post. Canal: Scaling social network-based
sybil tolerance schemes. In Proc. of ACM EuroSys, 2012.
13
[24] H. Yu, M. Kaminsky, P. B. Gibbons, and A. Flaxman. Sybilguard: Defending against sybil attacks via
social networks. In Proc. of ACM SIGCOMM, 2006.
[25] R. Zhou and K. Hwang. Powertrust: A robust and scalable reputation system for trusted peer-to-peer
computing. IEEE TPDS, 18, 2007.
Appendix
We first state a lemma which will be used in our proof.
Lemma 7.1 ([4]). Let {N ′(t), t ≥ 0} denote a Poisson process with a rate parameter λ. Let tk denote the
arrival time of k-th event. Let f (t1, . . . , tnN ′(t′) = n) denote the conditional probability density function of
t1, . . . , tn given N ′(t′) = n. Then we have f (t1, . . . , tnN ′(t′) = n) = n!/(t′)n, where 0 < t1 < . . . < tn < t′.
Proof of Theorem 4.1: Note that each new seller advertise product quality honestly. In this scenario, each
transaction results in one positive feedback. Recall our reputation updating rule specified in Equation (2), we
have that the reputation score at time slot τ equals the number of transactions arriving within time slot 0 to
time slot τ − 1. Recall the definition of Tr in Equation (4), we have that Tr/d ∈ N. With these observations
and by some basic probability arguments, we have
E[Tr] =
τ d Pr[Tr = τ d] = d
τ Pr[Tr/d = τ ]
∞Xτ =1
Pr[Tr/d ≥ τ ] = d
(1 − Pr[Tr/d ≤ (τ − 1)])
∞Xτ =1
∞Xτ =1
= d
= d
∞Xτ =1
∞Xτ =1
(1 − Pr[r(τ − 1) ≥ rh]) = d
∞Xτ =1 1−
τ −2Xℓ=0
N (ℓ) ≥ rh! .
E[Tr] = d
(1 −
∞Xτ =1
∞Xk=rh
e−λ1Pba(τ −1)d (λ1Pba(τ − 1)d)k
k!
).
Note thatPτ −2
We have
ℓ=0 N (ℓ) is a random variable which follows a Poisson distribution with parameter λ1Pba(τ −1)d.
Evaluating the first order derivative on E[Tr] with respect to rh and λ1Pba respectively, one can easily obtain
the monotonous property of E[Tr].
Proof of Theorem 4.2: Applying similar derivation as Theorem 4.1, we have that the reputation score
at time slot τ equals the number of transactions arriving within time slot 0 to τ − 1. Note that Tw/d ∈ N.
Recall the definition in Equation (4) we have that Tr > Tw if and only if r(Tw/d) < rh. Using some basic
probability arguments, we have
τ =0
τ =0
Pd =
N (τ ) < rh(cid:21) .
Note thatPTw /d−1
We have
N (τ ) is a random variable which follows a Poisson distribution with parameter λ1PbaTw.
Pd = Pr[r(Tw/d) < rh] = Pr(cid:20)XTw /d−1
Pr
N (τ ) = k =
Tw/d−1Xτ =0
rh−1Xk=0
Proof of Theorem 4.3: Let bGs =P∞
τ =0 δτ uN (τ ). Then bGs is a random variable and Gs = E[bGs]. Based
on whether a seller ramps up or not, we divide E[bGs] into two parts, i.e.,
E[bGs] = Pr[Tr > Tw]E[bGsTr > Tw]
+ Pr[Tr ≤ Tw]E[bGsTr ≤ Tw]
Evaluating the first order derivative on Pd with respect to rh, Tw and λ1Pba respectively, one can easily
obtain the monotonous property of Pd.
e−λ1PbaTw (λ1PbaTw)k
.
rh−1Xk=0
k!
14
We next derive the above two terms individually.
We first derive Pr[Tr > Tw]E[bGsTr > Tw]. Note that each new seller advertise product quality honestly.
Hence each transaction earns one positive feedback. Recall that Tw/d ∈ N. Note that at time slot Tw/d
a seller drops out, i.e., there will be no transactions from time slot Tw/d. Let K denote the number of
transactions arriving within time slot 0 to Tw/d − 1. Then K satisfies 0 ≤ K ≤ rh − 1 since we are given
Tr > Tw. Note that K = k, where k ≤ rh − 1 implies that Tr > Tw and K is of a value larger than rh − 1
implies that Tr ≤ Tw. Then we have
k=0
Pr[Tr > Tw]E[bGsTr > Tw]
=Xrh−1
=Xrh−1
Pr [K = k, Tr > Tw] E[bGsTr > Tw, K = k]
Pr [K = k] E[bGsK = k].
k=0
N (τ ) is a random variable which follows a Poisson distribution with parameter
Observe that K =PTw /d−1
τ =0
λ1PbaTw. We then have
Pr[Tr > Tw]E[bGsTr > Tw]
e−λ1PbaTw (λ1PbaTw)k
=
k!
rh−1Xk=0
E[bGsK = k].
We next derive E[bGsK = k]. Let t1, . . . , tK denote the arrival time of transaction 1, . . . , K. Then t1, . . . , tK
satisfy 0 < t1 < . . . < tK < Tw. Let f (t1, . . . , tkK = k) denote the probability density function of t1, . . . , tk
given that K = k. By applying Lemma 7.1, we obtain that f (t1, . . . , tkK = k) = k!/(Tw)k. Then we have
E[bGsK = k]
=Z . . .
Z
Z
=Z . . .
0<t1<...<tk<Tw
dt1 . . . dtk
0<t1<...<tk<Tw
f (t1, . . . , tkK = k)E[bGst1, . . . , tk]
kXj=1
uδ⌈tj /d⌉dt1 . . . dtk,
k!
T k
w
where the first step follows that given t1, . . . , tk is equivalent to given K = k. The second step follows that
the payment of transaction 1 ≤ j ≤ k is forwarded to the seller in time slot ⌈tj/d⌉ because of the shipment
delay. Namely, in computing of the long term profit gains, the j-th transaction results in a discounted profit
j=1 uδ⌈tj /d⌉. Combining them all, we have
gain of uδ⌈tj /d⌉. Hence E[bGst1, . . . , tk] =Pk
rh−1Xk=0
Pr[Tr > Tw]E[bGsTr > Tw] =
Z . . .
k!
T k
w
0<t1<...<tk<Tw
Z
kXj=1
uδ⌈tj /d⌉dt1 . . . dtk.
e−λ1PbaTw (λ1PbaTw)k
k!
Based on our reputation updating rule specified in Equation (2), we obtain that trh < Tw implies Tr ≤ Tw.
Note that trh is a random variable. Let f (trh ) denote the probability density function of trh . Observe
e−λ1Pbat(λ1Pbat)k/k!. Performing the first order derivative on this term, we have
We now derive Pr[Tr ≤ Tw]E[bGsTr ≤ Tw]. Let trh denote the arrival time of the rh-th transaction.
that Pr[trh ≤ t] =P∞
f (trh) = λ1Pbae−λ1Pbatrh (λ1Pbatrh )rh−1/(rh − 1)!. Then it follows that
k=rh
Pr[Tr ≤ Tw]E[bGsTr ≤ Tw] =Z Tw
=Z Tw
λ1Pbae−λ1Pbatrh
0
0
(λ1Pbatrh )rh−1
(rh − 1)!
f (trh )E[bGstrh]dtrh
E[bGstrh ]dtrh.
We next derive E[bGstrh ]. Let t1, . . . , trh−1 denote the arrival time of the 1-st,. . . , (rh − 1)-th transaction.
Let f (t1, . . . , trh−1trh) denote the probability density function of t1, . . . , trh−1 given trh. Then applying
15
Lemma 7.1, we obtain that f (t1, . . . , trh−1trh ) = (rh − 1)!/(trh)rh−1. Then it follows that
f (t1, . . . , trh−1trh )
Z
0<t1<...<trh −1<trh
E[bGstrh ] =Z . . .
E[bGst1, . . . , trh−1, trh]dt1 . . . dtrh−1
Z
=Z . . .
(trh )rh−1 u(cid:16) rhXj=1
0<t1<...<trh−1<trh
(rh − 1)!
λ1Pba(d⌈trh /d⌉ − trh)+λ2Pbrd
δ⌈tj /d⌉ + δ⌈trh /d⌉
δ⌈trh /d⌉+1
1 − δ (cid:17)dt1 . . . dtrh−1.
We elaborate more on computing E[bGst1, . . . , trh−1, trh ]. The transactions arriving within time 0 to time
trh contribute Prh
j=1 uδ⌈tj /d⌉ to the long term profit gains. Consider the transactions arriving within time
trh to d⌈trh/d⌉. This time interval belongs to time slot ⌊trh/d⌋. Note that in this time slot the reputation
score is lower than rh. This means that the number of transactions arriving in this time interval follows a
Poisson distribution with parameter λ1Pba(d⌈trh /d⌉ − trh). In expectation, transactions arriving in this time
interval contribute δ⌈trh /d⌉λ1Pba(d⌈trh /d⌉ − trh)u. Consider transactions arriving from time slots ⌈trh/d⌉ to
∞. In these time slots, a seller's reputation satisfies the condition to be labeled as reputable. The number of
transactions arriving in each of these time slots follows a Poisson distribution with parameter λ2Pbrd. These
transactions, in expectation, contribute uλ2Pbrd δ⌈trh
in total. Summing these terms together we obtain
/d⌉+1
1−δ
E[bGst1, . . . , trh−1, trh]. We have
Pr[Tr ≤ Tw]E[bGsTr ≤ Tw]
=Z Tw
Z . . .
λ2Pbrd + δ⌈trh /d⌉λ1Pba(d⌈trh /d⌉ − trh )(cid:17)dt1 . . . dtrh−1
(trh )rh−1 u(cid:16) rhXj=1
λ1Pbae−λ1Pbatrh
(λ1Pbatrh)rh−1
0<t1<...<trh−1<trh
δ⌈tj /d⌉ +
(rh − 1)!
(rh − 1)!
δ⌈trh /d⌉+1
1 − δ
Z
dtrh
0
Combing them all, we prove this theorem.
Proof of Theorem 5.1: We first derive E[T I
d . Note that sellers advertise honestly. This means that
all transactions result in positive feedbacks. This implies that the insurance certificate expires at the end of
the duration time. Note that N (ℓ), the number of transactions at time slot ℓ = 0, 1, . . . , ∞ before a seller ramps
r ] and P I
the expected ramp up time E[T I
up, follows a Poisson distribution, and we denote its parameter by eλT (ℓ). Applying Equation (9), we have
eλT (ℓ) = λ2Pbrd for all ℓ = 0, 1, . . . , ⌊Td/d⌋−1, andeλT (⌊Td/d⌋) = λ2Pbr(Td −d⌊Td/d⌋)+λ1Pba(d⌊Td/d⌋+d−Td),
andeλT (ℓ) = λ1Pbad for all ℓ = ⌊Td/d⌋ + 1, . . . , ∞. Then with a similar derivation as Theorem 4.1 we obtain
Gs first. Let bGI
τ when a seller subscribes to an insurance. Then bGI
Let us now derive long term profit gains (Gs) and average per seller transaction gains (Ge). We derive
τ =0 δτ uN I(τ ), where N I (τ ) denotes the number of transactions arriving in time slot
s]. With a similar
r ]. Furthermore, with a similar derivation as Theorem 4.2 we obtain P I
d .
s is a random variable and GI
derivation as Theorem 4.3, we have
s = E[bGI
s =P∞
s] = Pr[T I
+ Pr[T I
E[bGI
sT I
r > Tw]E[bGI
r ≤ Tw]E[bGI
r > Tw]
sT I
r ≤ Tw]
(1) a seller ramps up (T I
t ∈ [Td, T I
Let bλT (t) denote the transaction's arrival rate at time t[0, ∞). Applying Equation (9), we have two cases:
r ≤ Tw), then we havebλT (t) = λ2Pbr for all t ∈ [0, min{Tw, Td}],bλT (t) = λ1Pba for all
r > Tw), then we havebλT (t) = λ2Pbr
for all t ∈ [0, min{Tw, Td}], bλT (t) = λ1Pba for all t ∈ [Td, Tw], and bλT (t) = 0 for all t ∈ [Tw, ∞]. With these
observations and using a similar derivation as Theorem 4.3, one can easily obtain analytical expressions for
sT I
the term Pr[T I
r ≤ Tw] respectively. Combing them
all we complete this theorem.
r ], andbλT (t) = λ2Pbr for all t ∈ [T I
r , ∞]; (2) a seller drops out (T I
r > Tw] and the term Pr[T I
sT I
r ≤ Tw]E[bGI
r > Tw]E[bGI
Proof of Theorem 5.2: We want to derive the reasonable price that an E-Commerce operator can charge for
the insurance. The marginal long term profit gain of an insured seller is GI
s − CI . Note that the marginal long
16
term profit gain without insurance is Gs. Thus sellers has the incentive to buy an insurance if the marginal
profit gain corresponds to buying an insurance is larger than the marginal profit gain without insurance, i.e.,
GI
s − CI > Gs, which yields CI < GI
s − Gs.
Note that sellers advertise honestly. Let N ′(Td) denote the total number of products sold in insurance
duration time. It is easy to see that N ′(Td) follows a Poisson distribution with parameter λ2PbrTd. The
worst case is that all buyers hold the product till the last minute of the clearing time Tc and then return it.
Using a Chernoof bound [18] argument, one can easily bound the shipment cost (at the worst case) as
Pr[N ′(Td)CS ≥ N ′CS] ≤ e−λ2Pbr Td (eλ2PbrTd)N ′
/N ′N ′
Setting N ′ = max{ln ǫ−1 − λ2PbrTd, e2λ2PbrTd} we have Pr[N ′(Td)CS ≥ N ′CS] ≤ ǫ. The clearing time
follows the shipment delay d.
17
|
1907.05636 | 2 | 1907 | 2019-07-25T12:46:04 | From Observability to Significance in Distributed Information Systems | [
"cs.MA",
"cs.AI",
"cs.DC",
"eess.SY",
"eess.SY"
] | To understand and explain process behaviour we need to be able to see it, and decide its significance, i.e. be able to tell a story about its behaviours. This paper describes a few of the modelling challenges that underlie monitoring and observation of processes in IT, by human or by software. The topic of the observability of systems has been elevated recently in connection with computer monitoring and tracing of processes for debugging and forensics. It raises the issue of well-known principles of measurement, in bounded contexts, but these issues have been left implicit in the Computer Science literature. This paper aims to remedy this omission, by laying out a simple promise theoretic model, summarizing a long standing trail of work on the observation of distributed systems, based on elementary distinguishability of observations, and classical causality, with history. Three distinct views of a system are sought, across a number of scales, that described how information is transmitted (and lost) as it moves around the system, aggregated into journals and logs. | cs.MA | cs | From Observability to Significance
in Distributed Information Systems
Mark Burgess
Aljabr Inc.
ChiTek-i AS
Abstract -- To understand and explain process be-
haviour we need to be able to see it, and decide
its significance,
i.e. be able to tell a story about its
behaviours. This paper describes a few of the modelling
challenges that underlie monitoring and observation of
processes in IT, by human or by software. The topic of
the observability of systems has been elevated recently
in connection with computer monitoring and tracing of
processes for debugging and forensics. It raises the issue
of well-known principles of measurement, in bounded
contexts, but these issues have been left implicit in
the Computer Science literature. This paper aims to
remedy this omission, by laying out a simple promise
theoretic model, summarizing a long standing trail of
work on the observation of distributed systems, based
on elementary distinguishability of observations, and
classical causality, with history. Three distinct views of
a system are sought, across a number of scales, that
described how information is transmitted (and lost) as
it moves around the system, aggregated into journals
and logs.
I.
INTRODUCTION
The desire to expose the chains of cause and effect
that lead to observable phenomena is irresistible, yet
we know that the ability to observe and interpret sys-
tems is both limited and compromised by availability,
relativity, signalling speeds, and even by the noise of
environment. In computer science, these limitations
have largely been set aside to prioritize other concerns,
but in the era of wide area cloud computing this will
become harder to do, as our ability to observe and
assess software behaviours is filtered through ever
more opaque layers of abstraction.
The state of the art in monitoring relies principally
on brute force data collection and graphical presen-
tation. There is surprisingly little discussion about
the semantics of the process [1] -- [6]. Only recently
has there been any serious interest in semantics for
distributed tracing [7] -- [9]. Uncertainties, accrued by
sensory instrumentation are left for human operators
to untangle on their own. With few exceptions, the
literature on logging and monitoring (prior to the
present wave of Machine Learning studies), singles
out the design of machinery to collect data, without
due consideration of relevance, accuracy, or semantics
of what is collected (some recent examples include
[10] -- [15]). This is regrettable, but not uncommon in
technological literatures. The machinery becomes an
end in and of itself, and its knowledge-related function
is subordinated to the prowess of its performance.
Interestingly, the collection of observational mea-
surements from a distributed system is related to the
far-more widely studied problem of data consensus,
in Computer Science [16], [17]. The latter considers
how we may distribute multiple copies of information
over a wide area, with integrity of order -- surely one
of the most frequently revisited problems tackled in
distributed systems. The fascination with consensus
stems from the attempt to cling onto approximate
determinism. The topic of observability, on the other
hand, (literally the ability to observe systems, as
contrasted with monitoring which is following what
you actually can see) is its approximate inverse: how
can we meaningfully integrate data from widespread
sources into a viewpoint consistent with a single
observer, also with integrity. It's an issue that has
hounded physics and engineering for centuries. Like
its dispersive counterpart, observability is a problem
dominated by relativistic issues of space and time. Un-
like consistency, it has not been studied with anything
like the same degree of care.
The 'measurement problem', as it is known in
physics, bedevils every corner of science in different
ways. We need to ask: is there a consistent viewpoint
that can be arrived at without doing such violence
to the system as to wipe out any other signal. In
this paper, I want to apply the language of Promise
Theory [18] and Semantic Spacetime [18] -- [23] to the
understanding and analysis of distributed systems on
any scale.
IT tends to favour action over understanding; this
has probably led to the neglect of detailed models
for process monitoring. The issues of how we cope
with preferential sampling (which may result in result
bias) and relative scaling of sample populations, for
instance,
is of serious importance, but completely
absent from monitoring literature. Statisticians talk
about biases and significance of populations, but the
normal state of affairs for any observer, watching in
band of the process (i.e. colloquially in 'realtime'), is
to see a small random sample of data whose larger
scale significance cannot easily be assessed without
long term studies. Decisions and value judgements
are inevitably based on samples that are small and
statistically inadequate, so this problem cannot be
9
1
0
2
l
u
J
5
2
]
A
M
.
s
c
[
2
v
6
3
6
5
0
.
7
0
9
1
:
v
i
X
r
a
argued away by Central Limit Theorems and the like.
Observability is a necessary but not sufficient criterion
for understanding.
between pairs of agents, which therefore serve as
documentation of intent. An offer promise with body
+bi made by Si to Rj is written:
None of this is not what monitoring software
purports to do -- rather it pretends to offer instant
insight, independent of scale. In this paper, I'll try
to set out some definitions as clearly as possible, with
a view to answering a few basic questions, especially:
can we feasibly collect enough information to en-
able reversible reconstruction of process history, thus
enabling forensic causal reconstruction of scenarios
past? I'll show that, if a sufficient level of information
is promised, and agents keep their promises with suffi-
cient fidelity, then trajectories can be traced reversibly
to causal roots. However, the reconstruction of an
agent's ealier state from its current state (the tracing
analogue of 'rollback'), or from promised data, should
be considered impossible in general [24].
I'll focus on the two necessary conditions for
assessing systems, from basic information theory [25],
[26]:
•
The ability to observe remote data with rea-
sonable fidelity.
The ability to aggregate and combine remote
observations with similar fidelity.
•
These points are never more important than in ex-
tended cloud computing, where data collection sys-
tems extend all the way out to the edge of user contact,
e.g. the Internet of Things.
II. NOTATION AND DEFINITIONS
Let's begin with some definitions for the purpose
of making more precise statements. Promise Theory
provides a useful language here, in terms of promises
(or impositions), and assessments. In a promise the-
oretic model, any system is a collection of agents.
Agents may be humans or machines, hardware or
software. Usually, agents will be active processes.
Agents represent internalized processes that can make
and keep generalized promises to one another [18].
The generic label for agents in Promise Theory
is Ai, where Latin subscripts i, j, k, . . . numbers dis-
tinguishable agents for convenience (these effectively
become coordinates for the agents). We shall often
use the symbols Si and Rj, instead, for agents to
emphasize their roles as source (initiator) and receiver
(reactive). So the schematic flow of reasoning is:
S offers (+ promises) data.
1)
2) R accepts (- promises) or rejects the data,
either in full or in part.
3) R observes and forms an assessment αR(.)
of what it receives.
+bi−−→ Rj,
Si
(1)
where the + refers to an offer of some information
or behaviour (e.g. a service). This is a part of Si's
autonomous behaviour, and the promise constrains
only Si. Rj may or may not accept this by making a
dual promise, marked −b to denote the orientation of
intent:
−bj−−→ Si.
Rj
(2)
If both of these promises are given, and kept, then
influence in the form of vital information about the
body b will pass from Si to Rj. In general, the offer
and acceptance may not match precisely, in which case
the propagated information will be the overlap (mutual
information)
b∩ = bi ∩ bj,
(3)
in the manner of mutual information [25], [26]. I'll
suppose that modern systems are cloud computing
systems. The elementary agents of cloud computing
are processes, any of which may express promises
about state and services. Processes are hosted at agent
locations Ai, Si, Ri etc.
I use the following nomenclature for message
agents M:
•
•
• Eγ is an event, for example Lγ ⊂ Eγ may
be a line of information reported in a log
or journal. Greek indices γ label information
agents successive packets, i.e. γ = 1, 2, 3, . . ..
{Eγ} or{Lγ} refers to a collection of such
events or lines.
Si, Ri ∈ Ai refer to processes running on
computers.
{Si} refers to a collection of sources, etc.
•
• Ai refers to a process checkpoint in some
kind of dataflow, which has its own interior
event log and counters. Checkpoints typically
make promises about their identity, location,
local counter values, and intent to pass on data
in the form of packets Pi, with some promised
order.
Pi refers to a data packet passed between
checkpoints agents. Packets typically make
promises about their identity, data content,
schema, and type.
•
Latin indices therefore label
indices label events at the same location.
locations, and Greek
This third and final stage is the moment at which data
can be said to arrive at the receiver.
The details of a physical network are not directly
relevant, but the topology of actual interactions be-
tween agents is. It depends on the promises made
III. DEFINING THE PROBLEM
A. What is intended and what is promised?
There are two ways in which we use data to
interrogate a number of processes:
2
•
•
Tracing: ('During') -- in band observation,
in which data are sampled intentionally and
recorded as a process unfolds to maintain
'situation awareness'. e.g. the ECG or life
monitor approach to medical monitoring.
Diagnosis: ('After') -- out of band forensic
reconstruction of a system using data one
can find after an incident, where intent to
comprehend kicks in only after the event has
occurred: e.g. the post mortem approach to
medical investigation.
Most users will
try to combine these approaches,
paying attention mainly when significant events occur.
The automation of alarms (usually based on simple-
minded absolute thresholds) tells human operators
when to pay attention, at which point they have to
rely on what has been traced. The promise to maintain
awareness is an expensive one, and we rely heavily on
our skills of reconstruction after the fact.
B. Three perspectives about scale and relativity
Distributed processes are composed of agents that
pass information in space and time (see figure 1). Mes-
sages or events propagate from one agent to another,
and we consider the arrival of such information to be
an advance in the 'state' of the distributed system,
which is what we mean by the proper time of the
process. Events that occur in parallel, as separate
logical locations, know nothing about one another --
they are causally disconnected and lead independent
lives. The time on the wall clock or system clock is
not a 'proper' time, as we'll see below1.
Fig. 1: Space and time as agent parallelism and serialism
respectively.
There are three kinds of story or explanation we
want to be able to tell about distributed systems (figure
2):
1)
2)
3)
The data traveller log. What a travelling data
packet experiences along its journey, e.g.
which software including version handled it
and in what order?
The checkpoint visitor log, from key sign-
posts around the data processing landscape.
The log of what each checkpoint along the
journey saw, i.e. which data packets passed
through the checkpoint and what happened
to them?
The map of combinatoric intent,
the
relationships between invariant elements and
i.e.
1In Einsteinian relativity, the term proper time is reserved for the
time experienced by an observer about its own states, so I keep to
this convention here.
concepts, including the topology of check-
the types of data
points and influences,
passed between them,
significant occur-
rences, and so forth; i.e. the semantics of
the data, software, and invariant qualities
and quantities that summarize the processes
within the system's horizon.
These viewpoints require separate data collections.
Present day logging systems focus almost entirely on
the second of these.
In Promise Theory, one reduces a system to a
collection of agents, their promises, and their assess-
ments. Agents include the checkpoints from which
data emerge and are collected. A second layer of
agents comprises the data packets that are transmitted.
The promises made by these agents include commu-
nication, data compression, speed, and integrity. They
may include data formats and ordered protocols. We
equip different agents in the system with promises to
report the information available to them to observers.
I shall not be concerned with matters of authorization
and permission in this paper, but rather focus on the
difficulties experienced by those who are promised
information.
C. Diagnosis
It's up to an observer to infer something about
the state and history of a system, based on what
is observed. This is not as straightforward as soft-
ware systems have come to assume, especially as
cloud computing pushes the limits of observability.
At some point, this reconstruction involves a form of
reasoning -- not necessarily rigid logical reasoning, but
at least a process of joining dots into an acceptable
story.
Fig. 2: 3 Views. Travelling passport documents, versus logs
of entry and exit from a checkpoint, versus the map of
checkpoints and routes.
D. Diagnostic messages
Process tracing is a simpler problem than rea-
soning, because it can be constructed as a purely
Markov process -- at least in principle. Tracing is the
construction of a totally ordered path through a set
of agents. Reasoning, on the other hand,
involves
semantic relationships between clusters of agents that
may be considered to have an invariant meaning,
and it may combine several traces into a satisfactory
explanation. According to the definitions in [22], [23],
I'll simply define the following:
3
spaceliketimelikeTHE MAPcheckpointstravellingdata Definition 1 (Reasoning): Reasoning is a search
over a graph of ordered conceptual relationships.
This pragmatic and unconventional definition might
offend some logicians, but it's closer to what humans
call reasoning than a definition based on mathematical
logic. A few common issues crop up in diagnostics:
•
•
•
The predictability of agents' behaviours.
The distinguishability of agents and data mes-
sages.
Loss of information due to mixing of origin
sources.
Reordering of information due to latency.
•
The problem with the first two kinds of story in the
list, is the lack of a deterministic and universally de-
fined order between the transactions of 'event driven'
processes at separate source locations. The extent to
which we can write down spatially invariant orders,
process summaries, etc, which may be expected to
persist over a timescale useful for prediction, is the
essence of the difficulty in tracing causal history2.
IV. OBSERVABILITY OF MESSAGES
From the foregoing, we can define the concept of
observability between pairs of agents. It does not make
invariant sense to speak of 'observability' without
some qualifications, so we can be more precise:
Theorem 1 (Observability of X at S by R): A
range of promised set values X, sourced from an
agent S is observable by an agent R if and only if:
π+ : Si
π− : Rj
Xj
+Xi−−−→ Rj,
−Xj−−−→ Si.
⊆
Xi
(4)
(5)
(6)
Note that
the criterion for observability is not a
deterministic guarantee the ability to obtain a value on
demand. It is essentially a property of an information
channel, in the Shannon sense. There is only a finite
probability of all these promises being kept, which
makes observation a fundamentally non-deterministic
process. There are many impediments to keeping
promises in practice, not least of which the the law of
intermediaries3.
By assumption, agent's are autonomous and each
plays a role in the collaboration required to exchange
the information involved in monitoring. The defini-
tion of observability illuminates a basic dilemma in
monitoring: the autonomy of agents in any distributed
2The main approach to determining spacetime invariance in
science is by the use of statistics (aggregation or 'learning'): by
accumulating multiple samples, we hope to separate what is quickly
varying (fluctuation) from what is slowly varying (trend). Persistent
concepts are what remains during or after a process of learning has
separated these processes.
3See reference [18]. The law of intermediaries basically says
that intermediate agents cannot be relied upon to faithfully trans-
mit promises or intent, because all agents are fundamentally au-
tonomous.
process (i.e. their causal independence) means that
there is fundamental uncertainty about the process of
observation not just its outcome. Causal independence
is the very definition of a random variable. A source
of signals may believe it does all it can to ensure
correct transfer of information, and that any problems
lie in the delinquencies of the receiver; meanwhile,
the receiver believes it does all it can and trusts the
source and the network in between implicitly to report
with complete fidelity. The assessment of X by Rj
(denoted αj(π+)) is still a function of Rj's access and
capabilities at any given sample, and may be subject
to environmental interference.
A. Preliminaries about intent
Most
technologists believe that,
if they design
without 'bugs', they can achieve whatever outcome
they desire, given sufficient resources. This is not a
scalable view, so we need to be more cautious. In
the standard model of queueing theory [27], data are
produced by a source Si, at a rate λi messages per
unit time, and may be processed by a receiver Rj
(sometimes called a server) at a rate µj. The queue is
unstable and grows out of control as the traffic density
λi/µj → 1.
The terminology pulling and pushing data are
often used about attitudes to causal intent in commu-
nications. These lead to some confusion, so it's worth
mentioning them. I define these in accordance with
the same information channel principles.
Definition 2 (Push): A method of communication
in which a source agent S imposes its messages onto
a recipient R without invitation.
+M
S
R.
(7)
In data signalling, packets may be carried over a
wire, enter a network interface and be queued up for
sampling by a receiver, before the receiving process is
ready to accept them. This is imposition. The receiver
may then promise to sample (-) the messages from the
shared queue. The channel flow is thus controlled by
the sender.
Definition 3 (Pull): A method of communication
(sometimes called publish-subscribe) in which a re-
ceiver invites a source to provide a certain quota of
messages for sampling via an agreed channel.
−M−−→ S
R
which is then promised by the source
+M−−→ R
S
(8)
(9)
The flow along the channel is thus controlled by the
receiver.
Pull is always the fundamental 'last mile' stage of
a sampling operation; it may involve active polling
of the queue to match timescales that satisfy the
rigours of Nyquist's theorem. It optimizes message
transfer according to the downstream capabilities.
Push driven systems are sometimes associated with
4
reactive or event driven systems -- though this can be
misleading. Push and pull are effective on different
timescales. Push (notification) is useful in connection
with small signals when source data are sparse and
a receiver needs a short wake up message to collect
a package from the source, enabling it to conserve
resources. Push therefore provides non-redundant in-
formation when arrivals are sparse or infrequent, on
the timescale of the receiver's sampling [28]. Pull sys-
tems make more efficient use of queue processing, by
utilizing the information about autonomous capacity
to balance load.
B. Events and sampling
The concept of events plays a major role in the
language used for monitoring and data flow in IT, but
is seldom defined. Let's define it here in a way that
respects information theoretic transfer. Information is
only 'arrives' somewhere when it is sampled.
Definition 4 (Event): A discrete unit of process in
which an atomic change is observed or sampled.
We often imagine processes being driven by a flow of
events, like a stream4. Again, in terms of sampling,
this amounts to the following:
Definition 5 (Event or message driven agent):
An Event Driven Agent R makes a promise
conditionally on the sampling of message events M
from a S, with an average rate λ:
+EM
R
(10)
i.e. R can promise an observer O that it acknowledges
an event E on receipt of a message M. By Promise
Theory axioms, this assumes the prior promises:
O
+Mλ
S
R
R
or
S
−Mµ−−−−→ S
+Mλ−−−−→ R
(11)
(12)
Notice that by using the term sampling here, we do
not take a position on whether messages were imposed
by pushing from S to R, or whether R reached out to
S to pull the data. These distinctions are irrelevant to
the causal link that results from the message policy.
Data are not received until they are sampled by the
receiver. Note that there is no timescale implied by
the conditional promise in (10) -- the definition of
'immediate' or 'delayed' response is an assessment
to be made by the observer O.
The concept of reactive systems has been usurped
by a specific industry initiative [29], so it makes sense
to follow in the spirit of that:
Definition 6 (Reactive agent): An event driven
agent
that promises to keep its behaviour within
certain constraints relative to the sampling of events.
A simple example of these principles is to explain
data flow systems like Data Pipelines. These are scaled
combinations of reactive components.
Example 1: Pipelines and Petri Nets: Data pro-
cessing pipelines are hybrid networks of Event Driven
Agents, Reactive Agents, and Service Agents, that
promise to behave as an Event Driven Agent collec-
tively as well as component by component.
A simple fact of queueing theory, embodied as a
principle of autonomy in Promise Theory is to note
that: no amount of pressure or coercion will make an
agent process data faster than its maximum rate µ.
C. Missed and dropped samples
Observations inevitably get
lost
in any scien-
tific enterprise. In empirical science this contributes
to 'error bars' or uncertainties
in counting of
measurements -- but not usually to semantics of inter-
pretation. Interpretations are expected to be stable to
such small perturbations.
Reasoning in IT has its historical origins from
mathematical logic and precision: the avoidance of
doubt. But doubt is a central part of tolerance in sys-
tems. If we observe and inspect systems, we need to
do so in the framework of a stable intent that overrides
random fluctuations in measurement5. Monitoring and
measurement serve no actionable purpose unless there
is already a policy for behaviour in place. Ashby's
model of requisite complexity or 'good regulator'
in cybernetics [30], [31] summarizes how matching
information with information on the same level is
required when there is no intrinsic stability in a model
by which to compress such fluctuations.
V. THE TIME SERIES MODEL
Our received view of time -- as a river of events
that moves everything from past
the
same rate -- is a side effect of living in a rather slow
world, which is close to us, and which we see with
no perceptible delay. In IT, we cannot rely on this
privileged view, and we need to rethink time by going
back to the basics of how we measure it.
to future at
A. Events, clocks, and proper time
In an information theoretic sense, an event is an
observation of change in data sampled from a source
[25], [26]. In the Einsteinian sense, this signal is a tick
of a clock that an observer samples. When the tick
originates from within a process (e.g. a CPU kernel
tick), this defines a notion of 'proper time' for the
local process, indicating an advance in the state of
the process. When there are multiple agents involved,
working together, the language one often speaks of
'vector clocks' in IT, referring to Lamport [32].
Other agents, external to a ticking process, may
observe changes in it differently, either because they
lack access to observe the changes or because the sam-
pling of the changes require intermediary processes
like message passing to propagate the changes from
source to receiver. Thus the proper time experienced
4This illusion of a flow is maintained by a gradient of interme-
diate agents that accumulate samples in buffers.
5We see the effect of 'populism' in society today, when intentions
follow unstable polls instead of convictions.
5
by an agent may not correspond to the exterior time
generated by the sampling of remote events.
Example 2 (Thunder and lightning): When light-
ning strikes, observers in different contexts see and
hear it at different times. Observers very close, that
cannot sample faster than a certain rate, may not be
able to discriminate a difference between the flash and
the thundercrack. Light travels so fast the few agents
can detect a delay in the signal, so they conclude
that the flash occurs as 'the same time' (during the
same sample). But sounds travels more slowly, so
agents at different distances can discriminate the time
at which the sound reaches them. If they synchronize
their watches using light, they will measure different
times for the sound -- but the event happened due to
a process that took place in a single location, over
a tiny fraction of a second. What observers sample
is not always a high fidelity representation of what
happened at the source.
Using the language of Promise Theory [18], we
can define time from two perspectives. For conve-
nience, we'll make an identification between the con-
cept of an event Eγ and a line transaction Lγ in a
system log, or a data point recorded in a timeseries
database Dγ:
Lemma 1 (Events count time): The emission of
an event or 'log line' Eγ = {Lγ, Dγ} is a tick of
interior time clock.
This should be obvious, as events are changes that get
noticed. We can now define interior (proper) time and
exterior (relative) time:
Definition 7 (Interior time of process Si): An in-
dependent count of ticks originating from within a
process S, cannot be observed by any exterior agent
A?, unless promised and reported:
+ticki−−−→ A?.
Si
(13)
Interior time is the image of processes that originate
within the boundary of agent Si. At scale, we can con-
sider superagents of any scale, so interior time scales
and changes in meaning according to our definition of
local.
Definition 8 (Exterior time of process Si): An
independent count, by a remote receiver R, of
promised ticks (observed and aggregated from any
number of sources on a watchlist) that increases for
each sampled event arriving from a exterior process
source Si.
+ticki−−−→ R
−ticki−−−→ Si
+tickticki
−−−−−−→ A?.
Si
R
Si
(14)
(15)
(16)
Exterior time is attached to remote processes that may
originate on any scale. The recipient R that samples
events may itself be of any scale, with associated loss
of certainty about the definition of its interior clock
counters, but ideally R would use a single source from
an elementary agent, for precision.
On the timescales of computers, in our daily lives,
this sounds straightforward, but
the processes that
calibrate our normal idea of time (the system clocks)
are not faster than the sampling processes we are
trying to discriminate by. This leads to a breakdown in
the normal assumptions of universal time for all, and
forces precision agents to go back to basic definitions
of time in order to trace processes in band.
B. Clocks at different scales
We cannot avoid the effects that scaling has on
clocks. Even atomic clocks may not be considered
atomic, in the transactional meaning, on the scale of
subatomic processes. The lesson that Einstein taught
us is that processes need to embody their own clocks,
as single reference sources of truth.
Example 3 (System clock, e.g. Unix): The system
clock, provided by most operating systems derives
from a shallow hierarchy of exterior time services,
based on processes that promise approximate align-
ment. The clock timer C is an independent agent,
which promises a counter (UTC) to processes Pi,
C +UTC−−−→ Pi
−UTC−−−→ C
Pi
(17)
(18)
Using this as a conditional dependency, processes can
then promise timestamps based on the interior counter
+timestampUTC
−−−−−−−−−→ R.
Pi
(19)
Note that the coordination between duplicate redun-
dant clocks is weak. A time service like NTP, provided
by agent N, may be used to periodically align inde-
pendent clocks at a layer of the hierarchy above each
system clock:
C
C
N
Pi
−−−−−→ C
+UTCNTP
−UTCNTP
−−−−−→ N
+UTCUTCNTP
−−−−−−−−→ Pi
−UTC−−−→
C.
(20)
(21)
(22)
(23)
Each clock is independent, so it is only meaningful
to compare two timestamps from the same clock.
Moreover, the relationship between timestamps and
process ticks is indeterminate, since process ticks are
halted relative to the system clock during timesharing.
The use of timestamps in network protocols should
be considered unreliable, and only for round trip
comparisons.
Example 4 (Monotonic counters): Interior
process time can be obtained by incrementing a
counter by an atomic operation. A process P passes a
value v to a counter C, which is a persistent variable,
and promises to increment it as certain milestones
are passed:
+v−−→
−v−−→
C
P
+t=(v+1)v
−−−−−−−−→ P
C.
−t−−→
(24)
(25)
(26)
(27)
P
C
C
P
6
order at the source (figure 3). This requires the in-
troduction of a co-dependence between sender and
receiver, and a detailed explanation has been given in
[33]. Examples include the well-known TCP protocol,
and other more exotic variants.
This does not
imply that data are necessarily
observed in the same order by source and receiver.
Once data leave the agents that are entangled in this
way, the promise of order is not preserved, because
all agents are causally independent.
Lemma 5 (Promised Order Propagation): Data
exchanged without conditional sequence promises
may not be sampled in the same order as they were
promised.
The implication of this lemma is that predictable
coordination of sequences as invariant features be-
tween agents is expensive and unreliable: it does not
happen unintentionally, without chains of interdepen-
dent promises. This is a rather damning result for
monitoring that relies on timestamps for its depiction.
Order promises can be kept by labelled (sequence
numbering) or by waiting in lock-step for changes
one at a time. The independence of agents, and our
inability to make a promise on their behalf, means
that data passing through multiple intermediaries are
independent deliveries. Even a single agent cannot be
forced to deliver data in order, unless it has promised
(fully intends) to do so in advance, with full observ-
ability of the payload (and a receiver that can sample
at the Nyquist rate [22], [26]). Expecting clusters of
agents to preserve order, over possibly parallel routes
is even more unlikely, without prior intent. This can
be expressed by saying that unless there is a single
clock that determines when packets will be sampled
by a receiver, the order will not be preserved. The
default is that an incoming queue of samples serializes
them in a random order, without a surviving chain of
dependency.
We can arrange for such a single clock to be
authoritative (like a shared memory counter), but this
requires agents to make promises to abide by the
order, which in turn requires cooperation from end to
end throughout a channel, to preserve identity serially
and atomically (one agent sampled exclusively at a
time). Another approach to agreeing about time is
to bind clocks in lock-step to form a co-dependent
relationship between agents, known as entanglement
[33]. This is used, for example, in TCP's SYN-ACK
protocol. It promises synchronization at a possibly
unbounded cost in terms of interior time ticks.
Lemma 6 (Promised Order Propagation): The
intended order for events originating from more than
one Si may only be promised by interior cooperation
at the source, and assessed uniquely by an agent R
with observational capacity according to the Nyquist
law. Each rescaling of aggregated time ordering
introduces new uncertainty according to an observer's
clock.
There is no unique intent, for a collection of au-
tonomous agents, unless the multiple sources subor-
Fig. 3: A single agent, with a reference clock can be scaled
into a superagent provided the agents within promise to
coordinate their behaviours. Thus a collective formed from
independent sources can act as a single reliably ordered
source, but at a cost growing like N 2 in the number of
agents.
If independent agents need to coordinate their
clocks, they can build on a single source of truth,
by appointment to the role (see figure 3), essentially
transforming interior time into exterior time.
Lemma 2 (Interior consensus of clocks): The
promise to share interior time Ti from Si,
to an
agent Rj, with interior clocks Ti is equivalent to the
problem of data consensus between clock ticks.
This suggests that clock synchronization by voting in
band ('realtime') will lead to a significant delay in the
rate of time that can be promised as agreed ticks by
an entire superagent. This increased 'mass' of agent
clocks will slow the rate observable by an outside
sampler.
Lemma 3 (Aggregation of clocks): The aggrega-
tion of multiple sources of interior time Ti from Si,
by an agent Rj, with or without consensus, is not one
to one with the interior time of the receiver.
The proof of this may be seen from the Law of
Intermediate Agents [18], which tells us that, if there
are agents in between the source and receiver (which
is nearly always the case),
then no promises are
transferred automatically. We need a chain of delivery
promises to form an expectation of what we are
seeing. The outcome of this is that every agent may
see a different arrival order, assuming that
it can
distinguish between data transmissions. This is a well-
known result. Conversely:
Lemma 4 (Promised Order Propagation): The
order of a sequence of data, from a single agent, can
be promised by virtue of a single clock or counter.
this does not
Note that
imply that order will be
preserved, only that there is a set of promises between
sender and receiver that can be made to transfer
the order information (e.g. by numbering packets).
This is well known in 'reliable' data communications,
like TCP. Without proof, let's acknowledge that the
relative order of data can indeed be transferred reliably
between a sender and a receiver, if there is a promised
7
observer1A2A3A4clockA dinate themselves by cooperation to a single agreed
order, but any attempt
to coordinate between the
agents (and thus act as single superagent making a
common promise) would result in a change in the ticks
observed by R, unless the sampling resolution of R
is much less than the exterior time needed to assess
interior latency of interactions for agreement.
The conclusion of these extended remarks is that
there is no single clock by which to define the order
of events between different hosts. This is esssentially
because unrelated processes have no common time.
The whole idea is meaningless. What observers often
seek is a picture according to their own sense of time
(observer time) that integrates different processes into
a picture of the moment as they perceive it. Alas,
that impression cannot easily be reconstructed later,
even perhaps with a detailed 'post mortem', as it relies
on anchoring to out of band processes that were not
measured. If we introduce a single source of time for
a collection of hosts, by forming a superagent (with
all necessary interior cooperation), each agent within,
we can define a single reference time, but it is not the
proper time of any process.
C. Rule of thumb about time
Example 5 (Common assumptions of system time):
A commonly held belief is that, in interactions like
network protocols, we might define time in a number
of way.
•
•
•
•
The 'actual' time: there is a single source of
truth, by international convention, which is
the official value of UTC. This time standard
exists, but is only obtainable with latencies
that render it approximate. Through a hierar-
chy of services, like NTP, local system clocks
promise to approximate this time and to count
independently on their own at approximately
the same rate. These rates cannot be verified,
so in practice the closest we can obtain is the
current value of the local system clock, which
belongs to localhost.
The observed time: This is a timestamp ren-
dered by sampling the system clock, so it is
relative to localhost's assumed time standard
and has no significance beyond the agent that
sampled it. The observed time may not be
monotonic, for example if clock drift correc-
tions occur in between samples of the clock.
System time may therefore go backwards or
forwards at random, over extended processes.
The publication time: Timestamps may be
shared between processes, or recorded, incur-
ring additional processing delays. The reso-
lution of a timestamp may be quite low, al-
lowing processes to absorb processing delays,
but publication times are always later than the
timestamps they promise, e.g. the timestamp
when a log entry is written is always later
than the timestamp of the log entry.
The receipt or sampling time: If timestamps
are shared between agents, e.g. in recording
data in a log, or transmitting data across a
network, the published timestamp belongs to
the sender S's clock, and the receipt time
belongs to the receiver R's clock. These two
times are causally independent and their com-
parison is strictly meaningless. If all clocks
promise approximate alignment,
the differ-
ence between published and received times-
tamps may promise an accuracy whose un-
certainty is approximated by the Pythagorean
average of the uncertainties of the two times-
tamps at S and R.
Clearly, no protocol (except NTP) passes information
about its clock time uncertainties, so network time
falls foul of the Intermediate Agent law.
VI. THE ROLE OF TIMESCALES FOR
PREDICTABILITY
The purpose of monitoring is to be able to explain
behaviour and even predict problems in advance.
Without predictability, monitoring is little more than
somewhat arcane entertainment [1] -- [6]. One assumes
that, by learning about
the past or by building a
relationship with system behaviours in band, we are
able to predict something about the future behaviour.
This, in turn, assumes a stability under the repetition
of patterns.
Definition 9 (Predictability): A system that has
stable and repeated observable behaviour, on a
timescale much greater than the sampling rate, may
be called predictable.
It's, of course, paradoxical that the time when most
users want to monitor systems is when they are least
predictable and providing observations of no value.
A. Separation of timescales
We can make another observation about what
happens in interactions. The principle of separation
of timescales is a design principle for interacting
systems, based on the observation that dynamical in-
fluence causes timescales for change to mix. In earlier
work, I've referred to this as the most significant
principle for engineering -- more important than the
separation of concerns based on semantic (functional)
separation, such as data normalization or 'class',
which is the norm in Computer Science. Briefly, it
says:
Principle 1 (Separation of timescales):
robustly
processes
with
systems modularize
when
Functional
effectively
characteristic timescales are weakly coupled.
By 'robust', we refer to 'stability' [34]. This principle
makes a connection to the related problem of data
consensus, which is a strong coupling regime that
maintains data consistency over average timescales.
and
different
B. Dynamical coupling defined
The foregoing assertions can be justified by look-
ing at what coupling strength means for interacting
8
agents6. Phenomena that promise changes on very
different
timescales interact only weakly and can
therefore be treated as logically separate. By contrast,
agents that promise couplings on the same timescale
may influence one another and therefore belong to the
same class of phenomena. In terms of the foregoing
definitions, we can state the meaning of separation
more strongly, as a theorem:
Theorem 2 (Separation of causal influence): As
the ratio of timescales becomes large TR (cid:29) TS, the
effective coupling tends to zero
e → 0.
(28)
tends to zero (weak coupling).
To prove this, suppose a series of partially ordered
events at an agent S yields a series Eγ, γ = 1, 2, . . ..
Suppose a source agent S transmits the events, which
are aggregated into superagents E(n)(Eγ) of dimen-
sion n, by the receiver agent R,
S
R
−−−−−−−−−→ R
+E1,E2,...En
−E1,E2,...En
−−−−−−−−−→ S
+αEE1,E2,...En
−−−−−−−−−−−→ A?
(29)
(30)
(31)
so that the dimension of the information is reduced
by a factor of n by R:
R
α(E1, E2, . . . En) → R
(cid:12)(cid:12)(cid:12)E1, E2, . . . En
(cid:12)(cid:12)(cid:12)αE (E1, E2, . . . En)
(cid:12)(cid:12)(cid:12) = n
(cid:12)(cid:12)(cid:12) = 1
(32)
(33)
(34)
The average time between events, as assessed by R's
clock, may be denoted
TS (cid:39) 1
TR (cid:39) n.
(35)
(36)
So R assesses S's timescale to be 1 and its own
timescale to be n:
TR ≥ TS.
(37)
Thus the average interarrival times for the queue in
(31) λR ∼ 1/TR, etc, satisfy:
λR ≤ λS
(38)
and the effective influence, in fraction of messages
received compared to messages sent is expressed by
a coupling constant:
e ∼ λR
λS
≤ 1.
(39)
In a strongly coupled system e → 1, timescales
converge to the shortest timescale of the interacting
parts, making systems busier and more work intensive.
The utility of this observation is that, if one separates
causally independent parts of a system into super-
agents that make weaker promises to one another,
6The increasingly pervasive language of Complex Adaptive Sys-
tems leads to assertions about strong and weak coupling, but we
need to be able to define those things to use them.
9
that
any observed correlation between phenomena,
exceeds expectation, can be considered coincidental
or potentially faulty. This can be detected by a change
in the proper time event rate, measured by some agent
within a system. This principle therefore has signif-
icance to the use of observation for detecting faults
and design flaws in systems. It tends to maximize
the signal to noise ratio between promised and non-
promised behaviour [34].
C. Coupling strength, memory, and consensus
The concept of knowledge is already more uncer-
tain in a distributed system than in a local system
with random processes. Lamport's papers about seek-
ing the homogeneity of data sources is effectively a
monitoring problem in reverse. A collection of agents
monitors one or more sources and tries to equilibrate
the knowledge they promise. Data consensus is a con-
ditional promise of policy-determined values (called
quora), based on inputs reported from sources Si:
Si
R
R
+Eγ−−−→
−Eγ−−−→
R
Si
+Quorum(Eγ ) Eγ
−−−−−−−−−−−→ A?.
(40)
(41)
(42)
This strong coupling, represented by strong depen-
dency on data from a complete network of depen-
dencies, demonstrated that time and order of events
are fundamental obstacles in a system of distributed
computers in which observation has a finite latency
(usually agents that are spatially separated) [16], [17],
[32]. The topic touches on the relativity of simultane-
ity, and how to make sense of differing views about
what causes what.
The relationship with time is revealed by the 'FLP
result' [35], which exposed the essential impossibil-
ity of consistent distributed knowledge in an uncer-
tain 'asynchronous' environment. In an asynchronous
message-passing system, source or delivery agents
may delay messages indefinitely, duplicate them, or
deliver them out of order. In other words, there is no
fixed upper bound on how long a message will take
to be received. A consensus policy promises:
•
•
All trusted nodes promise the same result (a
non-local agreement).
All trusted nodes will eventually promise a
result.
Some approaches to working around the limitations of
asynchronicity play with strong synchrony promises in
order to eliminate these uncertainties [33].
In an asynchronous
interaction, each agent's
proper time may be used to define 'timeouts' to receiv-
ing data to keep a process from waiting for ever for
strong dependencies. Timeouts are a workaround that
weakens the effective coupling strength of an interac-
tion, by effectively measuring latency in interior time.
There is no unambiguous meaning to a timeout, except
the presence of a potential fault. Latency (round trip
time) is the only covariant measure in a relativistic
system, because it's one of the few measures that has
a purely local meaning.
In Promise Theory, the intermediate agent theorem
is the analogue of that result: it says that whenever
you rely on agents that are not yourself, to acquire
or deliver information, it no longer promises what its
originator intended. And if a remote agent promises
something, but doesn't promise it to you too, all bets
are off and there is a quadratically growing cost of
verifying. In monitoring, we are not usually interested
in a majority view, rather we are interested in what
happened specifically at what we believe was the
certain place and time of origin (though this is also
subject to uncertainty). We are sometimes interested
in a statistical view (which is not
the same as a
consensus view, because it admits and even measures
the statistical uncertainty of variations. If certain nodes
lie (sometimes called Byzantine behaviour) we want
to know about it, not merely cover it up.
There is clearly overlap in the concepts of dis-
tributed information, but monitoring seeks a picture
of actuality, rather than a cover-up operation to brush
uncertainty under the rug of consensus. Software
Engineering therefore has a conflict of intent with
monitoring: it wants to assure complete dependability
(promises always kept) by invoking protocols 'in
band', whereas monitoring is trying to expose when
promises are not kept, to 'out of band' human ob-
servers. These are the issues we need to deal with in
describing observability.
Readers may feel that the problem of distributed
ordering has been solved by distributed consensus
systems like Paxos and followers [16], [17], but this is
not the case. Consensus systems do not promise the
source order of observations, but rather an average
order by which observations are reported, which is a
policy decision.
D. Memory processes
It should be clear that memory is required to
stabilize values from multiple sources. To integrate
results from several sources, and to replace then with
an agreed result requires temporary memory, over the
necessary clock ticks of proper time -- at least as much
memory as there are source dependencies for each
outcome. The role of memory and its reliability also
play a role, but I don't want to discuss that here. In
most IT systems, memory unreliability is negligible.
VII. THE TIME-SERIES MODEL
Given a stream of trusted values reported by agent
interactions, the usual response is to try to build a
timeline for a system as a movie recording of past
history, using a panoramic lens. This throws us a
number of questions about how often one should
sample data.
A. Sampling rate
The naive view in the industry is that one should
collect as often as possible. Basic information theory
constrains our ability to extract information from data.
Many engineers feel that the virtues of fast sampling
are indisputable, just as the citing of many decimal
places leads to increased accuracy, but neither are true
(for the same reason). Excessive use of high resolution
sampling is a senseless arms race (a watched pot that
never boils). Continuous high density sampling of a
non-existent signal is not helpful. Nyquist's theorem
tells us that we can only know fully of changes that
occur half as fast as the rate at which we sample.
Shannon's theorem tells us that our information about
the system only increases when changes are observed.
Regarding the order of events, it's common to
rely on an independent clock service, located within
a network to try to synchronize clocks to some
calibrated count, and then rely on the homogeneity
of manufacturing in chip-sets that count time at a
more or less similar rate. Traditional clocks services
count time in seconds, but this sampling rate is much
too infrequent
to distinguish processes in modern
processes, where nanosecond timing discriminations
are becoming.
Physics tells us that relying on the counting of
an exterior agent is futile when clocks are located in
regions of very different gravitation, or when they are
moving with respect to one another. This already has
to be corrected for in satellite systems. The same effect
applies if agents are in virtual motion with respect to
one another. Only the interior proper time of a process
can be relied upon for comparisons.
Local measures of observables have to be aggre-
gated into coarse grains in order to measure them
against one another. Histograms of observational dis-
tributions are usually the best we can offer in terms
of observability. But distributions only tells the past
probability of behaviours in fairly static cases. When
we most want to know about a system, that's when
it's hardest to understand. 7
Time-like changes are normally assumed to be in-
stances of what may potentially be significant events.
The result is that human operators get excited by
graphical traces that suddenly rise or fall -- which has
an undoubtedly hypnotic appeal, but means nothing
without a larger context. Sliding windows are often
used to detect gradient changes in time series. En-
semble averages are used and even forced in data
distribution and consensus processes (see figure 4).
In other words, aggregation over time (not space) is
a necessary part of the learning that provides context
for prediction.
Principle 2 (The sampling rate): The
sampling
rate for a variable should typically be about half the
auto-correlation time for a variables in order to detect
meaningful stable variation.
This is the timescale suitable for learning. For the
purpose of anomaly detection, one might see a sudden
change in timescale as a result of an unexpected
7This is the paradox of weather forecasting: when nothing is
changing, we can predict the weather easily; but, when everything
is in flux, prediction is hopeless. [36].
10
coupling. Faster sampling could then be introduced
on suspicion of a transition in behaviour -- just as bi-
ological heart rates and attention spans quicken under
stress. Recording and storing reams of data that are
zero or constant cannot be in anyone's interest. Such
data is compressible. It contains no new information.
The potential problem with that approach is that the
cost of sampling is not free; the impact of sampling
on the system may become significant. Some authors
have advocated such adaptive sampling [37]. One then
has the decision about which part of the sampling
process to scale back: the act of measuring on each
local process has one cost, the act of aggregating the
samples in some central repository has another cost.
Neither of these is easily controllable, since multi-
tasking operating systems make the sharing decisions
to allocate cores, interrupts, network transmissions,
and tasks quite opaquely. It may be difficult to assess
which is the greater evil: uncertainty due to adaptive
sampling or uncertainty due to oversampling or un-
dersampling.
Fig. 4: Aggregation of observations from multiple sources
can happen at any node in a distributed process. When causal
influences come together, in this way, the confluence point
becomes an effective observer of the sources that feed into
it. Observers are not only human!
B. Shared resource counters (kernel metrics)
The consequence of lemma 6 is that scaling of ob-
servations causes not only a reduction of information
transmission rate, but possibly a loss of information
about the origins of shared assessments. Resources
counters, computed from the aggregation of data (like
most of the kernel resource metrics typically recorded
in popular monitoring tools), erase details that belong
to their higher resolution origins, in an unrecoverable
way (see section VIII). There are many such recorded
values in timesharing computer systems, because they
are useful mainly to the sharing agent. The fact that
they are shared with separate processes is a mixed and
slightly misleading blessing. It's sad to see so much
effort expended in sharing noise to observers desperate
for insight. I think we can do better.
Shared measures erase the information about data
origin, and thus such collective phenomena cannot be
traced backwards to an appointed cause. For example,
measuring the load average for a computer cannot
determine which programs caused a spike in the load
[38]. This is not, on the other hand, a reason to retain
every individual data characteristic for ever, because
the opposite is also true: some data cannot be exposed
without computing those high level functions.
11
C. Instrumentation of processes
Our symbolic representations of processes are al-
gorithms expressed as program 'code'. In a distributed
system, programmers have begun to instrument work-
loads for interprocess communication using service
meshes and centralized logging services. These do not
reveal behaviours interior to the processes, but may
provide traffic patterns by which to hypothesize about
process behaviours and intentions.
For process tracing, one needs to be in the inside
of the process, where the interior time ticks. From the
foregoing discussion, it's clear that the system clock
service is not the correct measure of time for process
diagnostics. The proper time of a process is measured
by the number of program counter transitions or
program steps incurred by its execution. This count
is significant because each increment
is the result
of an explicit causal jump instruction. The program
counter's value can be promised at each instruction
in code for debuggers to trace, forming a causal set
of correspondences between program locations and
increments. However, attempting to share these proper
times between processes is meaningless.
The system UTC clock, even an approximate local
copy of it, comes from an independent agent, and al-
though shared between many processes its increments
do not correspond to single channels of causation.
Rather they represent only the implicit increase of en-
tropy in collecting from all processes. When processes
are timeshared, they may be halted and interleaved in
complex ways.
D. Significant Events
As discussed above, the major drawback in time-
series thinking, for a distributed system, is that there
is no unique meaning to the order of transactions
originating from different sources when they are ag-
gregated from different locations. Each observer in
the universe sees events from their own perspective.
The lightning bolt and the thunder arrive earlier for
some than for others, because different processes
propagate information at different rates, over different
routes, and with different latent delays. Consensus is
expensive, heavy handed, and its goals are different
to the goals of observation.
Moreover, we have no uniform metric for time
other than the exterior clock, which has issues of its
own -- it doesn't represent local causation except for
the process that generates it. The question we want
to answer is: what was the reason for an event E?
i.e. can I infer the condition, state, or quality of the
system from this event, based on what I have observed
beforehand? Times are not causes.
A better approach, based in interior instrumen-
tation is to create a semantic counter that traces a
distributed process that we can trace backwards to
causal origins. Samples can be taken when something
is found to be significant within the context of the
process itself. This is the proper passage of signifi-
cant times. Regular sampling of processes is not an
SRSCCCS efficient way to record them because processes may
be busy or idle, etc. This is what system logging
enables -- but the opportunity is usually squandered
from a lack of a proper model.
Definition 10 (Significant event): An
Metric coordinates (clock times and numbered
locations) are not helpful when we have no invariant
measuring devices to define them by. The alternative
is to use descriptive labels, or semantic coordinates.
event
marker, provided by a source process, that signals
either an intentional change or an unintended
deviation from expected state.
Anomalies and faults fall into this. Processes that are
not able to keep their promises may also be significant
events.
When seeking the 'root cause' of an event, we
to go back to prior events that were
really want
significant. The rationale behind this is that,
in a
stable system (one that we expect to be predictable),
when all is as expected, unexpected events are most
likely to be caused by previous significant changes
and anomalies. Clearly, for every event there is a
prior one, until we reach the very beginning of time
(the system 'Big Bang'). However, we also have
episodic boundary conditions that act as 'Little Bangs'
for more constrained universes. Boundary conditions
are semantically special events that we attach special
significance to -- they are the origins of causation. We
aren't interested in every intermediate change, only in
prior events that make a splash.
We want to create a causal chain, a journal, some-
thing like a linked list. Instead of selecting data by
voting, we can select values based on their perceived
importance to outcomes of interest. This shifts the
focus of policy from the intermediate aggregator of
data to the observer: the observer is now expected to
have a specific question it wants answered, rather than
voyeuristically consuming data for entertainment.
E. From metric to semantic significance
The use of 'signposts' for labelling paths, as
semantic coordinates [39], traces back before maps
and calendars were invented by human civilizations.
Instead of imagining idealized coordinate grid sys-
tems, perhaps unmeasurable, signposts relate events
to less regular but highly recognizable things that we
can observe (the big tree at the river, Mount Fuji, the
Matterhorn, the year of the flood, the eclipse of the
moon, etc). This provides anchors for accessing mem-
ory and plausible anomalies that might have exerted
causal influence. We use these events as boundary
conditions on episodic sub-processes.
Descriptive naming (semantic coordinate assign-
ment) is thus more useful than ordinal naming (nu-
merical coordinates). The checkpoints and paths that
participate in processes may not be invariants, and
the numerical value of coordinates is irrelevant8. We
8One works hard to make this point in the physics of relativity
where coordinate systems get in our way of understanding space-
time processes
may need to identify a repeated pattern to some
degree of approximation in order to exemplify a
general concept from which lasting knowledge can be
derived. Anomalies that do not recur become effective
invariants,
in memory, because they are rare and
worth remembering. Featureless invariants (like empty
space) are indeed the most invariant of all, but have
such high entropy as to contain no information of
significance.
The significance or meaning of a signal is the
heuristic inverse of the (incompressible) information
within it. The more information we need to charac-
terize a room, the less stands out about it. If there is
one part that dominates, the rest is negligible -- hence
the principle of signalling significance.
Lemma 7 (Significance vs information):
Maximum entropy distributions contain no significant
events:
they are causally random, and all events
are observationally equivalent. Minimum entropy
distributions have the maximum significance, as they
imply strong correlation.
Entropy plays a subtle role in statistical distributions,
and therefore in ability to infer meaning from data.
F. Reversibility versus traceability
We want to be able to trace our knowledge of
a system back to know the cause of an effect. The
intended outcome is programmed into it, but there
are also unintended outcomes caused by environment
leaking into causal pathways. Because of the culture
of 'rollback' thinking in IT, which originates from
database transaction semantics, IT often muddles the
concept of traceability with reversibility9.
We must distinguish between the ability for an
observer to trace backwards from a sequence of obser-
vations to reconstruct the cause of a significant event,
and the ability to roll the state of a system back to
what it once was. For example, it's possible for an
observer to trace the source(s) of a river, but it is not
possible to reverse the river and roll it back to an
earlier state.
lost
In the former case, the enabling condition is for
in the chain of
no origin information to get
unfolding events (see figure 5). In the latter case, the
necessary condition for being able to undo a causal
sequence is that agents of the system itself have to
promise the inverse of every promise in the forward
direction conditionally on an undo condition -- this
is an additional set of promises pointing backwards
along the path, which is much more than an observer
being able to trace knowledge of promises backwards.
The necessary condition is insufficient even to promise
the result: agent processes must also be isolated from
external interference, else the precise inverse opera-
tions may be deflected off course by noise [24].
9Reversibility is something of a misunderstood concept in dy-
namics, especially as applied to IT system behaviours. The apparent
reversibility of the machinery is sometimes used as an argument
against causality, but the argument is built on a misunderstanding
that ignores the boundary conditions.
12
G. Partial order of agents and events
As a process propagates by passing messages, the
messages separate earlier times from later times on
the process's own clock or counter. Suppose this is
reflected in a sequence of messages, or lines in a log.
In a chain of lines Lγ belonging to a single source
Si, we can define a countable metric distance between
lines by the total ordering of the sequence, also in the
language of promises:
L1
+precedes
−−−−−→ L2
+precedes
−−−−−→ . . . Ln
−−−−−→ Ln+1. (43)
The observational binding is equivalent a more clas-
sical ordering relation <:
+precedes
L1 < L2 < . . . Ln < Ln+1.
where
S < R ↔
S
R
−−−−−→ R
+precedes
−precedes
−−−−−→ S
(44)
(45)
We need to be cautious about the interpretation of
these promises. Each agent
is making a separate
promise, but (by the law of agent autonomy) these
agents cannot make the assessment or promise it to
an observer who happens to be watching all of them.
Each agent can make its own promise available to the
observer, but it's up to the observer to order them
in the final
instance. This ordering may be come
mixed with other orderings as data are aggregated.
The promises indicate that the relationships are con-
Fig. 5: Causally ordered change in a process, and informa-
tion observed about the process are two distinct things. As
long as the observation of the process retains the order of the
process, inferences about causality can be made, regardless
of whether the system itself could be reversed. You can
trace the source of the Nile, but you can't make the river
flow backwards.
sidered persistent or even invariant by the promisers --
not merely local assessments made on the basis of
spurious data. However, if at any time the events being
ordered become indistinguishable, they can no longer
be ordered. This can happen when data are aggregated
without complete labelling.
Ordering reduces to the existence of (local) con-
ditional promises, whose scope may extend to other
observers in a scope σ [18]. The n-th agent in a
13
(cid:88)
sequence; by the axioms of Promise Theory, we must
have a chain of the form (figure 5):
An
An
An+1
+XnXn−1
−−−−−−−→ An+1
−Xn−1
−−−−−→ An+1
−Xn−−−→
An
(46)
(47)
(48)
(49)
And, in general, we may consider a general scope for
the above promises:
An
+XnXn−1
−−−−−−−→
+σ
An+1
(50)
Given such a set of promises, we can define a measure
of observable distance between agents An and Am
by assessment. Again, we note that interior relativity
makes distance an assessment by one agent about the
relationship between itself and one or two others.
H. Translation operator and Noether's theorem
If the agents were sufficiently homogeneous, we
could consider an operator interpretation for ∆, as the
generator of a translation on a set of states realized
by the positions (like ladder operations):
∆Ai(cid:105) → Ai+1(cid:105).
(51)
In fact, for every kind of promise there would be a
separate propagator, like a complete basis:
∆ =
cτ ∆τ .
(52)
τ
The problem with this kind of interpretation is that
is suggests the existence of a god's eye view once
again. It takes the existence of a privileged observer
to be able to order and rank the states in this way.
In classical physics, the continuity of the energy
function with respect to spacetime is what generates
conserved quantities like energy and momentum, thus
allowing these quantities to be used consistently as
counters for behavioural descriptions. We can see,
from the Promise Theory, that this conclusion also
follows from a privileged god's eye view of spacetime
locations.
This tells us that it is the assumption of continuity
by the observer that rationalizes the use of counting
metrics, including jumps and changes in metric be-
haviour. If source observability does not reveal dis-
continuities in the assumptions amongst independent
sources, the observer will not be able to discern that
information merely by monitoring.
If we assume the conservation of b as an axiom,
the ordering of b-influence must follow paths auto-
matically, even when the agents make unsynchronized
(asynchronous) promises, like a first order Markov
process. In order to explain conservation and causal
order over non-local regions, we need to extend the
promises to be conditional on non-local neighbouring
patches. Ordering information itself needs to propa-
gate.
observerAA2A3A4X X01X X12X X231 I. Causality
The principle of causality can be stated simply
by saying that earlier events are followed by later
events, at a given point of action, as a result of the
transmission of some information that we may call an
influence.
Definition 11 (Causality): A causal graph is a
complete graph of conditional promises. We say that
S causes influences X at R with cause c iff:
+XS c
−−−−−→ R
−XR−−−→ S,
S
R
(53)
(54)
and XS ⊂ XR.
If every transfer of influence in a system obeys
this property, then we can define the system to be
reversible:
J. Traceability (inference)
Lemma 8 (Traceability): If an observer has com-
plete information about promise causality, a process
graph may be called reversible, i.e. for every pair
S
R
+XS c
−−−−−→ R
−XR−−−→ S,
(55)
(56)
provided XS ⊆ XR. We can infer origin by using
complementarity to interpret a reversal of causal trac-
ing: X R = c and c = XS, such that
+X S c
−−−−−→ S
−X R−−−→ R.
(57)
(58)
R
S
If a set of agents An precedes another set An+1
by a promise
An
An+1
+XnXn−1
−−−−−−−→ An+1
−Xn−−−→
An.
(59)
(60)
Traceability requires that O be in the scope of this
chain, and that it assumes reversible semantics for
X, as 'is followed by' (which is automatically inter-
pretable as 'follows').
K. Reversibility (causation)
This may not point out a unique 'root' cause, but
it will point out the causal sets that act as source
(spacelike hypersurfaces) of the process.
Lemma 9 (Reversibility): requires, much stronger:
•
•
A complete chain of prior origin data be
available across the graph.
There should be no acausal
loops in the
process, else there may be branch alternatives
(eigenstates) or divergent unstable behaviour.
Example 6 (Service lookup thunder and lightning):
Consider the order of a process (figure 6) described
in the following promises:
S
−dns−−−→
+dns−−−→
A
+relay(dns)dns
−−−−−−−−→ R
−relay(dns)
−−−−−−→ A
A
S
A
R
(63)
(64)
(65)
(66)
Fig. 6: Causal order may be different from clock time. It is
generated by prerequisite dependencies: either by underlying
topology or by constraint. Agents can only trust directly
agents that they are in scope of (in practice, their direct
neighbours), as they have no calibrated information about
the promises of agents.
Agent A promises to listen for a DNS lookup
(a query, i.e. an invitation to reply). As long as this
promise exists, it can be considered to be polling S
for a response. S promises to provide DNS data, but it
hasn't specified what or when. If there is a fortuitous
match between the two, data will be passed from S
to A. A, in turn, promises R to pass on the data it
receives from A. It does a better job of promising
conditionally, so it will only pass on fresh data from
S when a reply is received, because the promise in
(65) is conditional. R, in turn, promises to accept data,
which can only happen after they arrive. The effect in
each interaction is to order data, but we don't know,
from these promises, how many times data get passed
between A and S, nor do we know how much latency
is experienced by any of the agents.
The conditional promises (dependency) represent
causal ordering. We can't say anything about the rela-
tive order of promise keeping unless it is constrained
in some fashion. Often we rely on incidental or ad hoc
serialization at a single observation point (a queue) to
define the ticks of our process clock. The problem is
that this serialization does not represent an invariant
of the process, so it's unreliable.
An
An−1
+Inv(Xn)Inv(Xn−1)
−−−−−−−−−−−−−→ An−1
An.
−Inv(Xn)
−−−−−−→
(61)
(62)
This holds for any agent (or superagent) An.
The condition for forensic back-tracing of a system
state (detection of cause) is that
L. Metric distances
Order is important, when it can be distinguished
because it allows us
to measure intervals. We
sometimes use intervals as significant measurements,
though Einstein pointed out
intervals are not
invariants, they are only 'covariant', changing with
the system of measures we establish.
that
14
−DNSASR−dns+dns+DNS dns Principle 3 (Distance semantics): Distance is an
assessment made by an observer with two comple-
mentary interpretations: distance suggests what might
lie in between the bounds of the interval, or it suggests
a measure of how similar two agents are, with respect
to location in some criterion 'space'.
The distance between two events Lγ and Lγ+β is
related to the ability of an observer to trace and count
the number of similar events in between. The distance
between events in a journal may contain implicit
information about what happens in between, but it
is not a substitute for the information itself. Metric
distance is therefore a counter that pays just enough
attention to agent properties to discriminate between
them on the basis of label, and be able to count, but not
necessarily enough to classify agents meaningfully. If
events follow on as nearest neighbours this tells us
something; if the same pattern is suddenly interleaved
by more lines this could be an indication of an
anomaly.
A histogram is a classification of multiple events
that get counted and form a distribution. The order of
the classes may (or may not) express a metric policy
about how near or far events are when they fall into
one of the classes. There is no a priori order to these
classes, but there might be a distance. It's therefore
an assessment policy of an observer to ensure proper
classification according to a model presumed by the
observer.
The proliferation of logs in IT systems means that
they get receive disproportionate focus, in the hope of
extracting far more than they are usually capable of
representing. The variety and standard of logging is
very poor indeed, in my view. What happens to order
and distance relationships in logs after aggregation?
There are many tools that imply log aggregation is a
good way to bring together all logs into one location,
but there is little discussion around the significance or
usefulness of the result [40] -- [42].
Aggregation of agents Lγ into superagents {Lγ}
may preserve or discard the order and interval dis-
tances between lines. Sometimes, data are not
in-
tentionally numbered by the sender and order is
assumed by the order to transmission (e.g. in UDP
transmissions). In that case, message packets may
become reordered by network redirection, or loss.
Some messages could be also be lost. Let's refer to
the cases by the common terminology
•
•
Reliable: promises all packets delivered in
order.
Unreliable: ad hoc, no promises about order
or loss.
In either case the latency between transmission and
final arrival is uncertain. Consensus of data is easy,
because the data are point to point and there is only a
single source and a single receiver for each message.
One way of trying to work around the law of
intermediate agents is to build up the notion of en-
tanglement between processes [33]. This takes sev-
eral cycles of mutual interaction between a pair of
15
agents, on some scale, as well as a small cache of
local memory. Entanglement can transform partially
reliable transmission of influence into fully reliable
transmission at the expense of some added sub-cycles
in the interaction.
The aggregation of messages without reference
to the agent and the interior timeline that generated
them implies that causal origins can never be traced
backwards. Timestamps have no value, because they
are unrelated to the process causation. We can thus
show that a log may preserve the reverse tracing of
causal history, but does not imply reversibility of state.
We can trace a story back to where events played
significant roles in the timelines of processes, but
we can't necessarily reconstruct the states of those
processes.
VIII. AGGREGATION OF SOURCE DATA
It's time to look more carefully about what aggre-
gation of data means. In the context of causality [40],
it matters both where and when signals come together,
and to what degree information is lost by mixing.
Distinguishability plays, again, a central role here. For
example, the commonly used metrics of load average,
CPU percentage, memory usage etc. The behaviour of
a process depends on the behaviour of the platform,
which in turn depends on the behaviours of the guest
processes [38]. These are measures of different scales,
since a platform is an aggregation of processes, and
so on.
When unexpected behaviour (signpost behaviour)
is observed in an aggregate variable, the culprit may
not even be in the same process. The relevant question
may seem to be: can we obtain information about
which process may have been responsible? But that is
not the right question, because it could be the accumu-
lation of many processes leading to an exhaustion of
resources which actually impacts the process we are
monitoring -- by undermining its critical dependencies.
When this is the case, we might be more interested
in why scheduling policies resulted in such a conflu-
ence of demand. Obviously, there are many layers of
decision behind such stress concentrations.
The causal connection between these cannot be
inferred with any certainty from quantitative mea-
surement however. One would rather expect to see
a process log fail to allocate memory from within
the privileged context of the process itself. Today, we
wrap processes in containers that are quite opaque.
In fact, processes are equally opaque when viewed
through kernel metrics, because there is unfortunately
little or no causal connection between the changes and
any particular process of interest.
A. Sampling resolution (timescales again)
We need to know when data belong together and
when they should be considered separate. For any col-
lector, this is a policy decision, but it can be informed
by the physics of the system. There is information
in the order, content, and volume of data. If that
information is squandered, it may be unrecoverable.
B. Erosion of metric significance
Experiments show that there is little correlation
between commonly collected quantitative metrics and
actual process semantics [43]. This is a historical
artifact that comes from the fact that observables were
designed for timesharing, not for process monitoring.
Measuring kernel metrics is something analogous to
watching the weather to plan for a crop. In some
cases, a change in a distant place may may trigger
outcomes that result from arbitrary choices in code
elsewhere. For example, if one sets an arbitrary thresh-
old for a value, in a conditional statement, the unusual
process weather originated elsewhere may push the
conditional over the limit unexpectedly and lead to
a discontinuous branch of behaviour. This is why
understanding relativity is so important in reasoning,
and why cloud computing is especially susceptible to
relativistic effects.
Entropy of mixing does not usually increase re-
lentlessly in IT systems, because new information is
being added in the form of semantic labels (boundary
conditions) all the time, e.g. when a particular set of
images is identified as belonging to the same person,
we name the set as the person; or when a sequence of
command instructions leads to the same failure mode,
we name the histories with the name of the failure
mode. This new information adds context.
As data scale, some information is lost, and new
information is added. Aggregating data and integrating
over time, throwing away time information, but build-
ing maps of invariant relationships. The map of what
remains distinguishable grows as more data are added,
because the number of possible storylines grows as
new invariants are added.
Lemma 10 (New data at all scales): Origin data
are lost by coarse graining, but combinatoric selec-
tions of aggregates leads to a new degree of freedom:
distinguishable routes or paths through the composite
variables.
Each story has its own semantics: the loss of event in-
dexing leads to the addition of fewer semantically sta-
ble storylines. Distances that require distinct labelling
become meaningless, but new emergent distinctions
lead to new possibilities for classification. If sufficient
information is retained to point backwards along to
causal signposts, specific paths can be traced without
muddying the global picture.
C. Learning and coarse graining defined
We can track the different scales of a system
by seeking to separate invariants from microscopic
local changes, using the principle of separation of
timescales. Learning over sequences (in a timelike
direction) effectively form Bayesian processes that can
act as aggregate state discriminators [44], [45].
Example 7 (Learning): The collection of data to
train a statistical algorithm may take weeks or months,
and involve large amounts of data that are compressed
into a composite form, irreversibly. The composite ag-
gregate is used as part of an algorithm that recognizes
images on a timescale of seconds. These processes
can be naturally decoupled.
The degree of separation of timescales corresponds
to what we call supervised learning (highly separated
timescales for learning and using) and unsupervised
learning (where timescales for learning and using are
approximately the same).
Even a single unique episode may eventually be-
come viewed as an invariant if it is not repeated and
hence is never challenged, but its significance may be
limited. We can only know this by learning over time.
The significance of concepts grows by the frequency
by which they become repeated. Thus learning and
garbage collection of insignificant concepts is needed
to prevent all the information from becoming noise.
For full episodic reconstruction, the invariant con-
nections that generate process stories need to integrate
with one another, like a linked list, using the a map
of invariant concepts as glue. The invariants represent
the aspects of processes that are not specific to a
single source. See the earlier work on characterizing
spacetime semantics [20] -- [23], based on earlier ex-
periments [46].
All causality is a representation of Shannon's basic
model of an information channel. The distinguisha-
bility of information is the key to following and
tracing processes, but where does the significance
of information lie? The significance of information
(associated with labels to it) is necessarily diluted by
scale, or the entropy of signal aggregation.
D. The Mashed Potato Theorem
Mixing of signals leads to loss of traceability.
Suppose you are at a restaurant and you receive some
mashed vegetables. You first assume that it's potato,
because that is common, but something doesn't taste
quite right. There are some other vegetables mixed in.
Closer inspection reveals some orangy colour (perhaps
carrots, or sweet potato, etc). How could you know
what was in the potato without accurate knowledge
from the source?
The loss of distinguishability (entropy of mixing)
tells us that we can't easily discern the content of
mashed potato without the recipe because we cannot
separate (classify) the parts of the signal.
Theorem 3 (Loss of distinguishability): Let Σ be
an alphabet of class categories that are distinguishable
by a set of source agents Si. Data aggregated from Si
without complete causal labelling σ ∈ Σ, from the
source cannot be separated into its original categories
Cσ with certainty, i.e. the promise
+Cσdata∩
−−−−−−→ A?
πcateg : R
(67)
is kept with equal probability for all Cσ.
To see this, we note that
the aggregation of data
16
involves promises:
=
+dataS−−−−→ R
−dataR−−−−→ S
dataS ∩ dataR.
πdata : S
πlisten : R
data∩
(68)
(69)
(70)
and we take data to be a collection of line signals
data ∼ {Lγ}. In order for the conditional promise
(67) to be causal, the promise of data in (68) have
to be a reversible function of the Cσ. But if data are
indistinguishable, then the Cσ must also be indistin-
guishable, thus
Cσ = Cτ ,
(71)
thus, the probability of discerning a signal Cσ, pσ ≡
P (Cσ) = pσ, and the entropy
∀σ, τ
pσ log pσ → max(Sent).
(72)
Sent = −(cid:88)
σ
If one rescales all the Cσ = Cτ into a single category
to indicate that all such categories are the same,
then the entropy is zero, Sent = 0, indicating that
the information per transmission, in the mixture, is
actually trivial. Thus, we must keep all labels from dif-
ferent sources and classifiers in order to retain useful
information. This does not depend on the amount of
data (or mashed potato). Moreover, if data are passed
on, the dependence of a reconstruction by (machine)
learning is unstable to the datasets10. The definition
of entropy in computer processes has been examined
recently to address its semantics [48].
E. Separation of concerns
The practical question remains: how should one
separate variant and invariant data when designing
systems? This depends partly on the structure of
intent and observations. Programmers are trained to
recognize what values are variable and abstract them
into parameters to invariant functions.
If we think about how we formulate stories, as hu-
mans, we embed variable fragments of causal history
(episodes) into larger assemblies of more invariant
concepts, which provide context for reasoning. This
is how experiences are organized around conceptual
models. I'll come back to this in section X on model
extraction.
Example 8 (Logging text compression): As
a
simple example, consider the generation of a log
message from a typical format string in code:
•
•
A separate format string is an invariant class
of messages. It can be replaced by a single
numerical value and looked up in a hash table
to compress data.
Standard data format can record format string
and variables in an indexed structure with
named members.
• Message significance level or priority (policy)
- imposed + or -?
10I won't consider this here, but see the signals lemma in [47].
17
•
Variable Substitutions in the format string are
variants with respect to the message. Some
of these may be invariants too (the name of
a host or function), while other data have no
long term significance (the time or date).
If we compare these points to a Unix syslog message,
the glog library, and many more examples, it's clear
that syslog satisfies none of these promises. Lines of
text are basically random.
F. Retaining semantic context for events
Concepts are the result of dimensional reduction
over contextual learning sets Ci from a number of
sources Si [22]. In invariant cases, the context can be
learnt by accumulation of evidence over time, because
it doesn't change. In general, significance may be
assessed based on a number of contextual sets Ci,
so when an alert message M is reported to an agent
R, this is in fact a conditional promise that depends
on the context:
+MC1,...,Cσ ∀i
−−−−−−−−−−−→ R
S
(73)
This means, by the conditional promise law, that the
promises supplying this context to S:
S
Ai
S
+Ci−−−→ S
−Ci−−−→ Ai,
−Ci−−−→ R,
(74)
(75)
(76)
are not available to R. The context is lost. This means
R has to trust the alert and its significance as a random
variable. This is no problem if the goal is to bring an
unrecognized condition to the attention of an operator.
However,
is to perform contextualized
reasoning on an aggregate scale, the graph of invariant
context also needs to be promised:
+Ci−−−→ R
−Ci−−−→ Ai,
(77)
(78)
and the conditional dependencies also need to be
captured:
if the goal
Ai
R
−(MCi, ∀i)
−−−−−−−−→ S
R
(79)
If we didn't apply this idempotently only to sparsely
occurring invariant concepts, the cost of the aggre-
gation would rise sharply. An expedient separation of
scales allows the context to be contained at the sources
as 'smart sensors' [22], [23].
Context can be framed and localized using names-
paces. Namespaces also provide unifying labels that
can usually be treated as invariants in information sys-
tems. Aggregation of messages, without cataloging,
indexing, or other labelling leads to ensemble entropy:
the irreversible loss of structural
information and
contextual semantics.
Transporting too much context is a questionable
idea. If the environment in which context originates
is lost, then the meaning of the context is also lost,
and the ability to reconstruct scenarios based on it
•
•
•
Same text (signal) as the same interpretation.
Information is encapsulated as transactions to
show partial order.
Every significant transaction needs to point to
its previous significant event.
These principles have been embodied in a proof of
concept implementation [49].
B. Dropping hints
To record useful events, from within the mean-
ingful context of a process, processes need an API
that constrains authors to produce information that
can be consumed later. For example, in the Koalja
history package, based on the principles in this paper,
significant events can be marked with signposts [49].
H.SignPost(&ctx,"Milestone 1...")
H.SignPost(&ctx,"Milestone 2...")
...
H.SignPost(&ctx,"Commence testing")
And these signposts can be detailed, using the four
spacetime semantic relations from [22], [23]:
H.SignPost(&ctx,"code signpost X").
Intent("open file X").
ReliesOn(H.NR("/etc/passed","file")).
FailedBecause("xxx").
PartOf(H.NR("main","coroutine"))
We shall explain this point further in a sequel [47].
X. MODEL EXTRACTION
If we pursue a concrete strategy of separating
timescales and extracting invariants from the chaff
of noisy variation, we can expect to infer causal and
conceptual relationships over long aggregate times by
learning. Learning is a process that happens across
several timescales, as noted in [22]. In modern Ma-
chine Learning parlance, we would say that acquiring
and stabilizing training data is a long term process,
while recognition and classification is a short term
process. Monitoring tries to achieve both processes as
an unsupervised in band single-scale process, so it has
to deal with the instabilities in band too.
The spacetime model in [22], [23] allows us to
define a partial ordering of semantics, represented as
agents in a virtual knowledge space. The future and
past cones are generated by the first two spacetime
semantic relations: for generalization or scope and
causal order. Ordered relationships are the most im-
portant ones because they tell the stories we seek.
Data may arrive in incidental order, for a variety of
reason that involve causal mixing. We need to extract
the intended order of system cooperation from the
incidental or unintended order of side channels that
muddle behaviour.
Fig. 7: As data get propagated farther from their initial con-
text, their original meaning is degraded, unless all context is
transported with them. Each of the rings may represent an
intermediate agent that may or may only promise to forward
data selectively or after distortion.
becomes of largely forensic interest. System designers
need to find expedient ways to compress context and
filter it: what can remain local at the source, and what
can be aggregated and assigned wider meaning? Thus
is remains a policy decision to balance the cost of
preservation against the actionable usefulness of doing
so.
IX. HISTORIES: LOGS AND JOURNALS
Now aware of the issues around sequentialism and
observability in distributed systems, we can tackle
the first two story types in section III-B. Logging of
process conditions may well be the most popular and
common approach to tracing in computer programs.
Isolated single-agent logs are simple serial queues,
or time-series databases, of varying degrees of so-
phistication, for keeping informative messages about
what transpires in a process. This is no longer true
for log aggregation unless complete causal linkage is
preserved.
A. Causal linkage
In most shared logging services, messages are
imposed by multiple process agents Si onto a queue
and are strongly ordered by a single receiver R.
Si
+L
R.
R accepts requests indiscriminately
−L−−→ Si.
R
(80)
(81)
Individual processes can voluntarily write their own
logs but this is not a common practice because the end
goal of logging, in modern practice, is to aggregate all
messages as 'big data' to be trawled.
Modern logging services,
like Prometheus etc,
provide more nuanced semantics with structured data
formats that can incorporate key-values; but
these
trust data to be useful. They are abused greatly by
programmers, who tend to dump any and all data into
a stream without regard for meaning or consequence,
in the hope of sorting it all out later. Logs need to
promise invariant causation:
18
ContextUsage New process timeline for ( myApp_name21.2.3 ) originally started as pid 17778
code signpost X
Beginning of test code
[dns lookup: 123.456.789.123]
[btw: example code]
[remarked: : look up a name]
[btw: xxx]
[coroutine: main]
MainLoop start
[function: main]
[intent: : open file X]
[file: /etc/passed]
[remarked: : Start process]
[go package: cellibrium]
root --> NOW,delta
Comment indented by subtime
Unix clock context
------------------------------------------------------------------------------------------
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:04 +0200 CEST
2019-06-03 13:40:07 +0200 CEST
2019-06-03 13:40:10 +0200 CEST
2019-06-03 13:40:13 +0200 CEST
2019-06-03 13:40:16 +0200 CEST
2019-06-03 13:40:19 +0200 CEST
2019-06-03 13:40:19 +0200 CEST
0 -->
->
1 -->
->
->
->
->
2 -->
->
->
->
->
->
3 -->
3 go>
->
->
->
->
->
->
->
3 go>
->
->
6 -->
->
7 -->
->
6 go>
->
->
->
->
->
6 go>
10 -->
1,1
1,2
2,1
2,2
2,3
2,4
2,5
3,1
3,2
3,3
3,4
3,5
3,6
4,1
5,1
4,2
5,2
5,3
5,4
5,3
5,4
4,3
6,1
6,2
6,3
7,1
7,2
8,1
8,2
9,1
9,2
9,3
9,3
9,3
9,3
10,1
11,1
The end!
Show the signposts
Run ps command
TEST1---------
[btw: Commence testing]
End of sideline concept test
[btw: Commence testing]
Starting Kubernetes deployment
[btw: Commence testing]
[remarked: : file read failed]
[remarked: : Finished ps command]
Commence testing
[remarked: : Starting kubernetes pod]
[remarked: : File drop in pipeline]
[remarked: : Querying data model]
[remarked: : Submit transformation result]
[btw: /bin/ps -eo user,pcpu,pmem,vsz,stime,etime,time,args]
[btw: Testing suite 1]
[intent: : read whole file of data]
[file: file://URI]
[system error message: open file://URI: no such file or directory]
[remarked: : Possibly anomalous CPU spike for this virtual CPU]
[anomalous CPU spike: CPU 22117.000000 > average 22115.000000]
A sideline to test some raw concept mapping
Fig. 8: An expanded view of a compressed local checkpoint log for a process. The first column is system clock timestamp
which provides approximate time of day context for relating events to human scales. The next fields use interior monotonic
counters that increment on SignPost events, even through concurrent coroutines. Each event points back to the preceding
event, to give actual causal history. Explanations of the event semantics are shown on the right.
A. Invariant sequences form explanations
In order to generate a map of invariants, we need
to not only identify them, but consider what they can
promise about one another, so that we may position
them in relative terms. In [22], it was shown that we
can plausibly define four kinds of semantic relation-
ship based on elementary spacetime considerations.
These ought to apply between pairs of agents in any
distributed system.
on the other hand, is a promise of belonging to a
common class. These two promises therefore represent
assembly versus classification of agents.
The extent to which we have the ability to localize
a causal influence is the essence of 'root cause' anal-
ysis: the ability to contain the process within a virtual
boundary which itself can make promises on a new
level. This is part of the motivation for virtualization
and containerization.
•
•
•
•
i) Contains: localization in spacetime (scope
of containment or ordering by scale)
ii) Follows: order, causation (Markov pro-
cesses or order by influence)
iii) Expresses:
tribute at any scale)
iv) Near to: measure, distance (assessment of
distance at any scale)
local distinction (scalar at-
Notice that order is distinct from distance, i.e. di-
rection and proximity are different concepts. These
concepts are not clearly distinguished in a vector
space.
Agents may promises exemplifiers, symbolic and
metric discriminators, or regional classifiers.
The relationship between a discriminator iii) and
a classifier iv) is subtle, and is easy to promise incon-
sistently. The essence of expressing an attribute is to
label the type, which can be combined with something
else to form a union of different promises (a kind
of semantic chemistry). A promise of containment,
19
B. Promising semantic maps
The four spacetime semantic relationships, de-
scribed in [22], [23] may be assigned between pairs of
concepts, originating by signal (+) or by inference (-),
entirely at the behest of an observer, and according to
the following 'selection rules':
1)
Distinguishability: Descriptive properties
that distinguish, describe, and embellish the
name of a concept are EXPRESS promise
types. These are scalar promises used to ex-
plain attributes that may form compositions
of attributes for aggregate 'hub concepts'.
For example: a banana may express the
colour yellow, ripeness, and sweetness. It
does not express fruit or Del Monte.
2) Generalization: membership in classes and
informal categories use the CONTAINS
promise type. These express subordination
to one or more umbrella concepts, and su-
perordination to instances and exemplars of
the named concept. Generalization is strictly
transitive.
ships between concepts, these assessments
may form the basis for a shared local co-
ordinate system.
The semantics of these relationships are not au-
tomatically orthogonal to one another, so we have to
maintain the incompatibility of the types by assign-
ment. The local promises are mutually incompatible,
which is to say the no two agents may promise more
than one of the three types CONTAINS, FOLLOWS,
EXPRESSES (or their inverses11.
•
•
•
EXPRESSES
CONTAINS.
CONTAINS
FOLLOWS.
FOLLOWS
EXPRESSES.
is
incompatible
with
is
is
incompatible
with
incompatible
with
The semantics are easily illustrated with an example.
The concepts blue and yellow are expressed by objects
that combine them as part of their identity: e.g. a
blue and yellow pattern, like the Swedish national
flag, and green paint may be composed of blue and
yellow paint, but neither the Swedish flag nor green
are generalizations of blue and yellow. The concept
of colour, on the other hand does not express blue or
yellow, but generalizes them as members.
The promise of proximity is slightly different:
•
NEAR is potentially compatible with any of
the above, since it is an informal assessment
of non-locality.
The assessment of proximity between agents
may seem to imply something about the or-
thogonal semantics above, but this is ambigu-
ous (see figure 12). For example, because
proximity is a type of relationship, not a
standardized metric constraint, relations may
vary in their interpretation:
A EXPRESSES 'close to B'
A EXPRESSES 'close to C'
(82)
(83)
'close to B' FOLLOWS 'close to C' (84)
Together these might suggest that A, B, C all
lie in a certain region and that there must
therefore be a category (dotted line in figure
12) that generalizes all of them. That kind of
inference is dangerous, because it is based on
coarse inference,
These selection rules can be applied in order to join
similar objects into hubs. Each observed instance
maps to a hub that can be broken down into atomic
concepts by expression. Containment promises are
generally learned on a much longer timescale (e.g.
11We must be cautious and pay attention to the Promise Theory
principle that just because one agent promises to contain another or
be followed by another, it does not imply that the agent concerned
agrees with this, and may not promise it. For example, firewalls
may create a one-way glass effect that prevents the inverse from
being implemented)
20
Fig. 9: The four kinds of promise that spacetime can
express: i) containment, ii) succession, iii) local attributes,
and iv) proximity. Although we can distinguish different
sub-types of these four, it's hypothesize that the four are
necessary and sufficient for describing observable phenom-
ena.
3)
4)
i.e.
follow-up concepts
Generalization is not as in taxonomy: a con-
cept may have any number of generaliza-
tions,
there is no unique typology to
concepts. The utility of recognition lies in
the overlapping nature of classes [50].
For example, a banana is generalized by fruit
and desserts, and has instances such as Del
Monte. It does not express these as attributes.
Dependency: promises of dependency --
prerequisite or
are
FOLLOWS type promises. They may link
concepts of any type into some meaningful
order, by any interpretation of the observer.
For example,
the beginning precedes the
end. "One" precedes "two" which precedes
"three", etc. Dependency is usually transi-
tive, but may contain loops (in feedback
cycles).
Similarity: the degree of similarity between
two concepts is an assessment that may be
promised by any observer,
to represent a
degree of similarity or closeness. This is
represented by promises of type NEAR. This
is an ad hoc assessment and should not be
taken too seriously.
In the case where several agents form an
agreement about the metric distance relation-
i)iv)iii)ii) <begin NON-LOCAL CAUSE>
(program start) --b(precedes)--> "MainLoop start"
(MainLoop start) --b(precedes)--> "Beginning of test code"
(Beginning of test code) --b(precedes)--> "code signpost X"
(code signpost X) --b(precedes)--> "Run ps command"
(code signpost X) --b(precedes)--> "TEST1---------"
(code signpost X) --b(precedes)--> "Commence testing"
(Commence testing) --b(precedes)--> "The end!"
(Commence testing) --b(precedes)--> "[remarked: : Possibly anomalous CPU spike for this CPU]"
(Commence testing) --b(precedes)--> "A sideline to test some raw concept mapping"
(Commence testing) --b(precedes)--> "Starting Kubernetes deployment"
(Commence testing) --b(precedes)--> "The end!"
(Commence testing) --b(precedes)--> "[remarked: : Possibly anomalous CPU spike for this CPU]"
(Commence testing) --b(precedes)--> "A sideline to test some raw concept mapping"
(Commence testing) --b(precedes)--> "Starting Kubernetes deployment"
(The end!) --b(precedes)--> "Show the signposts"
(A sideline to test some raw concept mapping) --b(precedes)--> "End of sideline concept test"
(The end!) --b(precedes)--> "Show the signposts"
(A sideline to test some raw concept mapping) --b(precedes)--> "End of sideline concept test"
(code signpost X) --b(may determine)--> "[dns lookup: 123.456.789.123]"
(TEST1---------) --b(may determine)--> "[file: file://URI]"
(TEST1---------) --b(may determine)--> "[file: file://URI]"
<end NON-LOCAL CAUSE>
Fig. 10: An excerpt of a map of invariants, generated by a search. Invariants are accumulated from distributed and concurrent
processes and their relationships are classified by the four spacetime types. The pathways through these relationships tell
different kinds of stories. The excerpt shown involves expansive reasoning: combining generalization and causality.
Fig. 11: The propagation cones indicate the past and future
as 'timelike' trajectories generated by the causal relation-
ship, and semantic scale or scope of meaning accumulated
in 'spacelike' directions around the average time axis in
rings of increasing generalization.
added by human expertise), and causal dependency
promises are added by processes that generate them
or observe them.
C. Storytelling from spacetime semantics
Once constructed,
the graph may be parsed to
generate stories, or automated reasoning. A reason-
ing process may be viewed as an expansive search
along alternating (+) axes (causal outcomes that are
related by generalization or exemplification by specific
instance), and tempered by elimination by relevance
criteria (-).
•
Starting from a topic of interest, we follow
promises of type FOLLOWS independently
in the forward and backward directions, to
explore the causal cone (figure 13).
21
Fig. 12: The assessment of proximity between agents may
seem to imply something about the orthogonal semantics
above, but this is ambiguous.
•
Arriving at each new concept, we follow
promises to generalize and specialize the con-
cept to find all links arising from the col-
lective generalized concept, and follow these
along different story paths. In other words,
we multiply the number of stories by concep-
tual associations that imply examples of the
same idea -- expanding the scope of meaning
without going off the rails.
Such promised relationships cannot be easily found
by in band machine learning techniques: this is an
orthogonal and complementary method, but learning
may form the basis for collapsing experience into an
arrangement of similar concepts on longer timescales
(see figure 13).
The algorithmic rules for parsing stories from the
concept graph come in several forms. A conceptual
reasoning search might be bounded at the start (re-
tarded), at the end (advanced) or at both ends (causal).
The first two are a form of brainstorming that ends
with a single concept: 'tell me all about X'. The latter
scopePast (retarded)Future (advanced)causality case asks for a specific explanation of 'Y given X'.
There is no unique path for any search, in general.
The paths most frequently trodden, i.e. have the most
frequently observed transitions, or most frequently
searched for concepts, become 'classical paths' and
may be favoured, someone analogous to a PageRank
search [51], [52].
Loops in causal relationships may be significant,
so we should detect them. Some loops may be errors
of identification, others may be cyclic reasoning (e.g.
self-consistent ideas, like eigenvalue problems).
Fig. 13: The scope of knowledge about spacelike informa-
tion is accumulated as memory from past events propagated
into a model of the present.
XI. MODELS, SHARDING, IDEMPOTENCE, AND
FORGETTING
From sampling of data at the edge of a network,
to actionable insight, there is a chain of reasoning to
monitoring that starts with observability and ends with
the deletion of irrelevant and antiquated data:
Data collection,
Stability or convergence to fixed points,
1)
2)
3) Model extraction,
4)
5)
Classification into buckets,
Controlled forgetting.
Each of these steps plays an important role. Data
collection provides the basic observability to trace sys-
tems at different scales and tell stories about them that
bring valued insights. The convergence of phenomena
to fixed points is an incredibly important principle in
dynamics, but one that receives too little attention12.
When systems fly all over the place, they are not
telling us anything significant. It's only when they
converge onto repeated patterns or stable attractors
that we can build on them as part of a reliable. Models
need to expose those differences. Today, there is a fas-
cination with using machine learning to try to expose
such fixed points, but the technique is only possible if
there is sufficient stability. When monitoring reaches
a level of maturity in IT, we will place as much value
in attending to semantics as we do in recording noise
12This is a principle that I have reiterated many times since
CFEngine [53] to stabilize and guide us towards invariant meaning.
22
today. Data that have the same semantics need not be
recorded twice. Once one has identified the invariants
of a system, these can be made idempotent. Model
identification classifies inputs into discrete alphabets.
Repeated symbols can be compressed, and if they are
repeated no harm need be done if they have fixed point
semantics.
For example, it is not a problem if we accidentally
collect the same data twice as long as they map to the
same place. Storing the same data twice is idempotent
unless we are counting frequencies. Frequency count-
ing can be made idempotent by labelling intervals.
The general principle is that we should engineer data
sources to permit convergence. Promise Theory re-
veals the mutual responsibility for information transfer
between sender and receiver.
Principle 4 (Convergent data): The safest way to
avoid data inconsistency is to design messages in such
a way that repeated messages always map to the same
location and update them without breaking a promise.
Fixed points lead to stable models, which lead
to efficient
indexing of knowledge. This helps in
scaling storage (e.g.
in sharding), and it helps in
fault tolerance. When data are recorded around proper
index points, it doesn't matter if data get delivered
multiple times (idempotence) or even out of order:
everything will find its proper place in the end. Model
extraction tells us how to compress the data into an
alphabet or catalogue of meaningful and significant
ideas, and therefore separate into buckets or shards.
Finally, perhaps the most important issue of all is
how to forget what is no longer of value. Keeping
data and even models around forever is a senseless
squandering of resources and an irresponsible and
unsustainable use of technology. One wonders how
many of the photos now being eagerly accumulated
in the cloud will be preserved in ten years' time. The
same is true of monitoring data that were collected last
week. If we don't understand the timescales, context,
and relevance of data,
then we have no business
collecting it, because it cannot tell us anything of
value. A policy for forgetting can usually be built into
a definition of context from the start, e.g. through finite
windows and sliding sets, running averages, and so on
[54].
Example 9 (Model based collection): In
CFEngine, weekly data were mapped idempotently
to a finite number of buckets marked by 5 minute
intervals
throughout a week, based on a prior
measurement survey using autocorrelations. After a
weekly period, the buckets would wrap around, like
a clockface and data would update the corresponding
image in the map. This approach was able to
promise a limited stability of expectations, as well
as automated forgetting (constant weight gradient
of
temporal history), and thus effective garbage
collection
What's remarkable is how few of these issues
actually get any attention in the literature. One hears
arguments like 'its cheap to keep all data forever' --
which smacks of the sudden realization of global
scopePast (retarded)Future (advanced)causality warming or the plastic crisis. Every time we advocate
increasing something, we need to think about
the
balancing garbage collection process.
XII. SUMMARY AND CONCLUSIONS
The collection of accurate data is not in question.
Today, there is increasing interest capturing 'digital
twin' representations of agents in the real world,
with every detail available for possible inspection. No
one could resist the idea of such knowledge, unless
it invades their privacy. The question in technology
monitoring is rather whether every detail should be
centralized and whether data can be compressed with-
out loss.
In this summary of what can be observed about
distributed systems, we see that tracing events back
to 'root cause' is an ill-defined problem, but tracing
back a significant likely cause is indeed possible with
careful labelling (especially of time). This kind of
labelling is not commonly provided in current tooling.
Assuming access to data, we might hypothesize that
a complete monitoring system would promise to:
•
•
•
•
•
•
Separate timescales.
Identify the alphabet of system invariants
Capture local histories of instances, in the
context in which they happen.
Identify significant events at different scales
and measure their invariance.
Tools for reconstructing and backtracing of
histories from local data.
Tools for generating semantic past-future
cones for causal reasoning.
Today, many IT monitoring systems transmit raw data
in large quantities to a central point for analysis,
without attempting to alphabetize the data before
transmission. In effect, by ignoring the existence of a
model (summarized by an alphabet of non-overlapping
signals) one is repeatedly sending the same model over
the network again and again, wastefully, and to no
gain. If we can classify observations at their source,
and condense them into an alphabet of signals, a vast
data compression can be accomplished for both faster
recognition and potentially greater semantic content.
That will be the subject for a sequel [47].
It should be clear that nothing about the ability
to trace systems enables full reversibility of state,
which should be considered difficult to impossible,
depending on scale [24] -- so ultimately monitoring
may be of little value. More value could be captured
by building intrinsic stability into systems in the first
place.
The elephant in the monitoring system is an essen-
tial attitude in the industry concerning the purpose of
monitoring. Systems are only sustainable, knowable,
and predictable when they seek stability -- not when
they labour under the burden of intrusive inspection.
For a lot of practitioners there is a conflict of interest
here. If we seek to measure all that is random or
unstable, by oversampling and consuming resources
wastefully, it will be neither stable nor sustainable.
Consensus protocols, for instance, promise semantic
stability, and are popular (if somewhat over-used)
in software engineering13. They draw attention to a
preoccupation with semantics in software engineering,
i.e. a desire for stable qualitative outcomes, at the ex-
pense of quantitative delay. Software engineers seem
not to trust concepts like intrinsic dynamical stability
i.e. systems that promise to converge to predictable
quantitative outcomes. Monitoring tends to treat all
software as adversarial, and we put more faith in ill-
designed monitoring than in an initial software design.
This is a paradox that will
inevitably lead to big
surprises and catastrophic events.
By focusing on essentials, there are many issues
I've not had time to mention in this paper. I hope to
return to some of these in future work.
Acknowledgements: I am especially grateful to
William Louth for discussions. I'd also like to thank
Nicolas Charles, Simon Lucy, Colm MacC´arthaigh,
and Adrian Cockcroft for helpful comments and ref-
erences.
REFERENCES
[1] P. Hoogenboom and J. Lepreau. Computer system perfor-
mance problem detection using time series models. Pro-
ceedings of the USENIX Technical Conference, (USENIX
Association: Berkeley, CA), page 15, 1993.
[2] C Hogan. Metrics for management. Proceedings of the Ninth
Systems Administration Conference (LISA IX) (USENIX As-
sociation: Berkeley, CA, page 125, 1995.
J.L. Hellerstein. An approach to selecting metrics for
detecting performance problems in information systems. Per-
formance Evaluation Review, 24:266, 1996.
[3]
[4] A. Gonzalez Prieto and R. Stadler. Adaptive distributed
ACM SIGCOMM
monitoring with accuracy objectives.
workshop on Internet Network Management (INM 06), Pisa,
Italy, 2006.
[5] R. Sekar, T. Bowen, and M. Segal. On preventing intru-
sions by process behaviour monitoring. Proceedings of the
workshop on intrusion detection and network monitoring,
USENIX, 1999.
[6] D. Dasgupta and S. Forest. An anomaly detection algorithm
inspired by the immune system. Artifical immune systems
and their applications, page 262, 1998.
[7] The open tracing initiative. opentracing.io, 2017.
[8] W. Louth. ObservabilityŁŁtraces and trees. Medium article
(@autoletics), 2019.
[9] W. Louth. ObservabilityŁŁtraces and tags. Medium article
(@autoletics), 2019.
[10] M.I. Seltzer and C. Small. Self-monitoring and self-adapting
operating systems. Proceedings of the Sixth workshop on
Hot Topics in Operating Systems,Cape Cod, Massachusetts,
USA. IEEE Computer Society Press, 1997.
J. Cradley Chen, Y. Endo, D. Mazieres, A. Dias, M. Seltzer,
and M.D. Smith. The measured performance of personal
computer operating systems. ACM transactions on computing
systems and Proceedings of the 15th ACM symposium on
Operating System Principles, 1995.
[11]
13The question of whether to invest in promising an expensive
and late consensus over a coarse grain of space, or whether to
expose its divergences as a feature remains a policy choice -- one
that currently aligns with opposite poles of Dev and Ops.
23
[12] P. A. Porras and P. G. Neumann. EMERALD: Event mon-
itoring enabling responses to anomalous live disturbances.
In Proc. 20th NIST-NCSC National Information Systems
Security Conference, pages 353 -- 365, 1997.
[13] H. Abdu, H. Lutfiya, and M. Bauer. A model for adaptive
monitoring configurations. Proceedings of the VI IFIP/IEEE
IM conference on network management, page 371, 1999.
[14] A. Balliu, D. Olivetti, O. Babaoglu, M. Marzolla,
A. Sirbu.
A big data analyzer
arXiv:1509.00773v1 [cs.DC], 2015.
, and
large trace logs.
for
[15] Paul Barford and Mark Crovella. Generating representative
web workloads for network and server performance evalua-
tion. In SIGMETRICS '98/PERFORMANCE '98: Proceed-
ings of the 1998 ACM SIGMETRICS joint international con-
ference on Measurement and modeling of computer systems,
pages 151 -- 160, New York, NY, USA, 1998. ACM Press.
[16] Leslie Lamport.
Paxos Made Simple.
32(4):51 -- 58, December 2001.
SIGACT News,
[17] Diego Ongaro and John Ousterhout.
In search of an un-
In Proceedings of the
derstandable consensus algorithm.
2014 USENIX Conference on USENIX Annual Technical
Conference, USENIX ATC'14, pages 305 -- 320, Berkeley,
CA, USA, 2014. USENIX Association.
J.A. Bergstra and M. Burgess. Promise Theory: Principles
and Applications. χtAxis Press, 2014.
[18]
[19] Mark Burgess. An approach to understanding policy based
on autonomy and voluntary cooperation. In IFIP/IEEE 16th
international workshop on distributed systems operations and
management (DSOM), in LNCS 3775, pages 97 -- 108, 2005.
[20] M. Burgess. Spacetimes with semantics (i).
http://arxiv.org/abs/1411.5563, 2014.
[21] M. Burgess. Spacetimes with semantics (ii).
http://arxiv.org/abs/1505.01716, 2015.
[22] M. Burgess. Spacetimes with semantics (iii).
http://arxiv.org/abs/1608.02193, 2016.
[23] M. Burgess. A spacetime approach to generalized cognitive
reasoning in multi-scale learning.
https://arxiv.org/abs/1702.04638, 2017.
[24] M. Burgess and A. Couch. On system rollback and totalized
J. Log.
fields: An algebraic approach to system change.
Algebr. Program., 80(8):427 -- 443, 2011.
[25] C.E. Shannon and W. Weaver. The mathematical theory of
communication. University of Illinois Press, Urbana, 1949.
[26] T.M. Cover and J.A. Thomas. Elements of Information
Theory. (J.Wiley & Sons., New York), 1991.
[27] Leonard Kleinrock. Queueing Systems: Computer Applica-
tions, volume 2. John Wiley & Sons, Inc., 1976.
[28] R. Badonnel and M. Burgess. Service load balancing with
autonomic servers: Reversing the decision making process.
In Resilient Networks and Services, Second International
Conference on Autonomous Infrastructure, Management and
Security, AIMS 2008, Bremen, Germany, July 1-3, 2008,
Proceedings, pages 92 -- 104, 2008.
J. Bon´er, D. Farley, R. Kuhn, and M. Thompson. The reactive
manifesto. https://www.reactivemanifesto.org/.
[29]
[30] W.R. Ashby. Design for a brain. J. Wiley & Sons, 1952.
[31] W.R. Ashby. An introduction to cybernetics.
J. Wiley &
Sons, 1956.
[32] Leslie Lamport. Time, clocks, and the ordering of events in
a distributed system. Commun. ACM, 21(7):558 -- 565, July
1978.
[33] P. Borrill, M. Burgess, A. Karp, and A. Kasuya. Spacetime-
entangled networks (i) relativity and observability of stepwise
consensus. arXiv:1807.08549 [cs.DC], 2018.
[34] M. Burgess. A Treatise on Systems: Volume 2: Intentional
systems with faults, errors, and flaws. in progress, 2004-.
[35] M.J. Fischer, N.A. Lynch, and M.S. Paterson. Impossibility
of distributed consensus with one faulty process. J. ACM,
32(2):374 -- 382, April 1985.
[36] M. Burgess.
In Search of Certainty: the science of our
information infrastructure. Xtaxis Press, 2013.
[37] White paper. The new rules of sampling. Technical report,
Honeycomb.com, 2019.
[38] A. Cockcroft. Utilization is virtually useless as a metric! In
Proceedings of Int. CMG Conference, pages 557 -- 562, 2006.
[39] M. Burgess. Smart Spacetime. χtAxis Press, 2019.
[40] M. Dam and R. Stadler. A generic protocol for network state
aggregation. RVK 05, Linkping, Sweden, June 14-16, 2005.
[41] M. Burgess and M. Disney. Understanding scalability in
In Lec-
network aggregation with continuous monitoring.
ture Notes on Computer Science, Proc. 18th IFIP/IEEE
Distributed Systems: Operations and Management (DSOM
2007), volume (submitted). Springer, 2007.
[42] F. Wuhib, M. Dam, R. Stadler, and A. Clemm. Robust
monitoring of network-wide aggregates through gossiping.
In 10th IFIP/IEEE International Symposium on Integrated
Management (IM 2007), 2007.
[44]
[43] M. Burgess, H. Haugerud, T. Reitan, and S. Straumsnes.
Measuring host normality. ACM Transactions on Computing
Systems, 20:125 -- 160, 2001.
J. Pearl. Probabilistic Reasoning in Intelligent Systems:
Networks of Plausible Inference. Morgen Kaufmann, San
Francisco, 1988.
J. Pearl. Causality. Cambridge University Press, Cambridge,
2000.
[45]
[46] A. Couch and M. Burgess. Compass and direction in topic
maps. (Oslo University College preprint), 2009.
[47] M. Burgess and W. Louth. Preserving the significance of
distributed observations. unpublished, 2019.
[48] H. Zenil, N.A. Kiani, and J. Tegn´er. The thermodynamics
of network coding and an algorithmic refinement of the
principle of maximum entropy. Entropy, 21(560), 2019.
[49] Aljabr
Inc.
Koalja
history
package.
https://github.com/AljabrIO/ koalja-operator/tree/master/pkg/history.
[50] M. Burgess. A tiny overview of cfengine: convergent main-
tenance agent. In Proceedings of the 1st International Work-
shop on Multi-Agent and Robotic Systems, MARS/ICINCO,
2005.
[51] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank
citation ranking: Bringing order to the web. Technical report,
Stanford University, Stanford, CA, 1998.
J. Bjelland, M. Burgess, G. Canright, and K. Eng-Monsen.
Eigenvectors of directed graphs and importance scores:
dominance, t-rank, and sink remedies. Data Mining and
Knowledge Discovery, 20(1):98 -- 151, 2010.
[52]
[53] M. Burgess. A site configuration engine. Computing systems
(MIT Press: Cambridge MA), 8:309, 1995.
[54] M. Burgess.
Two dimensional
time-series for anomaly
detection and regulation in adaptive systems. Lecture Notes
in Computer Science, IFIP/IEEE 13th International Work-
shop on Distributed Systems: Operations and Management
(DSOM 2002), 2506:169, 2002.
24
|
0904.3669 | 1 | 0904 | 2009-04-23T11:54:55 | Collaborative systems and multiagent systems | [
"cs.MA"
] | This paper presents some basic elements regarding the domain of the collaborative systems, a domain of maximum actuality and also the multiagent systems, developed as a result of a sound study on the one-agent systems. | cs.MA | cs | (cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:4)(cid:1)(cid:11)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:16)(cid:17)(cid:7)(cid:1)(cid:18)(cid:13)(cid:5)(cid:7)(cid:1)(cid:11)(cid:18)(cid:1)(cid:12)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:11)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:2)(cid:3)(cid:3)(cid:4)(cid:5)(cid:19)(cid:7)(cid:1)(cid:24)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:1)(cid:8)(cid:16)(cid:10)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:6)(cid:19)(cid:7)(cid:1)(cid:27)(cid:15)(cid:28)(cid:1)(cid:29)(cid:13)(cid:14)(cid:6)(cid:1)(cid:30)(cid:19)(cid:15)(cid:1)(cid:31)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:1)(cid:1)(cid:2)(cid:2) (cid:3)(cid:3) (cid:3)(cid:3)(cid:4)(cid:4)(cid:5)(cid:5)(cid:2)(cid:2)(cid:6)(cid:6)(cid:4)(cid:4) (cid:7)(cid:7) (cid:8)(cid:8)(cid:9)(cid:9)(cid:10)(cid:10) (cid:11)(cid:11) (cid:12)(cid:12)(cid:13)(cid:13) (cid:12)(cid:12) (cid:7)(cid:7)(cid:10)(cid:10)(cid:14)(cid:14) (cid:12)(cid:12) (cid:11)(cid:11)(cid:4)(cid:4)(cid:15)(cid:15)(cid:16)(cid:16) (cid:11)(cid:11)(cid:14)(cid:14)(cid:17)(cid:17) (cid:3)(cid:3) (cid:7)(cid:7) (cid:8)(cid:8)(cid:4)(cid:4)(cid:18)(cid:18)(cid:10)(cid:10)(cid:15)(cid:15) (cid:7)(cid:7) (cid:11)(cid:11) (cid:12)(cid:12)(cid:13)(cid:13) (cid:12)(cid:12) (cid:7)(cid:7)(cid:10)(cid:10)(cid:14)(cid:14) (cid:12)(cid:12) (cid:11)(cid:11)
(cid:19)(cid:12)(cid:12)(cid:8)(cid:12)(cid:7)(cid:20)(cid:21)(cid:6)(cid:2)(cid:22)(cid:20)(cid:11)(cid:19)(cid:3)(cid:8)(cid:15)(cid:11)(cid:23) (cid:17)(cid:15)(cid:7)(cid:10)(cid:4)(cid:15)(cid:17)(cid:24)(cid:11)(cid:21)(cid:25)(cid:26)(cid:11)
(cid:27)(cid:15)(cid:12)(cid:7)(cid:6)(cid:17)(cid:28)(cid:7)(cid:2)(cid:6)(cid:11)(cid:1)(cid:6)(cid:8)(cid:12)(cid:7)(cid:8)(cid:15)(cid:4)(cid:11)(cid:29)(cid:22)(cid:10)(cid:3)(cid:8)(cid:4)(cid:11)(cid:30)(cid:2)(cid:22)(cid:6)(cid:4)(cid:15)(cid:24)(cid:11)(cid:21)(cid:25)(cid:26)(cid:11)(cid:1)(cid:4)(cid:15)(cid:16)(cid:8)(cid:16)(cid:4)(cid:7)(cid:10)(cid:11)
(cid:31) (cid:8)(cid:5)(cid:8)(cid:12)(cid:28)(cid:17)(cid:12)!(cid:11)"(cid:15)(cid:8)(cid:9)(cid:10)(cid:6)(cid:12)(cid:8)(cid:7)(cid:13)(cid:11)(cid:2)(cid:22)(cid:11) (cid:8)(cid:14)(cid:8)#(cid:2)(cid:4)(cid:6)(cid:4)(cid:24)(cid:11)$(cid:2)(cid:14)(cid:4)(cid:15)(cid:8)(cid:4)(cid:11)
"#$%& (cid:2)(cid:29)(cid:7)(cid:1) ’(cid:26)(cid:16)(cid:9)(cid:4)(cid:9)(cid:6)(cid:4)(cid:1) (cid:25)(cid:9)(cid:6)((cid:10)(cid:3)(cid:15)(cid:17)(cid:1) (cid:16))(cid:15)(cid:6)*(cid:4)(cid:1) (cid:6)(cid:5)(cid:6)(cid:14)(cid:6)(cid:3)(cid:15)(cid:6)(cid:1) +(cid:6)(cid:1) ,(cid:4)((cid:17)(cid:1) (cid:16)(cid:26)(cid:1)
(cid:25)(cid:9)(cid:10)*(cid:10)(cid:9)(cid:6)(cid:1) (cid:5)(cid:4)(cid:1) +(cid:13)(cid:14)(cid:6)(cid:3)(cid:10)(cid:26)(cid:5)(cid:1) (cid:19)(cid:10)(cid:19)(cid:15)(cid:6)(cid:14)(cid:6)(cid:5)(cid:13)(cid:9)(cid:1) (cid:16)(cid:13)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)-(cid:1) (cid:26)(cid:3)(cid:1) +(cid:13)(cid:14)(cid:6)(cid:3)(cid:10)(cid:26)(cid:1) +(cid:6)(cid:1)
(cid:14)(cid:4).(cid:10)(cid:14)(cid:17)(cid:1) (cid:4)(cid:16)(cid:15)(cid:26)(cid:4)(cid:5)(cid:10)(cid:15)(cid:4)(cid:15)(cid:6)-(cid:1) /(cid:10)(cid:1) (cid:16)(cid:26)(cid:1) (cid:25)(cid:9)(cid:10)*(cid:10)(cid:9)(cid:6)(cid:1) (cid:5)(cid:4)(cid:1) (cid:19)(cid:10)(cid:19)(cid:15)(cid:6)(cid:14)(cid:6)(cid:5)(cid:6)(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)-(cid:1)
+(cid:6)(*(cid:13)(cid:5)(cid:15)(cid:4)(cid:15)(cid:6)(cid:1) 1(cid:3)(cid:1) (cid:26)(cid:9)(cid:14)(cid:4)(cid:1) (cid:26)(cid:3)(cid:26)(cid:10)(cid:1) (cid:19)(cid:15)(cid:26)+(cid:10)(cid:26)(cid:1) (cid:15)(cid:6)(cid:14)(cid:6)(cid:10)(cid:3)(cid:10)(cid:16)(cid:1) (cid:4)(cid:19)(cid:26)(cid:25)(cid:9)(cid:4)(cid:1) (cid:19)(cid:10)(cid:19)(cid:15)(cid:6)(cid:14)(cid:6)(cid:5)(cid:13)(cid:9)(cid:1) (cid:16)(cid:26)(cid:1)
(cid:26)(cid:3)(cid:1)(cid:19)(cid:10)(cid:3)0(cid:26)(cid:9)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:7)(cid:1)
(cid:1)
(cid:1)
%(cid:20)(cid:11)(cid:1)(cid:2)(cid:3)(cid:3)(cid:4)(cid:5)(cid:2)(cid:6)(cid:4)(cid:7)(cid:8)(cid:9)(cid:10)(cid:11)(cid:4)(cid:12)(cid:11)&’(cid:2)(cid:8)(cid:15)(cid:7)(cid:3)(cid:13)(cid:11)((cid:8)(cid:7)(cid:25)!(cid:11)
(cid:1)
(cid:11)(cid:3)(cid:1)(cid:13)(cid:9)+(cid:6)(cid:9)(cid:1)(cid:15)(cid:13)(cid:1)(cid:26)(cid:3)+(cid:6)(cid:9)(cid:19)(cid:15)(cid:4)(cid:3)+(cid:1)(cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1)(cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)(cid:10)(cid:15)(cid:1)(cid:10)(cid:19)(cid:1)(cid:3)(cid:6)(cid:16)(cid:6)(cid:19)(cid:19)(cid:4)(cid:9)2(cid:1)(cid:15)(cid:13)(cid:1)(cid:12)(cid:13)(cid:16)(cid:26)(cid:19)(cid:1)(cid:13)(cid:3)(cid:1)(cid:15)(cid:28)(cid:13)(cid:19)(cid:6)(cid:1)
(cid:10)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)3(cid:28)(cid:10)(cid:16)(cid:28)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:25)(cid:4)(cid:9)(cid:15)(cid:3)(cid:6)(cid:9)(cid:19)-(cid:1) (cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)(cid:3)0(cid:1) (cid:10)(cid:3)(cid:1) (cid:19)(cid:13)(cid:5)*(cid:10)(cid:3)0(cid:1)
(cid:25)(cid:9)(cid:13),(cid:5)(cid:6)(cid:14)(cid:19)(cid:7)(cid:1) (cid:29)(cid:28)(cid:6)(cid:9)(cid:6)(cid:1) (cid:10)(cid:19)(cid:1) (cid:3)(cid:13)(cid:1) +(cid:13)(cid:26),(cid:15)(cid:1) (cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:4)(cid:3)(cid:1) (cid:10)(cid:14)(cid:25)(cid:9)(cid:13)*(cid:6)(cid:14)(cid:6)(cid:3)(cid:15)(cid:1) (cid:13)(cid:12)(cid:1) (cid:10)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)
3(cid:13)(cid:26)(cid:5)+(cid:1),(cid:6)(cid:1)(cid:4)(cid:25)(cid:25)(cid:9)(cid:13)(cid:25)(cid:9)(cid:10)(cid:4)(cid:15)(cid:6)(cid:7)(cid:1)(cid:29)(cid:28)(cid:6)(cid:1)(cid:26)(cid:19)(cid:6)(cid:1)(cid:13)(cid:12)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:11)(cid:3)(cid:15)(cid:6)(cid:9)(cid:3)(cid:6)(cid:15)(cid:1)(cid:28)(cid:4)(cid:19)(cid:1),(cid:6)(cid:6)(cid:3)(cid:1)(cid:10)(cid:3)(cid:16)(cid:9)(cid:6)(cid:4)(cid:19)(cid:10)(cid:3)0(cid:1) (cid:5)(cid:6)(cid:4)+(cid:10)(cid:3)0(cid:1)(cid:15)(cid:13)(cid:1)
,(cid:6)(cid:15)(cid:15)(cid:6)(cid:9)(cid:1) (cid:16)(cid:13)(cid:14)(cid:14)(cid:26)(cid:3)(cid:10)(cid:16)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) ,(cid:6)(cid:15)3(cid:6)(cid:6)(cid:3)(cid:1) (cid:25)(cid:6)(cid:13)(cid:25)(cid:5)(cid:6)-(cid:1) (cid:3)(cid:13)(cid:1) (cid:14)(cid:4)(cid:15)(cid:15)(cid:6)(cid:9)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:25)(cid:5)(cid:4)(cid:16)(cid:6)(cid:1) (cid:15)(cid:28)(cid:6)2(cid:1) (cid:12)(cid:10)(cid:3)+(cid:1) (cid:15)(cid:28)(cid:6)(cid:10)(cid:9)(cid:1)
(cid:19)(cid:6)(cid:5)*(cid:6)(cid:19)(cid:7)(cid:1) (cid:8)(cid:15)(cid:10)(cid:5)(cid:5)-(cid:1) (cid:14)(cid:4)(cid:3)2(cid:1) (cid:26)(cid:19)(cid:6)(cid:9)(cid:19)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:12)(cid:4)(cid:16)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:5)(cid:10)(cid:14)(cid:10)(cid:15)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1) (cid:13)(cid:12)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:16)(cid:13)(cid:14)(cid:14)(cid:26)(cid:3)(cid:10)(cid:16)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1)
(cid:14)(cid:6)(cid:4)(cid:3)(cid:19)(cid:1) (cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:6).(cid:10)(cid:19)(cid:15)(cid:1) (cid:19)(cid:13)(cid:1) (cid:12)(cid:4)(cid:9)(cid:7)(cid:1) (cid:29)(cid:28)(cid:6)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:12)(cid:4)(cid:16)(cid:6)(cid:19)(cid:1) (cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:16)(cid:4)(cid:3)(cid:1) ,(cid:6)(cid:1) +(cid:10)(cid:9)(cid:6)(cid:16)(cid:15)(cid:5)2(cid:1) (cid:14)(cid:4)(cid:3)(cid:10)(cid:25)(cid:26)(cid:5)(cid:4)(cid:15)(cid:6)+(cid:1) (cid:4)(cid:9)(cid:6)(cid:1)
(cid:16)(cid:13)(cid:3)(cid:19)(cid:10)+(cid:6)(cid:9)(cid:6)+(cid:1) (cid:4)(cid:1) 3(cid:4)2(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) (cid:4)(cid:3)2(cid:1) (cid:26)(cid:19)(cid:6)(cid:9)(cid:1) (cid:15)(cid:13)(cid:1) ,(cid:6)(cid:1) (cid:4),(cid:5)(cid:6)(cid:1) (cid:15)(cid:13)(cid:1) (cid:4)(cid:16)(cid:16)(cid:6)(cid:19)(cid:19)(cid:1) (cid:4)(cid:1) (cid:16)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:7)(cid:1) (cid:11)(cid:15)(cid:1) (cid:10)(cid:19)(cid:1)
(cid:16)(cid:13)(cid:3)(cid:19)(cid:10)+(cid:6)(cid:9)(cid:6)+(cid:1)(cid:15)(cid:28)(cid:4)(cid:15)(cid:1)(cid:4)(cid:3)2(cid:15)(cid:28)(cid:10)(cid:3)0(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:26)(cid:19)(cid:6)(cid:9)(cid:1)(cid:3)(cid:6)(cid:6)+(cid:19)(cid:1)(cid:10)(cid:19)(cid:1)4(cid:26)(cid:19)(cid:15)(cid:1)(cid:4)(cid:1)(cid:16)(cid:5)(cid:10)(cid:16)5(cid:1)(cid:4)3(cid:4)2-(cid:1)(cid:15)(cid:28)(cid:6)(cid:9)(cid:6)(cid:1)(cid:10)(cid:19)(cid:1)(cid:3)(cid:13)(cid:1)(cid:3)(cid:6)(cid:6)+(cid:1)
(cid:13)(cid:12)(cid:1) (cid:19)(cid:25)(cid:6)(cid:16)(cid:10)(cid:12)2(cid:10)(cid:3)0(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:16)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:1) (cid:28)(cid:13)3(cid:1) (cid:15)(cid:13)(cid:1) +(cid:13)(cid:1) (cid:4)(cid:1) (cid:16)(cid:6)(cid:9)(cid:15)(cid:4)(cid:10)(cid:3)(cid:1) (cid:15)(cid:28)(cid:10)(cid:3)0(cid:1) ,(cid:26)(cid:15)(cid:1) (cid:13)(cid:3)(cid:5)2(cid:1) (cid:15)(cid:13)(cid:1) (cid:4)(cid:19)5(cid:1) 3(cid:28)(cid:4)(cid:15)(cid:1)
(cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:15)(cid:28)(cid:10)(cid:3)0(cid:1) (cid:10)(cid:19)(cid:7)(cid:1) 6(cid:26)(cid:15)(cid:1) (cid:15)(cid:28)(cid:10)(cid:19)(cid:1) (cid:10)(cid:19)(cid:1) (cid:26)(cid:3)(cid:12)(cid:13)(cid:9)(cid:15)(cid:26)(cid:3)(cid:4)(cid:15)(cid:6)(cid:5)2(cid:1) (cid:4)(cid:1) (cid:19)(cid:28)(cid:4)(cid:5)(cid:5)(cid:13)3(cid:1) (cid:15)(cid:9)(cid:26)(cid:15)(cid:28)-(cid:1) (cid:4)(cid:19)(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) (cid:14)(cid:4)(cid:3)2(cid:1)
(cid:4)(cid:25)(cid:25)(cid:5)(cid:10)(cid:16)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:26)(cid:19)(cid:6)(cid:9)(cid:19)(cid:1)+(cid:13)(cid:3)7(cid:15)(cid:1)(cid:28)(cid:4)*(cid:6)(cid:1)(cid:15)(cid:13)(cid:1)3(cid:9)(cid:10)(cid:15)(cid:6)(cid:1)(cid:4)(cid:3)2(cid:1)(cid:16)(cid:13)+(cid:6)(cid:1)(cid:13)(cid:9)(cid:1)(cid:16)(cid:13)(cid:14)(cid:14)(cid:4)(cid:3)+(cid:19)-(cid:1)(cid:4)(cid:19)(cid:1)(cid:15)(cid:28)(cid:6)2(cid:1)+(cid:13)(cid:1)
3(cid:28)(cid:6)(cid:3)(cid:1) (cid:10)(cid:15)(cid:1) (cid:16)(cid:13)(cid:14)(cid:6)(cid:19)(cid:1) (cid:15)(cid:13)(cid:1) (cid:14)(cid:13)(cid:9)(cid:6)(cid:1) (cid:16)(cid:13)(cid:14)(cid:25)(cid:5)(cid:6).(cid:1) (cid:25)(cid:9)(cid:13),(cid:5)(cid:6)(cid:14)(cid:19)(cid:7)(cid:1) (cid:29)(cid:28)(cid:6)(cid:1) (cid:26)(cid:19)(cid:6)(cid:9)(cid:1) (cid:3)(cid:6)(cid:6)+(cid:19)(cid:1) (cid:15)(cid:13)(cid:1) 8(cid:15)(cid:6)(cid:5)(cid:5)9(cid:1) (cid:15)(cid:28)(cid:6)(cid:1)
(cid:16)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:1) (cid:28)(cid:13)3(cid:1) (cid:15)(cid:13)(cid:1) (cid:19)(cid:13)(cid:5)*(cid:6)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:25)(cid:9)(cid:13),(cid:5)(cid:6)(cid:14)(cid:7)(cid:1) (cid:2)(cid:5)(cid:15)(cid:28)(cid:13)(cid:26)0(cid:28)(cid:1) ,(cid:6)(cid:10)(cid:3)0(cid:1) (cid:16)(cid:13)(cid:14)(cid:25)(cid:5)(cid:6).(cid:1) (cid:15)(cid:13)(cid:13)(cid:5)(cid:19)-(cid:1) (cid:26)(cid:19)(cid:26)(cid:4)(cid:5)(cid:1)
(cid:10)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:3)(cid:13)(cid:15)(cid:1) (cid:4)3(cid:4)(cid:9)(cid:6)(cid:1) (cid:13)(cid:12)(cid:1) 3(cid:28)(cid:4)(cid:15)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:26)(cid:19)(cid:6)(cid:9)(cid:19)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:15)(cid:9)2(cid:10)(cid:3)0(cid:1) (cid:15)(cid:13)(cid:1) +(cid:13)(cid:1) ,2(cid:1)
(cid:26)(cid:19)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:28)(cid:6)(cid:14)-(cid:1) (cid:4)(cid:3)+(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:3)(cid:13)(cid:15)(cid:1) (cid:13)(cid:12)(cid:1) 0(cid:9)(cid:6)(cid:4)(cid:15)(cid:1) (cid:28)(cid:6)(cid:5)(cid:25)(cid:1) (cid:10)(cid:3)(cid:1) (cid:19)(cid:13)(cid:5)*(cid:10)(cid:3)0(cid:1) (cid:25)(cid:9)(cid:13),(cid:5)(cid:6)(cid:14)(cid:19)(cid:7)(cid:1) (cid:11)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1)
(cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:19)(cid:26)(cid:25)(cid:25)(cid:13)(cid:19)(cid:6)+(cid:1) (cid:15)(cid:13)(cid:1) (cid:19)(cid:13)(cid:5)*(cid:6)(cid:1) (cid:25)(cid:9)(cid:13),(cid:5)(cid:6)(cid:14)(cid:19)(cid:1) 3(cid:10)(cid:15)(cid:28)(cid:13)(cid:26)(cid:15)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:26)(cid:19)(cid:6)(cid:9)(cid:1) (cid:19)(cid:25)(cid:6)(cid:16)(cid:10)(cid:12)2(cid:10)(cid:3)0(cid:1) (cid:6)(cid:4)(cid:16)(cid:28)(cid:1)
(cid:19)(cid:15)(cid:6)(cid:25)(cid:1)(cid:15)(cid:13)(cid:1),(cid:6)(cid:1)(cid:6).(cid:6)(cid:16)(cid:26)(cid:15)(cid:6)+(cid:7)(cid:1)(cid:1)
(cid:1)
(cid:30) !(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:4)(cid:1)(cid:11)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:16)(cid:17)(cid:7)(cid:1)(cid:18)(cid:13)(cid:5)(cid:7)(cid:1)(cid:11)(cid:18)(cid:1)(cid:12)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:11)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:2)(cid:3)(cid:3)(cid:4)(cid:5)(cid:19)(cid:7)(cid:1)(cid:24)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:1)(cid:8)(cid:16)(cid:10)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:6)(cid:19)(cid:7)(cid:1)(cid:27)(cid:15)(cid:28)(cid:1)(cid:29)(cid:13)(cid:14)(cid:6)(cid:1)(cid:30)(cid:19)(cid:15)(cid:1)(cid:31)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:29)(cid:13)(cid:1) (cid:9)(cid:6)(cid:4)(cid:5)(cid:10)((cid:6)(cid:1) (cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) ,(cid:6)(cid:15)3(cid:6)(cid:6)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:4)(cid:3)+(cid:1) (cid:15)(cid:28)(cid:6)(cid:10)(cid:9)(cid:1) (cid:26)(cid:19)(cid:6)(cid:9)(cid:19)-(cid:1) (cid:10)(cid:3)(cid:1)
(cid:13)(cid:9)+(cid:6)(cid:9)(cid:1) (cid:15)(cid:13)(cid:1) (cid:13),(cid:15)(cid:4)(cid:10)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:3)(cid:6)(cid:6)+(cid:6)+(cid:1) (cid:10)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:15)(cid:13)(cid:1) ,(cid:6)(cid:1) (cid:4),(cid:5)(cid:6)(cid:1) (cid:15)(cid:13)(cid:1) (cid:19)(cid:13)(cid:5)*(cid:6)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:25)(cid:9)(cid:13),(cid:5)(cid:6)(cid:14)(cid:19)(cid:1) (cid:13)(cid:12)(cid:1)
(cid:15)(cid:28)(cid:6)(cid:1) (cid:26)(cid:19)(cid:6)(cid:9)(cid:19)-(cid:1) (cid:10)(cid:15)(cid:1) (cid:10)(cid:19)(cid:1) (cid:10)(cid:14)(cid:25)(cid:13)(cid:19)(cid:6)+(cid:1) (cid:15)(cid:13)(cid:1) (cid:26)(cid:19)(cid:6)(cid:1) (cid:19)(cid:13)(cid:14)(cid:6)(cid:1) (cid:4)(cid:9)(cid:15)(cid:10)(cid:12)(cid:10)(cid:16)(cid:10)(cid:4)(cid:5)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:16)(cid:6)(cid:1) ,(cid:4)(cid:19)(cid:6)+(cid:1)
(cid:15)(cid:6)(cid:16)(cid:28)(cid:3)(cid:13)(cid:5)(cid:13)0(cid:10)(cid:6)(cid:19)-(cid:1)(cid:9)(cid:6)(cid:12)(cid:6)(cid:9)(cid:9)(cid:10)(cid:3)0(cid:1)(cid:3)(cid:13)(cid:15)(cid:1)(cid:13)(cid:3)(cid:5)2(cid:1)(cid:15)(cid:13)(cid:1)(cid:10)(cid:3)+(cid:10)*(cid:10)+(cid:26)(cid:4)(cid:5)(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1),(cid:26)(cid:15)(cid:1)(cid:15)(cid:13)(cid:1)(cid:14)(cid:4)(cid:3)2(cid:1)
(cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) 3(cid:13)(cid:9)5(cid:10)(cid:3)0(cid:1) (cid:15)(cid:13)0(cid:6)(cid:15)(cid:28)(cid:6)(cid:9)-(cid:1) (cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)(cid:3)0(cid:7)(cid:1) (cid:29)(cid:28)(cid:10)(cid:19)(cid:1)
(cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:1) (cid:10)(cid:19)(cid:1) (cid:4)(cid:3)(cid:1) (cid:10)(cid:14)(cid:25)(cid:13)(cid:9)(cid:15)(cid:4)(cid:3)(cid:15)(cid:1) (cid:25)(cid:4)(cid:9)(cid:15)(cid:1) (cid:13)(cid:12)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:15)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)-(cid:1) (cid:4)(cid:3)+(cid:1)
(cid:15)(cid:28)(cid:6)(cid:1) (cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:1) (cid:16)(cid:4)(cid:3)(cid:1) (cid:13)(cid:3)(cid:5)2(cid:1) ,(cid:6)(cid:1) (cid:26)(cid:3)+(cid:6)(cid:9)(cid:19)(cid:15)(cid:13)(cid:13)+(cid:1) (cid:4)(cid:12)(cid:15)(cid:6)(cid:9)(cid:1) (cid:26)(cid:3)+(cid:6)(cid:9)(cid:19)(cid:15)(cid:4)(cid:3)+(cid:10)(cid:3)0(cid:1) (cid:15)(cid:28)(cid:6)(cid:1)
(cid:10)(cid:3)+(cid:10)*(cid:10)+(cid:26)(cid:4)(cid:5)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:7)(cid:1) (cid:11)(cid:15)(cid:1) (cid:10)(cid:19)(cid:1) (cid:10)(cid:14)(cid:25)(cid:13)(cid:9)(cid:15)(cid:4)(cid:3)(cid:15)(cid:1) (cid:15)(cid:13)(cid:1) 5(cid:3)(cid:13)3(cid:1) (cid:28)(cid:13)3(cid:1) (cid:14)(cid:26)(cid:16)(cid:28)(cid:1) (cid:13)(cid:12)(cid:1) (cid:4)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:1) (cid:10)(cid:19)(cid:1)
(cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1) (cid:4)(cid:3)+(cid:1) (cid:3)(cid:13)(cid:15)(cid:1) (cid:16)(cid:13)(cid:13)(cid:9)+(cid:10)(cid:3)(cid:4)(cid:15)(cid:6)+(cid:1) (cid:13)(cid:9)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)*(cid:6)(cid:7)(cid:1) (cid:29)(cid:28)(cid:6)(cid:1) +(cid:10)(cid:19)(cid:15)(cid:10)(cid:3)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) ,(cid:6)(cid:15)3(cid:6)(cid:6)(cid:3)(cid:1)
(cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1)(cid:4)(cid:3)+(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)*(cid:6)(cid:1)(cid:10)(cid:19)(cid:1)(cid:15)(cid:28)(cid:4)(cid:15)(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)*(cid:6)(cid:1)(cid:13)(cid:3)(cid:5)2(cid:1)(cid:9)(cid:6)(cid:12)(cid:6)(cid:9)(cid:19)(cid:1)(cid:15)(cid:13)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1)(cid:26)(cid:25)(cid:13)(cid:3)(cid:1)
(cid:4)(cid:3)(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1) (cid:15)(cid:28)(cid:10)(cid:3)0(cid:1) (cid:13)(cid:9)(cid:1) (cid:25)(cid:6)(cid:9)(cid:19)(cid:13)(cid:3)-(cid:1) 3(cid:28)(cid:10)(cid:5)(cid:6)(cid:1) (cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1) (cid:9)(cid:6)(cid:12)(cid:6)(cid:9)(cid:19)(cid:1) (cid:15)(cid:13)(cid:1) 3(cid:13)(cid:9)5(cid:10)(cid:3)0(cid:1) 4(cid:13)(cid:10)(cid:3)(cid:15)(cid:5)2(cid:1) 3(cid:10)(cid:15)(cid:28)(cid:1)
(cid:15)(cid:28)(cid:6)(cid:1)(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:19)(cid:7)(cid:1)(cid:1)(cid:1)
(cid:1) (cid:11)(cid:3)(cid:1)(cid:13)(cid:9)+(cid:6)(cid:9)(cid:1)(cid:15)(cid:13)(cid:1),(cid:6)(cid:1)(cid:4),(cid:5)(cid:6)(cid:1)(cid:15)(cid:13)(cid:1),(cid:26)(cid:10)(cid:5)+(cid:1)(cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)(cid:10)(cid:15)(cid:1)(cid:10)(cid:19)(cid:1) (cid:3)(cid:6)(cid:16)(cid:6)(cid:19)(cid:19)(cid:4)(cid:9)2(cid:1)(cid:15)(cid:13)(cid:1)
(cid:10)+(cid:6)(cid:3)(cid:15)(cid:10)(cid:12)2(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:16)(cid:4)(cid:25)(cid:4),(cid:10)(cid:5)(cid:10)(cid:15)(cid:10)(cid:6)(cid:19)(cid:1)(cid:15)(cid:28)(cid:4)(cid:15)(cid:1)(cid:4)(cid:9)(cid:6)(cid:1)(cid:15)(cid:13)(cid:1),(cid:6)(cid:1)(cid:4)++(cid:6)+(cid:1)(cid:15)(cid:13)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:10)(cid:3)+(cid:10)*(cid:10)+(cid:26)(cid:4)(cid:5)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) (cid:19)(cid:13)(cid:1)(cid:15)(cid:28)(cid:4)(cid:15)(cid:1)
(cid:15)(cid:28)(cid:6)2(cid:1),(cid:6)(cid:16)(cid:13)(cid:14)(cid:6)(cid:1)(cid:4),(cid:5)(cid:6)(cid:1)(cid:15)(cid:13)(cid:1)3(cid:13)(cid:9)5(cid:1)3(cid:10)(cid:15)(cid:28)(cid:1)(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:7)(cid:1)(cid:29)(cid:28)(cid:6)(cid:1)(cid:14)(cid:13)+(cid:6)(cid:5)(cid:10)(cid:3)0(cid:1)(cid:13)(cid:12)(cid:1)(cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1)
(cid:28)(cid:4)(cid:19)(cid:1) ,(cid:6)(cid:16)(cid:13)(cid:14)(cid:6)(cid:1) (cid:4)(cid:1) (cid:25)(cid:9)(cid:10)(cid:13)(cid:9)(cid:10)(cid:15)2(cid:1) (cid:4)(cid:19)(cid:1) +(cid:26)(cid:9)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:5)(cid:4)(cid:19)(cid:15)(cid:1) +(cid:6)(cid:16)(cid:4)+(cid:6)(cid:1) (cid:19)(cid:13)(cid:14)(cid:6)(cid:1) (cid:10)(cid:14)(cid:25)(cid:13)(cid:9)(cid:15)(cid:4)(cid:3)(cid:15)(cid:1) (cid:25)(cid:9)(cid:13)0(cid:9)(cid:6)(cid:19)(cid:19)(cid:1)
3(cid:10)(cid:15)(cid:28)(cid:10)(cid:3)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:4)(cid:9)(cid:15)(cid:10)(cid:12)(cid:10)(cid:16)(cid:10)(cid:4)(cid:5)(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:12)(cid:10)(cid:6)(cid:5)+(cid:1)(cid:28)(cid:4)(cid:19)(cid:1),(cid:6)(cid:6)(cid:3)(cid:1)(cid:4)(cid:16)(cid:28)(cid:10)(cid:6)*(cid:6)+(cid:7)(cid:1)
(cid:1)
(cid:1)
)(cid:20)(cid:11)(cid:23) (cid:17)(cid:3)(cid:7)(cid:8)(cid:4)(cid:18)(cid:10)(cid:15)(cid:7)(cid:11)(cid:30)(cid:13)(cid:12)(cid:7)(cid:10)(cid:14)(cid:12)(cid:11)((cid:8)(cid:7)(cid:25)(cid:8)(cid:15)(cid:11)(cid:4)(cid:15)(cid:11)(cid:8)(cid:15)(cid:7)(cid:10)(cid:3)(cid:3)(cid:8)(cid:18)(cid:10)(cid:15)(cid:7)(cid:11)(cid:10)(cid:15)(cid:9)(cid:8)(cid:6)(cid:2)(cid:15)(cid:14)(cid:10)(cid:15)(cid:7)(cid:11)
(cid:1)
& (cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) ’(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:10)(cid:19)(cid:1) (cid:4)(cid:15)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:19)(cid:6)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:13)(cid:12)(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:4)(cid:3)+(cid:1)
& (cid:4)(cid:16)(cid:28)(cid:10)(cid:3)(cid:6)(cid:1) ’(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0-(cid:1) (cid:15)3(cid:13)(cid:1) (cid:19)(cid:26),+(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:19)(cid:1) (cid:13)(cid:12)(cid:1) (cid:4)(cid:9)(cid:15)(cid:10)(cid:12)(cid:10)(cid:16)(cid:10)(cid:4)(cid:5)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:16)(cid:6)(cid:7)(cid:1) : (cid:6)(cid:10)(cid:19)(cid:19)(cid:1) (cid:28)(cid:4)(cid:19)(cid:1)
+(cid:6)(cid:12)(cid:10)(cid:3)(cid:6)+(cid:1)& (cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)’(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1)(cid:4)(cid:19)(cid:1)8(cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1)(cid:15)(cid:28)(cid:4)(cid:15)(cid:1)(cid:10)(cid:19)(cid:1)+(cid:13)(cid:3)(cid:6)(cid:1),2(cid:1) (cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1)(cid:4)(cid:3)+(cid:1)
3(cid:28)(cid:10)(cid:16)(cid:28)(cid:1)(cid:10)(cid:19)(cid:1)(cid:13)(cid:3)(cid:5)2(cid:1)(cid:25)(cid:13)(cid:19)(cid:19)(cid:10),(cid:5)(cid:6)(cid:1)+(cid:26)(cid:6)(cid:1)(cid:15)(cid:13)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:25)(cid:9)(cid:6)(cid:19)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:13)(cid:12)(cid:1)(cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)9(cid:7)(cid:1)
(cid:29)(cid:9)(cid:4)+(cid:10)(cid:15)(cid:10)(cid:13)(cid:3)(cid:4)(cid:5)(cid:1) & (cid:4)(cid:16)(cid:28)(cid:10)(cid:3)(cid:6)(cid:1) ’(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:6)(cid:16)(cid:28)(cid:3)(cid:13)(cid:5)(cid:13)0(cid:10)(cid:6)(cid:19)(cid:1) (cid:26)(cid:19)(cid:26)(cid:4)(cid:5)(cid:5)2(cid:1) (cid:10)(cid:14)(cid:25)(cid:5)2(cid:1) (cid:4)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)
(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:10)(cid:19)(cid:1) (cid:15)(cid:9)2(cid:10)(cid:3)0(cid:1) (cid:15)(cid:13)(cid:1) (cid:14)(cid:4).(cid:10)(cid:14)(cid:10)((cid:6)(cid:1) (cid:19)(cid:13)(cid:14)(cid:6)(cid:1) (cid:26)(cid:15)(cid:10)(cid:5)(cid:10)(cid:15)2(cid:1) (cid:12)(cid:26)(cid:3)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1) 3(cid:10)(cid:15)(cid:28)(cid:13)(cid:26)(cid:15)(cid:1) (cid:28)(cid:4)*(cid:10)(cid:3)0(cid:1) (cid:4)(cid:3)2(cid:1)
5(cid:3)(cid:13)3(cid:5)(cid:6)+0(cid:6)(cid:1) (cid:4),(cid:13)(cid:26)(cid:15)(cid:1) (cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) 3(cid:10)(cid:15)(cid:28)(cid:10)(cid:3)(cid:1) (cid:10)(cid:15)(cid:19)(cid:1) (cid:6)(cid:3)*(cid:10)(cid:9)(cid:13)(cid:3)(cid:14)(cid:6)(cid:3)(cid:15)(cid:7)(cid:1) (cid:29)(cid:28)(cid:6)(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)
(cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:9)(cid:6)(cid:12)(cid:6)(cid:9)(cid:1) (cid:15)(cid:13)(cid:1) +(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:19)(cid:1) 3(cid:10)(cid:15)(cid:28)(cid:1) (cid:14)(cid:4)(cid:3)2(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) (cid:4)(cid:3)+(cid:1) (cid:15)(cid:28)(cid:6)(cid:9)(cid:6)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:16)(cid:13)(cid:3)(cid:19)(cid:10)+(cid:6)(cid:9)(cid:6)+(cid:1) (cid:19)(cid:13)(cid:14)(cid:6)(cid:1)
(cid:14)(cid:6)(cid:16)(cid:28)(cid:4)(cid:3)(cid:10)(cid:19)(cid:14)(cid:19)(cid:1)(cid:12)(cid:13)(cid:9)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1)(cid:13)(cid:12)(cid:1)(cid:10)(cid:3)+(cid:6)(cid:25)(cid:6)(cid:3)+(cid:6)(cid:3)(cid:15)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)7(cid:1),(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:7)(cid:1)(cid:1)
(cid:29)(cid:28)(cid:6)(cid:1)(cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)+(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:1) (cid:9)(cid:6)(cid:12)(cid:6)(cid:9)(cid:19)(cid:1)(cid:15)(cid:13)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)+(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:19)(cid:1)3(cid:28)(cid:6)(cid:9)(cid:6)(cid:1) (cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1)
(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1)(cid:4)(cid:9)(cid:6)(cid:1)(cid:10)(cid:3)*(cid:13)(cid:5)*(cid:6)+(cid:1)(cid:4)(cid:3)+(cid:1)(cid:14)(cid:6)(cid:16)(cid:28)(cid:4)(cid:3)(cid:10)(cid:19)(cid:14)(cid:19)(cid:1)(cid:12)(cid:13)(cid:9)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:10)(cid:3)+(cid:6)(cid:25)(cid:6)(cid:3)+(cid:6)(cid:3)(cid:15)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)7(cid:1),(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:26)(cid:9)(cid:19)(cid:1)
(cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:28)(cid:4)*(cid:6)(cid:1) (cid:15)(cid:13)(cid:1) ,(cid:6)(cid:1) (cid:16)(cid:13)(cid:3)(cid:19)(cid:10)+(cid:6)(cid:9)(cid:6)+(cid:7)(cid:1) (cid:29)(cid:28)(cid:6)(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:10)(cid:3)(cid:16)(cid:5)(cid:26)+(cid:6)(cid:1) (cid:4)(cid:3)2(cid:1)
(cid:19)(cid:10)(cid:15)(cid:26)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1)3(cid:28)(cid:6)(cid:9)(cid:6)(cid:1)(cid:4)(cid:3)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:10)(cid:19)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1)(cid:15)(cid:13)(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:1)3(cid:10)(cid:15)(cid:28)(cid:1)(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)-(cid:1)(cid:6)*(cid:6)(cid:3)(cid:1)(cid:10)(cid:12)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)
(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) (cid:28)(cid:4)*(cid:6)(cid:1) (cid:4)(cid:1) (cid:19)(cid:15)(cid:4)(cid:15)(cid:10)(cid:16)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:7)(cid:1) (cid:29)(cid:28)(cid:6)(cid:1) (cid:14)(cid:4)(cid:10)(cid:3)(cid:1) (cid:6).(cid:25)(cid:5)(cid:4)(cid:3)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) (cid:16)(cid:13)(cid:3)(cid:19)(cid:10)+(cid:6)(cid:9)(cid:10)(cid:3)0(cid:1)
(cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:19)(cid:10)(cid:15)(cid:26)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1)3(cid:28)(cid:6)(cid:3)(cid:1)(cid:4)(cid:1)(cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:10)(cid:19)(cid:1)(cid:4)(cid:16)(cid:15)(cid:26)(cid:4)(cid:5)(cid:5)2(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1)(cid:10)(cid:19)(cid:1)
(cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:10)(cid:15)(cid:19)(cid:1) *(cid:6)(cid:9)2(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:6)+(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:1) (cid:10)(cid:19)(cid:1) ,(cid:6)(cid:10)(cid:3)0(cid:1) (cid:13)(cid:12)(cid:15)(cid:6)(cid:3)(cid:1) (cid:26)(cid:19)(cid:6)+(cid:1) (cid:4)(cid:19)(cid:1) (cid:4)(cid:1) ,(cid:4)(cid:19)(cid:10)(cid:19)(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) (cid:4)(cid:1) (cid:14)(cid:13)(cid:9)(cid:6)(cid:1)
(cid:16)(cid:13)(cid:14)(cid:25)(cid:5)(cid:6).(cid:1)(cid:4)(cid:3)+(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)*(cid:6)(cid:1),(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:7)(cid:1)(cid:2)(cid:5)(cid:15)(cid:28)(cid:13)(cid:26)0(cid:28)(cid:1)(cid:4)(cid:1)(cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:10)(cid:19)(cid:1),(cid:6)(cid:10)(cid:3)0(cid:1)(cid:10)(cid:3)*(cid:13)(cid:5)*(cid:6)+(cid:1)
(cid:10)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:25)(cid:9)(cid:13)(cid:16)(cid:6)(cid:19)(cid:19)-(cid:1) (cid:10)(cid:15)(cid:19)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:1) (cid:10)(cid:19)(cid:1) (cid:13)(cid:3)(cid:5)2(cid:1) (cid:13),*(cid:10)(cid:13)(cid:26)(cid:19)(cid:1) 3(cid:28)(cid:6)(cid:3)(cid:1) (cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1)
(cid:25)(cid:9)(cid:6)(cid:19)(cid:6)(cid:3)(cid:15)-(cid:1)(cid:4)(cid:3)+(cid:1)+(cid:6)(cid:15)(cid:6)(cid:9)(cid:14)(cid:10)(cid:3)(cid:6)(cid:19)(cid:1)(cid:15)(cid:28)(cid:6)(cid:19)(cid:6)(cid:1)(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1)(cid:15)(cid:13)(cid:1)(cid:25)(cid:4)(cid:9)(cid:15)(cid:10)(cid:16)(cid:10)(cid:25)(cid:4)(cid:15)(cid:6)(cid:1)(cid:15)(cid:13)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:16)(cid:13)(cid:5)(cid:5)(cid:4),(cid:13)(cid:9)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1)
(cid:1)(cid:30)(cid:27)(cid:22)(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:4)(cid:1)(cid:11)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:16)(cid:17)(cid:7)(cid:1)(cid:18)(cid:13)(cid:5)(cid:7)(cid:1)(cid:11)(cid:18)(cid:1)(cid:12)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:11)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:2)(cid:3)(cid:3)(cid:4)(cid:5)(cid:19)(cid:7)(cid:1)(cid:24)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:1)(cid:8)(cid:16)(cid:10)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:6)(cid:19)(cid:7)(cid:1)(cid:27)(cid:15)(cid:28)(cid:1)(cid:29)(cid:13)(cid:14)(cid:6)(cid:1)(cid:30)(cid:19)(cid:15)(cid:1)(cid:31)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:4)(cid:3)+(cid:1) (cid:4)+*(cid:6)(cid:9)(cid:19)(cid:10)(cid:4)(cid:5)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:19)(cid:10)(cid:15)(cid:26)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:7)(cid:1) (cid:11)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:16)(cid:4)(cid:19)(cid:6)(cid:1) 3(cid:28)(cid:6)(cid:3)(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:10)(cid:19)(cid:1)
(cid:28)(cid:4)(cid:25)(cid:25)(cid:6)(cid:3)(cid:10)(cid:3)0(cid:1),2(cid:1)(cid:5)(cid:4)2(cid:6)(cid:9)(cid:10)(cid:3)0(cid:1),(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:19)-(cid:1)(cid:4)(cid:5)(cid:5)(cid:1)(cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1)(cid:5)(cid:6)*(cid:6)(cid:5)(cid:19)(cid:1)3(cid:28)(cid:10)(cid:16)(cid:28)(cid:1)(cid:10)(cid:14)(cid:25)(cid:5)2(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1)
3(cid:10)(cid:15)(cid:28)(cid:1) (cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:4)(cid:16)(cid:15)(cid:26)(cid:4)(cid:5)(cid:5)2(cid:1) (cid:16)(cid:13)(cid:3)(cid:15)(cid:9)(cid:10),(cid:26)(cid:15)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:13)(cid:1) (cid:4)(cid:3)+(cid:1) (cid:10)(cid:3)(cid:16)(cid:5)(cid:26)+(cid:6)+(cid:1) (cid:10)(cid:3)(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)
(cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:7)(cid:1)
(cid:8)(cid:13)(cid:1) (cid:12)(cid:4)(cid:9)-(cid:1) (cid:14)(cid:13)(cid:19)(cid:15)(cid:1) (cid:13)(cid:12)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:4)(cid:5)0(cid:13)(cid:9)(cid:10)(cid:15)(cid:28)(cid:14)(cid:19)(cid:1) (cid:28)(cid:4)*(cid:6)(cid:1) ,(cid:6)(cid:6)(cid:3)(cid:1) +(cid:6)*(cid:6)(cid:5)(cid:13)(cid:25)(cid:6)+(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) (cid:4)(cid:1)
(cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:7)(cid:1) (cid:29)(cid:28)(cid:6)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:13)(cid:12)(cid:1) (cid:4)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:10)(cid:19)(cid:1) (cid:12)(cid:13)(cid:16)(cid:26)(cid:19)(cid:6)+(cid:1) (cid:13)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) 3(cid:4)2(cid:1) (cid:4)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)
(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:10)(cid:14)(cid:25)(cid:9)(cid:13)*(cid:6)(cid:19)(cid:1) (cid:10)(cid:15)(cid:19)(cid:1) (cid:13)3(cid:3)(cid:1) (cid:16)(cid:28)(cid:4)(cid:9)(cid:4)(cid:16)(cid:15)(cid:6)(cid:9)(cid:10)(cid:19)(cid:15)(cid:10)(cid:16)(cid:19)(cid:7)(cid:1) & (cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:10)(cid:19)(cid:1) (cid:13)(cid:26)(cid:15)(cid:1) (cid:13)(cid:12)(cid:1)
;(cid:26)(cid:6)(cid:19)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:4)(cid:19)(cid:1) (cid:5)(cid:13)(cid:3)0(cid:1) (cid:4)(cid:19)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) +(cid:13)(cid:6)(cid:19)(cid:1) (cid:3)(cid:13)(cid:15)(cid:1) (cid:4)(cid:12)(cid:12)(cid:6)(cid:16)(cid:15)(cid:1) (cid:13)(cid:9)(cid:1) (cid:10)(cid:19)(cid:1) (cid:3)(cid:13)(cid:15)(cid:1)
(cid:4)(cid:12)(cid:12)(cid:6)(cid:16)(cid:15)(cid:6)+(cid:1) ,2(cid:1) (cid:4)(cid:1) (cid:3)(cid:6)(cid:10)0(cid:28),(cid:13)(cid:9)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:7)(cid:1) (cid:2)(cid:5)(cid:15)(cid:28)(cid:13)(cid:26)0(cid:28)(cid:1) (cid:4)(cid:3)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:10)(cid:19)(cid:1) (cid:3)(cid:13)(cid:15)(cid:1) *(cid:6)(cid:9)2(cid:1) (cid:4)3(cid:4)(cid:9)(cid:6)(cid:1) (cid:13)(cid:12)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1)
(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)7(cid:1) (cid:6).(cid:10)(cid:19)(cid:15)(cid:6)(cid:3)(cid:16)(cid:6)-(cid:1)(cid:10)(cid:15)(cid:1)3(cid:10)(cid:5)(cid:5)(cid:1)(cid:16)(cid:13)(cid:3)(cid:19)(cid:10)+(cid:6)(cid:9)(cid:1)(cid:15)(cid:28)(cid:6)(cid:14)(cid:1)(cid:4)(cid:19)(cid:1)(cid:25)(cid:4)(cid:9)(cid:15)(cid:1)(cid:13)(cid:12)(cid:1)(cid:10)(cid:15)(cid:19)(cid:1) (cid:6)(cid:3)*(cid:10)(cid:9)(cid:13)(cid:3)(cid:14)(cid:6)(cid:3)(cid:15)(cid:1)(cid:4)(cid:3)+(cid:1)
(cid:15)(cid:28)(cid:6)(cid:10)(cid:9)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:1) 3(cid:10)(cid:5)(cid:5)(cid:1) ,(cid:6)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)0(cid:9)(cid:4)(cid:15)(cid:6)+(cid:1) (cid:10)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:6)+(cid:1) (cid:28)2(cid:25)(cid:13)(cid:15)(cid:28)(cid:6)(cid:19)(cid:10)(cid:19)(cid:7)(cid:1) (cid:24)(cid:13)(cid:13)(cid:9)+(cid:10)(cid:3)(cid:4)(cid:15)(cid:10)(cid:3)0(cid:1) (cid:4)(cid:1)
0(cid:9)(cid:13)(cid:26)(cid:25)(cid:1),(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:1)(cid:19)(cid:6)(cid:6)(cid:14)(cid:19)(cid:1)(cid:15)(cid:13)(cid:1),(cid:6)(cid:1)(cid:25)(cid:13)(cid:19)(cid:19)(cid:10),(cid:5)(cid:6)(cid:1),2(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1)(cid:13)(cid:12)(cid:1)(cid:4)(cid:1)(cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:7)(cid:1)(cid:1)
<(cid:3)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1)(cid:19)(cid:10)+(cid:6)-(cid:1)(cid:15)(cid:28)(cid:6)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1)(cid:13)(cid:12)(cid:1)(cid:4)(cid:1)(cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)3(cid:10)(cid:5)(cid:5)(cid:1)(cid:3)(cid:13)(cid:15)(cid:1)(cid:4)(cid:5)3(cid:4)2(cid:19)(cid:1) (cid:5)(cid:6)(cid:4)+(cid:1)
(cid:15)(cid:13)(cid:1) ,(cid:6)(cid:19)(cid:15)(cid:1) (cid:9)(cid:6)(cid:19)(cid:26)(cid:5)(cid:15)(cid:19)(cid:1) 3(cid:10)(cid:15)(cid:28)(cid:10)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:13)(cid:1) (cid:10)(cid:3)(cid:16)(cid:9)(cid:6)(cid:4)(cid:19)(cid:6)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:6)(cid:12)(cid:12)(cid:10)(cid:16)(cid:10)(cid:6)(cid:3)(cid:16)2(cid:7)(cid:1)
(cid:29)(cid:28)(cid:6)(cid:9)(cid:6)(cid:1)(cid:10)(cid:19)(cid:1)(cid:4)(cid:1)+(cid:10)(cid:12)(cid:12)(cid:6)(cid:9)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:13)(cid:12)(cid:1)(cid:16)(cid:13)(cid:3)(cid:19)(cid:16)(cid:10)(cid:13)(cid:26)(cid:19)(cid:3)(cid:6)(cid:19)(cid:19)(cid:1)(cid:13)(cid:12)(cid:1)(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)-(cid:1)(cid:4)(cid:3)+(cid:1)(cid:16)(cid:13)(cid:13)(cid:9)+(cid:10)(cid:3)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)-(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)
;(cid:26)(cid:6)(cid:19)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:4)(cid:9)(cid:10)(cid:19)(cid:10)(cid:3)0(cid:1) (cid:10)(cid:12)(cid:1) (cid:28)(cid:10)0(cid:28)(cid:6)(cid:9)(cid:1) (cid:5)(cid:6)*(cid:6)(cid:5)(cid:1) (cid:5)(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) 3(cid:10)(cid:5)(cid:5)(cid:1) (cid:10)(cid:14)(cid:25)(cid:5)(cid:10)(cid:16)(cid:10)(cid:15)(cid:5)2(cid:1) (cid:5)(cid:6)(cid:4)+(cid:1) (cid:15)(cid:13)(cid:1) (cid:28)(cid:10)0(cid:28)(cid:6)(cid:9)(cid:1)
(cid:25)(cid:6)(cid:9)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:3)(cid:16)(cid:6)(cid:19)(cid:7)(cid:1)
=(cid:10)(cid:19)(cid:15)(cid:9)(cid:10),(cid:26)(cid:15)(cid:6)+(cid:1)(cid:2)(cid:9)(cid:15)(cid:10)(cid:12)(cid:10)(cid:16)(cid:10)(cid:4)(cid:5)(cid:1)(cid:11)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)>=(cid:2)(cid:11)?(cid:1)(cid:10)(cid:19)(cid:1)(cid:4)(cid:1) (cid:19)(cid:26),+(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:1)(cid:13)(cid:12)(cid:1)(cid:4)(cid:9)(cid:15)(cid:10)(cid:12)(cid:10)(cid:16)(cid:10)(cid:4)(cid:5)(cid:1)
(cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:16)(cid:6)(cid:1) >(cid:2)(cid:11)?-(cid:1) (cid:16)(cid:13)(cid:3)(cid:16)(cid:6)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:16)(cid:13)(cid:3)(cid:19)(cid:10)(cid:19)(cid:15)(cid:1) (cid:10)(cid:3)(cid:1) (cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1) (cid:10)(cid:3)+(cid:6)(cid:25)(cid:6)(cid:3)+(cid:6)(cid:3)(cid:15)(cid:1)
(cid:6)(cid:3)(cid:15)(cid:10)(cid:15)(cid:10)(cid:6)(cid:19)(cid:1) 3(cid:28)(cid:10)(cid:16)(cid:28)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:1) (cid:10)(cid:3)(cid:1) (cid:4)(cid:1) (cid:16)(cid:6)(cid:9)(cid:15)(cid:4)(cid:10)(cid:3)(cid:1) +(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:7)(cid:1) (cid:2)(cid:15)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) ,(cid:6)0(cid:10)(cid:3)(cid:3)(cid:10)(cid:3)0-(cid:1) =(cid:2)(cid:11)(cid:1) 3(cid:4)(cid:19)(cid:1)
+(cid:10)*(cid:10)+(cid:6)+(cid:1)(cid:10)(cid:3)(cid:15)(cid:13)(cid:1)(cid:15)3(cid:13)(cid:1)(cid:19)(cid:26),+(cid:10)(cid:19)(cid:16)(cid:10)(cid:25)(cid:5)(cid:10)(cid:3)(cid:6)(cid:19)@(cid:1)+(cid:10)(cid:19)(cid:15)(cid:9)(cid:10),(cid:26)(cid:15)(cid:6)+(cid:1)(cid:25)(cid:9)(cid:13),(cid:5)(cid:6)(cid:14)(cid:1)(cid:19)(cid:13)(cid:5)*(cid:10)(cid:3)0(cid:1)>=A(cid:8)?-(cid:1)(cid:12)(cid:13)(cid:16)(cid:26)(cid:19)(cid:6)+(cid:1)
(cid:13)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:14)(cid:4)(cid:3)(cid:4)0(cid:6)(cid:14)(cid:6)(cid:3)(cid:15)(cid:1) (cid:13)(cid:12)(cid:1) (cid:10)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) 3(cid:10)(cid:15)(cid:28)(cid:10)(cid:3)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:15)(cid:28)(cid:4)(cid:15)(cid:1) (cid:16)(cid:13)(cid:3)(cid:15)(cid:4)(cid:10)(cid:3)(cid:1) (cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1)
(cid:19)(cid:26),(cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) 3(cid:13)(cid:9)5(cid:10)(cid:3)0(cid:1) (cid:15)(cid:13)0(cid:6)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) (cid:4)(cid:1) (cid:16)(cid:6)(cid:9)(cid:15)(cid:4)(cid:10)(cid:3)(cid:1) 0(cid:13)(cid:4)(cid:5)(cid:1) (cid:4)(cid:3)+(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)
>& (cid:2)(cid:8)?-(cid:1) (cid:9)(cid:6)(cid:12)(cid:6)(cid:9)(cid:9)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:13)(cid:1) (cid:14)(cid:4)(cid:3)(cid:4)0(cid:10)(cid:3)0(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) ,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:1) (cid:13)(cid:12)(cid:1) (cid:4)(cid:1) (cid:16)(cid:13)(cid:5)(cid:5)(cid:6)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:13)(cid:12)(cid:1) (cid:10)(cid:3)+(cid:6)(cid:25)(cid:6)(cid:3)+(cid:6)(cid:3)(cid:15)(cid:1)
(cid:6)(cid:3)(cid:15)(cid:10)(cid:15)(cid:10)(cid:6)(cid:19)(cid:1)(cid:13)(cid:9)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:7)(cid:1)(cid:1)
& (cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) >& (cid:2)(cid:8)?(cid:1) (cid:9)(cid:6)(cid:25)(cid:9)(cid:6)(cid:19)(cid:6)(cid:3)(cid:15)(cid:1) (cid:4)(cid:1) (cid:19)(cid:26),+(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:1) (cid:13)(cid:12)(cid:1) (cid:4)(cid:9)(cid:15)(cid:10)(cid:12)(cid:10)(cid:16)(cid:10)(cid:4)(cid:5)(cid:1)
(cid:10)(cid:3)(cid:15)(cid:6)(cid:5)(cid:5)(cid:10)0(cid:6)(cid:3)(cid:16)(cid:6)(cid:1) 3(cid:28)(cid:10)(cid:16)(cid:28)(cid:1) (cid:13)(cid:12)(cid:12)(cid:6)(cid:9)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:25)(cid:9)(cid:10)(cid:3)(cid:16)(cid:10)(cid:25)(cid:5)(cid:6)(cid:19)(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) ,(cid:26)(cid:10)(cid:5)+(cid:10)(cid:3)0(cid:1) (cid:16)(cid:13)(cid:14)(cid:25)(cid:5)(cid:6).(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:15)(cid:28)(cid:4)(cid:15)(cid:1)
(cid:10)(cid:14)(cid:25)(cid:5)2(cid:1) (cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1)(cid:4)(cid:3)+(cid:1)(cid:14)(cid:6)(cid:16)(cid:28)(cid:4)(cid:3)(cid:10)(cid:19)(cid:14)(cid:19)(cid:1)(cid:12)(cid:13)(cid:9)(cid:1)(cid:16)(cid:13)(cid:13)(cid:9)+(cid:10)(cid:3)(cid:4)(cid:15)(cid:10)(cid:3)0(cid:1)(cid:10)(cid:3)+(cid:6)(cid:25)(cid:6)(cid:3)+(cid:6)(cid:3)(cid:15)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)7(cid:1)
,(cid:6)(cid:28)(cid:4)*(cid:10)(cid:13)(cid:9)(cid:7)(cid:1) (cid:2)(cid:19)(cid:1) (cid:15)(cid:28)(cid:6)(cid:9)(cid:6)(cid:1) (cid:10)(cid:19)(cid:1) (cid:3)(cid:13)(cid:1) 0(cid:6)(cid:3)(cid:6)(cid:9)(cid:4)(cid:5)(cid:5)2(cid:1) (cid:4)(cid:16)(cid:16)(cid:6)(cid:25)(cid:15)(cid:6)+(cid:1) +(cid:6)(cid:12)(cid:10)(cid:3)(cid:10)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)-(cid:1) (cid:10)(cid:15)(cid:1) (cid:10)(cid:19)(cid:1)
(cid:16)(cid:13)(cid:3)(cid:19)(cid:10)+(cid:6)(cid:9)(cid:6)+(cid:1)(cid:15)(cid:28)(cid:4)(cid:15)(cid:1)(cid:4)(cid:3)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:10)(cid:19)(cid:1)(cid:4)(cid:3)(cid:1)(cid:6)(cid:3)(cid:15)(cid:10)(cid:15)2-(cid:1)(cid:4)(cid:1) (cid:19)(cid:13)(cid:9)(cid:15)(cid:1)(cid:13)(cid:12)(cid:1) (cid:9)(cid:13),(cid:13)(cid:15)-(cid:1)3(cid:10)(cid:15)(cid:28)(cid:1)0(cid:13)(cid:4)(cid:5)(cid:19)-(cid:1)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1)(cid:4)(cid:3)+(cid:1)
5(cid:3)(cid:13)3(cid:5)(cid:6)+0(cid:6)(cid:1)(cid:10)(cid:3)(cid:1)(cid:4)(cid:1)(cid:16)(cid:6)(cid:9)(cid:15)(cid:4)(cid:10)(cid:3)(cid:1)+(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)-(cid:1)(cid:5)(cid:13)(cid:16)(cid:4)(cid:15)(cid:6)+(cid:1)3(cid:10)(cid:15)(cid:28)(cid:10)(cid:3)(cid:1)(cid:4)(cid:1)(cid:16)(cid:6)(cid:9)(cid:15)(cid:4)(cid:10)(cid:3)(cid:1)(cid:6)(cid:3)*(cid:10)(cid:9)(cid:13)(cid:3)(cid:14)(cid:6)(cid:3)(cid:15)(cid:7)(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:30)(cid:27)(cid:30)(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:4)(cid:1)(cid:11)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:16)(cid:17)(cid:7)(cid:1)(cid:18)(cid:13)(cid:5)(cid:7)(cid:1)(cid:11)(cid:18)(cid:1)(cid:12)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:11)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:2)(cid:3)(cid:3)(cid:4)(cid:5)(cid:19)(cid:7)(cid:1)(cid:24)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:1)(cid:8)(cid:16)(cid:10)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:6)(cid:19)(cid:7)(cid:1)(cid:27)(cid:15)(cid:28)(cid:1)(cid:29)(cid:13)(cid:14)(cid:6)(cid:1)(cid:30)(cid:19)(cid:15)(cid:1)(cid:31)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:8)(cid:6)(cid:3)((cid:13)(cid:9)(cid:19)(cid:1)
#(cid:12)(cid:12)(cid:6)(cid:16)(cid:15)(cid:13)(cid:9)(cid:19)(cid:1)
B(cid:13)(cid:4)(cid:5)(cid:19)(cid:1)
(cid:2)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1)
=(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:1)
C(cid:3)(cid:13)3(cid:5)(cid:6)+0(cid:6)(cid:1)
(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:6)(cid:7)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:1)(cid:8)(cid:5)(cid:7)(cid:4)(cid:9)(cid:1)(cid:10)(cid:11)(cid:10)(cid:9)(cid:7)(cid:12)(cid:1)
(cid:2)0(cid:6)(cid:3)(cid:15)(cid:1)
(cid:1)
(cid:1)
=(cid:26)(cid:6)(cid:1) (cid:15)(cid:13)(cid:1) (cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)7(cid:1) (cid:16)(cid:13)(cid:14)(cid:25)(cid:5)(cid:6).(cid:10)(cid:15)2-(cid:1) (cid:15)(cid:28)(cid:6)(cid:9)(cid:6)(cid:1) (cid:28)(cid:4)*(cid:6)(cid:1) (cid:15)(cid:13)(cid:1) ,(cid:6)(cid:1) (cid:12)(cid:13)(cid:26)(cid:3)+(cid:1)
(cid:19)(cid:13)(cid:5)(cid:26)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1) (cid:12)(cid:13)(cid:9)(cid:1) (cid:26)(cid:19)(cid:10)(cid:3)0(cid:1) & (cid:4)(cid:16)(cid:28)(cid:10)(cid:3)(cid:6)(cid:1) ’(cid:6)(cid:4)(cid:9)(cid:3)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:6)(cid:16)(cid:28)(cid:3)(cid:13)(cid:5)(cid:13)0(cid:10)(cid:6)(cid:19)(cid:1) (cid:15)(cid:13)(cid:1) (cid:14)(cid:4)(cid:3)(cid:4)0(cid:6)(cid:1) (cid:15)(cid:28)(cid:10)(cid:19)(cid:1)
(cid:16)(cid:13)(cid:14)(cid:25)(cid:5)(cid:6).(cid:10)(cid:15)2(cid:7)(cid:1)6(cid:6)(cid:12)(cid:13)(cid:9)(cid:6)(cid:1) (cid:19)(cid:15)(cid:26)+2(cid:10)(cid:3)0(cid:1)(cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:4)(cid:5)(cid:15)(cid:6)(cid:9)(cid:3)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1) (cid:19)(cid:28)(cid:13)(cid:26)(cid:5)+(cid:1),(cid:6)(cid:1)(cid:15)(cid:4)5(cid:6)(cid:3)(cid:1)(cid:10)(cid:3)(cid:15)(cid:13)(cid:1)
(cid:16)(cid:13)(cid:3)(cid:19)(cid:10)+(cid:6)(cid:9)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)-(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:4)(cid:5)(cid:15)(cid:6)(cid:9)(cid:3)(cid:4)(cid:15)(cid:10)*(cid:6)(cid:1),(cid:6)(cid:10)(cid:3)0(cid:1)(cid:16)(cid:6)(cid:3)(cid:15)(cid:9)(cid:4)(cid:5)(cid:10)((cid:6)+(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)3(cid:10)(cid:15)(cid:28)(cid:1)(cid:4)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:7)(cid:1)
(cid:29)(cid:28)(cid:6)(cid:1)(cid:16)(cid:6)(cid:3)(cid:15)(cid:9)(cid:4)(cid:5)(cid:10)((cid:6)+(cid:1)(cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)(cid:28)(cid:4)*(cid:6)(cid:1)(cid:4)(cid:1)(cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:15)(cid:28)(cid:4)(cid:15)(cid:1)(cid:10)(cid:19)(cid:1)(cid:14)(cid:4)5(cid:10)(cid:3)0(cid:1)(cid:4)(cid:5)(cid:5)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)+(cid:6)(cid:16)(cid:10)(cid:19)(cid:10)(cid:13)(cid:3)(cid:19)-(cid:1)
3(cid:28)(cid:10)(cid:5)(cid:6)(cid:1) (cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:19)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1) (cid:13)(cid:3)(cid:5)2(cid:1) (cid:28)(cid:6)(cid:5)(cid:25)(cid:10)(cid:3)0(cid:7)(cid:1) (cid:8)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1) (cid:14)(cid:26)(cid:19)(cid:15)(cid:1) ,(cid:6)(cid:1) (cid:4)(cid:16)(cid:16)(cid:6)(cid:25)(cid:15)(cid:6)+(cid:1) (cid:4)(cid:19)(cid:1)
(cid:16)(cid:6)(cid:3)(cid:15)(cid:9)(cid:4)(cid:5)(cid:10)((cid:6)+(cid:1)(cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)(cid:12)(cid:9)(cid:13)(cid:14)(cid:1)(cid:4)(cid:1)+(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:1)(cid:15)(cid:28)(cid:4)(cid:15)(cid:1)(cid:4)(cid:5)(cid:19)(cid:13)(cid:1)(cid:4)(cid:5)(cid:5)(cid:13)3(cid:19)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:4)(cid:25)(cid:25)(cid:9)(cid:13)(cid:4)(cid:16)(cid:28)(cid:7)(cid:1)
(cid:2)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:1)(cid:16)(cid:4)(cid:3)(cid:1)(cid:28)(cid:4)*(cid:6)(cid:1) (cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1)(cid:10)(cid:3)(cid:25)(cid:26)(cid:15)(cid:1)+(cid:4)(cid:15)(cid:4)(cid:1)(cid:4)(cid:3)+(cid:1) (cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1)(cid:4)(cid:16)(cid:15)(cid:13)(cid:9)(cid:19)(cid:7)(cid:1)6(cid:26)(cid:15)(cid:1)(cid:10)(cid:12)(cid:1)
(cid:6)(cid:4)(cid:16)(cid:28)(cid:1) (cid:6)(cid:3)(cid:15)(cid:10)(cid:15)2(cid:1) (cid:10)(cid:19)(cid:1) (cid:15)(cid:9)(cid:4)(cid:3)(cid:19)(cid:14)(cid:10)(cid:15)(cid:15)(cid:10)(cid:3)0(cid:1) (cid:25)(cid:6)(cid:9)(cid:16)(cid:6)(cid:25)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1) (cid:15)(cid:13)3(cid:4)(cid:9)+(cid:19)(cid:1) (cid:4)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1) (cid:16)(cid:13)(cid:9)(cid:6)(cid:1) (cid:25)(cid:9)(cid:13)(cid:16)(cid:6)(cid:19)(cid:19)(cid:1) (cid:4)(cid:3)+(cid:1)
(cid:9)(cid:6)(cid:16)(cid:6)(cid:10)*(cid:6)(cid:19)(cid:1)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1)(cid:12)(cid:9)(cid:13)(cid:14)(cid:1)(cid:10)(cid:15)-(cid:1)(cid:15)(cid:28)(cid:6)(cid:3)(cid:1)(cid:15)(cid:28)(cid:6)(cid:9)(cid:6)(cid:1)(cid:10)(cid:19)(cid:1)(cid:13)(cid:3)(cid:5)2(cid:1)(cid:13)(cid:3)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1)(cid:9)(cid:6)(cid:25)(cid:9)(cid:6)(cid:19)(cid:6)(cid:3)(cid:15)(cid:6)+(cid:1),2(cid:1)(cid:15)(cid:28)(cid:6)(cid:1)(cid:16)(cid:13)(cid:9)(cid:6)(cid:1)
(cid:25)(cid:9)(cid:13)(cid:16)(cid:6)(cid:19)(cid:19)(cid:7)(cid:1) : (cid:10)(cid:15)(cid:28)(cid:10)(cid:3)(cid:1) (cid:4)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:25)(cid:9)(cid:13)(cid:16)(cid:6)(cid:19)(cid:19)-(cid:1) (cid:10)(cid:15)(cid:1) (cid:10)(cid:19)(cid:1) (cid:14)(cid:13)+(cid:6)(cid:5)(cid:10)(cid:3)0(cid:1) (cid:10)(cid:15)(cid:19)(cid:6)(cid:5)(cid:12)-(cid:1) (cid:14)(cid:13)+(cid:6)(cid:5)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:28)(cid:6)(cid:1)
(cid:6)(cid:3)*(cid:10)(cid:9)(cid:13)(cid:3)(cid:14)(cid:6)(cid:3)(cid:15)(cid:1)(cid:4)(cid:3)+(cid:1)(cid:10)(cid:15)(cid:19)(cid:1)(cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1)3(cid:10)(cid:15)(cid:28)(cid:1)(cid:10)(cid:15)(cid:7)(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)(cid:30)(cid:27)(cid:21)(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:4)(cid:1)(cid:11)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:16)(cid:17)(cid:7)(cid:1)(cid:18)(cid:13)(cid:5)(cid:7)(cid:1)(cid:11)(cid:18)(cid:1)(cid:12)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:11)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:2)(cid:3)(cid:3)(cid:4)(cid:5)(cid:19)(cid:7)(cid:1)(cid:24)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:1)(cid:8)(cid:16)(cid:10)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:6)(cid:19)(cid:7)(cid:1)(cid:27)(cid:15)(cid:28)(cid:1)(cid:29)(cid:13)(cid:14)(cid:6)(cid:1)(cid:30)(cid:19)(cid:15)(cid:1)(cid:31)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
#(cid:3)*(cid:10)(cid:9)(cid:13)(cid:3)(cid:14)(cid:6)(cid:3)(cid:15)(cid:1)
(cid:8)(cid:6)(cid:3)((cid:13)(cid:9)(cid:1)
#(cid:12)(cid:12)(cid:6)(cid:16)(cid:15)(cid:13)(cid:9)(cid:1)
B(cid:13)(cid:4)(cid:5)(cid:19)(cid:1)
(cid:2)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1)
=(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:1)
C(cid:3)(cid:13)3(cid:5)(cid:6)+0(cid:6)(cid:1)
(cid:1)
(cid:2)0(cid:6)(cid:3)(cid:15)(cid:1)
(cid:8)(cid:6)(cid:3)((cid:13)(cid:9)(cid:1)
#(cid:12)(cid:12)(cid:6)(cid:16)(cid:15)(cid:13)(cid:9)(cid:1)
(cid:1)
B(cid:13)(cid:4)(cid:5)(cid:19)(cid:1)
(cid:2)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1)
=(cid:13)(cid:14)(cid:4)(cid:10)(cid:3)(cid:1)
C(cid:3)(cid:13)3(cid:5)(cid:6)+0(cid:6)(cid:1)
(cid:1)
(cid:2)0(cid:6)(cid:3)(cid:15)(cid:1)
(cid:2)(cid:3)(cid:4)(cid:5)(cid:8)(cid:7)(cid:1)(cid:13)(cid:14)(cid:6)(cid:9)(cid:3)(cid:8)(cid:5)(cid:7)(cid:4)(cid:9)(cid:1)(cid:10)(cid:11)(cid:10)(cid:9)(cid:7)(cid:12)(cid:1)
(cid:1)
(cid:1)
(cid:29)(cid:28)(cid:6)(cid:1)(cid:14)(cid:26)(cid:5)(cid:15)(cid:10)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)(cid:1)(cid:4)(cid:9)(cid:6)(cid:1)+(cid:10)(cid:12)(cid:12)(cid:6)(cid:9)(cid:6)(cid:3)(cid:15)(cid:1)(cid:12)(cid:9)(cid:13)(cid:14)(cid:1)(cid:15)(cid:28)(cid:6)(cid:1) (cid:19)(cid:10)(cid:3)0(cid:5)(cid:6)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)-(cid:1)
(cid:4)(cid:3)+(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) +(cid:10)(cid:12)(cid:12)(cid:6)(cid:9)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1) (cid:16)(cid:13)(cid:3)(cid:19)(cid:10)(cid:19)(cid:15)(cid:19)(cid:1) (cid:10)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:6).(cid:10)(cid:19)(cid:15)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1) (cid:13)(cid:12)(cid:1) (cid:19)(cid:6)*(cid:6)(cid:9)(cid:4)(cid:5)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) 3(cid:28)(cid:10)(cid:16)(cid:28)(cid:1) (cid:4)(cid:9)(cid:6)(cid:1)
(cid:14)(cid:13)+(cid:6)(cid:5)(cid:10)(cid:3)0(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) 0(cid:13)(cid:4)(cid:5)(cid:19)(cid:1) (cid:4)(cid:3)+(cid:1) (cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:19)(cid:1) ,(cid:6)(cid:15)3(cid:6)(cid:6)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:14)(cid:19)(cid:6)(cid:5)*(cid:6)(cid:19)(cid:7)(cid:1) 6(cid:6)(cid:15)3(cid:6)(cid:6)(cid:3)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1)
(cid:15)(cid:28)(cid:6)(cid:9)(cid:6)(cid:1) (cid:10)(cid:19)(cid:1) (cid:26)(cid:19)(cid:26)(cid:4)(cid:5)(cid:5)2(cid:1) (cid:4)(cid:1) +(cid:10)(cid:9)(cid:6)(cid:16)(cid:15)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)-(cid:1) ,(cid:26)(cid:15)(cid:1) (cid:4)(cid:5)(cid:15)(cid:28)(cid:13)(cid:26)0(cid:28)(cid:1) (cid:15)(cid:28)(cid:10)(cid:19)(cid:1) (cid:10)(cid:3)(cid:15)(cid:6)(cid:9)(cid:4)(cid:16)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) (cid:10)(cid:19)(cid:1) (cid:19)(cid:6)(cid:6)(cid:3)(cid:1) (cid:4)(cid:19)(cid:1)
(cid:19)(cid:15)(cid:10)(cid:14)(cid:26)(cid:5)(cid:10)(cid:1) (cid:13)(cid:12)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:6)(cid:3)*(cid:10)(cid:9)(cid:13)(cid:3)(cid:14)(cid:6)(cid:3)(cid:15)-(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:16)(cid:13)(cid:14)(cid:14)(cid:26)(cid:3)(cid:10)(cid:16)(cid:4)(cid:15)(cid:10)(cid:13)(cid:3)(cid:1) ,(cid:6)(cid:15)3(cid:6)(cid:6)(cid:3)(cid:1) (cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:1) (cid:10)(cid:19)(cid:1) +(cid:10)(cid:19)(cid:15)(cid:10)(cid:3)(cid:16)(cid:15)(cid:1)
(cid:12)(cid:9)(cid:13)(cid:14)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) (cid:6)(cid:3)*(cid:10)(cid:9)(cid:13)(cid:3)(cid:14)(cid:6)(cid:3)(cid:15)(cid:7)(cid:1) (cid:11)(cid:3)(cid:1) (cid:16)(cid:4)(cid:19)(cid:6)(cid:1) (cid:13)(cid:12)(cid:1) (cid:15)(cid:28)(cid:6)(cid:19)(cid:6)(cid:1) (cid:19)2(cid:19)(cid:15)(cid:6)(cid:14)(cid:19)-(cid:1) (cid:15)(cid:28)(cid:6)(cid:1) +2(cid:3)(cid:4)(cid:14)(cid:10)(cid:16)(cid:1) (cid:13)(cid:12)(cid:1) (cid:15)(cid:28)(cid:6)(cid:1)
(cid:6)(cid:3)*(cid:10)(cid:9)(cid:13)(cid:3)(cid:14)(cid:6)(cid:3)(cid:15)(cid:1)(cid:16)(cid:4)(cid:3)(cid:1),(cid:6)(cid:1)+(cid:6)(cid:12)(cid:10)(cid:3)(cid:6)+(cid:1),2(cid:1)(cid:13)(cid:15)(cid:28)(cid:6)(cid:9)(cid:1)(cid:4)0(cid:6)(cid:3)(cid:15)(cid:19)(cid:7)(cid:1)
(cid:1)
(cid:1)
$(cid:10)(cid:22)(cid:10)(cid:6)(cid:10)(cid:15)(cid:28)(cid:10)(cid:12)(cid:11)
D& (cid:10)(cid:15)!EF(cid:1) (cid:1) & (cid:10)(cid:15)(cid:16)(cid:28)(cid:6)(cid:5)(cid:5)(cid:1)(cid:29)(cid:7)(cid:1)(cid:20)(cid:1)(cid:13)(cid:8)(cid:15)(cid:16)(cid:3)(cid:4)(cid:7)(cid:1)(cid:17)(cid:7)(cid:8)(cid:18)(cid:4)(cid:3)(cid:4)(cid:5)-(cid:1)& (cid:16)B(cid:9)(cid:4)3(cid:20)G(cid:10)(cid:5)(cid:5)-(cid:1)(cid:30)!!E(cid:7)(cid:1)
(cid:1)
DA(cid:13)(cid:25)(cid:22)(cid:27)F(cid:1) A(cid:13)(cid:25)(cid:1) G(cid:7)(cid:31)(cid:7)(cid:1) (cid:20)(cid:1) (cid:2)(cid:3)(cid:10)(cid:9)(cid:7)(cid:12)(cid:7)(cid:1) (cid:3)(cid:4)(cid:9)(cid:7)(cid:6)(cid:3)(cid:5)(cid:7)(cid:4)(cid:9)(cid:7)(cid:1) (cid:19)(cid:4)(cid:1) (cid:20)(cid:18)(cid:21)(cid:22)(cid:6)(cid:7)(cid:12)(cid:7)(cid:1) (cid:23)(cid:7)(cid:1) (cid:15)(cid:6)(cid:8)(cid:10)(cid:3)(cid:24)(cid:3)(cid:15)(cid:8)(cid:18)(cid:7)-(cid:1)
& (cid:6)+(cid:10)(cid:4)(cid:14)(cid:10)(cid:9)(cid:4)(cid:1)A(cid:26),(cid:5)(cid:10)(cid:19)(cid:28)(cid:10)(cid:3)0(cid:1)(cid:24)(cid:13)(cid:7)-(cid:1)(cid:24)(cid:5)(cid:26)4(cid:20)H(cid:4)(cid:25)(cid:13)(cid:16)(cid:4)-(cid:1)(cid:21)(cid:22)(cid:22)(cid:27)(cid:1)
(cid:1)
(cid:1)
DI(cid:26)(cid:10)! F(cid:1)
(cid:1)
(cid:1)
I(cid:26)(cid:10)(cid:3)(cid:5)(cid:4)(cid:3)(cid:1) J(cid:7)"(cid:7)(cid:1) (cid:20)(cid:1) (cid:25)(cid:26)(cid:27)(cid:28)(cid:27)(cid:29)(cid:1) (cid:30)(cid:18)(cid:21)(cid:5)(cid:18)(cid:8)(cid:12)(cid:10)(cid:1) (cid:24)(cid:21)(cid:18)(cid:1) (cid:13)(cid:8)(cid:15)(cid:16)(cid:3)(cid:4)(cid:7)(cid:1) (cid:17)(cid:7)(cid:8)(cid:18)(cid:4)(cid:3)(cid:4)(cid:5)-(cid:1) & (cid:13)(cid:9)0(cid:4)(cid:3)(cid:1)
C(cid:4)(cid:26)(cid:12)(cid:14)(cid:4)(cid:3)(cid:3)-(cid:1)(cid:8)(cid:4)(cid:3)(cid:1)& (cid:4)(cid:15)(cid:6)(cid:13)-(cid:1)(cid:24)(cid:4)(cid:5)(cid:10)(cid:12)(cid:13)(cid:9)(cid:3)(cid:10)(cid:4)(cid:7)(cid:1)(cid:30)!! (cid:1)
(cid:30)(cid:27) (cid:1)
(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:4)(cid:1)(cid:11)(cid:3)(cid:12)(cid:13)(cid:9)(cid:14)(cid:4)(cid:15)(cid:10)(cid:16)(cid:17)(cid:7)(cid:1)(cid:18)(cid:13)(cid:5)(cid:7)(cid:1)(cid:11)(cid:18)(cid:1)(cid:12)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:11)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:2)(cid:3)(cid:3)(cid:4)(cid:5)(cid:19)(cid:7)(cid:1)(cid:24)(cid:13)(cid:14)(cid:25)(cid:26)(cid:15)(cid:6)(cid:9)(cid:1)(cid:8)(cid:16)(cid:10)(cid:6)(cid:3)(cid:16)(cid:6)(cid:1)(cid:8)(cid:6)(cid:9)(cid:10)(cid:6)(cid:19)(cid:7)(cid:1)(cid:27)(cid:15)(cid:28)(cid:1)(cid:29)(cid:13)(cid:14)(cid:6)(cid:1)(cid:30)(cid:19)(cid:15)(cid:1)(cid:31)(cid:4)(cid:19)(cid:16)(cid:7)(cid:1)(cid:20)(cid:1)(cid:21)(cid:22)(cid:22)(cid:23)(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)
(cid:1)(cid:30)(cid:27)(cid:27)(cid:1)
|
1811.09306 | 1 | 1811 | 2018-11-22T21:04:52 | Second-Order Agents on Ring Digraphs | [
"cs.MA",
"eess.SY",
"math.OC"
] | The paper addresses the problem of consensus seeking among second-order linear agents interconnected in a specific ring topology. Unlike the existing results in the field dealing with one-directional digraphs arising in various cyclic pursuit algorithms or two-directional graphs, we focus on the case where some arcs in a two-directional ring graph are dropped in a regular fashion. The derived condition for achieving consensus turns out to be independent of the number of agents in a network. | cs.MA | cs | Second-Order Agents on Ring Digraphs
Sergei Parsegov
Pavel Chebotarev
Skolkovo Institute of Science and Technology;
V.A. Trapeznikov Institute of Control Sciences
V.A. Trapeznikov Institute of Control Sciences
of Russian Academy of Sciences;
of Russian Academy of Sciences
V.A. Kotelnikov Institute of Radioengineering and Electronics
Moscow, Russia
[email protected]
of Russian Academy of Sciences
Moscow, Russia
[email protected]
8
1
0
2
v
o
N
2
2
]
A
M
.
s
c
[
1
v
6
0
3
9
0
.
1
1
8
1
:
v
i
X
r
a
Abstract -- The paper addresses the problem of consensus
seeking among second-order linear agents interconnected in a
specific ring topology. Unlike the existing results in the field
dealing with one-directional digraphs arising in various cyclic
pursuit algorithms or two-directional graphs, we focus on the case
where some arcs in a two-directional ring graph are dropped in
a regular fashion. The derived condition for achieving consensus
turns out to be independent of the number of agents in a network.
Index Terms -- ring digraph, cyclic pursuit, second-order agents
I. INTRODUCTION
Simple averaging control laws based on local interactions
have paved the way to a new class of models in modern control
theory and, more widely,
in the interconnected dynamical
systems theory. Such systems consist of a large (as usual)
number of identical subsystems and are supposed to achieve
certain global goals. The subsystems, or agents, are coupled
in some way and therefore share an amount of common
information. During the last 15 years,
the complexity of
models of networked systems increased significantly starting
from a simple continuous-time consensus model (suprisingly
first proposed by sociologists in 1964 and rediscovered much
later), see [1] for details. It
to analyze the
evolution of these models considering separately three main
entities comprising a dynamical network: agent complexity,
interaction structure complexity, and link complexity. Recent
results and challeging problems in this field may be found
in the lecture course [2] and monographs [3] and [4]. In this
paper, we focus on specific interaction structures and analyze
their properties in the case of second-order agents.
is convenient
Control problems related to networks with specific com-
munication patterns play an important
role in coopera-
tive/decentralized control. Within this trend, it is supposed
that each agent
interacts with a predefined number of its
indexed neighbors. A pioneering work on consensus where
a simple averaging rule was proposed and thoroughly studied
was published in 1878 by J.G. Darboux [5]. Although this
problem dealt with the evolution of planar polygons, it turned
out to be the first theoretical result on discrete-time cyclic
The results were obtained with support of the Russian Science Foundation,
project no. 16-11-00063 granted to IRE RAS.
pursuit. This class of algorithms has a long history (see, e.g.,
[6], [7], [8], [9], and references therein), and has a wide
range of applications including but not
limited to numer-
ous formation control tasks as patrolling, boundary mapping,
etc. The cyclic pursuit is a strategy where agent i pursues
its neighbor i − 1 , while agent 1 pursues agent n, thus
the topology of communication is a Hamiltonian cycle. The
extensions to hierarchical structures are considered in [10]
and [11]; [12] addresses the case of heterogeneous agents;
geometrical problems related to cyclic pursuit-like algorithms
are investigated in [13] and [14]. Some pursuit algorithms
utilize the rotation operator in order to follow the desired
trajectories, see [15] and references therein. Another group of
strategies/protocols are based on two-directional topologies,
i.e., each agent i has the relative information of the neighbors
i − 1 and i + 1 (with 0 ≡ n). For example, in [16] the
agents are interconnected by a two-directional ring. The row
straightening problems studied in [17], [18], [19] also imply
symmetric communications except for the fixed "anchors"
(the endpoints of the segment). Theoretical motivation behind
studying regular network structures is that for some cases, this
may lead to the closed-form computation of the spectra of the
corresponding Laplacian matrices.
In what follows, we study the problem of reaching consen-
sus for second-order agents with velocity damping (friction).
The models of such kind naturally arise, e.g.,
in energy
systems [20] and formation control [2], [21], [22]. The com-
munication topology studied in the present paper is a digraph
G with a specific structure: it has n = 2m vertices, m ≥ 3,
and contains a Hamiltonian cycle supplemented by the inverse
cycle, where every second arc is dropped, see Fig. 1. In some
sense, this "intermediate" digraph with regular structure lies
between the two-directional ring and the Hamiltonian cycle
appearing in cyclic pursuit algorithms. The obtained results
fill the gap between undirected and directed ring topologies
for this special case: the loss of arcs may result in instability
of the whole system and therefore a challenging problem is to
derive a suitable condition on the tunable damping parameter
that guarantees convergence to consensus irrespective of the
number of agents in the network. Alternatively, one may
consider the opposite transformation, that is addition of arcs
to an unidirected communication structure in order to reduce
relatively high damping coefficients of agents.
The paper is organized as follows. The next section presents
the notations used in the paper. Section 3 introduces some
mathematical preliminaries needed for further exposition and
discusses the statement of the problem. The main theoretical
result is formulated and proved in Section 4. Finally, the results
of numerical simulation and conclusions are given.
II. NOTATIONS
are reserved for indices;
The following notation will be used throughout the paper:
√−1 denotes the imaginary unit, the letters i and k
• j :=
• the unit vector is defined by 1n := [1, 1, . . . , 1](cid:62) ∈ Rn;
• ⊗ denotes the Kronecker product of two matrices.
Below we consider some definitions and auxiliary lemmas
needed for further discussion and give an exact problem
formulation.
III. PRELIMINARIES AND PROBLEM STATEMENT
Throughout the paper, we consider a group of n identical
agents with a directed communication topology and suppose
that n = 2m, where m ≥ 3. Each agent obeys the second-
order dynamics of the form
i ∈ 1 : n,
vi = −γvi + ui,
(1)
where xi, vi, ui ∈ R are the position coordinate of the ith
agent, its velocity and control input, respectively; γ > 0
denotes the damping coefficient.
The state-space representation of system (1) writes as
(cid:26) xi = vi,
ξi = Aξi + Bui,
(cid:21)
(cid:20)0
1
0 −γ
where ξi = [xi, vi]
(cid:62), A =
, and B = [0, 1]
Suppose that the agents interact through the topology de-
picted in Fig. 1. Thus, each agent knows the relative distances
(2)
(cid:62).
These conditions turn out to be independent on the number of
agents comprising the network system.
The following definitions and results are used in the further
considerations.
A. Graph Theory
We suppose that the communication network is represented
by a fixed, directed graph G = (V,E), where V = {1, . . . , n}
denotes the vertex (or node) set and E is the set of arcs. We
also assume that the graph is simple, i.e., it has no self-loops,
and no multiple arcs. Let us define the Laplacian matrix of an
unweighted digraph G as follows:
Definition 1 ([23]). The Laplacian matrix of a digraph G is
the matrix L = (lik) ∈ Rn×n in which, for k (cid:54)= i, lik =
−1 whenever (i, k) ∈ E(G), otherwise lik = 0. The diagonal
k(cid:54)=i lik, i, k ∈ V(G).
Lemma 1 ([24], [25]). The Laplacian matrix L of a directed
graph G has at least one zero eigenvalue with 1n as a corre-
sponding right eigenvector and all nonzero eigenvalues have
positive real parts. Furthermore, zero is a simple eigenvalue
of L if and only if G has a spanning converging tree, i.e., it
has at least one vertex accessible from all other vertices.
entries of L are of the form lii = −(cid:80)
Obviously,
the digraph G depicted in Fig. 1 is simple,
unweighted, and contains a spanning converging tree.
The system of n agents (1) with feedback protocol (3) obeys
the following dynamics:
x + γ x = −Lx,
(5)
where x = [x1, . . . , xn](cid:62) and L is the Laplacian matrix
associated with the dependency digraph G:
0 −1
0
0
0
0
···
0
0
···
0
0
2 −1
···
...
...
...
...
...
0 −1
1
0
0
0 −1
2 −1
0
0 −1
1
0
0
2 −1
−1
1
0 −1
...
...
···
···
···
this closed-loop network system can be
(6)
L =
0
0
0
Also note that
.
equivalently described by
Fig. 1. The interaction topology between agents (1).
ξ = F ξ,
where F ∈ R2n×2n has the form
between itself and the nearest one or two indexed neighbors.
Let K ∈ R1×2 be the matrix K = [1, 0],
therefore the
decentralized control protocol has the form
(cid:26)K(ξi+1 − ξi) + K(ξi−1 − ξi), i is odd,
i ∈1 : n. (3)
ui =
K(ξi−1 − ξi),
i is even,
The main objective of the paper is to establish conditions for
a network of second-order agents (1), (2) governed by protocol
(3) that guarantee consensus in the sense of
t→∞(cid:107)ξi(t) − ξk(t)(cid:107) = 0, ∀i, k ∈ 1 : n.
lim
(4)
F = I ⊗ A − L ⊗ BK,
and I ∈ Rn×n denotes the identity matrix.
Although the consensus conditions (5), (7) can be verified in
a straightforward way, this may be computationally expensive
in the case of a large number of agents. The framework
presented below allows to reduce the problem to a couple
of simpler ones using the notion of consensus region. The
criterion for scalar agents was first proposed by Polyak and
Tsypkin in [26]; similar results were obtained later in [28] and
[27]. Some other extensions may be found in [3] and [4].
(7)
(8)
B. Consensus Region
Suppose that the consensus problem is studied for a net-
worked dynamical system described by the equation
φ(s)x = −Lx,
dt, and L is the
where φ(s) is a scalar polynomial, s := d
Laplacian matrix of the dependency digraph containing a
spanning converging tree.
Definition 2 ([4], [26], [27]). The Ω-region of the function
φ(s) is the set of points λ on the complex plane for which the
function φ(s) − λ has no zeros in the closed right half-plane:
Ω = {λ ∈ C : φ(s) − λ (cid:54)= 0 whenever Re(s) ≥ 0}.
Note that Ω is precisely the region of parameters λ such
that the matrix A − λBK is Hurwitz stable. The function
φ(s) is sometimes referred to as the generalized frequency
variable [27], [28].
Lemma 2 ( [4], [26], [27]). The system described by (2)
reaches consensus under protocol (3) if and only if
λi ∈ Ω,
i ∈ 2 : n,
where λi, i ∈ 2 : n, are the nonzero eigenvalues of −L.
The details of determining the consensus region may be
found in [26]; in the case of φ(s) = s2 + γs, γ > 0, this
region has form of the interior of a parabola in the complex
plane: φ(jω) = −ω2 + jγω, −∞ < ω < ∞.
C. Cyclotomic Equation
Lemma 3 ([29]). The roots of the cyclotomic equation
zm − 1 = 0
are the de Moivre numbers
zk = ej 2πk
m , k ∈ 0 : m − 1.
They form a regular polygon with each vertex lying on the
unit circle in the complex plane.
D. Cassini Ovals
Definition 3 ([30]). The Cassini curve (or the Cassini ovals)
is a quartic curve defined as the set of points in the plane such
that the product of the distances (denoted by b2) to two fixed
points (a, 0) and (−a, 0) is constant:
[(x − a)2 + y2][(x + a)2 + y2] = b4.
(9)
IV. MAIN RESULTS
Based on the above problem formulation and preliminary re-
sults, in this section we derive analytic conditions of achieving
consensus for network systems (5), (7). These conditions turn
out to be independent of the number of agents.
First, we study the spectrum of the Laplacian matrix de-
fined by (6) and obtain the curves which are the loci of its
eigenvalues.
(cid:113)
and its roots are
λ(1,2)
k
= 1.5 ± 0.5
Let the kth root of unity zk = ej 2πk
m − 1. Then
5 + 4ej 2πk
(12)
m .
m be ak + jbk, k ∈ 0 :
(cid:113)
√
Lemma 4. The eigenvalues of Laplacian matrix (6) have the
form
λ(1,2)
k
= 1.5 ± 0.5
5 + 4ej 2πk
m
and lie on the Cassini ovals defined by
√
[(x −
5)2 + y2] = 24,
5)2 + y2][(x +
where x = 2(x − 3/2), y = 2y.
Proof. According to Theorem 4 in [31] our graph is essentially
cyclic (i.e., the spectrum of L contains non-real eigenvalues)
and the characteristic polynomial of L is of the form
(10)
(Z2(λ))m − 1,
where Zn(λ) = (λ − 2)Zn−1(λ) − Zn−2(λ) is a modified
Chebyshev polynomial of the second kind with the initial
conditions Z0(λ) ≡ 1, Z1(λ) ≡ λ − 1, see [31] for details.
Armed with the knowledge gained from Theorem 4 in [31]
and Lemma 3 we find Z2 = λ2 − 3λ + 1, the characteristic
polynomial of L is
(λ2 − 3λ + 1)m − 1,
(11)
a2
k + b2
k = 1.
(13)
The 2m roots of (11) can be found from the equation
λ2 − 3λ + 1 − ak − jbk = 0, k = 0, 1, . . . , m − 1.
Let λ = x + jy. Then from
x + jy = 1.5 ± 0.5(cid:112)5 + 4ak + j4bk
we find ak = (x − 3/2)2 − y2 − 5/4 and bk = 2yx − 3y.
Taking into account (13) one arrives at
((x − 3/2)2 − y2 − 5/4)2 + 4y2(x − 3/2)2 = 1.
The last equation can be simply rewritten in the form of (10).
The eigenvalues of L with n = 16 and n = 44 are shown in
Fig. 2. The egg-shaped Cassini ovals can be recognized easily.
Corollary 1. The eigenvalues of −L are
(cid:113)
λ(1,2)
k
= −1.5 ± 0.5
5 + 4ej 2πk
m
(14)
and the equation of the corresponding Cassini curve is
((x + 3/2)2 − y2 − 5/4)2 + 4y2(x + 3/2)2 − 1 = 0.
(15)
Remark 1. According to consensus criterion [4], [26], [27]
for the second-order agents (1), the consensus is reached if
and only if all the nonzero eigenvalues of −L lie in the Ω-
region of φ(s) = s2 + γs. Checking this criterion requires
Fig. 2. The eigenvalues of L with n = 16 and n = 44.
computation of the spectrum. Below we propose a condition
independent of the number of agents. It is sufficient yet simple
and not that much conservative.
Theorem 1. The system of interconnected second-order agents
(7) reaches a consensus in the sense of (4) for all γ >
7 .
Proof. Since φ(jω) = −ω2 + jγω,
the equation can be
rewritten as y2 = −γ2x. Therefore, the consensus condition
can be reduced to the condition that the Cassini ovals described
by (15) in Corollary 1 belong to the interior of the parabola
y2 = −γ2x without intersection (except for the one at the
origin), see Fig. 3. Substituting this into (15) we arrive at the
(cid:113) 6
Fig. 4. Ω-region bounded by y2 = −γ2x and the Cassini ovals, γ = 0.7.
Remark 2. In this paper, we limit ourselves to the case of
m ≥ 3 due to the fact that for m = 1, the considered graph
does not exist and the case of m = 2 results in a real spectrum
of L [31] which implies reaching consensus for any γ > 0.
Remark 3. By virtue of the consensus region approach, a
similar condition for second-order agents whose dependency
digraph is a Hamiltonian cycle (Fig. 5) can be derived. In this
case, the corresponding Laplacian matrix is a circulant matrix
(see, e.g., [32] for details). Its eigenvalues are located on a unit
circle centered at (−1, j0). Recall that the consensus region
Fig. 5. Diraph of the cyclic pursuit.
Fig. 3. Ω-region bounded by y2 = −γ2x and the Cassini ovals, γ = 2.
equation in x with parameter γ:
x(x3 + (6− 2γ2)x2 + (γ4 − 6γ2 + 11)x + 6− 7γ2) = 0. (16)
The zero root is out of our interest, thus let us study the
properties of the cubic polynomial
x3 + (6 − 2p)x2 + (p2 − 6p + 11)x + 6 − 7p = 0,
(17)
where p = γ2. The case that determines the required condition
corresponds to one intersection, and therefore one real negative
root and a pair of roots with the same sign of their real parts.
According to Vieta's formulas, a product of the roots satisfies
x1x2x3 = 7p − 6 and is negative. Therefore, p > 6/7 implies
the absence of any intersection except for the one at (0, 0).
The consensus condition independent of the number of
agents follows immediately.
Fig. 6. Ω-region bounded by y2 = −γ2x and the unit circle, γ = 2.
defined by the transfer function of agent (1) is the interior
of a parabola. It can be verified that the equation for the
intersection writes as x2 + 2x − γ2x = 0, which implies that
the absence of intersection (except for the one at the origin)
holds true for γ >
2. This demonstrates that the consensus
√
condition for second-order agents (1) in cyclic pursuit can be
derived easier as compared to, e.g., [7].
The consensus condition for the case of undirected (or, in
terms of digraph, two-directional) ring topology is even sim-
pler. Since the spectrum of the Laplacian matrix is completely
real, the eigenvalues λ2, . . . , λn all belong to the Ω-region for
any γ > 0 providing consensus in the sense of (4).
V. NUMERICAL EXAMPLES
criterion.
First, we demonstrate the low conservatism of the proposed
The sufficient condition of Theorem 1 guarantees conver-
gence to consensus in the sense of (4) whenever γ >(cid:112)6/7.
with dampings γ ≤ (cid:112)6/7: if all the nonzero eigenvalues
However, small groups of agents may still reach consensus
λi, i ∈ 2 : n, of −L lie inside the Ω-region,
then the
conditions of the general criterion formulated in Lemma 2 are
satisfied (which implies that the corresponding eigenvalues λi,
i ∈ 2 : 2n, of the system matrix F represented by (8) lie in the
open left half-plane of the complex plane). This can be verified
as follows. Let us set γ =(cid:112)6/7 and compute the eigenvalues
of F for various numbers of agents n = 6, 8, 10, . . .. Taking
the maximum of their real parts we observe that this maximum
approaches zero as the number of agents increases, see Fig. 7.
Fig. 7. The values of max
i∈2:2n
(logarithmic scale); the damping coefficient is γ =(cid:112)6/7.
Re(λi) subject
to n = 6, 8, 10, . . . , 60
Now, to illustrate the dynamics of convergence to consensus,
consider a formation of n = 50 identical second-order agents
(1) linked by the topology depicted in Fig. 1. Let the initial
state of each agent be randomly generated within [0, 10].
System (7) exhibits stable behavior of the transients with
a damping coefficient γ = 2, see Fig. 8. Computation of the
eigenvalues of F yields max
i∈2:2n
stability margin γ =(cid:112)6/7 ≈ 0.926 presented by Theorem 1
In the case of smaller gain γ = 0.95 close enough to the
Re(λi) = −0.0032.
we still observe similar behavior, see Fig. 9.
Fig. 8. The trajectories of the system (5) with n = 50 agents, γ = 2.
Fig. 9. The trajectories of the system (5) with n = 50 agents, γ = 0.95.
The dynamics becomes different when the consensus con-
dition is violated. For example, for γ = 0.9 the transients
Re(λi) = 0.000633. Fig. 10 presents the results
diverge; max
i∈2:2n
of simulations for this unstable case.
Table 1 illustrates the dependence of the actual consensus
margin γ on the number of agents n.
TABLE I
CONSENSUS MARGINS
n
γ
10
0.8195
20
0.8999
30
0.9149
40
0.9195
50
0.9218
60
0.9230
Therefore,
the simulation results confirm the theoretical
conclusions of Theorem 1 and demonstrate low conservatism
of the provided sufficient condition of consensus for a large
number of agents.
[9] J. A. Marshall, M. E. Broucke and B. A. Francis,
"Formations of
Vehicles in Cyclic Pursuit," IEEE Transactions On Automatic Control,
vol. 49, no. 11, pp. 1963 -- 1974, 2004.
[10] S. L. Smith, M. E. Broucke, and B. A. Francis, "A Hierarchical Cyclic
Pursuit Scheme for Vehicle Networks," Automatica, vol. 41, no. 6, pp.
1045 -- 1053, 2005.
[11] D. Mukherjee and D. Ghose,"Generalized Hierarchical Cyclic Pursuit,"
Automatica, vol. 71, pp. 318 -- 323, 2016.
[12] A. Sinha and D. Ghose, "Generalization of Linear Cyclic Pursuit with
Application to Rendezvous of Multiple Autonomous Agents," IEEE
Transactions on Automatic Control, vol. 51, no. 11, pp. 1819 -- 1824,
2006.
[13] A. N. Elmachtoub and C. F. van Loan, "From Random Polygon to
Ellipse. An Eigenanalysis," SIAM Rev., vol. 52, no. 1, pp. 151 -- 170,
2010.
[14] P. S. Shcherbakov, "Formation Control: The Van Loan Scheme and Other
Algorithms," Autom. Remote Control, vol. 72, no. 10, pp. 2210 -- 2219,
2011.
[15] J. L. Ramirez-Riberos, M. Pavone, E. Frazzoli, and D. W. Miller, "Dis-
tributed Control of Spacecraft Formations via Cyclic Pursuit. Theory
and Experiments," AIAA J. Guidance, Control, Dynamics, vol. 33, no.
5, pp. 1655 -- 1669, 2010.
[16] D. Mukherjee and D. Zelazo, "Robust Consensus of Higher Order
Agents over Cycle Graphs," in Proc. of the 58th Israel Annual Con-
ference on Aerospace Sciences, pp. 1072 -- 1083, 2018.
[17] I. A. Wagner and A. M. Bruckstein, "Row Straightening via Local
Interactions," Circuits Syst. Signal Process., vol. 16, no. 2, pp. 287 --
305, 1997.
[18] Ya. I. Kvinto and S. E. Parsegov, "Equidistant Arrangement of Agents
on Line: Analysis of the Algorithm and Its Generalization," Autom.
Remote Control, vol. 73, no. 11, pp. 1784 -- 1793, 2012.
[19] A. V. Proskurnikov and S. E. Parsegov, "Problem of Uniform Deploy-
ment on a Line Segment for Second-Order Agents," Autom. Remote
Control, vol. 77, no. 7, pp. 1248 -- 1258, 2017.
[20] D. Goldin, "Double Integrator Consensus Systems with Application to
Power Systems," in Proc. 4th IFAC Workshop NecSys-2013, pp. 206 --
211, 2013.
[21] W. Ren and Y. C. Cao, Distributed Coordination of Multi-Agent Net-
works, Springer, London, 2011.
[22] W. Ren. "On Consensus Algorithms for Double-Integrator Dynamics,"
IEEE Trans. Autom. Control, vol. 53, no. 6, pp. 1503 -- 1509, 2008.
[23] P. Chebotarev and R. Agaev, "Forest Matrices around the Laplacian
Matrix," Linear Algebra and Its Applications, vol. 356, pp. 253 -- 274,
2002.
[24] R. P. Agaev and P. Y. Chebotarev, "The Matrix of Maximum Out Forests
of a Digraph and Its Applications," Autom. Remote Control, vol. 61, no.
9, pp. 1424 -- 1450, 2000.
[25] R. P. Agaev and P. Yu. Chebotarev, "Spanning Forests of a Digraph
and Their Applications," Autom. Remote Control, vol. 62, no. 3, pp.
443 -- 466, 2001.
[26] B. T. Polyak and Y. Z. Tsypkin,"Stability and Robust Stability of
Uniform Systems," Autom. Remote Control, vol. 57, no. 11, pp. 1606 --
1617, 1996.
[27] S. Hara, H. Tanaka, and T. Iwasaki, "Stability Analysis of Systems with
Generalized Frequency Variables," IEEE Trans. Autom. Control, vol. 59,
no. 2, pp. 313 -- 326, 2014.
[28] S. Hara, T. Hayakawa, and H. Sugata, "Stability Analysis of Linear
Systems with Generalized Frequency Variables and Its Applications to
Formation Control," in Proc. IEEE Conf. Decision Control, pp. 1459 --
1466, 2007.
[29] O. Neumann, "Cyclotomy: From Euler through Vandermonde to Gauss,"
in Leonhard Euler: Life, Work and Legacy, R. E. Bradley and C. E. San-
difer, Eds. Amsterdam: Elsevier, 2007, pp. 323 -- 362.
[30] J. D. Lawrence, A Catalog of Special Plane Curves, New York: Dover,
1972.
[31] R. P. Agaev and P. Yu. Chebotarev, "Which Digraphs with Ring Structure
are Essentially Cyclic?" Adv. Appl. Math. vol. 45, pp. 232 -- 251, 2010.
[32] D. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas, Prince-
ton University Press, 2009.
Fig. 10. The trajectories of the system (5) with n = 50 agents, γ = 0.9.
VI. CONCLUSIONS
The contribution of the paper is threefold:
• First, the specific communication ring topology was in-
vestigated; it was discovered that the eigenvalues of the
corresponding Laplacian matrix lie on the Cassini ovals;
• Second, the criterion for reaching consensus was pro-
posed that relies solely on the value of the damping
coefficient of a single agent and neglects the number of
them;
• Third, the theoretical results were supplemented by nu-
merical experiments. It was shown that for a large number
of interacting agents, the sufficient condition presented by
Theorem 1 is quite close to the necessary one.
The possible extensions include establishing consensus con-
ditions for second-order agents distributed on other regular
topologies together with the control protocols that use solely
local measurements. These problems will be the subject of
continuing research.
REFERENCES
[1] A. V. Proskurnikov and R. Tempo, "A Tutorial on Modeling and Analysis
of Dynamic Social Networks. Part I," Annual Reviews in Control, no.
43, pp. 65 -- 79, 2017.
[2] F. Bullo, Lectures on Network Systems (With contributions by J. Cort´es,
F. Dorfler, and S. Mart´ınez), 2018, http://motion.me.ucsb.edu/book-lns/
[3] F. L. Lewis, H. Zhang, K. Hengster-Movric, and A. Das, Cooperative
Control of Multi-Agent Systems: Optimal and Adaptive Design Ap-
proaches, London: Springer, 2014.
[4] Z. Li and Z. Duan, Cooperative Control of Multi-Agent Systems: A
Consensus Region Approach, CRC Press, 2017.
[5] J. G. Darboux, "Sur un Probl`eme de G´eom´etrie ´El´ementaire," Bulletin
des Sciences Math´ematiques et Astronomiques, vol. 2, no. 1, pp. 298 --
304, 1878.
[6] P. J. Nahin, Chases and Escapes: The Mathematics of Pursuit and
Evasion, Princeton University Press, 2007.
[7] B. R. Sharma, S. Ramakrishnan, and M. Kumar, "Cyclic Pursuit in a
Multi-Agent Robotic System with Double-Integrator Dynamics under
Linear Interactions," Robotica, vol. 31, no. 7, pp. 1037 -- 1050, 2013.
[8] Y. Elor and A. M. Bruckstein, "Uniform Multi-Agent Deployment on a
Ring," Theor. Comput. Sci., vol. 412, no. 8 -- 10, pp. 783 -- 795, 2011.
|
1204.1581 | 1 | 1204 | 2012-04-07T00:10:33 | A new approach of designing Multi-Agent Systems | [
"cs.MA",
"cs.AI"
] | Agent technology is a software paradigm that permits to implement large and complex distributed applications. In order to assist analyzing, conception and development or implementation phases of multi-agent systems, we've tried to present a practical application of a generic and scalable method of a MAS with a component-oriented architecture and agent-based approach that allows MDA to generate source code from a given model. We've designed on AUML the class diagrams as a class meta-model of different agents of a MAS. Then we generated the source code of the models developed using an open source tool called AndroMDA. This agent-based and evolutive approach enhances the modularity and genericity developments and promotes their reusability in future developments. This property distinguishes our design methodology of existing methodologies in that it is constrained by any particular agent-based model while providing a library of generic models | cs.MA | cs | (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
A new approach of designing Multi-Agent Systems
With a practical sample
Sara Maalal
Team of Systems‟ Architecture, Laboratory of computing,
Systems and Renewable Energy
National and High School of Electricity and Mechanic
ENSEM BP 8118, Oasis
Casablanca, Maroc
Abstract—Agent technology is a software paradigm that permits
to implement large and complex distributed applications [1]. In
order to assist analyzing, conception and development or
implementation phases of multi-agent systems, we’ve tried to
present a practical application of a generic and scalable method
of a MAS with a component-oriented architecture and agent-
based approach that allows MDA to generate source code from a
given model. We’ve designed on AUML the class diagrams as a
class meta-model of different agents of a MAS. Then we
generated the source code of the models developed using an open
source tool called AndroMDA. This agent-based and evolutive
approach enhances the modularity and genericity developments
and promotes their reusability in future developments. This
property distinguishes our design methodology of existing
methodologies in that it is constrained by any particular agent-
based model while providing a library of generic models [2].
Keyword- Software agents; Multi-agents Systems (MAS); Analysis;
Software design; Modeling; Models; Diagrams; Architecture;
Model Driven Architecture (MDA); Agent Unified Modeling
Language (AUML); Agent Modeling Language (AML).
I.
INTRODUCTION
Currently the computer systems are increasingly complex,
often distributed over several sites and consist of software
interacting with each other or with humans. The need for model
human behavior in specific computer programs has prompted
officials to use technology that affected the last decade and
whose movements are very remarkable. In this context,
designing multi-agent systems (MAS) is complex because they
require the inclusion of several parts of the system which can
often be approached from different angles. We must identify
and analyze all system problems to find models for multi-
agents to implement and integrate them into a coherent system.
This is the software engineering and well justifies the use of a
method of analysis, design and development of multi-agents
systems [2].
This paper describes a practical example of a new generic
model designed for modeling multi-agent systems and based on
a class diagram, defining the different types of agents and
meeting our needs for development and testing of MAS
applications.
Malika Addou
Team of Systems‟ Architecture Laboratory of computing,
Systems and Renewable Energy
Hassania School of Public
Works EHTP BP 8108, Oasis
Casablanca, Maroc
II. MULTI-AGENT SYSTEMS
A. Definitions
- An agent is a computer system within an environment and
with an autonomous behavior made for achieving
the
objectives that were set during its design [3].
- A multi-agents system is a system that contains a set of
agents that interact with communication s protocols and are able
to act on their environment. Different agents have different
spheres of influence, in the sense that they have control (or at
least can influence) on different parts of the environment.
These spheres of influence may overlap in some cases; the fact
that they coincide may cause dependencies reports between
agents [4].
The MAS can be used in several application areas such as
e-commerce, economic systems, distributed
information
systems, organizations...
B. Types of agent
Starting from the definitions cited above, we can identify
the following agent types [5]:
The reactive agent is often described as not being
"clever" by itself. It is a very simple component that
perceives the environment and is able to act on it. Its
capacity meets mode only stimulus-action that can be
considered a form of communication.
The cognitive agent is an agent more or less intelligent,
mainly characterized by a symbolic representation of
knowledge and mental concepts. It has a partial
representation of the environment, explicit goals, it is
capable of planning their behavior, remember his past
actions, communicate by sending messages, negotiate,
etc..
The intentional agent or BDI (Belief, Desire and
Intention) is an intelligent agent that applies the model
of human intelligence and human perspective on the
world using mental concepts such as knowledge,
beliefs, intentions, desires, choices, commitments. Its
behavior can be provided by the award of beliefs,
desires and intentions.
www.ijacsa.thesai.org
148 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
The rational agent is an agent that acts in a manner
allowing it to get the most success in achieving the
tasks they were assigned. To this end, we must have
measure of performance,
if possible objective
associated with a particular task that the agent should
run.
The AAII methodology was developed based on the
experience accumulated during the construction of BDI
systems. In this methodology, we have a set of
templates that, when they have been fully elaborated,
define the specifications of agents such as desires,
beliefs and intentions [9].
The adaptive agent is an agent that adapts to any
changes that the environment can have. He is very
intelligent as he is able to change its objectives and its
knowledge base when they change.
The communicative agent is an agent that is used to
communicate information to all around him. This
information can be made of his own perceptions as it
may be transmitted by other agents.
Figure 1. Types of agents
III. THE DESIGN METHODOLOGIES – STATE OF THE ART
Building high quality software for real-world applications is
a difficult task because of the large number and the flexibility
of components but also because of the complexity of
interconnections required. The role of software engineering is
precisely that of providing methodologies that can facilitate
control of this complexity. A methodology by definition can
facilitate the process of engineering systems. It consists of
guides that cover the entire lifecycle of software development.
Some are technical guides; others are managing the project [6].
We‟ll name “method” the approach to use a rigorous
process for generating a set of models that describe various
aspects of software being developed using a well- defined
notation.
To this end, several software engineering paradigms have
been proposed, such as object-oriented design patterns, various
software architectures. These paradigms fail especially when it
concerns the development of complex distributed systems for
two reasons: the interactions between the various entities are
defined in a too rigid way and there is no mechanism complex
enough to represent the organizational structure system [7]. The
paradigm of agents and multi-agent systems can be a good
answer
to
these problems, because
the agent-oriented
approaches significantly increase our ability to model, design
and build complex distributed systems [8].
There are many methodologies for analysis and design of
multi-agent systems. We cite below some examples of existing
methodologies [2]:
The first version of Gaia methodology, which modeled
agents from the object-oriented point of view, was
revisited 3 years later by the same authors in order to
represent a MAS as an organized society of individuals
[10]. In fact, the agent entity, which is a central
element of the meta-model of Gaia, can play one or
more roles. A role is a specific behavior to be played
by an agent (or kind of agents), defined in term of
permissions, responsibilities, activities, and interactions
with other roles. When playing a role, an agent updates
its behavior in terms of services that can be activated
according to some specific pre- and post- conditions. In
addition, a role is decomposed in several protocols
when agents need to communicate some data. The
environment abstraction specifies all the entities and
resources a multi-agent system may interact with,
restricting the interactions by means of the permitted
actions [1].
The Gaia methodology gives the possibility to design
MAS using an organizational paradigm and to traverse
systematically the path that begins by setting out the
demands of the problem and to lead to a fairly detailed
and immediate implementation [9]. Gaia permits to
design a hierarchical non-overlapping structure of
agents with a limited depth. From the organizational
point of view, agents form teams as they belong to a
unique organization, they can explicitly communicate
with other agents within the same organization by
means of collaborations, and organizations can
communicate between them by means of interactions.
If
inter-organization communication
is omitted,
coalitions and congregations may also be modeled [1].
However, this methodology is somewhat limited since
we can describe MAS with different architectures of
agents [9].
The main contribution of MESSAGE was
the
definition of meta-models for specification of the
elements that can be used to describe each of the
aspects that constitute a multi-agent system (MAS)
from five viewpoints: organization, agents, goals/tasks,
interactions and domain. MESSAGE adopted the
Unified Process and centered on analysis and design
phases of development [11].
INGENIAS starts from the results of MESSAGE and
provides a notation to guide the development process
of a MAS from analysis to implementation [12] [13].
It is both a methodology and a set of tools for
development of multi-agent systems (MAS). As a
methodology, it tries to integrate results from other
from
proposals and considers
the MAS
five
complementary viewpoints: organization,
agent,
www.ijacsa.thesai.org
149 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
tasks/goals,
is
It
interactions, and environment.
supported by a set of tools for modeling (graphical
editor), documentation and code generation (for
different
INGENIAS
platforms). The
agent
methodology does not explicitly model social norms,
although
they are
implicit
in
the organizational
viewpoint. Organizational dynamics are not considered
i.e., how agents can join or leave the system, how they
can form groups dynamically, what their life-cycle is,
etc [14]. The authors have developed an agent-oriented
software tool called INGENIAS Development Kit
(IDK) [15]. It allows to edit consistent models
(according
to
INGENIAS specification) and
to
generate documented code in different languages such
as JADE [16], Robocode, Servlets or Gracias Agents
[1].
Multi-agent systems Software Engineering (MaSE) is a
start-to-end methodology that covers from the analysis
to the implementation of a MAS [17]. The main goal of
MaSE is to guide a designer through the software life-
cycle
from a documented specification
to an
implemented agent system, with no dependency of a
particular MAS architecture, agent architecture,
programming language, or message-passing system.
AUML (Agent Unified Modeling Language) is an
evolving standard for a design methodology to support
MAS. It is based on the UML methodology used with
object oriented systems. This notation was proposed to
adapt the UML‟s one in order to describe the agent-
oriented modeling [18].
AUML provides tools for:
Specification protocol of interaction between
agents,
Representation of the internal behaviour of an
agent,
Specification of roles, package interface agent,
mobility, etc [2].
The Agent Modeling Language (AML) is a semiformal
visual modeling language for specifying, modeling and
documenting systems that incorporate concepts drawn
from multi-agents systems (MAS) theory [19].
ASPECS
for
(Agent-oriented Software Process
Engineering Complex Systems) provides a holonic
perspective to design MAS [20]. Considering that
complex systems
typically exhibit a hierarchical
configuration, on the contrary to other methodologies,
it uses holons instead of atomic entities. Holons, which
are agents recursively composed by other agents,
permit to design systems with different granularities
until the requested tasks are manageable by individual
entities.
The goal of the proposed meta-model of ASPECS is to
gather the advantages of organizational approaches as
well as of those of the holonic vision in the modeling
of complex system [1].
All these methodologies presented above are still quite
recent. They are mainly focused on the analysis phase, whereas
design and implementation phases are missing or are redirected
to agent-oriented methodologies, which do not offer enough
tools to model organizational concepts. Therefore, there is still
a gap between analysis and design, which must be specified
clearly, correctly and completely [14].
Finally, the maturity of methodologies can be analyzed by
the number of systems that have adopted them. Most of
analyzed methodologies have associated applications that show
their feasibility. These methodologies have been applied in
different fields such as medical informatics [21], manufacturing
[20] [22], and e-commerce [23]. MaSE and INGENIAS are the
most used ones. Unfortunately, the number of real world
applications that use agent-oriented methodologies is still low
[1].
IV. THE MDA APPROACH
The MDA (Model Driven Architecture) proposes a
methodological framework and architecture for systems
development that focuses first on the functionality and
application behavior, without worrying about the technology
implemented. The
the application will be
with which
implementation of
the application goes
through
the
transformation of business models in specific models to a target
platform (Fig.2). One research was done in this area as the
dissertation of Jarraya T. [24]
Figure 2. The MDA approach
of automation,
process independent
business
The
of need, is
from the
which comes
expression
described as a "CIM"
(Computation Independent
functional analysis,
Model). The detailed
the heart of
the process is
concentrated
in
the "PIM" (Platform
Independent Model), which, as its name suggests, is strictly
independent of the
technical architecture
and the target
language.
the
The "PSM" (Platform Specific Model) is
model
for engineering design obtained by transformation
of PIM by projection on the target technical architecture. It is
this model that is based on code generation [5].
The benefits to businesses on the MDA are primarily:
www.ijacsa.thesai.org
150 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
The fact that architectures based on MDA are ready
for technological developments.
The ease of integrating applications and systems
around a shared architecture
Broader interoperability for not being tied to a
platform.
One of the main tools of MDA, we have AndroMDA who
takes as its input a business model specified in the Unified
Modeling Language (UML) and generates significant portions
of the layers needed to build, for example, a Java application
[25]. AndroMDA's ability to automatically translate high-level
business specifications into production quality code results in
significant time savings when implementing Java applications.
The diagram below maps various application layers to, for
examples, Java technologies supported by AndroMDA [5].
providing a platform independent way for clients to
access
their functionality. AndroMDA can even
generate business processes and workflows for the
jBPM workflow engine (part of the JBoss product
line).
Data Access Layer: AndroMDA
the
leverages
tool
mapping
popular
object-relational
called Hibernate to generate the data access layer for
applications. AndroMDA does this by generating Data
Access Objects (DAOs) for entities defined in the
UML model. These data access objects use the
Hibernate API to convert database records into objects
and vice-versa. AndroMDA also supports Enterprise
Java Beans EJB3/Seam [26] for data access layer (pre-
release).
Data
Stores: Since
AndroMDA
generated
applications use Hibernate to access the data, you can
use any of the databases supported by Hibernate.
The generation process of AndroMDA is as follows [5] :
Figure 3. Application layers suppor ted by AndroMDA
Presentation Layer: AndroMDA currently offers two
technology options to build web based presentation
layers: Struts and JSF.
It accepts UML activity
diagrams as input to specify page flows and generates
Web components that conform to the Struts or JSF
frameworks.
Business Layer: The business layer generated by
AndroMDA consists primarily of services that are
configured using
the Spring Framework. These
services are implemented manually in AndroMDA-
generated blank methods, where business logic can be
defined. These generated services can optionally be
front-ended with EJBs, in which case the services
must be deployed in an EJB container (e.g.,JBoss).
Services can also be exposed as Web Services,
Figure 4. Generation process of AndroMDA
Preparation of the project in MagicDraw
Preparing use cases
Preparation of class diagram
Preparation of state charts
Code Generation
Generating the database
Deploy the application
V. PROPOSED APPROACH
Our approach is based on model driven architecture (MDA)
which aims to establish the link between the existing agent
architectures and models or meta-model multi-agent systems
that we build based on AUML. Our idea is to offer a design
methodology based on agents AUML notation for establishing
a generic class diagram that the designer can use to design his
system [3]. This diagram is considered as a meta-model which
www.ijacsa.thesai.org
151 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
is not generated by any tool and must be defined by the
modeler himself.
Figure 5. An AUML generic cla ss diagram for a MAS
Our approach has a lot of benefits, it allows:
Reducing costs and development times for new
applications.
Improving quality of applications.
Reducing complexity of application development.
Ability to generate all the necessary components
described.
Modularity and reusability of the developments.
Coercion by the MDA model.
Generating a library of generic models.
A. Description of the AUML generic Class Diagram
The diagram is conceived in three layers, each one is
represented by a relationship between classes: A first part
which is a relation between agent and its environment, a second
part of specialisation of the agent class, and at the last part, a
specialisation of the cognitive agent class [3].
1- The first part
The first part consists of two important classes:
Perceptions. Attributes can be all the information that
an environment should have, plus the following
common information:
Deterministic when the next state of the
environment is determined in a unique way
by the current state and action of the agent,
so the environment is deterministic. If the
outcome is uncertain (especially if, as a result
of action of the agent, the environment can
evolve in different ways), we are in the non-
deterministic case.
Static if the environment cannot change its
state without the intervention of the agent.
The environment is dynamic if its state can
change without the action of the agent in the
time interval between two perceptions of the
agent.
Continuous if any portion of an environment
state to another requires passing through a
sequence of intermediate states, otherwise
the environment is discrete.
Perception is a section where the designer should
determinate all environment perceptions, example:
number of agents.
Environment contains several functions allowing to
start running, to perceive information from agents
linked to it and to modify its state after each action
from
those agents,
that
is respectively Run(),
Perceive() and ModifState().
Agent is the main class on the diagram that allows the
designer
to express all agent properties. The
constructor of Agents takes three sections: Roles,
Attributes
agent
are
and Perception. Roles
functionalities. Attributes are all information that an
agent should possess. And finally Perception which is
a section where the designer should determinate all
agents‟ perceptions about his environment or the other
agents.
Agent contains several functions who allows starting
running and perceiving information from environment
or agents linked to it and to execute all its actions, that
is respectively Run(), Perceive() and Act().
The first part consists also of two important association
classes:
-Action, between agent and his environment.
-Interaction, between agents.
Action is an association class between agent and
environment. It lists all possible actions that an agent
can execute on his environment.
- Environment,
- Agent
Environment is an important class on the diagram
because it influences all the system. Environment‟s
data is represented by two sections, Attributes and
Interaction is a reflexive association class between
agents. Agent can request
information by
the
getInformation() function and send it by the inform()
function. Agent may also deal with some constraints
www.ijacsa.thesai.org
152 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
the function
inform by
that
to
is possible
it
informaboutConstraintes(). The
acceptance
of
partnership is added also to the main functionalities
of Agent by the function acceptPartnerShip().
2- The second part
The second part represents a specialisation relation of the
Agent class. It consists of three important classes:
- Reactive agent,
- Cognitive agent,
- Communicative agent.
Reactive agent is a type of agent. It possesses the
same properties of the Agent class.
Cognitive agent is another specialization of the Agent
class. In this class, the designer should determinate the
representations of the agent that he must have during
its execution. The class possesses also one important
function “Decide()” where agent can decide to execute
an action or not according to his goals.
Communicative agent is the last specialization of the
Agent
class. Like Cognitive
agent
class,
Communicative agent class has representations but
possesses a different function called “Communicate()”
where agent must use to communicate his information
to the other agents.
3- The third part
The third part represents a specialization relation of the
Cognitive agent class. It consists of three important classes:
- Adaptive agent,
- Intentional agent,
- Rational agent.
Adaptive agent is a type of cognitive agent. It
possesses the same properties of the Agent class, the
knowledge base and the “Decide()” function. As
mentioned in the types of agent section above, an
adaptive agent is able to change its objectives and its
knowledge base as and when these changes. This
functionality
the
by
expressed
is
“Change_information()” function.
Intentional agent or BDI Agent is designed from the
"Belief-Desire-Intention” model. It is a type of
cognitive agent. In the same case of Adaptive Agent
class, this class possesses the same properties of the
Agent class, the knowledge base and the “Decide()”
function.
In this class, the designer should determinate the
agent‟s beliefs represented by the Beliefs section. The
beliefs of an agent are the information that the agent
has on the environment and other agents that exist in
the same environment. Beliefs may be incorrect,
incomplete or uncertain, and because of that, they are
different from knowledge of the agent, which is
information still true. Beliefs can change over time as
the agent by its ability to perceive or interact with
other agents, collects more information.
The designer should also determinate the agent‟s
intentions represented by the Intentions section. The
intentions of an agent are the actions it has decided to
do to accomplish his goals.
To choose the correct agent‟s beliefs from the
incorrect
ones,
this
class
offers
the
“Revise_beliefs(Pres, Belief)” function which is based
on the agent‟s knowledge base and his beliefs. Then,
the “Generate_desires(Belief, int)” function comes to
generate all the agent‟s desires that he may be able to
accomplish at once. The desires of an agent
representing all things the agent would like to see
made. An agent may have conflicting desires, in
which case he must choose between her desires a
subset that is consistent. This subset consists of his
desires is identified with the beliefs and the intentions
of the agent.
Another function comes after that, the “Filter(Belief,
Generate_desires, int)” which filters all those elements
above and gives the consistent beliefs, desires and
intentions of the intentional agent.
Finally, the agent can select his actions according to
this
the
by
them
execute
and
filtration
“Actions_selection(Filter)” function.
Rational agent is the last specialisation of the
Cognitive Agent class. Like Intentional Agent class,
Rational Agent class has the Beliefs and the Intentions
sections but possesses
just one function called
“Mesure_performance(Percept, Belief)” where agent
must use to execute his actions as efficient as possible.
This function is based both on his perceptions and his
beliefs.
B. The generic UML Class Diagram
This generic AUML class diagram was subsequently
converted into a generic class diagram based on UML notation.
This transformation will allow the designer to easily use
AndroMDA to generate the source code equivalent to its UML
diagram [1].
The passage from AUML to UML was performed by
following the steps below:
1. Keep the same titles of classes and associations
which constitute the AUML diagram.
2. Assign roles, perceptions, intentions, beliefs and
representations of each agent, and any possible
additional attributes, in the attributes part of the
UML class.
3. Combine all methods or
functions
operations part of the UML class.
in
the
We can obtain, in the end, the following result shown in
Fig. 6:
www.ijacsa.thesai.org
153 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
Figure 6. An UML generic class diagram for a MAS
Our approach can present one desadvantage. It is the
complexity of generating a good code source by AndroMDA.
The model developed at the design phase, should be reliable in
order to build the application and realize its implementation
without errors [5].
V. APPLICATION EXAMPLE
A. Description
Our proposed AUML class diagram was used for design of
one multi-agent system for a Chat Application. This example is
designed as follows [5]:
Three reactive agents: These agents will be the chatters,
the interest that these are reactive agents relies on the fact
that an agent doesn't react before the declaration of the
name of the receiver by the user of the application.
Therefore an agent will react to get ready to catch the
name and the message and to send it to the appropriate
person. He will react also to clear the sent and the received
message from their area in his interface.
We can respectively obtain the following AUML and UML
diagrams corresponding to this example, shown in the Figures
7 and 8:
B. Realization
To validate our model for this example, we‟ve tried to
download AndroMDA with all the required dependencies
(including all profiles referenced by models). Then, we
generated our project « ChatAgents » by running « mvn
org.andromda.maven.plugins:andromdaapp-maven
plugin:3.4-SNAPSHOT:generate ». The
result of
command is as follows:
this
Figure 7. AUML Class diagram for a chat application
Figure 8. UML C lass diagram for a chat application
When we examine the various folders and files created by
the andromdapp plug-in, we will notice files called pom.xml in
various folders under ChatAgents. These files make up several
Maven projects. In fact, the ChatAgents directory contains a
hierarchy of Maven projects as shown below [5].
ChatAgents: This is the master project that controls
the overall build process and common properties.
mda: The mda project is the most important sub-
project of the application. It houses the ChatAgents
UML model under the src/main/uml directory. The
mda project is also where AndroMDA is configured
to generate
the
to assemble
the files needed
application.
common: The common sub-project collects resources
and classes that are shared among other sub-projects.
These include value objects and embedded values.
www.ijacsa.thesai.org
154 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
classes that use the Spring framework, optionally
making use of Hibernate and/or EJBs under the hood.
These include entity classes, data access objects,
hibernate mapping files, and services.
web: The web sub-project collects those resources
and classes that make up the presentation layer.
app: The app sub-project collects those resources and
classes that are required to build the .ear bundle.
By opening the file “ChatAgents.xml” in MagicDraw, we
will be
able to build various graphs of our model to
generate then
entire
the
of
code
source
the
application. Note that AndroMDA can't read MagicDraw 17
models directly. Therefore, you can export it to another file
format: EMF-UML2.
After import of AndroMDA profiles to use for our
application, we designed our class diagram as shown in Fig.10
as follows [5]:
The result of exporting our “ChatAgents” model to EMF-
UML2
the
in
located
is
format
folder C:/ChatAgents/mda/src/main/uml in explorer. Below his
content:
ChatAgents.xml: the MagicDraw 17 model file.
ChatAgents.uml: ChatAgents model in EMF/UML2
format. It's the file that will be processed by
AndroMDA.
10 files ending with .profile.uml:
profiles used by ChatAgents.uml
the different
Following the definition of our model, the generation
of application code is achieved by executing the command
"mvn install", the result appears as in the figure [5].
Thus,
the class “Chat.java” is created and can be
easily accessed and modified by the developer where he has the
ability to implement its operations in the generated code.
We conducted this implementation and got the final result.
Figure 9 : ChatAgents project generation
ChatAgents
-- mda
-- common
-- core
-- web
+-- app
core: The core sub-project collects resources and
VI. CONCLUSION AND FUTURE SCOPE
The purpose of this paper is to demonstrate the feasibility of
our approach to analyze, design and implement multi-agent
systems. With AUML modeling and MDA, we can generate all
the necessary components described by the class meta-model
that we proposed. Which leads us to obtain a generic design
based on SOA more or less reusable components using one of
the most MDA tools used in development is AndroMDA [27].
In the future, we would like to model another application
sample of our model but in a more complex form using
cognitive or adaptive agents and in other platforms like C++,
Web services, etc. It will help us to validate the efficacy of our
proposed approach and lead us to consider it as a generic
approach which can be adopted by every type of information
system and used for any real world application.
www.ijacsa.thesai.org
155 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
Figure 10 : Cla ss diagram built on MagicDraw 17
Figure 11 : Code generation a fter definition model
Figure 12. Chat applica tion with three agents
ACKNOWLEDGMENT
I would like to thank to my advisor Ms. M. Addou, Phd. for
his invaluable guidance and many useful suggestions during
my work on this paper. I would also like to express my
gratitude to all those who gave me the possibility to complete
this paper.
www.ijacsa.thesai.org
156 P a g e
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 2, No. 11, 2011
REFERENCES
[1] D. Isern, D.Sanchez, A.Moreno, “Organizational structures supported by
agent-oriented methodologies” , The journal of Systems and Software,
vol. 84, n. 2, Oxford, UK: Elsevier, 2011, pp. 169 -184.
[2] S. Maalal, M. Addou, “A Model Design of Multi -Agent Systems”,
Proceedings of the 2nd Edition of the IEEE International Conference on
Multimedia Computing and Systems ICMCS‟11 , Ouarzazate Morocco,
p. 674, 2011.
[3] M. Wooldr idge, Intelligent Agents, Multi agent systems, In The MIT
Press, “A modern Approach to D istributed Artificial Intelligence”,
(England Ma ssachutts London: MIT Press Cambridge, 1995, p. 27 -78)
[4] M. Wooldridge, An Introduction to Multi -Agent Systems, Wiley &
Sons, 2000.
[5] S. Maalal, M. Addou, “A practical application of a method of designing
multi-agent systems based on the AUML language and the MDA
approach”, Proceedings of the Fourth Work shop on Informa tion
Technologies and Communication WOTIC‟11, Ca sablanca, Morocco,
p.104, 2011.
[6] O. Shehory, A. Sturm, “Evaluation of modeling techniques for agent -
bases systems”, Proceedings of the 5th International Conference on
Autonomous Agents, pp.624-631, 2001.
[7] N. R. Jennings, “On agent-ba sed software engineering” , Artificial
Intelligence, vol. 117, pp. 277-296, 2000.
[8] M. Wooldr idge, N. R. Jennings, “Intelligent agent: Theory and practice”,
The Knowledge Engineer ing Review, Vol. 10, n. 2, pp. 115 -152, 1995.
[9] A. M. Florea, D . Kayser, S. Pentiuc, A. El Fallah Segrounichi,
Intelligents agents, Agents Intelligents, Politechnica University of
Bucharest, 2002.
[10] L. Cernuzzi, T. Juan, L.Ster ling, F. Zambonelli, “The Gaia methodology:
basic concepts and extensions”, Methodologies and Software
Engeneering for Agent Systems, US: Springer , pp.69 -88, 2004.
[11] J. Pavón, , J . Gómez-Sanz., “Agent Oriented Software Engineer ing with
INGENIAS”, Proceedings of the international Central and Eastern
European conference on Multi -Agent Systems CEEMAS‟03 , pp.394-
403, 2003.
[12] R. Fuentes-Fernández, I. García -Magariñio, A.M. Gómez-Rodríguez,
J.C. González-Moreno, “A
technique for defining agent -oriented
engineering processes with tool support” , Artificial Intelligence, vol.23,
pp.432-444.
[13] J. Pavón, , J.J. Gómez-Sanz., R. Fuentes, „The INGENIAS methodology
and tools” in Agent-oriented Methodologies, B . Henderson -Sellers and
P. G iorgini Eds. Idea Group, 2005, pp. 236–276.
[14] E. Argente, V. Julian, V. Botti, “Multi-agent system development ba sed
on organizations”, Electronic Notes in Theoretical Computer Science,
vol.150, pp.55 -71, 2006.
[15] IDK
(INGENIAS
http://sourceforge.net/projects/ingenia s/
Development
Kit),
[16] JADE (Java Agent DEvelopment Framework), http://jade.tilab.com/.
[17] S.A. DeLoach, “The MaSE methodology”, in Methodologies and
Software Engineering for Agent Systems, F. Bergenti, M.P Gleizes, F.
Zambonelli, Eds. The Agent-oriented Software Engineering Handbook.
Kluwer Academic Publishers, 2004, pp. 107–125
[18] S. Lynch, K. Rajendran, “Design Diagrams for Multi -agents Systems”,
Proceedings of the 16th Annual Workshop of the Psychology of
Programming Interest Group PPIG‟04 , pp. 66-78, 2004.
[19] R. Cervenka, I. Trencansky, “Agent Modeling Language (AML) : A
Comprehensive Approach to Modeling MAS”, Informatica, vol. 29, n. 4,
pp. 391-400, 2005.
[20] M. Cossentino, N . Gaud, V. Hilaire, S.Galland, A. Koukam, „ASPECS:
An Agent-oriented Software Process for Engineer ing Complex Systems:
How to design agent societies under a holonic perspective”, 2010.
[21] D. Isern, C . Gómez-Alonso, A. Moreno, “Methodological development
of a multi-agent system in the healthcare domain”, Commun, SIWN 3,
pp. 65–68, 2008.
[22] A. Giret, V. Botti, S. Valero, “MAS methodology for HMS”, In the
Second International Conference on Indu str ial Applications of Holonic
and Multi-Agent Systems HoloMAS, Springer -Ver lag, Copenhagen,
Denmark, pp. 39–49, 2005.
[23] J. Ferber, O . Gutknecht, F. Michel, “From agents to organizations: an
organizational view of multi -agent systems” , in Spr inger -Verlag Berlin
Heidelberg, P. G iorgini, J. Müller , J . Odell, Eds 2003, in the 4th
International Workshop on Agent -oriented Software Engineer ing IV
(AOSE) , Melbourne, Australia, pp. 214–230, 2003.
[24] T . Jarraya, Re-u se of interaction protocols and Career-or iented models
for multi-agents development, Réutilisation des protocoles d‟interaction
et Démarche or ientée modèles pour le développement multi -agents ,
Ph.D. Thesis, Dept. Computer Engineering, University of Reims
Champagne Ardenne, France, 2006.
[25] N. Bhatia,
“Getting Started with AndroMDA
(www.andromda.org, 2010).
for
Java”
[26] JBoss Seam (http://www.jboss.com/products/seam/).
[27] S. Maalal, M. Addou, “A Model Design of Multi -Agents Systems”, in
the
Informa tion and
International Conference on Models of
Communication Systems MICS‟10, Rabat, Morocco, 2010, unpublished.
AUTHORS PROFILE
Sara M aalal wa s born in Rabat the Morocco‟s capita l in 1985. She received
his professional ma ster in Computer Engineering and Internet (3I), Option:
Security Network s and Systems, in 2008 from the Faculty of science of
HASSAN II University, Ca sablanca, Morocco . In 2010 she joined the system
architecture team of the National and H igh School of Electr icity and Mechanic
(ENSEM: Ecole Nationale Supérieure d‟Electricité et de Mécanique),
Casablanca, Morocco.
Her actual main research interests concern Designing and modeling Multi -
Agent Systems.
Ms. Maalal is actually a Software Engineer in a Moroccan multinational
society called Hightech Payment Systems (HPS) which has always proved
itself a s a leading payment solutions provider .
Malika Addou received her Ph.D . in Ar tificial Intelligence from University of
Liege, Liege, Belgium, in 1992. She got her engineer degree in Compu ter
Systems from the Mohammadia School of Engineers (EMI
: Ecole
Mohammadia des ingénieurs), Rabat, Morocco in 1982. She is Professor of
Computer Science at the Hassania School of Public Work s (EHTP : Ecole
Hassania des Travaux Publics), Ca sablanca, since 1982.
Her research focuses on Software Engineering (methods and technologies for
design and development), on Information Systems (D istributed Systems) and
on Artificial Intelligence (especially Multi-Agent Systems technologies) .
www.ijacsa.thesai.org
157 P a g e
|
1611.08951 | 1 | 1611 | 2016-11-28T01:10:54 | Distributed Estimation for Adaptive Networks Based on Serial-Inspired Diffusion | [
"cs.MA"
] | Distributed estimation and processing in networks modeled by graphs have received a great deal of interest recently, due to the benefits of decentralised processing in terms of performance and robustness to communications link failure between nodes of the network. Diffusion-based algorithms have been demonstrated to be among the most effective for distributed signal processing problems, through the combination of local node estimate updates and sharing of information with neighbour nodes through diffusion. In this work, we develop a serial-inspired approach based on message-passing strategies that provides a significant improvement in performance over prior art. The concept of serial processing in the graph has been successfully applied in sum-product based algorithms and here provides inspiration for an algorithm which makes use of the most up-to-date information in the graph in combination with the diffusion approach to offer improved performance. | cs.MA | cs |
DISTRIBUTED ESTIMATION FOR ADAPTIVE NETWORKS BASED ON
SERIAL-INSPIRED DIFFUSION
Cornelius T. Healy and Rodrigo C. de Lamare
CETUC - Pontifical Catholic University of Rio de Janeiro - PUC-RJ
Email: [email protected], [email protected]
Rio de Janeiro, Brazil
ABSTRACT
Distributed estimation and processing in networks modeled by
graphs have received a great deal of interest recently, due to the
benefits of decentralised processing in terms of performance and
robustness to communications link failure between nodes of the
network. Diffusion-based algorithms have been demonstrated to be
among the most effective for distributed signal processing problems,
through the combination of local node estimate updates and sharing
of information with neighbour nodes through diffusion. In this work,
we develop a serial-inspired approach based on message-passing
strategies that provides a significant improvement in performance
over prior art. The concept of serial processing in the graph has
been successfully applied in sum-product based algorithms and
here provides inspiration for an algorithm which makes use of the
most up-to-date information in the graph in combination with the
diffusion approach to offer improved performance.
Index Terms -- Diffusion networks, wireless sensor networks,
distributed processing.
1. INTRODUCTION
Distributed signal processing is an important tool for problems
which may be modeled by a graph of nodes working to estimate
a parameter of interest as it allows computations to be carried lo-
cally at the individual nodes, avoiding the need for a centralised
processing unit and thus offering robustness to scenarios where the
communication links to that central node are subject to channel ef-
fects. To achieve this each node makes use of its local observations
in combination with the estimates produced at neighbour nodes to
produce an improved estimate of the parameters of interest. This
improved estimate is then shared with all neighbours of the node,
leading to propagation of the information through the network.
The distributed estimation problem has been considered in terms
of incremental [1], consensus [2] and diffusion [3, 4, 5] strategies.
The incremental strategy in general demands a computationally
costly operation for identifying a path through the graph nodes upon
which to operate, while the diffusion based strategies have been
demonstrated to be superior to those based on consensus in terms of
convergence, performance and stability [6].
Recent work on the diffusion strategies has included sparsity-
aware approaches of [7, 8, 9, 10, 11, 12, 13], which exploit the
knowledge that the parameter vector to be estimated may be sparse.
Related work on improving the combiners in the information diffu-
sion stage of the algorithm has been reported in [14, 15]. The effect
of the network topology on the diffusion strategies and how it may
be exploited has been considered in [16, 17, 18, 19, 20]. A number
of studies have considered imperfect communications links between
nodes in the network [21, 22]. Significant work has also been carried
out on analysis of the diffusion strategies [23, 24].
In this paper, the schedule of node estimate updates is consid-
ered as a source of performance improvement. This is motivated
by the observation that such an approach in the case of the sum-
product algorithm operating on a bipartite graph, as for the decoding
of LDPC codes, offers significantly faster convergence at almost no
cost in terms of additional complexity. This approach was termed se-
rial, shuffled or layered scheduling in the literature [25, 26]. Further
improvements were found through more advanced update schedules
[27, 28, 29, 30, 31]. Effectively, after each individual node update,
the newly updated messages are made available to the neighbour
nodes of the updated node, ensuring those neighbours compute their
own updates with more up-to-date and accurate information. As the
diffusion approach is based on the sharing of information with neigh-
bour nodes in the graph of the network, this concept of serialisation
translates well. In particular, we develop a serial-inspired (SI) least-
mean square (LMS) , which we denote SI-LMS and can exploit the
schedule of node updates to obtain improved performance. In the
proposed SI-LMS algorithm, serialisation is introduced through the
inclusion of an additional diffusion combination which has access
to the most recently updated estimates in the graph. The proposed
SI-LMS algorithm offers significant improvements in convergence
speed, as is demonstrated by the simulation study provided in this
paper.
This paper is organized as follows. Section 2 provides the prob-
lem statement and introduces the diffusion strategy for system iden-
tification using the LMS algorithm. In Section 3 the proposed algo-
rithm is developed and described in detail, along with pseudocode
representation. Section 4 provides the numerical simulation results,
and Section 5 concludes the paper.
Notation: Throughout this paper, lowercase letters such as x
indicate scalars, lowercase boldface letters such as x denote column
vectors and uppercase boldface letters such as A denote matrices.
The superscript i denotes that A(i) is the realization of A at the time
index i, likewise for scalars as in d(i). Subscripts are used to identify
the node or nodes in the graph with which a value is associated.
2. PROBLEM STATEMENT AND THE DIFFUSION
STRATEGY
Consider a network modeled by a graph with N nodes as depicted in
Fig. 1. The value d(i)
k is the scalar observation at time instant i for
the node k in the graph, and the observation is related to the input
k
Nk
Fig. 1. An example of a connected graph with neighbourhood Nk of
node k shown.
signal x(i)
k by
d(i)
k = ωH
0 x(i)
k + n(i)
k ,
i = 1, 2, . . . , N,
(1)
where the input signal vector x(i)
for node k at time index i is an
k
M × 1 vector. The value n(i)
is the noise sample at node k and
k
time index i and has zero mean and variance σ2
v,k. The goal of the
distributed estimation problem is to estimate the value of ω0 based
on the knowledge at the nodes in the network of the observations
d(i)
k , the input signal vectors x(i)
k and the relation in (1) through use
of that local knowledge and the ability to share information with
neighbours in the network graph.
The diffusion strategy for distributed estimation involves a pro-
cess of local adaptation with the information available using for
example the LMS estimate update, followed by information sharing
with neighbour nodes involving a weighted sum of the estimates
across the neighbourhood of each node. This process leads to
diffusion of information through the fully connected graph. The
adaptation and combination steps of the diffusion strategy can be
performed in either order, leading to in one case the adapt -- then --
combine (ATC) diffusion strategy and in the other the combine --
then -- adapt (CTA) diffusion strategy [3]. The two steps of ATC
diffusion are described by
iterate
iterate
{d(i)
k , xk,i}
ω(i−1)
k
A.F.
ψ(i)
k
Node k
ψ(i)
k
ψ(i)
a
ψ(i)
d
(a) Produce output ψ(i)
k .
(b) Produce output ω(i)
k .
Fig. 2. (a) Filter adaptation with local observations and combined
estimate. (b) Information diffusion through weighted combination
of neighbour estimates.
3. PROPOSED SERIAL-INSPIRED ALGORITHM
In this section, the proposed SI-LMS algorithm is introduced. In the
case of sum-product type algorithms operating on bipartite graphs,
the standard message update schedule is to activate all nodes of one
type and then to activate all nodes of the other type. It was demon-
strated that improvements in error rate convergence of the algorithms
may be found at no increased computational complexity if the nodes
in the graph are updated in a serial fashion, also termed shuffled or
layered schedule [25, 26]. This improvement in convergence be-
haviour is derived from the fact that the most recently updated mes-
sages in the graph may be used to improve the next updates within
an iteration. This observation has motivated the investigation of the
diffusion LMS algorithm and resulted in the development of the al-
gorithm proposed in this paper. Essentially, the process of informa-
tion diffusion is used to improve the LMS adaptation through the use
of new estimates at neighbour nodes as soon as they are available.
The MSE cost function at node k takes the form
Jk(ω) = E[d(i)
k − ωH x(i)
k 2],
which through expansion and rearrangement results in
Jk(ω) = Jk,min + ω − ω02
Rx,k ,
(6)
(7)
ψ(i)
k = ω(i−1)
k
+ µkx(i)
k [d(i)
k
k − ω(i−1)H
cklψ(i)
,
l
x(i)
k ]∗,
(2)
(3)
where Jk,min is the value of Jk(ω) evaluated at ω = ω0. The local
cost function at node k when information sharing with neighbours is
allowed becomes:
ω(i)
k = X
l∈Nk
where the values ckl are known as the combination coefficients and
provide the weighting in the combination step of the respective algo-
rithms and wk represents the parameter estimator. They are related
to the topology of the graph, being nonzero only if node k and node
l are neighbours and additionally must satisfy the constraint:
ckl = 1, l ∈ Nk∀k.
(4)
X
l
There are a number of rules specifying the combination coefficients
to be found in the literature, including the uniform , Metropolis , rel-
ative degree and Laplacian rules. The Metropolis combiner is given
by:
1
max(Nl,Nk) ,
ckl =
ckl = 1 − Pl∈Nk/k ckl,
ckl = 0,
if l /∈ Nk
if l ∈ Nk, l 6= k
if l ∈ Nk, l = k
(5)
Fig. 2 provides the block diagram for the ATC diffusion algo-
rithm which implements (2) - (3).
J local
k
(ω) = X
l∈Nk
al,kJl(ω).
(8)
The global cost function is simply the sum across all nodes l of this
local cost function. Rewritten in terms of the node of interest and
its neighbours, we have the global function from the perspective of
node k:
J global
k
(ω) = J local
k
(ω) + X
l∈Nk\{k}
J local
l
(ω).
(9)
Applying the steepest-descent method we arrive at the recursion for
updating the parameters of the estimator:
ω(i)
k = ω(i−1)
k
− µk[∇ωJ global
k
(ω(i−1)
k
)]∗,
(10)
(ω(i−1)
) is the gradient of J global
where ∇ωJ global
(ω) with respect
to ω evaluated at ω(i−1)
. Through the use of the expanded versions
of (6) and through a number of approximations detailed in [4, 32],
k
k
k
k
the update recursion may be reformulated as:
ω(i)
k = ω(i−1)
k
+ µk X
l∈Nk
al,k(rdx,l − Rx,lω(i−1)
k
)
+ µk X
l∈Nk\{k}
bl,k(ω0 − ω(i−1)
k
),
(11)
k
k x(i)
k x(i)H
] and rdx,l = E[d(i)∗
where Rx,l = E[x(i)
k ]. In (11) the
previous estimate is corrected by a filter adaptation term and an in-
formation diffusion term. In the development of the ATC and CTA
algorithms, the two correction terms are applied successively. In the
proposed algorithm, the information diffusion term will be applied
both before and after the adaptation of the estimator, and as in the
development of those algorithms the best available estimate will be
used to substitute for both ω0 and ω(i−1)
in (11). In the proposed
algorithm these best estimates are improved upon the previously pre-
sented works through the observation that the estimates which have
been updated at neighbour nodes are available for use immediately,
and so the first diffusion correction term is applied in a serial fash-
ion. In particular, we employ instantaneous estimates of Rx,l and
rdx,l to obtain the recursion for the proposed SI-LMS algorithm:
k
iterate
Node k
ω(i−1)
k
ω(i−1)
a
{d(i)
k , xk,i}
iterate
Node k
ψ(i)
k
A.F.
φi
k
φ(i)
k
ω(i−1)
b
φ(i)
c
φ(i)
d
φ(i)
a
φ(i)
d
(a) Produce output ψ(i)
k .
(b) Produce output φ(i)
k .
(c) Produce output ω(i)
k .
Fig. 3. (a)Weighted sum of updated and prior estimates at the neigh-
bours. (b) Filter adaptation with local observations and combined
estimate. (c) Final weighted combination of neighbour estimates.
• Nk extra multiplications
• Nk extra additions
With the complexity cost across the network being N times each
of these. Note that in addition to the cost in terms of increased com-
plexity, the proposed SI-LMS algorithm requires that the nodes in
the graph share their updated estimates with their neighbours as soon
as they are produced by the LMS adaptation. Thus the SI-LMS al-
gorithm also incurs an extra communications cost of
ω(i)
k = ω(i−1)
k
+ µk X
l∈Nk
al,kx(i)
k [d(i)
k − ω(i−1)H
k
x(i)
k ]∗
• N (Nk − 1) transmissions of a vector of dimension M × 1
when compared to the ATC-LMS diffusion algorithm [3].
+ µk X
l∈Nk\{k}
bl,k(ω0 − ω(i−1)
k
),
(12)
4. SIMULATION RESULTS
Given the previous development, the steps of the proposed SI-LMS
algorithm are as follows:
1. Combine prior and new estimates available from neighbour-
hood nodes, including the node of interest, by weighted sum.
2. Adapt the parameters according to the chosen rule, using lo-
cal observations at the node and combined estimate from the
first step. Make new estimate available to neighbour nodes.
3. Combine estimates available from neighbourhood.
The SI-LMS algorithm is presented as
ψ(i)
k = ck,kω(i−1)
k
+ X
ℓ∈Nk:ℓ≥k
ck,ℓω(i−1)
ℓ
+ X
m∈Nk:m<k
φ(i)
k = ψ(i)
k + µkx(i)
ω(i)
p = X
q∈Np
k x(i)
k ]∗,
k [d(i)
k − ψ(i)H
ap,qφ(i)
q ,
ck,mφ(i)
m ,
(13)
(14)
(15)
Fig. 3 provides a block diagram of the proposed SI-LMS al-
gorithm, with blocks for the initial serial information diffusion, the
adaptation for the estimator and the final diffusion combination, re-
In Alg. ?? the pseudocode for the proposed SI-LMS
spectively.
algorithm is provided. This gives the details of the algorithm.
3.1. Computational Cost and Bandwidth Requirements of the
Proposed SI-LMS Algorithm
The proposed SI-LMS algorithm provides improvements in conver-
gence speed through the use of an additional combination of new and
old estimates prior to the LMS estimate adaptation. This additional
weighted sum prior to the adaptation step (line 4 in Algorithm ??)
comprises the only additional cost of the proposed algorithm. Thus,
in terms of complexity per node, the proposed algorithm costs
In this section, the simulation study for the proposed algorithm is
presented. Its performance, in terms of mean-square error (MSE),
is compared to the diffusion ATC algorithm [3]. Fig. 4 provides
the network graph topology, showing the network considered has
N = 20 nodes. We adopted the Metropolis combining rule [32].
The unknown parameter vector to be estimated has length M = 5.
Two cases are considered for the input signal, one with the signal
variance equal at all nodes in the network, another with varying sig-
nal variances. The noise of (1) is modeled by white complex circu-
lar Gaussian random variables with zero mean, with signal variances
that are arbitrary. Two cases are considered, the first in which the
signal variances are the same across all nodes in the network, and
another where they are allowed to vary. The variances are provided
in Fig. 5. The variances of Fig. 5(a) and 5(b) correspond to the
simulation environment for the results of Fig. 6 while the variances
of Fig. 5(c) and 5(d) correspond to the simulation environment for
the results of Fig. 7. The step size at all nodes in the network is
µk = 0.01 for both Fig. 6 and Fig. 7. Additional results for the
scenario with different variances at the nodes in the network are pro-
vided for a larger step size of µk = 0.05 in Fig. 8. The results
provided are averaged over 100 independent runs.
Figs. 6 and 7 demonstrate that the proposed SI-LMS algorithm
outperforms the standard-form ACT-LMS algorithm in speed of con-
vergence, with approximately a 40% reduction in the number of it-
erations required to converge. Fig. 8 shows that the performance
improvements of the proposed SI-LMS algorithm are consistent.
5. CONCLUSION
In this paper, a diffusion-based SI-LMS algorithm has been pre-
sented, which exploited the most recent estimates available in the
network graph to improve the convergence of the estimates through-
out the network. This was achieved through the inclusion of an ad-
ditional information diffusion step, which is carried out in a serial
i
t
e
a
n
d
r
o
o
c
−
y
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
Different signal powers − N=20,M=5
ATC LMS
SI−LMS
5
0
−5
−10
−15
−20
)
B
d
(
E
S
M
0.7
0.8
0.9
1
−25
0
50
100
200
150
250
Iteration number
300
350
400
0.5
0.6
0.4
x−coordinate
Fig. 4. The topology of the network for the results of Figs. 5 to 7.
11
10
9
8
7
11
10
9
8
7
)
B
d
(
k
,
2 u
σ
)
B
d
(
k
,
2 u
σ
Signal Power
5
10
(a)
15
20
5
10
15
20
Node index
(c)
−15
)
B
d
(
k
,
2 v
σ
−20
−25
−15
−20
−25
)
B
d
(
k
,
2 v
σ
Noise power
5
10
(b)
15
20
5
10
15
20
Node index
(d)
Fig. 5. The details of signal and noise power at the nodes in the
network of Fig. 4 considered in the results of Figs. 6 and 7, respec-
tively.
Same signal powers − N=20,M=5
ATC LMS
SI−LMS
5
0
−5
−10
−15
−20
)
B
d
(
E
S
M
−25
0
50
100
200
150
250
Iteration number
300
350
400
Fig. 6. The network MSE of the network with topology of Fig. 4 and
the signal and noise variance parameters provided in of Fig. 5(a) and
5(b).
manner. A discussion of the costs of the proposed SI-LMS algo-
rithm in terms of increased computation required at the nodes in the
Fig. 7. The network MSE of the network with topology of Fig. 4 and
the signal and noise variance parameters provided in of Fig. 5(c) and
5(d).
Different signal powers − N=20,M=5
ATC LMS
SI−LMS
0
−5
−10
−15
−20
)
B
d
(
E
S
M
−25
1
10
20
30
50
40
60
Iteration number
70
80
90
100
Fig. 8. The network MSE of the network with topology of Fig. 4
and the variance parameters provided in of Fig. 5(c) and 5(d) for the
case when a larger step size is used in the adaptive algorithms.
graph and additional necessary communication of estimates required
for information diffusion was provided, demonstrating that the pro-
posed algorithm is not prohibitively costly considering the benefits
offered. Numerical results justify the proposed SI-LMS algorithm
and illustrate its performance advantages.
6. REFERENCES
[1] D. P. Bertsekas, "A new class of incremental gradient methods
for least squares problems," SIAM J. on Optimization, vol. 7,
no. 4, pp. 913 -- 926, Apr. 1997.
[2] J.N. Tsitsiklis and M. Athans, "Convergence and asymptotic
agreement in distributed decision problems," IEEE Trans. Au-
tom. Control, vol. 29, no. 1, pp. 42 -- 50, Jan 1984.
[3] C. G. Lopes and A. H. Sayed, "Diffusion least-mean squares
over adaptive networks: Formulation and performance analy-
sis," IEEE Trans. Signal Process., vol. 56, no. 7, pp. 3122 --
3136, July 2008.
[4] F. S.Cattivelli and A. H. Sayed, "Diffusion lms strategies for
[19] S. Xu, R.C. de Lamare, and H.V. Poor, "Adaptive link selec-
tion strategies for distributed estimation in diffusion wireless
networks," in Proc. IEEE ICASSP, May 2013, pp. 5402 -- 5405.
[20] S. Xu, R. C. de Lamare, and H. V. Poor, "Adaptive link se-
lection strategies for distributed estimation in wireless sensor
networks," EURASIP Journal on Advances in Signal Process-
ing, 2015.
[21] X. Zhao, S-Y Tu, and A.H. Sayed,
"Diffusion adaptation
over networks under imperfect information exchange and non-
stationary data," IEEE Trans. Signal Process., vol. 60, no. 7,
pp. 3460 -- 3475, July 2012.
[22] A. Rastegarnia, W.M. Bazzi, A. Khalili, and J.A. Chambers,
"Diffusion adaptive networks with imperfect communications:
link failure and channel noise," IET Signal Processing, vol. 8,
no. 1, pp. 59 -- 66, Feb 2014.
[23] X. Zhao and A.H. Sayed, "Performance limits for distributed
estimation over lms adaptive networks," IEEE Trans. Signal
Process., vol. 60, no. 10, pp. 5107 -- 5124, Oct 2012.
[24] J. Chen and A.H. Sayed, "On the learning behavior of adap-
tive networks part i: Transient analysis," IEEE Trans. Inform.
Theory, vol. 61, no. 6, pp. 3487 -- 3517, June 2015.
[25] J. Zhang and M. Fossorier, "Shuffled belief propagation de-
coding," in Asilomar Conf. Signals, Systems and Computers,
Nov 2002, vol. 1, pp. 8 -- 15 vol.1.
[26] D.E. Hocevar, "A reduced complexity decoder architecture via
layered decoding of LDPC codes," in IEEE Workshop on Sig-
nal Processing Systems, SIPS, Oct 2004, pp. 107 -- 112.
[27] C.T. Healy and R.C. de Lamare, "Knowledge-aided informed
dynamic scheduling for LDPC decoding," in IEEE Conf. on
Communication Workshop, ICCW, June 2015, pp. 2212 -- 2217.
[28] Cornelius T Healy and Rodrigo C de Lamare, "Design of ldpc
codes based on multipath emd strategies for progressive edge
growth," IEEE Transactions on Communications, vol. 64, no.
8, pp. 3208 -- 3219, 2016.
[29] A.G.D. Uchoa, C.T. Healy, and R.C. de Lamare, "Iterative de-
tection and decoding algorithms for mimo systems in block-
fading channels using ldpc codes,"
IEEE Trans. Vehicular
Tech., 2016.
[30] R. C. de Lamare and R. Sampaio-Neto,
"Minimum mean-
squared error iterative successive parallel arbitrated decision
feedback detectors for ds-cdma systems," IEEE Transactions
on Communications, vol. 56, no. 5, pp. 778 -- 789, 2008.
[31] R. C. de Lamare, "Adaptive and iterative multi-branch mmse
decision feedback detection algorithms for multi-antenna sys-
tems," IEEE Transactions on Wireless Communications, vol.
12, no. 10, pp. 5294 -- 5308, 2013.
[32] A.H. Sayed,
"Diffusion adaptation over networks,"
in E-
Reference Signal Processing, R. Chellapa and S. Theodoridis,
Eds. Elservier, 2013.
distributed estimation," IEEE Trans. Signal Process., vol. 58,
no. 3, pp. 1035 -- 1048, Mar. 2010.
[5] J. Chen and A. H. Sayed, "Diffusion adaptation strategies for
distributed optimization and learning over networks," IEEE
Trans. Signal Process., vol. 60, no. 8, pp. 4289 -- 4305, Aug
2012.
[6] SY Tu and A. H. Sayed, "Diffusion strategies outperform con-
sensus strategies for distributed estimation over adaptive net-
works," Signal Processing, IEEE Transactions on, vol. 60, no.
12, pp. 6217 -- 6234, Dec 2012.
[7] Y. Liu, C. Li, and Z. Zhang,
"Diffusion sparse least-mean
squares over networks," IEEE Trans. Signal Process., vol. 60,
no. 8, pp. 4480 -- 4485, Aug 2012.
[8] S. Chouvardas, K. Slavakis, Y. Kopsinis, and S. Theodoridis,
"A sparsity promoting adaptive algorithm for distributed learn-
ing," IEEE Trans. Signal Process., vol. 60, no. 10, pp. 5412 --
5425, Oct 2012.
[9] R. C. de Lamare and R. Sampaio-Neto, "Adaptive reduced-
rank processing based on joint and iterative interpolation, dec-
imation, and filtering," IEEE Transactions on Signal Process-
ing, vol. 57, no. 7, pp. 2503 -- 2514, 2009.
[10] Rui Fa, R. C. de Lamare, and Lei Wang, "Reduced-rank stap
schemes for airborne radar based on switched joint interpola-
tion, decimation and filtering algorithm," IEEE Transactions
on Signal Processing, vol. 58, no. 8, pp. 4182 -- 4194, 2010.
[11] Yang Zhaocheng, RC De Lamare, and Li Xiang, "L1 regu-
larized stap algorithm with a generalized sidelobe canceler ar-
chitecture for airborne radar," IEEE Transactions on Signal
Processing, vol. 60, no. 2, pp. 674 -- 686, 2012.
[12] S. Xu, R.C. de Lamare, and H.V. Poor,
"Distributed com-
pressed estimation based on compressive sensing," IEEE Sig-
nal Processing Letters, vol. 22, no. 9, pp. 1311 -- 1315, Sept
2015.
[13] T. G. Miller, S. Xu, R. C. de Lamare, and H. V. Poor,
"Distributed spectrum estimation based on alternating mixed
discrete-continuous adaptation," IEEE Signal Processing Let-
ters, vol. 23, no. 4, pp. 551 -- 555, 2016.
[14] N. Takahashi, I. Yamada, and A.H. Sayed, "Diffusion least-
mean squares with adaptive combiners: Formulation and per-
formance analysis," IEEE Trans. Signal Process., vol. 58, no.
9, pp. 4795 -- 4810, Sept 2010.
[15] S-Y Tu and A.H. Sayed, "Optimal combination rules for adap-
tation and learning over networks,"
in IEEE Workshop on
Comp. Adv. in Multi-Sensor Adaptive Process., CAMSAP, Dec
2011, pp. 317 -- 320.
[16] C.G. Lopes and A.H. Sayed, "Diffusion adaptive networks
in Proc. IEEE ICASSP, March
with changing topologies,"
2008, pp. 3285 -- 3288.
[17] P. Clarke and R. C. de Lamare, "Transmit diversity and re-
lay selection algorithms for multirelay cooperative mimo sys-
tems," IEEE Transactions on Vehicular Technology, vol. 61,
no. 3, pp. 1084 -- 1098, 2012.
[18] Tong Peng and Rodrigo C de Lamare,
"Adaptive buffer-
aided distributed space-time coding for cooperative wireless
networks," IEEE Transactions on Communications, vol. 64,
no. 5, pp. 1888 -- 1900, 2016.
|
1902.06039 | 4 | 1902 | 2019-04-11T02:54:29 | PT-ISABB: A Hybrid Tree-based Complete Algorithm to Solve Asymmetric Distributed Constraint Optimization Problems | [
"cs.MA",
"cs.DS"
] | Asymmetric Distributed Constraint Optimization Problems (ADCOPs) have emerged as an important formalism in multi-agent community due to their ability to capture personal preferences. However, the existing search-based complete algorithms for ADCOPs can only use local knowledge to compute lower bounds, which leads to inefficient pruning and prohibits them from solving large scale problems. On the other hand, inference-based complete algorithms (e.g., DPOP) for Distributed Constraint Optimization Problems (DCOPs) require only a linear number of messages, but they cannot be directly applied into ADCOPs due to a privacy concern. Therefore, in the paper, we consider the possibility of combining inference and search to effectively solve ADCOPs at an acceptable loss of privacy. Specifically, we propose a hybrid complete algorithm called PT-ISABB which uses a tailored inference algorithm to provide tight lower bounds and a tree-based complete search algorithm to exhaust the search space. We prove the correctness of our algorithm and the experimental results demonstrate its superiority over other state-of-the-art complete algorithms. | cs.MA | cs |
PT-ISABB: A Hybrid Tree-based Complete Algorithm to Solve
Asymmetric Distributed Constraint Optimization Problems
Yanchen Deng
College of Computer Science,
Chongqing University
Chongqing, China
[email protected]
Ziyu Chen∗
College of Computer Science,
Chongqing University
Chongqing, China
[email protected]
Dingding Chen
College of Computer Science,
Chongqing University
Chongqing, China
[email protected]
Xingqiong Jiang
College of Computer Science,
Chongqing University
Chongqing, China
[email protected]
Qiang Li
College of Electrical Engineering,
Chongqing University
Chongqing, China
[email protected]
ABSTRACT
Asymmetric Distributed Constraint Optimization Problems (AD-
COPs) have emerged as an important formalism in multi-agent
community due to their ability to capture personal preferences.
However, the existing search-based complete algorithms for AD-
COPs can only use local knowledge to compute lower bounds,
which leads to inefficient pruning and prohibits them from solving
large scale problems. On the other hand, inference-based complete
algorithms (e.g., DPOP) for Distributed Constraint Optimization
Problems (DCOPs) require only a linear number of messages, but
they cannot be directly applied into ADCOPs due to a privacy
concern. Therefore, in the paper, we consider the possibility of
combining inference and search to effectively solve ADCOPs at an
acceptable loss of privacy. Specifically, we propose a hybrid com-
plete algorithm called PT-ISABB which uses a tailored inference
algorithm to provide tight lower bounds and a tree-based complete
search algorithm to exhaust the search space. We prove the correct-
ness of our algorithm and the experimental results demonstrate its
superiority over other state-of-the-art complete algorithms.
KEYWORDS
ADCOP; Complete algorithms; Search; Inference
ACM Reference Format:
Yanchen Deng, Ziyu Chen, Dingding Chen, Xingqiong Jiang, and Qiang
Li. 2019. PT-ISABB: A Hybrid Tree-based Complete Algorithm to Solve
Asymmetric Distributed Constraint Optimization Problems. In Proc. of the
18th International Conference on Autonomous Agents and Multiagent Systems
(AAMAS 2019), Montreal, Canada, May 13 -- 17, 2019, IFAAMAS, 9 pages.
1 INTRODUCTION
Distributed Constraint Optimization Problems (DCOPs) [28] are a
fundamental framework in multi-agent systems where agents co-
operate with each other to optimize a global objective. DCOPs have
∗Corresponding author.
Proc. of the 18th International Conference on Autonomous Agents and Multiagent Systems
(AAMAS 2019), N. Agmon, M. E. Taylor, E. Elkind, M. Veloso (eds.), May 13 -- 17, 2019,
Montreal, Canada. © 2019 International Foundation for Autonomous Agents and
Multiagent Systems (www.ifaamas.org). All rights reserved.
been successfully deployed in many real world applications includ-
ing smart grids [6], radio frequency allocation [15], task scheduling
[25], etc.
Algorithms for DCOPs can generally be classified into two cate-
gories, i.e., complete algorithms and incomplete algorithms. Search-
based complete algorithms like SBB [10], AFB [7], ConFB [16],
ADOPT [14] and its variants [9, 27] perform distributed searches
to exhaust the search space, while inference-based complete al-
gorithms including Action-GDL [26], DPOP [20] and its variants
[21, 22] use dynamic programming to optimally solve problems. In
contrast, incomplete algorithms including local search [12, 17, 29],
GDL-based algorithms [3, 4, 24, 30] and sampling-based algorithms
[5, 18] trade optimality for small computational efforts.
Asymmetric Distributed Constraint Optimization Problems (AD-
COPs) [8] are a notable extension to DCOPs, which can capture
ubiquitous asymmetric structures in real world scenarios [2, 13, 23].
That is, a constraint in an ADCOP explicitly defines the exact pay-
off for each participant instead of assuming equal payoffs for con-
strained agents. Solving ADCOPs is more challenging since algo-
rithms must evaluate and aggregate the payoff for each participant
of a constraint. ATWB and SABB [8] are asymmetric versions of
AFB and SBB based on an one-phase strategy in which the algo-
rithms systematically check each side of the constraints before
reaching a full assignment. Besides, AsymPT-FB [11] is another
search-based complete algorithm for ADCOPs, which implements
a variation of forward bounding on a pseudo tree. However, to
the best of our knowledge, there is no asymmetric adaptation of
inference-based complete algorithms for DCOPs (e.g., DPOP). That
is partially because these algorithms require the total knowledge
of each constraint to perform variable elimination optimally. In
other words, parent agents must surrender their private constraints
to eliminate their children variables, which is unacceptable in a
asymmetric scenario.
In this paper, we investigate the possibility of combining both
inference and search to efficiently solve ADCOPs at an acceptable
loss of privacy. Specifically, our main contributions are listed as
follows.
• We propose a hybrid tree-based complete algorithm for AD-
COPs, called PT-ISABB.1 The algorithm first uses a tailored
version of ADPOP [19] to solve a subset of constraints, and
the inference results stored in agents are served as look-up
tables for tight lower bounds. Then, a variant of SABB is
implemented on a pseudo tree to guarantee optimality.
• We theoretically show the completeness of our proposed
algorithm. Moreover, we also prove that the lower bounds
in PT-ISABB are at least as tight as the ones in AsymPT-FB
when its maximal dimension limit k = ∞ .
• We empirically evaluate our algorithm on various bench-
marks. Our study shows that PT-ISABB requires signifi-
cantly fewer messages and lower NCLOs than state-of-the-
art search-based complete algorithms including AsymPT-FB.
The experimental results also indicate that our proposed
algorithm leaks less privacy than AsymPT-FB when solving
complex problems.
2 BACKGROUND
In this section, we review the preliminaries including ADCOPs,
pseudo tree, DPOP and ADPOP.
2.1 Asymmetric Distributed Constraint
Optimization Problems
An asymmetric distributed constraint optimization problem can be
defined by a tuple ⟨A, X , D, F⟩ in which
• A = {a1, . . . , aq} is a set of agents
• X = {x1, . . . , xn} is a set of variables
• D = {D1, . . . , Dn} is a set of finite and discrete domains.
Each variable xi takes a value from Di
• F = { f1, . . . , fm} is a set of constraints. Each constraint
fi : Di1 ×· · ·× Dik → Rk
+ defines a set of non-negative costs
for every possible value combination of the set of variables
it is involved in
For the sake of simplicity, we assume that an agent only controls
a variable and all constraints are binary. Therefore, the term agent
and variable can be used interchangeably. Besides, for the constraint
between xi and xj, we denote the private cost functions for xi and
xj as fij and fji, respectively. Note that in the asymmetric setting,
fij does not necessarily equal to fji. A solution to an ADCOP is the
assignments to all the variables with the minimal aggregated cost.
An ADCOP can be visualized by a constraint graph in which the
vertexes denote the variables and the edges denote the constraints
between agents. Fig. 1 (a) visualizes an ADCOP with four agents
and four constraints.
2.2 Pseudo Tree
A pseudo tree is an ordered arrangement to a constraint graph
in which agents in different branches are independent and thus
search can be performed in parallel on these independent branches.
A pseudo tree can be generated by a depth-first traverse of the
constraint graph, which categorizes the constraints into tree edges
and pseudo edges. For an agent ai, we denote its parent as P(ai)
which is the ancestor connecting to ai through a tree edge, its
1The source code is available at https://github.com/czy920/DCOPSovlerAlgorithm_PTISABB.
(a)
(b)
Figure 1: An example of constraint graph and pseudo tree
pseudo parents as PP(ai) which is a set of ancestors connecting to
ai through pseudo edges, its children and pseudo children as C(ai)
and PC(ai) which are the sets of descendants connecting to ai via
tree edges and pseudo edges, respectively. For the sake of clarity, we
denote all the parents of ai as AP(ai) = PP(ai) ∪ {P(ai)}. We also
denote its separator, i.e., the set of ancestors that are constrained
with ai or its descendants, as Sep(ai) [21]. Fig. 1 (b) gives a possible
pseudo tree deriving from Fig. 1 (a).
2.3 DPOP and ADPOP
DPOP is an important inference-based algorithm that performs
dynamic programming on a pseudo tree, starting with a phase of
utility propagation. In the phase, each agent joins the received util-
ities from its children with its local utility, eliminates its dimension
by calculating the optimal utility for each assignment combination
of its separator, and propagates the reduced utility to its parent.
After that, a value propagation phase starts from the root agent. In
the phase, each agent chooses the optimal assignment according to
the utilities calculated in the previous phase and the assignments
from its parent, and broadcasts the extended assignments to its
children. The algorithm terminates when all agents have chosen
their optimal assignments.
Although DPOP only requires a linear number of messages to
solve a DCOP, its memory consumption is still exponential in in-
duced width, which prohibits it from solving more complex prob-
lems. Thus, Petcu et. al, proposed ADPOP which is an approximate
version of DPOP and allows the desired trade-off between solution
quality and computational complexity. Specifically, ADPOP imposes
a limit maxDim on the maximum number of dimensions in each
message. When the number of dimensions in an outgoing message
exceeds the limit, the algorithm drops a set of dimensions to stay
below the limit. That is, the algorithm computes an upper bound
and a lower bound by applying a maximal/minimal projection on
these dimensions. During the value propagation phase, agents can
make decisions according to the highest utilities in either upper
bounds or lower bounds.
3 PROPOSED METHOD
In this section, we present our proposed PT-ISABB, a two-phase
hybrid complete algorithm for ADCOPs. We begin with a motiva-
tion, and then present the details of the inference phase and the
search phase, respectively.
a minimal projection (line 12 - 13). Here, local_utili denotes the
combination of the constraints between agent ai and its parent and
pseudo parents enforced in ai side, i.e.,
local_utili =
fij
aj ∈AP(ai)
Note that in our algorithm we do not require the parent agents
to disclose their private functions to perform inference exactly.
In this way, a local utility table only involves the functions of
that agent and the privacies of its parent and pseudo parents are
therefore guaranteed. On the other hand, however, ignoring the
private functions of parents and pseudo parents leads to severe
inconsistencies when performing variable elimination. In other
words, we actually trade lower bound tightness for privacy. We try
to alleviate the problem by performing non-local elimination which
is elaborated as follows.
When ai receives a UTIL message from its child ac, it joins the
utility from ac with its corresponding private constraint function
and then eliminates the dimension xc for a more complete utility
Child_utilc
(line 4). Compared to DPOP and ADPOP, the elimina-
i
tion of each variable is postponed to its parent in the pseudo-tree.
Taking Fig. 1 (b) for example, the UTIL message from a3 to a2 is
given by
f32 + f31
if the maximal dimension limit k ≥ 3, and the elimination of x3 is
actually performed by a2. That is,
3
2 = min
x3
(f23 + f32 + f31)
Child_util
Then, ai initiates the search phase after receiving all UTIL messages
from its children if it is the root agent (line 6 - 8). Otherwise, it
propagates the joint utility to its parent (line 9 - 10).
3.3 Search Phase
The phase performs a branch-and-bound search on a pseudo tree
to exhaust the search space. Specifically, each branching agent de-
composes the problem into several subproblems, and each of its
children solves a subproblem in parallel. To detect and discard the
suboptimal solution, each agent maintains a upper bound for its sub-
problem and an lower bound for each value in its domain. Therefore,
each agent ai needs to maintain the following data structures.
• Srch_valc
records the assignment currently being explored
i
in the subtree rooted at ac ∈ C(ai). The data structure is
necessary because asynchronous search is carried out in
parallel in sub-trees based on different possible values of xi.
• hiдh_costi(di) is the cost for di ∈ Di between ai and its par-
ent and pseudo parents under the current partial assignment
(Cpa), which is initially set to the cost enforced in ai side.
That is,
hiдh_costi(di) =
fi j(di, Cpa(xj))
(1)
i (di) is the lower bound of child ac ∈ C(ai) for di ∈ Di,
• lbc
which is initially set to the utility under Cpa and di. That is,
(2)
i (Cpa[Sep(ac)], xi = di)
i (di) = Child_util c
lbc
aj ∈AP(ai)
where Cpa[Sep(ac)] is a slice to Cpa under Sep(ac), i.e.,
Cpa[Sep(ac)] = {(xj , dj)(xj , dj) ∈ Cpa ∧ xj ∈ Sep(ac)}
Figure 2: Pseudo code of inference phase
3.1 Motivation
The existing search-based complete algorithms for ADCOPs can
only use local knowledge to compute lower bounds, which leads
to inefficient pruning. More specifically, unassigned agents report
the best local costs under the given partial assignments to compute
lower bounds. Taking Fig. 1 as an example, in AsymPT-FB agent
a1 can receive LB_reports from a2 and a3. As a consequence, a1
can only be aware of the lower bounds of f21 and f31 and does
not have any knowledge about the remaining constraints (i.e., the
constraints between a2 and a3, a2 and a4). On the other hand, in-
ference algorithms like DPOP are able to aggregate and propagate
the global utility, but they are not applicable to ADCOPs due to
a privacy concern. For example, a3 needs to know both f13 and
f23 to optimally eliminate x3, which violates the privacies of a1
and a2. Thus, to overcome the pathologies, we propose a novel hy-
brid scheme to solve ADCOPs, which combines both inference and
search. Specifically, the scheme consists of the following phases.
• Inference phase: performing a bottom-up utility propaga-
tion with respect to a subset of constraints to build look-up
tables for lower bounds
• Search phase: using a tree-based complete search algorithm
for ADCOPs to exhaust the search space and guarantee opti-
mality
In this paper, we propose a tailored version of ADPOP for the
inference phase to avoid the severe privacy loss and exponential
memory consumption in DPOP. Furthermore, we implement SABB
on a pseudo tree for the search phase and propose an algorithm
called PT-ISABB. Although they both operate on pseudo trees, our
algorithm excels AsymPT-FB twofold. When its maximal dimension
limit k = ∞, the lower bounds in our algorithm are at least as tight
as the ones in AsymPT-FB (see Property 4.1 for detail). Moreover,
PT-ISABB avoids to perform forward bounding which is expensive
during the search phase.
3.2 Inference Phase
Fig. 2 gives the sketch of the inference phase for PT-ISABB. The
phase begins with leaf agents who send their local utilities to their
parents via UTIL messages (line 1 - 3). Particularly, if the number
of dimensions in the utility exceeds the limit k (line 11), we drop
the dimensions of the highest ancestors to stay below the limit by
Algorithm1:InferencephaseforaiWhenInitialization:1joini←copy(local_utili)2ifaiisaleafthen3SendUtil()WhenreceivedUTIL(child_util)fromac∈C(ai):4Child_utilci←minxcchild_util⊗fic5joini←joini⊗Child_utilci6ifaihasreceivedallUTILfromC(ai)then7ifaiistherootthen8startSearchphase9else10SendUtil()FunctionSendUtil():11ifdim(joini)>kthen12selectS⊂dim(joini)\xi,s.t.S=dim(joini)−k13joini←minSjoini14sendUTIL(joini)toP(ai)1 Figure 4: Pseudo code of search phase (auxiliary functions)
• Spai(di) is the optimal assignment to its subproblem under
Cpa when xi = di and is initially set to {xi = di}. Particu-
larly, Spai(d∗
i ) is the optimal solution if ai is the root agent,
where d∗
= arg mindi ∈Cmplt_vali lbi(di).
i
Fig. 3 and Fig. 4 give the pseudo codes of the search phase for
PT-ISABB. The phase begins with the root agent sending the first
element in its domain to its children (line 1 - 4). When an agent ai
receives a CPA message from its parent, it first stores the partial
assignment Cpa and upper bound ubi and then finds the first fea-
sible assignment (line 5 - 7), i.e., the first assignment di such that
lbi(di) < ubi (line 57 - 60). If such an assignment exists, ai sends
COST_REQ messages to its parent and pseudo parents to request
the private costs of other side for di (line 8 - 10, line 13). Otherwise,
it sends a BACKTRACK message with an infinity cost and an empty
subproblem assignment (line 64 - 65) to its parent to announce that
the given Cpa is infeasible (line 11 - 12).
When ai receives a COST message for di, it adds the other side
cost to hiдh_costi(di) (line 14). After receiving all the COST mes-
sages for di from its parent and pseudo parents, ai is able to de-
termine whether it should continue to explore di. If ai is a leaf
agent, it just updates the current upper bound and switches to the
next feasible assignment d′
(line 17 - 19) since the search space no
longer needs to be expanded. If such d′
i
exists, ai requests costs for
i
the assignment (line 20 - 21). Otherwise, it backtracks to its parent
by reporting the best cost and the best subproblem assignment ex-
plored under Cpa (line 22 - 23, line 66 - 68). If ai is not a leaf agent
and the current lower bound for di is still less than its upper bound,
it expands the search space by sending CPA messages to its children
who are going to explore di (line 25 - 27). Each message contains
an extended partial assignment (line 61) and an upper bound which
is the remainder after deducting the cost incurred by di and the
lower bounds of the other children from ai's upper bound (line 62).
Otherwise, di is proven to be suboptimal and the agent switches to
the next feasible assignment (line 28 - 30). If such an assignment
exists, ai requests costs for it (line 31 - 32). A backtrack takes place
if all children exhaust ai's domain (line 33 - 34).
Figure 3: Pseudo code of search phase (message passing)
When ai receives a BACKTRACK message from ac, it re-
places the initial lower bound with the actual cost reported
by ac (or ∞ if di is infeasible for ac given Cpa).
• lbi(di) is the lower bound for di ∈ Di, i.e.,
i (di)
lbc
lbi(di) = hiдh_costi(di) +
(3)
• Cmplt_valsi is the set of assignments for which ai has re-
ceived all BACKTRACK messages from its children, and is
initially set to ∅. Particularly, Cmplt_valsi = Di if ai is a leaf
agent.
• lb∗
is the best cost explored under Cpa, which is given by
i
(4)
min
lb∗
i =
ac ∈C(ai)
di ∈Cmpl t_valsi
Particularly, if Cmplt_valsi = ∅, lb∗
lbi(di)
= ∞.
i
Algorithm2:Searchphaseforai(messagepassing)WhenInitialization:1ifaiistherootthen2InitializeVariables()3di←thefirstelementinDi,Cpa←(xi,di)4sendCPA(Cpa,∞)toac,∀ac∈C(ai)WhenreceivedCPA(Cpa,ubi)fromP(ai):5store{Cpa,ubi}6InitializeVariables()7di←NextFeasibleAssignment(null)8ifdi(cid:44)nullthen9sendCOST_REQ(Cpa(xj),di)toaj,∀aj∈AP(ai)10Srch_valci←di,∀ac∈C(ai)11else12SendBacktrack()WhenreceivedCOST_REQ(di,dc)fromac∈C(ai)∪PC(ai):13sendCOST(fic(di,dc),dc)toacWhenreceivedCOST(cost,di)fromap∈AP(ai):14hiдh_costi(di)←hiдh_costi(di)+cost15ifaihasnotreceivedallCOSTfromAP(ai)fordithen16return17ifaiisaleafthen18ubi←min(ubi,lbi(di))19d′i←NextFeasibleAssignment(di)20ifd′i(cid:44)nullthen21sendCOST_REQ(Cpa(xj),d′i)toaj,∀aj∈AP(ai)22else23SendBacktrack()24else25iflbi(di)<ubithen26foreachac∈C(ai)∧Srch_valci=dido27SendCpa(di,ac)28else29d′i←NextFeasibleAssignment(di)30Srch_valci←d′i,∀ac∈C(ai)∧Srchci=di31ifd′i(cid:44)nullthen32sendCOST_REQ(Cpa(xj),d′i)toaj,∀aj∈AP(ai)33elseif∀ac∈C(ai),Srch_valci=nullthen34SendBacktrack()WhenreceivedBACKTRACK(lb∗c,Spac)fromac∈C(ai):35di←Srch_valci,lbci(di)←lb∗c,Spai(di)←Spai(di)∪Spac36ifaihasreceivedallBACKTRACKfromC(ai)fordithen37Cmplt_valsi←Cmplt_valsi∪{di}//marksdiascomplete38ubi←min(ubi,lbi(di))39d′i←NextFeasibleAssignment(di),Srch_valci←d′i40ifd′i(cid:44)nullthen41ifaihasreceivedallCOSTfromAP(ai)ford′ithen42SendCpa(d′i,ac)43elseifaihasnotrequestedcostsford′ithen44sendCOST_REQ(Cpa(xj),d′i)toaj,∀aj∈AP(ai)45elseif∀ac∈C(ai),Srch_valci=nullthen46ifaiistherootthen47sendTERMINATEtoac,∀ac∈C(ai)48terminate49else50SendBacktrack()WhenreceivedTERMINATEfromP(ai):51sendTERMINATEtoac,∀ac∈C(ai)52terminate1Algorithm2:Searchphaseforai(auxiliaryfunctions)FunctionInitializeVariables():53initializeCmplt_valsi54initializehiдh_costi(di),Spai(di),∀di∈Di55initializeSrch_valci,∀ac∈C(ai)56initializelbci(di),∀ac∈C(ai),di∈DiFunctionNextFeasibleAssignment(di):57d′i←theelementnexttodiinDi58whiled′i(cid:44)null∧lbi(d′i)≥ubido59d′i←thenextelementinDi60returnd′iFunctionSendCpa(di,ac):61TmpCpa←Cpa∪(xi,di)62ubc←ubi−hiдh_costi(di)−(cid:80)aj∈C(ai)∧j(cid:44)clbji(di)63sendCPA(TmpCpa,ubc)toacFunctionSendBacktrack():64ifCmplt_valsi=∅then65sendBACKTRACK(∞,∅)toP(ai)66else67d∗i←argmindi∈Cmplt_valsilbi(di)68sendBACKTRACK(lb∗,Spai(d∗i))toP(ai)1 When ai receives a BACKTRACK message for di from a child ac,
i (di) with the actual
it updates the corresponding lower bound lbc
cost lb∗
c reported by ac if Cpa and the assignment di is feasible
(otherwise lb∗
c = ∞), and merges the best assignments from ac (line
35). If ai has received all the BACKTRACK messages for di from
its children, it marks di as complete and updates the current upper
bound for its subproblem (line 36 - 38). ai also needs to determine
the next assignment d′
exists and
i
ai has received all the COST messages from its parent and pseudo
parents, it informs ac to explore d′
by sending a CPA message (line
40 - 42). Otherwise, ai requests costs for d′
if it has not been done
(line 43 - 44). If d′
does not exist and all children have exhausted
i
ai's domain, ai informs its children to terminate and terminates
itself if it is the root agent (line 45 - 48). Otherwise, it backtracks to
its parent (line 49 - 50).
for ac to explore (line 39). If d′
i
i
i
4 THEORETICAL RESULTS
4.1 Correctness
In this section, we first prove the termination and optimality, and
further establish the completeness of PT-ISABB.
Lemma 4.1. PT-ISABB will terminate after a finite number of iter-
ations.
Proof. Directly from the pseudo codes, the inference phase will
terminate since it only needs a linear number of messages. Thus, to
prove the termination, it is enough to show that the same partial
assignment cannot be explored twice in the search phase, i.e., an
agent will not receive two identical Cpas. Obviously, the claim holds
for the root agent since it does not receive any CPA message. For
an agent and a given Cpa from its parent, it will send several CPA
messages to each child. Since each of them contains the different
assignments of the agent (line 29, line 39, line 57 - 60), the Cpas sent
to the child are all different. Therefore, the termination is hereby
guaranteed.
□
Lemma 4.2. For an agent ai and a given Cpa, the cost incurred by
any assignment Spai to the subtree rooted at ai with the assignment
(xi = di) is no less than the corresponding lower bound lbi(di).
Proof. The lemma for a leaf is trivial since lbi(di) is set to the
cost of di which is obviously no greater than the cost of the feasible
assignment. We now focus on no-leaf agents. Recall that ai will
replace the original lower bound with the actual cost reported by
ac after receiving a BACKTRACK message for di from ac ∈ C(ai)
(line 35). Thus, to prove the lemma, it is sufficient to show that
i (di) is no greater than the actual cost
the initial lower bound lbc
of Spac ,∀ac ∈ C(ai), where Spac ⊂ Spai is the assignment to the
subtree rooted at ac.
Consider the induction basis, i.e., ai's children are leafs. For each
child ac ∈ C(ai), we have
cost(Spac) =
fjc(dj, dc) + fc j(dc, dj)
fic(di, xc) +
aj ∈AP(ac)
≥ min
xc
fc j(xc, dj)
aj ∈AP(ac)
≥ Child_util c
i (Cpa[Sep(ac)], xi = di) = lbc
i (di)
where dl is the assignment to xl in Cpa or Spac.The equation in
the second to the last step holds when the maximal dimension limit
k = ∞. Thus, the lemma holds for the basis.
Assume that the lemma holds for all ac ∈ C(ai). Next, we are
going to show the lemma holds for ai as well. For each child ac ∈
C(ai), we have
cost(Spac ) =
≥
fc j (dc , dj ) + fjc (dj , dc ) +
fc j (dc , dj ) + fjc (dj , dc ) +
fc j (xc , dj ) +
cost(Spac′)
lbc′
c (dc )
lbc′
c (xc )
fic (di , xc ) +
aj ∈AP(ac )
≥ min
xc
ac′∈C(ac )
ac′∈C(ac )
aj ∈AP(ac )
aj ∈AP(ac )
≥ Child_ut ilc
i (Cpa[Sep(ac )], xi = di ) = lbc
ac′∈C(ac )
i (di )
which establishes the lemma.
□
Lemma 4.3. For an agent ai and a given Cpa, any assignment to
the subtree rooted at ai with cost greater than ubi cannot be a part of
a solution with cost less than the global upper bound.
Proof. We will prove recursively, by showing that for a partial
assignment Spai to the subtree rooted at ai with cost(Spai) > ubi,
any partial assignment Spaj ⊃ Spai to the subtree rooted at aj
will have cost(Spaj) > ubj where aj = P(ai). Note that ubi could
be either an upper bound from aj via a CPA message (line 5) or
a result of updating the upper bound locally (line 18, line 38). ai
cannot backtrack by reporting Spai in the latter case since there
must exist a better partial assignment whose cost is ubi. If ubi is
received from aj, according to line 62, we have
ubj = ubi + hiдh_costj(dj) +
j (dj)
lbc
ac ∈C(aj)∧c(cid:44)i
Thus, cost(Spai) > ubi necessarily means that any partial assign-
ment Spaj ⊃ Spai will have cost(Spaj) > ubj.
□
Theorem 4.4. PT-ISABB is complete.
Proof. Immediately from Lemma 4.1, Lemma 4.2 and Lemma
4.3, the algorithm will terminate and all pruned assignments are
suboptimal. Thus, PT-ISABB is complete.
□
4.2 Lower bound tightness
Property 4.1. For an agent ai and a given Cpa, the initial lower
i (di) of ac ∈ C(ai) for di is at least as tight as the one in
bound lbc
AsymPT-FB when the maximal dimension limit k = ∞.
Proof. In AsymPT-FB, the lower bound for ac after receiving
all the LB_Reports from the subtree rooted at ac is given by the
sum of the best single side local costs of ac's descendants under
Cpa. That is,
Subtr eeLBc
i (di) =
aj ∈Desc(ac )
+ min
xc
al ∈AP(ac )
fjl (xj , dl )
al ∈Sep(ac )∩P P(aj )
min
xj
fcl (xc , dl )
where Desc(ac) is the set of the descendants of ac. For the sake of
clarity, we denote the vector of xc and its descendant variables as
i (di). Since k = ∞, the
i (di) = min
lbc
xc
i (di) ≥ SubtreeLBc
xc. Next, we will show lbc
inference phase does not drop any dimension. Thus, we have
(cid:169)(cid:173)(cid:173)(cid:171)
fjl (xj , xl ) +
fcl (xc , dl ) +
al ∈C(aj )
fcl (xc , xl ) + fic(di , xc)
fjl (xj , xl )(cid:170)(cid:174)(cid:174)(cid:172)
al ∈AP(aj )∩Desc(ai )
al ∈AP(aj )∩Sep(ac )
aj ∈Desc(ac )
fjl (xj , dl )
al ∈C(ac )
fjl (xj , dl )
aj ∈Desc(ac )
al ∈AP(aj )∩Sep(ac )
fcl (xc , dl )
+
+
+
al ∈AP(ac )
≥ min
xc
al ∈AP(ac )
Since xc ∈ xc and AP(aj) ⊃ PP(aj), the right-hand side of the
inequality in the last step can be further reduced. That is,
fjl (xj , dl )
i (di) ≥ min
lbc
xc\xc
aj ∈Desc(ac )
al ∈AP(aj )∩Sep(ac )
+ min
xc
al ∈AP(ac )
al ∈AP(ac )
+ min
xc
≥ min
xc\xc
≥
+ min
xc
aj ∈Desc(ac )
fcl (xc , dl )
fjl (xj , dl )
fjl (xj , dl )
aj ∈Desc(ac )
al ∈P P(aj )∩Sep(ac )
fcl (xc , dl )
min
xj
fcl (xc , dl )
al ∈Sep(ac )∩P P(aj )
(a)
(b)
al ∈AP(ac )
= Subtr eeLBc
i (di)
which concludes the property.
Figure 5: Performance comparison under different agent
numbers
□
and lbc
Sep(ai)+1
max
4.3 Complexity
i (di) for each child,
Since an agent ai stores Child_utilc
i
the overall space complexity in the worst case (i.e., k = ∞) is
+C(ai)Di) where dmax = maxaj ∈Sep(ai) Dj.
O(C(ai)d
Since it contains all the dimensions of Sep(ai) and itself, the size of
an UTIL message from ai is O(d
) when k = ∞. For a CPA
message, it consists of the assignment of each agent and an upper
bound. Thus, the size of a CPA message is O(A). Other messages
including COST_REQ, COST, BACKTRACK and TERMINATE carry
several scalars and thus they only require O(1) space.
Different than standard DPOP/ADPOP, PT-ISABB only requires
A − 1 messages in the inference phase since it does not have the
value propagation phase. Like any other search based complete algo-
rithm, the message number of the search phase grows exponentially
with respect to the agent number.
Sep(ai)+1
max
5 EXPERIMENTAL RESULTS
We empirically evaluate PT-ISABB with state-of-the-art search-
based complete algorithms for ADCOPs including SABB, ATWB
and AysmPT-FB on three configurations. To demonstrate the real
power of non-local elimination, we also consider SABB on a pseudo-
tree (PT-SABB) and the local elimination version of PT-ISABB (PT-
ISABB, local) with k = ∞. In the first ADCOP configuration, we set
the graph density to 0.25, the domain size to 3 and vary the agent
number from 8 to 18. The second configuration is ADCOPs with 8
agents and the domain size of 8. The graph density varies from 0.25
to 1. In the last configuration, we consider asymmetric MaxDCSPs
with 10 agent, the domain size of 10 and the graph density of 0.4,
and the tightness varies from 0.1 to 0.8. For each of the settings, we
generate 50 random instances and the results are averaged over all
instances. In our experiments, we use the number of non-concurrent
logical operations (NCLO) [16] to evaluate hardware-independent
runtime, in which the logical operations in the inference phase
are accesses to utility tables, and for the search phase and other
competitors they are constraint checks. Also, we use the message
number and the size of total information exchanged to measure the
network load. For asymmetric MaxDCSPs, we use entropy [1] to
quantify the privacy loss [8, 11]. The experiments are conducted
on an i7-7820x workstation with 32GB of memory and for each
algorithm we set the timeout to 2 minutes.
Fig. 5 gives the performance comparison on different agent num-
bers, and the average induced widths in the experiments are 1 ∼
6.84. It can be seen from the figure that all the algorithms suffer
from exponential overheads as the agent number grows. Among
them, our proposed PT-ISABB requires significant fewer messages
and lower NCLOs than the other competitors, which demonstrates
the superiority of the hybrid execution of inference and search.
On the other hand, although PT-ISABB (k = ∞, local) employs an
complete one-side inference to construct the initial lower bounds,
it is still inferior to PT-ISABB with k > 2, which demonstrate the
necessity of non-local elimination. Besides, it is worth noting that
81012141618Agent Number102103104105106107Message Number81012141618Agent Number102103104105106NCLOsATWBSABBAsymPT-FBPT-SABBPT-ISABB(k=1,local)PT-ISABB(k=2)PT-ISABB(k=4)PT-ISABB(k=6)PT-ISABB(k=1) (a)
(b)
(a)
(b)
Figure 6: Performance comparison under different graph
densities
Figure 7: Performance comparison under different tightness
PT-ISABB requires much fewer messages than AsymPT-FB even
when the maximal dimension limit k is small (e.g., k = 2). That
is because PT-ISABB does not rely on forward bounding which is
expensive in message-passing to compute lower bounds. Moreover,
the phenomenon also indicates that our algorithm can produce
tighter lower bounds even if the memory budget is relatively low.
Fig. 6 gives the results under different graph densities. The aver-
age induced widths here are 1 ∼ 6. Note that in this configuration,
the size of the search space does not change and the complexity
is reflected in the topologies. It can be concluded from the fig-
ure that all the tree-based algorithms exhibit great superiorities
when the graph density is low, and the advantages vanish as the
density grows. That is because those algorithms can effectively par-
allel the search processes on sparse problems. Dense problems, on
the other hand, usually result in pseudo trees with low branching
factors, making the tree-based algorithms require more messages
than SABB. Even so, our proposed PT-ISABB with large k still
outperforms SABB when the problems are fully connected, which
demonstrates the necessity of tighter lower bounds. Additionally,
the figure also indicates that PT-ISABB with different k performs
similarly on sparse problems, but the performances vary a lot on
dense problems. That is due to the fact that the induced width of a
pseudo tree is relatively small when solving a spare problem and
thus only a small set of dimensions is dropped during the inference
phase. Besides, it can be seen from the figure that although both
PT-ISABB (k = ∞, local) and PT-ISABB (k = ∞) perform complete
Table 1: The size of total information exchanged of each al-
gorithm under different graph densities (in KB)
inferences, the non-local elimination version requires lower NCLOs
and fewer message numbers in most of the settings. That is because
the non-local elimination version can provide tighter lower bounds
which result in efficient pruning, and thus the algorithm incurs
fewer constraint checks and messages in the search phase.
Table 1 presents the size of total information exchanged of each
algorithm under different densities. It can be seen from the table
that all the non-local elimination versions of PT-ISABB exhibit
great advantages over the other search-based competitors, except
that PT-ISABB (k = 2) is slightly inferior to AsymPT-FB when
solving the fully-connected problems. The phenomenon indicates
that although a message in the inference phase is generally larger
than the one in the search phase, the algorithms can still gain great
reductions on network traffics since the effective pruning in the
search phase greatly reduces the message number. Besides, it is
interesting to find that a large dimension limit (e.g., k = ∞) does
0.250.40.550.70.851Density102103104105106107Message NumberATWBSABBAsymPT-FBPT-SABBPT-ISABB(k=1,local)PT-ISABB(k=2)PT-ISABB(k=4)PT-ISABB(k=6)PT-ISABB(k=1)0.250.40.550.70.851Density102103104105106107NCLOsATWBSABBAsymPT-FBPT-SABBPT-ISABB(k=1,local)PT-ISABB(k=2)PT-ISABB(k=4)PT-ISABB(k=6)PT-ISABB(k=1)0.10.20.30.40.50.60.70.8Tightness102103104105106107Message Number0.10.20.30.40.50.60.70.8Tightness102103104105106107NCLOsATWBSABBAsymPT-FBPT-SABBPT-ISABB(k=1,local)PT-ISABB(k=2)PT-ISABB(k=4)PT-ISABB(k=6)PT-ISABB(k=1)AlgorithmDensity0.250.40.550.70.851ATWB91,273.2997,715.94175,594.67190,195.50266,654.30373,255.37SABB64,264.4468,342.30142,308.71164,670.95233,956.11329,344.13AsymPTFB168.071,174.823,161.299,483.7623,958.5257,261.91PT-SABB85.37990.773,260.3012,920.1035,221.34103,409.30PT-ISABB(k=∞,local)67.41581.412,137.578,653.5225,444.4573,442.96PT-ISABB(k=2)8.59343.041,661.957,426.9923,572.6869,207.66PT-ISABB(k=4)8.58143.45811.753,812.2514,755.5150,844.86PT-ISABB(k=6)8.57147.73897.093,199.699,734.1735,617.43PT-ISABB(k=∞)8.59147.55897.393,198.511,0435.5794,536.77 not necessarily result in the smallest traffic. That is because the size
of a message in the inference phase is exponential to the minimum
of the induced width and the dimension limit. Besides, it should be
noted that although PT-SABB requires more messages than ATWB
and SABB when solving fully-connected problems according to Fig.
6, it still incurs much smaller traffic due to the fact that the last
agent in ATWB and SABB needs to broadcast the reached complete
solution to all other agents once a new solution is constructed. In
contrast, agents in PT-SABB only back up the assignments to their
descendants, which are subsets of the complete solution, to their
parents via BACKTRACK messages.
Fig. 7 presents the results on asymmetric MaxDCSPs with differ-
ent tightness, and the average induced width is 3.92. This configu-
ration neither increases the search space nor affects the topologies,
but instead increases the difficulty of pruning. All the algorithms
except ATWB produces few messages when solving problems with
low tightness. That is because on these problems the algorithms
can find low upper bounds very quickly to prune most of the search
space. As the tightness grows, the number of prohibited combina-
tions increases and the algorithms can no longer find low upper
bounds promptly. As a result, the algorithms require much more
search efforts to exhaust the search space. Since they cannot ex-
ploit topologies to accelerate the search process, SABB and ATWB
perform poorly and can only solve the problems with tightness
up to 0.6. On the other hand, the tree-based algorithms divide a
problem to several smaller subproblems at each branching agent
and search the subproblems in parallel. Thus, those algorithms
exhibit better performances and solve all the problems. Among
them, our proposed PT-ISABB with k ≥ 4 incurs much smaller
overheads, which demonstrates the effectiveness of the inference
phase in computing tighter lower bounds. In other words, although
PT-ISABB only guarantees to produce lower bounds as tight as
the ones of AsymPT-FB when k = ∞ according to Property 4.1, it
requires less memory consumption to compute such lower bounds
in practice. Besides, it can be seen from the figure that PT-SABB
incurs smaller communication overheads than AsymPT-FB when
solving the problems with low tightness, which demonstrates for-
ward bounding is expensive in message-passing. Additionally, it can
be concluded that PT-ISABB algorithms with large k require much
more NCLOs than the other competitors when solving problems
with low tightness. That is no surprise since inference on problems
with large domain sizes would be more expensive, and a search-
based algorithm actually can find a feasible solution very quickly
even if the lower bounds are poor when solving these problems.
Fig. 8 gives privacy losses under different tightness. Privacy loss
in PT-ISABB comes from both the inference phase and the search
phase. Specifically, since the variable elimination is performed by
parents (i.e., line 4 of the inference phase), parents can easily figure
out which pairs of assignments are feasible with respect to the
constraints enforced in children sides from the zero entries of the
utilities from children. Thus, the inference phase would cause a
half privacy loss only on each tree edge in the worst case, which
is still better than leaking at least a half privacy if we directly use
DPOP to solve problems. Besides, the direct disclosure mechanism
of the search phase in which agents request their (pseudo) parents
to expose the private costs before expanding the search space also
leads to the privacy loss. However, the loss could be much reduced
Figure 8: Privacy losses under different tightness
via the effective pruning. It can be seen that our proposed algorithm
leaks more privacy than the other competitors when solving the
problems with low tightness. That is no surprise because these
problems usually have feasible solutions, which leads to the fact
that most of entries in a utility from a child are zero. That being said,
PT-ISABB with k ≥ 4 leaks less privacy than the other competi-
tors when solving the problems with high tightness. The reason is
twofold: parents can no longer infer the feasible assignment pairs as
the tightness grows, and the inference phase produces tight lower
bounds which lead to the effective pruning in the search phase.
Besides, it is worth mentioning that the local elimination version of
PT-ISABB performs better in terms of privacy preservation when
solving the problems with low tightness. That is because variables
are already eliminated before sending utilities to their parents. As a
result, parents can only know the best utilities they can achieve, but
cannot figure out the corresponding assignments to their children.
6 CONCLUSION
It is known that DPOP/ADPOP for DCOP cannot be directly ap-
plied to ADCOP due to a privacy concern. In this paper, we take
ADPOP into solving ADCOP for the first time by combining with a
tree-based variation of SABB, and present a two-phase complete
algorithm called PT-ISABB. In the inference phase, a non-local
elimination version of ADPOP is performed to solve a subset of
constraints and build look-up tables for the tighter lower bounds.
In the search phase, a tree-based variation of SABB is implemented
to exhaust the search space. The experimental results show that our
algorithms exhibit great superiorities over state-of-the-art search
based algorithms, as well as the local elimination version of PT-
ISABB. Also, our algorithms leak less privacy when solving complex
problems.
ACKNOWLEDGMENTS
The authors would like to thank the anonymous referees for their
valuable comments and helpful suggestions. This work is supported
by the Chongqing Research Program of Basic Research and Fron-
tier Technology under Grant No.:cstc2017jcyjAX0030, Fundamen-
tal Research Funds for the Central Universities under Grant No.:
2018CDXYJSJ0026, National Natural Science Foundation of China
under Grant No.: 51608070 and the Graduate Research and Innova-
tion Foundation of Chongqing, China under Grant No.: CYS18047
0.10.20.30.40.50.6Tightness00.10.20.30.40.50.60.7Privacy LossATWBSABBAsymPT-FBPT-SABBPT-ISABB(k=1,local)PT-ISABB(k=2)PT-ISABB(k=4)PT-ISABB(k=6)PT-ISABB(k=1) REFERENCES
[1] Ismel Brito, Amnon Meisels, Pedro Meseguer, and Roie Zivan. 2009. Distributed
constraint satisfaction with partially known constraints. Constraints 14, 2 (2009),
199 -- 234.
[2] David A Burke, Kenneth N Brown, Mustafa Dogru, and Ben Lowe. 2007. Sup-
ply chain coordination through distributed constraint optimization. In The 9th
International Workshop on DCR.
[3] Ziyu Chen, Yanchen Deng, and Tengfei Wu. 2017. An Iterative Refined Max-
sum_AD Algorithm via Single-side Value Propagation and Local Search. In Proc.
of the 16th Conference on AAMAS. 195 -- 202.
[4] Alessandro Farinelli, Alex Rogers, Adrian Petcu, and Nicholas R Jennings. 2008.
Decentralised coordination of low-power embedded devices using the max-sum
algorithm. In Proc. of the 7th AAMAS. 639 -- 646.
[5] Ferdinando Fioretto, William Yeoh, and Enrico Pontelli. 2016. A dynamic
programming-based MCMC framework for solving DCOPs with GPUs. In Interna-
tional Conference on Principles and Practice of Constraint Programming. Springer,
813 -- 831.
[6] Ferdinando Fioretto, William Yeoh, Enrico Pontelli, Ye Ma, and Satishkumar J
Ranade. 2017. A Distributed Constraint Optimization (DCOP) Approach to the
Economic Dispatch with Demand Response. In Proc. of the 16th Conference on
AAMAS. 999 -- 1007.
[7] Amir Gershman, Amnon Meisels, and Roie Zivan. 2009. Asynchronous forward
bounding for distributed COPs. Journal of Artificial Intelligence Research 34 (2009),
61 -- 88.
[8] Tal Grinshpoun, Alon Grubshtein, Roie Zivan, Arnon Netzer, and Amnon Meisels.
2013. Asymmetric Distributed Constraint Optimization Problems. Journal of
Artificial Intelligence Research 47 (2013), 613 -- 647.
[9] Patricia Gutierrez, Pedro Meseguer, and William Yeoh. 2011. Generalizing ADOPT
and BnB-ADOPT. In Proc. of the 22nd IJCAI. 554 -- 559.
[10] Katsutoshi Hirayama and Makoto Yokoo. 1997. Distributed partial constraint
satisfaction problem. In International Conference on Principles and Practice of
Constraint Programming. 222 -- 236.
[11] Omer Litov and Amnon Meisels. 2017. Forward bounding on pseudo-trees for
DCOPs and ADCOPs. Artificial Intelligence 252 (2017), 83 -- 99.
[12] Rajiv T Maheswaran, Jonathan P Pearce, and Milind Tambe. 2004. Distributed
Algorithms for DCOP: A Graphical-Game-Based Approach.. In Proceeding of
ISCA PDCS'04. 432 -- 439.
[13] Rajiv T. Maheswaran, Milind Tambe, Emma Bowring, Jonathan P. Pearce, and
Pradeep Varakantham. 2004. Taking DCOP to the Real World: Efficient Complete
Solutions for Distributed Multi-Event Scheduling. In Proc. of the 3rd AAMAS.
310 -- 317.
[14] Pragnesh Jay Modi, Wei-Min Shen, Milind Tambe, and Makoto Yokoo. 2005.
ADOPT: Asynchronous distributed constraint optimization with quality guaran-
tees. Artificial Intelligence 161, 1-2 (2005), 149 -- 180.
[15] Tânia L Monteiro, Guy Pujolle, Marcelo E Pellenz, Manoel C Penna, and
Richard Demo Souza. 2012. A multi-agent approach to optimal channel assign-
ment in wlans. In Wireless Communications and Networking Conference (WCNC).
2637 -- 2642.
[16] Arnon Netzer, Alon Grubshtein, and Amnon Meisels. 2012. Concurrent forward
bounding for distributed constraint optimization problems. Artif. Intell. 193
(2012), 186 -- 216.
[17] Steven Okamoto, Roie Zivan, and Aviv Nahon. 2016. Distributed breakout: beyond
satisfaction. In Proc. of the 25th IJCAI. 447 -- 453.
[18] Brammert Ottens, Christos Dimitrakakis, and Boi Faltings. 2017. DUCT: An Upper
Confidence Bound Approach to Distributed Constraint Optimization Problems.
ACM Transactions on Intelligent Systems and Technology (TIST) 8, 5 (2017), 69.
[19] Adrian Petcu and Boi Faltings. 2005. Approximations in distributed optimization.
In International Conference on Principles and Practice of Constraint Programming.
802 -- 806.
[20] Adrian Petcu and Boi Faltings. 2005. A Scalable Method for Multiagent Constraint
[21] Adrian Petcu and Boi Faltings. 2006. ODPOP: an algorithm for open/distributed
[22] Adrian Petcu and Boi Faltings. 2007. MB-DPOP: a new memory-bounded algo-
Optimization. In Proc. of the 19th IJCAI. 266 -- 271.
constraint optimization. In Proc. of the 21st AAAI. 703 -- 708.
rithm for distributed optimization. In Proc. of the 20th IJCAI. 1452 -- 1457.
[23] Sarvapali D Ramchurn, Perukrishnen Vytelingum, Alex Rogers, and Nick Jen-
nings. 2011. Agent-based control for decentralised demand side management in
the smart grid. In Proc. of the 10th AAMAS. 5 -- 12.
[24] Alex Rogers, Alessandro Farinelli, Ruben Stranders, and Nicholas R Jennings. 2011.
Bounded approximate decentralised coordination via the max-sum algorithm.
Artificial Intelligence 175, 2 (2011), 730 -- 759.
[25] Evan A Sultanik, Pragnesh Jay Modi, and William C Regli. 2007. On modeling
multiagent task scheduling as a distributed constraint optimization problem. In
Proc. of the 20th IJCAI. 1531 -- 1536.
[26] Meritxell Vinyals, Juan A Rodriguez-Aguilar, and Jesús Cerquides. 2009. General-
izing DPOP: Action-GDL, a new complete algorithm for DCOPs. In Proc. of The
8th AAMAS. 1239 -- 1240.
[27] William Yeoh, Ariel Felner, and Sven Koenig. 2010. BnB-ADOPT: An asyn-
chronous branch-and-bound DCOP algorithm. Journal of Artificial Intelligence
Research 38 (2010), 85 -- 133.
[28] William Yeoh and Makoto Yokoo. 2012. Distributed problem solving. AI Magazine
33, 3 (2012), 53.
[29] Weixiong Zhang, Guandong Wang, Zhao Xing, and Lars Wittenburg. 2005. Dis-
tributed stochastic search and distributed breakout: properties, comparison and
applications to constraint optimization problems in sensor networks. Artificial
Intelligence 161, 1-2 (2005), 55 -- 87.
[30] Roie Zivan and Hilla Peled. 2012. Max/min-sum distributed constraint opti-
mization through value propagation on an alternating DAG. In Proc. of the 11th
AAMAS. 265 -- 272.
|
1805.10109 | 1 | 1805 | 2018-05-25T12:39:27 | Resisting hostility generated by terror: An agent-based study | [
"cs.MA"
] | We aim to study through an agent-based model the cultural conditions leading to a decrease or an increase of discrimination between groups after a major cultural threat such as a terrorist attack. We propose an agent-based model of cultural dynamics inspired from the social psychological theories. An agent has a cultural identity comprised of the most acceptable positions about each of the different cultural worldviews corresponding to the main cultural groups of the considered society and a margin of acceptance around each of these most acceptable positions. An agent forms an attitude about another agent depending on the similarity between their cultural identities. When a terrorist attack is perpetrated in the name of an extreme cultural identity, the negatively perceived agents from this extreme cultural identity modify their margins of acceptance in order to differentiate themselves more from the threatening cultural identity. We generated a set of populations with cultural identities compatible with data given by a survey on groups' attitudes among a large sample representative of the population of France; we then simulated the reaction of these agents facing a threat. For most populations, the average attitude toward agents with the same preferred worldview as the terrorists becomes more negative; however, when the population shows some cultural properties, we noticed the opposite effect as the average attitude of the population becomes less negative. This particular context requires that the agents sharing the same preferred worldview with the terrorists strongly differentiate themselves from the terrorists' extreme cultural identity and that the other agents be aware of these changes. | cs.MA | cs | Resisting hostility generated by terror: an agent-based
study
Huet S. 1, Deffuant G. 1, Nugier A. 2, Streith M. 2, Guimond S.2
1 LISC, Irstea, 9 av. Blaise Pascal, 63178 Aubière, France – sylvie.huet and
[email protected]
2 LAPSCO, UCA, 17 rue Paul Collomp, 63037 Clermont-Ferrand, France
armelle.nugier, michel.streith and [email protected]
Abstract. We propose an agent-based model of cultural dynamics inspired from the social
psychological theories. An agent has a cultural identity made of most acceptable positions
about each of the different cultural worldviews corresponding to the main cultural groups of the
considered society and a margin of acceptance around each of these most acceptable positions.
An agent forms an attitude about another agent depending on the similarity between their cul-
tural identities. When a terrorist attack is perpetrated in the name of an extreme cultural identi-
ty, the agents which are perceived negatively from this extreme cultural identity modify their
margins of acceptance in order to differentiate themselves more from the threatening cultural
identity. We initialized agents' cultural identities with data given by a survey on groups' atti-
tudes among a large sample representative of the population of France, and then simulated
these agents facing a threat. While in most cases, agents' attitude become more negative toward
the agents belonging to the same group as the terrorists, we notice that when the population
shows some cultural properties, the opposite takes place: the average attitude of the population
for the cultural group to which the terrorists argued to belong becomes less negative. These
properties are identified and explained. They especially permit to non-terrorists agents assimi-
lated to terrorists to differentiate from them, and to other agents to perceive this change.
Keywords: Intergroup hostility, culture dynamics, Terror Management Theory,
self-opinion
Cultural worldviews are defined as "shared conceptions of reality" [1, 2]. They are an
important defense mechanism allowing people to cope with existential threats. This is
why people are motivated to maintain faith in them. Terror Management Theory
(TMT) has shown that a death fatality reminder, such as a terrorist attack, is a cultural
threat or a self-worth threat [3]. Such a threat generally leads to an increase in nega-
tive intergroup bias in order to defend one's cultural worldviews [4, 5]. However,
recent research showed that "increased prejudice and hostility are not an inevitable
response to existential threat" [6]. Some cultural properties, when becoming salient
simultaneously with the threat, increase perceived similarity of members of different
groups, and protect against an increase of the intergroup hostility [6, 7]. Understand-
ing when and why people react to a cultural or collective threat one way or the other
is a basic problem having widespread theoretical and practical implications. To deal
with this paradox, we study how the simulated change of cultural worldviews due to a
cultural threat impacts virtual groups' attitudes toward each other. Our agent-based
2
model helps to characterize cultural properties leading to acceptance or hostility be-
tween groups.
Agent-based model of culture dynamics have been seminally introduced by Axel-
rod [8]. The Axelrod model represents a culture as a set of traits and changes an un-
shared cultural trait by two agents to a shared one, with a probability depending on
their level of shared properties. Several variants have been studied [9], introducing
also a process leading traits to be more different [8, 10-13] instead of being shared.
However, none of these models consider the impact of an existential threat on the
cultural properties in relation with the self-worth and the group dynamics. This con-
clusion is also true for opinion dynamics model, very close from the cultural models.
Though some of them include the possibility of rejecting the other's opinion instead
conforming [12], or some rules for the evolution of the self-worth [14, 15], none of
them address cultural properties, self-worth and group dynamics in relation to each
other.
We model agents facing a cultural threat with dynamics of attitudes inspired by the
general principles of the TMT. An agent has a segment of tolerance for each of the
main cultural worldviews available in its environment and forms an attitude about the
other agents depending on the similarity of acceptance segments about these
worldviews. We assume that a terrorist attack is related to some extreme acceptance
segments and related attitudes about the worldviews that are perceived as a threat by
some agents, leading to differentiation from these extreme positions. These changes in
acceptance segments modify the attitudes that the agents have about each other. We
initialize the population of agents from aggregated data given by a representative
survey on groups' attitudes conducted in France in 2014 [16] and then we submit it to
a virtual threat. We generally observe an increase of hostility toward the group as-
similated to the terrorist's group except in some particular cases where, on the contra-
ry the hostility decreases. We study the evolution of the population in relation to its
initial properties and propose some explanations to these variations.
The next section presents the model as well as the material. Then section 2 shows
how the model is initialized and parameterized. Section 3 gives details about the evo-
lutions and the related cultural properties. We finally conclude and discuss our results.
1 Method and materials
1.1 The model
Hypothesis
This model is based on the idea of cultural worldview given by [3] and inspired by the
Social Judgement Theory (SJT) [19]. We now propose an overview of the concepts
that we use and their translation into the model.
Cultural worldview. We assume that K cultural worldviews are available in the
environment. A cultural worldview is a consistent set of concepts, beliefs, tradi-
3
tions or rituals organizing the world and agent behaviour. For instance we con-
sider Christian, Muslim and areligious worldviews.
Agent position about a worldview. Each agent has a most acceptable position
about each worldview which is defined on a continuous axis from -1 (very nega-
tive), to +1 (very positive). This most acceptable position expresses her attitude
about the considered worldview, i.e how much she likes/adheres, or dis-
likes/rejects it. It can be related to the most acceptable position of the SJT [19].
Note that an agent may have positive most acceptable positions about several
worldviews. For instance, an agent can have a high positive most acceptable posi-
tion for the areligious worldview, and also have a lower positive most acceptable
position for Christian or Muslim worldviews.
Lower and higher margin of acceptance, acceptance segment of an agent for a
worldview. In addition to her position, an agent has margins of acceptance around
it. The lower margin of acceptance is a segment of the [-1,1] axis which is lower
than the agent most acceptable position, and the higher margin of acceptance is a
segment defined similarly higher than the agent's most acceptable position. For
convenience, the lower margin of acceptance is called margin(l) in the following,
while the higher margin of acceptance is called margin(h). The segment defined
by the union of the lower and higher margins of acceptance and the most ac-
ceptable position is called the acceptance segment. The positions located in this
segment are assumed acceptable for the agent, while the positions located outside
are not acceptable and perceived negatively.
Cultural identity of an agent. The segments of acceptance for the K available
worldviews represent the cultural identity of an agent [17, 18]. They indeed ex-
press the agent's own positions about the different worldviews and the positions
that it considers acceptable.
Attitude of an agent about an acceptance segment for a worldview. We compute
an agent's attitude about an acceptance segment for a worldview applying a simi-
larity function between this segment and her own acceptance segment for this
worldview. The result is positive if the segments strongly overlap and more and
more negative when they are separated and far apart. This is supported by Social
judgement theory [19] which has shown "… the perceived distance depends on
the level of involvement and the width of the latitude of acceptance." [21, 22].
Attitude of an agent about a cultural identity. The attitude about a cultural identi-
ty is the average of the attitudes about the acceptance segments for the
worldviews. To summarize, the more an agent perceives the cultural identity of
another to be similar, i.e. in agreement to its own, the more its attitude about this
agent is positive [6], whereas perceived differences lead to negativity. Indeed,
"people exaggerate the value of those who share their worldview or who provide
positive evaluations and denigrate the value of those with diverging worldviews
or who provide negative evaluations." [20].
Cultural group of an agent. We assume that, if asked to declare her membership
in a cultural group (such as Christian, Muslim, areligious), the agent would an-
swer the one corresponding to the worldview for which his position is the high-
est.
4
We use this model to simulate how a terrorist attack may change the cultural identities
of the agents and, as a consequence, the opinions attitudes of agents about each other
(supposing that they are aware of all changes). We suppose that when perpetrating an
attack, terrorists stress their cultural identity, for instance an extreme positive position
for the Islamic worldview, extreme negative positions for the other worldviews and
very narrow margins of acceptance around all these positions. We assume that the
agents whose cultural identity is perceived negatively by the terrorists [17, 18], feel
threat for their identity [3] and decrease their margins of acceptance in order to dif-
ferentiate themselves from the terrorist cultural identity. Moreover, we assume that
they reduce their margins more strongly when they perceive the terrorist identity rela-
tively close to their own, (though far enough to be threatening) than when they are
already very different.
The cultural identities of agents
We consider a population of N agents, each agent i has a cultural identity defined
by:
K values between -1 and +1, corresponding to its most acceptable position 𝑎𝑘
𝑖 pro
or anti on each of the K cultural worldview present in the population. In the simu-
lated populations, for each agent there is at least one worldview for which the
agent has a positive most acceptable position. The identity group of an agent is de-
fined by the worldview for which it has the highest most acceptable position.
𝑖 , and lower 𝑏𝐾
Moreover, the agent i is also defined by margins of acceptance higher
𝑖 its most acceptable positions which define a segment of ac-
𝐵𝐾
ceptance indicating the acceptable positions for the agent. The segment going from
𝑖 is called mar-
𝑖 𝑡𝑜 𝐵𝐾
𝑎𝑘
gin(l). These segments are included in [-1;+1].
𝑖 is called margin(h). The segment going from 𝑏𝐾
𝑖 to 𝑎𝑘
Differences between cultural identities determine the attitude of agents about
each other
Attitudes of agents about each other are computed from the comparison of their
cultural identities. Each agent perceives its environment through its cultural identity
and its attitude about another agent depends also on its perception of the other agent's
cultural identity. Its attitude about its own cultural identity is at the maximum value:
1. An agent's attitude toward another agent's cultural identity is an average of the
attitudes its gives to the perceived segments of acceptance composing the other
agent's cultural identity. Indeed the agent perceives another agent's segments of ac-
ceptance as a discretized segment of positions 𝑎.
Then, considering a worldview K of an agent i, its related most acceptable position
𝐾(𝑎) about a position a
and margins of acceptance allow it to compute the attitude 𝜔𝑖
of the worldview K as follows:
a) 𝜔𝐾
𝑖 (𝑎) = 1, if 𝑎 < 𝐵𝐾
𝑖
𝑖 and 𝑎 > 𝑏𝐾
(1)
5
𝑖 , if 𝑎 ≤ 𝑏𝐾
𝑖
𝑖
𝑖 , if 𝑎 ≥ 𝐵𝐾
b) 𝜔𝐾
𝑖 (𝑎) =
c) 𝜔𝐾
𝑖 (𝑎) =
𝑒𝑦−1
𝑒𝑦+1
𝑒𝑦−1
𝑒𝑦+1
with 𝑦 = 1 +
with 𝑦 = 1 +
𝑖
𝑎−𝑎𝐾
𝑖 −𝑏𝐾
𝑎𝐾
𝑖 −𝑎
𝑎𝐾
𝑖 −𝑎𝐾
𝐵𝐾
𝑖 (𝑎) equals +1 if a is in the margins of acceptance of i. If not,
We suppose that 𝜔𝐾
𝑖 () is 0 for a at the bounds of the segment of acceptance and it decreases with the
𝜔𝐾
distance between a and its closest bound of the segment of acceptance, with an as-
ymptote at -1. Moreover, the decrease to -1 is faster for smaller margin of acceptance.
When the decrease is fast, the opinion of the agent can be almost the same (close to -
1) for different values of attitude a. This function behavior is illustrated in figure 1.
𝑖 (𝑎) related to each position a (in abscissa) for two agents
Fig. 1. Computation of the attitude 𝜔𝐾
with a similar segment of acceptance valued 1 for every a of their segment, but a different
width of margin(h) (see the orange arrow for the orange agent, and the blue arrow for the blue
agent). We see the agent having the smallest width, the blue one, going more quickly to highly
negative values than the orange agent for every a outside their segment
The axis of each worldview K is divided regularly into d values 𝑎𝑝(from -1 to +1)
and agent i computes its attitude about the K acceptance segment of agent j, its aver-
age attitude for the positions 𝑎𝑝 located in the acceptance segment of the agent j.
More formally, the opinion of i about j 's acceptance segment in worldview K is given
by:
∑
𝑖𝑗 = 2
𝜔𝐾
𝑑
𝑝=1
𝑖 (𝑎𝑝)max (𝜔𝐾
𝜔𝐾
𝑑
∑ max (𝜔𝐾
𝑝=1
𝑗 (𝑎𝑝), 0)
𝑗 (𝑎𝑝), 0)
(2)
𝑖𝑗
Note, that 𝜔𝐾
Finally the overall attitude of i about j is the average attitude over the #K different
is maximal when the two acceptance segments are identical.
acceptance segments designing totally the cultural identity of an agent:
𝜔𝑖𝑗 =
∑
𝑖𝑗
𝜔𝐾
#𝐾
𝐾=1
#𝐾
(3)
6
Qualitatively, the agents with large margins of acceptance tend to have a positive
attitude of most of the others, whereas agents with small margins of acceptance are
very selective and have very negative attitudes about many others.
Impact of a threat on agent's cultural identity
A terrorist attack is modeled by a scenario of messages in the media, conveying the
cultural identity of terrorists. The most acceptable positions of terrorists are assumed
to be very positive for one worldview and very negative about the others. Moreover
their margins of acceptance are assumed to be very small.
The agents of the population perceiving as negative the attitude of terrorist about
their cultural identity are assumed to perceive a "threat" and modify their margins of
acceptance. More precisely, let the most acceptable positions of terrorist be defined
𝑞), the agents compute 𝜔𝑞𝑖 q the attitude of terrorist q, about
by values (𝑎𝐾
them. If this attitude is positive, i is not scared and does not react. But if 𝜔𝑞𝑖 is nega-
tive, i modifies its margins of acceptance away from the acceptance segments of q as
follows.
𝑞 , 𝑏𝐾
𝑞, 𝐵𝐾
If 𝜔𝑞𝑖 < 0, the intensity of the margin of acceptance modification 𝜇 is:
𝜇 = 𝛼
(4)
𝑒𝜔𝑞𝑖
− 1
𝑒𝜔𝑞𝑖 + 1
where 𝛼 is a positive number smaller than 1. The value of 𝜇 is close to -1 when 𝜔𝑞𝑖
is very negative.
For 𝐾 equal to the main acceptance segment of the terrorist, i.e. the aggressor, the
is modified as follows (with t equals
𝑞
𝑖 } which is the closest to 𝑎𝐾
𝑖 ∈ {𝑏𝐾
𝑖 , 𝐵𝐾
bound 𝑏𝐾
to the time):
𝑖 (𝑡 + 1) = 𝑏𝐾
𝑏𝐾
𝑖 (𝑡) + 𝜇(𝑏𝐾
𝑖 (𝑡) − 𝑎𝐾
𝑖 − 𝜀
𝑖
𝑖 (𝑡) − 𝑎𝐾
𝑏𝐾
𝑖 ))
𝑖 (𝑡) − 𝑎𝐾
𝑏𝐾
(5)
where 𝜀 is a small positive number representing the smallest possible margin of ac-
ceptance width (parameter of the model).
This change in the margins of acceptance results in a more negative attitude about
the aggressor, in accordance with experimental observations [20].
1.2 Data source
A survey on group's attitudes in France has been used to initialize the model [16].
We wanted our virtual groups composing our populations to respect the measured in-
group and intergroup attitudes. The surveyed sample includes 1000 people, repre-
sentative of the French population. People answered to the question « what is your
general attitude about the following groups?». The groups are A (for Areligious), M
(for Muslims), and C (for Christians). They have to answer using a five-point scale
ranging from "strongly unfavorable" to "strongly favorable". Averages and standard-
deviations for in-group attitudes and every out-group attitudes have been computed
and normalized between -1 to +1 for modeling purpose. Figure 2 shows the groups'
attitudes.
7
Fig. 2. Attitudes of each group M (for Muslims), C (for Christians), A (for Areligious) toward
each other (averages and standard-deviations in parenthesis for each relation computed from the
representative survey of the French population)
2
Initialisation of the virtual population and parametrization of
the model
2.1 Building virtual initial populations
Our population of virtual agents, have a cultural identity defined by their ac-
ceptance segment for 3 cultural worldviews C, M and A. The worldview for which the
agent has the highest most acceptable position defines the agent's cultural group.
We assume that each group includes agents of two types:
Agents with exclusive identities: these agents have their acceptance segment in the
positive side for the worldview 𝐾 ∈ {𝑀, 𝐶, 𝐴} defining their group, and on the neg-
ative side for the other worldviews. Amongst these agents, we identify the extrem-
ists which have most acceptable positions close to 1 or -1 and narrow acceptance
segments.
Agents with inclusive identities: these agents have one or two most acceptable
which are positive and the corresponding margin of acceptance in the positive side
and the one or two others close to zero, with a margin of acceptance also mostly in
the positive side
We assume that each group (C, M or A) of the initial population includes x% of in-
clusive and y% of exclusive agents having their highest most acceptable position for
the worldview defining their group. The values x and y are determined to fit indicators
extracted from the survey.
Thus the set of values defining the 3 acceptance segments of each one of the 6
"prototypical" agents, as well as every x and y associated to each group are deter-
mined by an optimization minimizing the distance between indicators computed from
the virtual population and references. Indeed, to build a virtual initial population close
M (3.6%)C (60.6%)A (35.8%)0.52 (0.48)0.14 (0.45) 8
enough to the aggregated indicators (averages and standards deviations) computed
from the data of the survey, we minimize the sum of the absolute distances to the
references of the corresponding indicators measured for the virtual populations.
From the optimization results, we keep the 120 best virtual populations having rel-
ative average errors (over all the distance to the references) going from 5% to 7%, and
maximum errors going from 21 % to 44 %.
2.2
Parameterizing the evolution of the population.
Once we have virtual populations, we want to simulate their evolution facing a set
of "threat" messages conveying the cultural identity of a terrorist.
The tested scenario comprises seven consecutive similar terrorists' messages. The
terrorists have very narrow acceptance segments with a very positive most acceptable
position for the worldview M defining their group and very negative most acceptable
positions about the two other worldviews (C and A). Thus they tend to have a very
negative attitude about all other cultural identities.
The following section presents how the agents change their margins of acceptance
when they receive the messages. The parameters of the dynamics take the following
value: α=0.5; ε=0.05 and d = 400.
3
Evolution of virtual group attitudes facing "threats"
We simulate the evolution of our 120 populations facing a set of seven consecutive
"threat" messages. At each step, due to the high level of mediatization and related
discussions after a terrorist act, agents are considered as perfectly informed about the
other's current acceptance segments. We investigate the result to identify how the
average attitude of the population about the cultural identities of the inclusive and
exclusive agents of group M, which is also the group of the terrorists, evolves.
They show that in most of the selected populations, the average attitude about
group M cultural identities decreases, while for some of them it increases on the con-
trary. The attitudes about the other groups C and A remain rather stable on average.
The challenge is thus to understand why and when a population becomes more posi-
tive toward M after facing a threat coming from agents from this group.
From the analysis of the results shown in figure 3, we obtain more information about
which agent decreases or increases its attitude about group M. We see at first that the
majority of each group decreases its attitude toward inclusive and exclusive M (in
white). Complementary, we see also these are mainly the agents of group C which
generally become more positive, about inclusive M (in red) only, but also sometimes
about exclusive M (if C are inclusives) (in dark blue). Moreover, we observe that
agents of group A can also sometimes increase their attitude about inclusive M (in
red), but never about exclusive M. We observe also a minority of M agents increases
their attitude about exclusive M and decrease it for inclusive M (in light blue).
We then focus on agent's cultural properties explaining these observations. During
the simulations, agents change their higher margin of acceptance in their attitudinal
9
segment on the M worldview. In populations increasing their average attitude about
group M, we notice that:
M agents decrease strongly their higher margin of acceptance on the M worldview;
Non-M (especially C) agents decrease moderately their higher margin of ac-
ceptance on the M worldview.
Fig. 3. Distribution in percentage of all the agents' types of change of attitude over all the 120
initial populations and times
Going further in the observation of data, we establish that the following set of condi-
tions is met altogether in these populations at the time 0 and/or later due to the impact
of the threat:
For the M agents: the attitude toward M worldview tends to be lower and _ the
margin(h – for higher)) of acceptance larger - than in the average population.
Moreover, a large margin(l - for lower) of acceptance is necessary to get an in-
crease of attitude from others. Indeed, these properties lead to decrease strongly the
margin(h) of acceptance when facing an M extremist cultural identity. The
margin(l) of acceptance has to be large enough to ensure an overlap of its
acceptance segment with the ones of non-M agents, overlap which is necessary for
a positive attitude.
For the non-M agents: the attitude about M worldview is higher, the margin(h) of
acceptance is smaller and the margin(l) of acceptance is larger than in the average
populations. The small margin(h) of acceptance (condition 1) indeed ensures that
this margin is weakly modified by the threat. Their high attitude and their large
margin(l) of acceptance (condition 2) ensure an overlap of their acceptance
segment with the one of the group M agents (mainly inclusive).
The next figures illustrate why these conditions lead to increase the attitude about
the group with the same main worldview as the terrorists. We consider a standard case
10
when the expected outcome of decreasing average opinion about group M takes place.
In the following figures, the attitudinal function on the M worldview of a group A or
C exclusive or inclusive is represented in orange while the attitudinal function of a
group M inclusive or exclusive agent is represented in blue. The left panel represents
the attitudinal funtion before the aggression and the right panel this function after the
aggression.
Figures 4 illustrate what occur when there isn't a small margin(h) of acceptance of
non-M blue agent (this is generally a property of inclusive agents). We clearly see on
the figures that despite the increasing part of the overlap in the inclusive M
acceptance segment, the decreasing of the margin(h) of acceptance of the inclusive A
is such as it perceives the inclusive M as very much further than before.
Figures 5 illustrate what occur when there isn't a large margin(h) of acceptance
for blue M agents (this is generally a property of exclusive agents). We observe that
the attitude of the inclusive C for the exclusive M is not changed by the change of M
since the average view of C for M remains -1.
Figures 6 illustrate what occur when the condition 2 is not met. We have then a
large margin(h) of acceptance for both M and non-M agents. Then, when the two
agents (C and M) reject the extremist's view, the non-M agent becomes less tolerant
(from 0.29 to -0.01 on the graphs)!
Fig. 4. Evolution of the attitude of inclusive A (in orange) for inclusive M (in blue): on the left
before the terrorist attack, on the right, after the agression: the attitude decreases (from 0.23 to
0.14) due to the decreasing of the margin(h) of acceptance of the inclusive A and despite the
part of the overlap over the inclusive M acceptance segment has increased. Indeed, being less
inclusive(h), it sees the part of the inclusive M's segment external to it strongly more negatively
than before.
11
Fig. 5. Evolution of the attitude of exclusive C (in orange) for exclusive M (in blue): on the left
before the aggression, on the right, after the agression: the attitude does not change despite the
change of M margin(h) of acceptance
Fig. 6. Evolution of the attitude of exclusive C (in orange) for exclusive M (in blue): on the left
before the attack, on the right, after the agression: most of the time, when the two agents (C and
M) reject the extremist's view, the non-M agent becomes less tolerant (from 0.29 to -0.01 on
the graphs).
4
Conclusion - Discussion
The modeled populations increasing their attitude about the group of main
worldview M to which a terrorist group is assimilated, show specific initial
acceptance segments about this worldview: (1) all agents tend to have a large
margin(l) of acceptance; (2) group C and group A agents have a small margin(h) of
acceptance and group M agents have a large margin(h) of acceptance; (3) group C and
A agents have a high attitude about worldview M .
In other words, an average increase of attitude about the group of the terrorists
takes place if:
this group is mainly composed from very inclusive agents with a moderately
positive attitude for their own group worldview. They also should have a large
margin of acceptance for critical attitudes about their group worldview.
the members of the other groups have a relatively high positive attitude about the
terrorist group worldview, and their view is not affected by the attack. Moreover,
12
they are tolerant, inclusive for attitudes different from theirs about the terrorist
group worldview.
The large margin of acceptance for mean or mild attitudes about the terrorist group
worldview increases the overlap between acceptance segments of the agents of the
population. Such overlaps can be related to the feeling of wide similarity and
solidarity observed in the large French demonstrations that took place after the
Charlie Hebdo terrorist attacks. These conclusions are in line with the results of [6, 7].
Indeed, we observe that some cultural properties, when becoming salient simultane-
ously with the threat, may increase perceived similarity of members of different
groups, and protect against an increase of the intergroup hostility.
It should be underlined that, in the model, this solidarity takes place because
the agents of the same group as the terrorists reject strongly the radical attitude
of the terrorists, and that the other agents are aware of their evolution. The next
step of our work is to come back to experiment to check the relevance of the model
behaviour.
Another step regarding the study of the model is a larger study of the sensitivity of
the result to the distribution of inclusive versus exclusive agents in each group.
Indeed, we have seen inclusive agents are more enclined to become less negative to M
than the other agents. Then the proportion of such type of agent in the population is
very important to observe an increase of the attitude. A variation of this proportion
can probably change drastically the result.
Acknowledgements
The authors acknowledge funding from the "Mission for Interdisciplinarity" of the
CNRS.
References
1.
2.
3.
4.
Greenberg, J. and J. Arndt, Terror management theory, in Handbook of theories of
social psychology, P.A.M. Van Lange, A.W. Kruglanski, and E. Higgins, Editors.
2013, Sage: London, England. p. 398-415.
Pyszczynski, T., J. Greenberg, and S. Solomon, A dual-process model of defense
against conscious and unconscious death-related thoughts: An extension of terror
management theory. Psychological Review, 1999. 106(4): p. 835-845.
Greenberg, J., T. Pyszczynski, and S. Solomon, The Causes and Consequences of a
Need for Self-Esteem: A Terror Management Theory, in Public Self and Private Self,
R.F. Baumeister, Editor. 1986, Springer, New York, NY: Springer Series in Social
Psychology.
Greenberg, J. and S. Kosloff, Terror Management Theory: Implications for
Understanding Prejudice, Stereotyping, Intergroup Conflict, and Political Attitudes.
Social and Personality Psychology Compass, 2008. 2(5): p. 1881-1894.
13
Vyver, J.V.d., et al., Boosting Belligerence:How the July 7, 2005, London Bombings
Affected Liberals' Moral Foundations and Prejudice. Psychological Science, 2016.
27(2): p. 169-177.
Motyl, M., et al., Subtle priming of shared human experiences eliminates threat-
induced negativity toward Arabs, immigrants, and peace-making. Journal of
Experimental Social Psychology, 2011. 47(6): p. 1179-1184.
Nugier, A., et al., The Psychological Effects of Terrorism are Moderated by Cultural
Worldviews. International Review of Social Psychology, 2016. 29(1): p. 7.
Axelrod, R., The dissemination of culture a model with local convergence and global
polarization. Journal of Conflict Resolution, 1997. 41(2): p. 203-226.
Castellano, C., S. Fortunato, and V. Loreto, Statistical physics of social dynamics.
Reviews of Modern Physics, 2009. 81(2): p. 591-646.
Schaller, M. and C.S. Crandall, The Psychological Foundations of Culture. 2004,
Mahwah, New Jersey - London: Lawrence Erlbaum Associates. 384.
Radillo-Díaz, A., L.A. Pérez, and M. del Castillo-Mussot, Axelrod models of social
influence with cultural repulsion. Physical Review E, 2009. 80(6): p. 066107.
Flache, A., et al., Models of Social Influence: Towards the Next Frontiers. Journal of
Artificial Societies & Social Simulation, 2017. 20(4): p. 31.
Flache, A. and M.W. Macy, Local Convergence and Global Diversity: From
Interpersonal to Social Influence. Journal of Conflict Resolution, 2011. 55(6): p. 970-
995.
Deffuant, G., S. Huet, and T. Carletti. The Leviathan Model: Absolute Dominance,
Generalised Distrust, Small Worlds and Other Patterns Emerging from Combining
Vanity with Opinion Propagation. in JASSS, 2013, 16, (1) 5..
Friedkin, N. and E. Johnsen, Social influence and opinions. Journal of Mathematical
Psychology, 1990. 15(3-4): p. 193-205.
Guimond, S., The effects of integration policies on intergroup relations and well-
being in France and in Québec, F.r.s.t. ANR, Editor 2016, Laboratoire de
Psychologie Sociale et Cognitive (LAPSCO): Clermont-Ferrand.
Tajfel, H. and J.C. Turner, The social identity theory of intergroup behavior., in
Psychology of intergroup relations, S. Worchel and W.G. Austin, Editors. 1986,
Nelson-Hall: Chicago. p. 7-24.
Turner, J.C. and K.J. Reynolds, Self-categorization theory., in Handbook of theories
of social psychology, P.A.M. Van Lange, A.W. Kruklanski, and T.E. Higgins,
Editors. 2013, Sage: London. p. 399-417.
Sherif, M. and C.I. Hovland, Social Judgment. Assimilation and contrast effects.
Communication and attitude change. 1961: New Haven and London: Yale University
Press.
Pyszczynski, T., S. Solomon, and J. Greenberg, Chapter One - Thirty Years of Terror
Management Theory: From Genesis to Revelation, in Advances in Experimental
Social Psychology, J.M. Olson and M.P. Zanna, Editors. 2015, Academic Press. p. 1-
70.
Petty, R., J. Cacioppo, and C. Haugtvedt, Ego-involvement and Persuasion: An
appreciative look at the Sherif's contribution to the study of self-relevance and
attitude change, in Social judgment and intergroup relations: Essays in honor of
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
14
22.
Muzafer Sherif, D. Granberg and G. Sarup, Editors. 1992, Springer/Verlag: New
York. p. 147-175.
Eagly, A.H. and K. Telaak, Width of the latitude of acceptance as a determinant of
attitude change. Journal of Personality and Social Psychology, 1972. 23(3): p. 388-
397.
|
1804.04746 | 3 | 1804 | 2018-09-06T16:58:16 | Analytically Modeling Unmanaged Intersections with Microscopic Vehicle Interactions | [
"cs.MA"
] | With the emergence of autonomous vehicles, it is important to understand their impact on the transportation system. However, conventional traffic simulations are time-consuming. In this paper, we introduce an analytical traffic model for unmanaged intersections accounting for microscopic vehicle interactions. The macroscopic property, i.e., delay at the intersection, is modeled as an event-driven stochastic dynamic process, whose dynamics encode the microscopic vehicle behaviors. The distribution of macroscopic properties can be obtained through either direct analysis or event-driven simulation. They are more efficient than conventional (time-driven) traffic simulation, and capture more microscopic details compared to conventional macroscopic flow models. We illustrate the efficiency of this method by delay analyses under two different policies at a two-lane intersection. The proposed model allows for 1) efficient and effective comparison among different policies, 2) policy optimization, 3) traffic prediction, and 4) system optimization (e.g., infrastructure and protocol). | cs.MA | cs | Analytically Modeling Unmanaged Intersections
with Microscopic Vehicle Interactions
Changliu Liu and Mykel J. Kochenderfer
8
1
0
2
p
e
S
6
]
A
M
.
s
c
[
3
v
6
4
7
4
0
.
4
0
8
1
:
v
i
X
r
a
Abstract -- With the emergence of autonomous vehicles, it is
important to understand their impact on the transportation
system. However, conventional traffic simulations are time-
consuming. This paper introduces an analytical traffic model for
unmanaged intersections accounting for microscopic vehicle in-
teractions. The macroscopic property, i.e., delay at the intersec-
tion, is modeled as an event-driven stochastic dynamic process,
whose dynamics encode the microscopic vehicle behaviors. The
distribution of macroscopic properties can be obtained through
either direct analysis or event-driven simulation. They are more
efficient than conventional (time-driven) traffic simulation, and
capture more microscopic details compared to conventional
macroscopic flow models. We illustrate the efficiency of this
method by delay analyses under two different policies at a two-
lane intersection. The proposed model allows for 1) efficient
and effective comparison among different policies, 2) policy
optimization, 3) traffic prediction, and 4) system optimization
(e.g., infrastructure and protocol).
I. INTRODUCTION
With the emergence of autonomous vehicles, it is important
to understand how the microscopic interactions of those
autonomous vehicles affect the delay of the macroscopic
traffic flow, especially at unmanaged intersections.
The literature contains many traffic models that can support
the analysis of delay and congestion [1]. There are two major
types of traffic models: 1) microscopic simulation models
where every car is tracked and 2) macroscopic flow models
where traffic is described by relations among aggregated
values such as flow speed and density, without distinguishing
its constituent parts. The major advantage of microscopic
simulation models is the precise description of inter-vehicle
interactions. Such models have been widely adopted in
evaluating the performance of autonomous vehicles [2].
However, it can be time-consuming to obtain the micro-macro
relationships by simulation. Only "point-wise" evaluation can
be performed in the sense that a single parametric change in
vehicle behavior requires new simulations. To gain a deeper
understanding of the micro-macro relationships, an analytical
model is desirable.
Macroscopic flow models provide a tractable mathematical
structure with few parameters to describe interactions among
vehicles. Those models usually come in the form of partial
differential equations. However, it remains challenging to
model intersections. Existing methods introduce boundary
constraints to represent intersections [3], [4]. However, these
models can only tolerate simple first-in-first-out (FIFO)
C. Liu and M. Kochenderfer are with the Department of Aeronau-
tics and Astronautics, Stanford University, CA 94305 USA (e-mail:
changliuliu, [email protected]).
(a)
(b)
Fig. 1: Intersection scenario. (a) Road topology. (b) Conflict graph.
(a)
(b)
(c)
Fig. 2: The time of occupancy at the intersection. (a) The desired
time of occupancy. (b) The actual time of occupancy under the
FIFO policy. (c) The actual time of occupancy under the FO policy.
policies at intersections. To consider other kinds of policies,
the vehicles need to be treated as particles that interact among
one another, which has not been captured by existing flow
models.
This paper introduces an analytical stochastic continuous-
time, discrete-event traffic model. We use it to describe
delays at unmanaged intersections under different microscopic
vehicle behaviors. The model considers microscopic inter-
actions and is analytical, which absorbs the advantages of
both the microscopic simulation models and macroscopic
flow models. With this model, we can better understand how
the microscopic behavior design of a single vehicle affects
the macroscopic transportation system. In addition to direct
analysis, we can perform event-driven simulation under this
model, which is more efficient than conventional time-driven
traffic simulation. The policies under consideration are not
required to have closed-form solutions.
The remainder of the paper is organized as follows.
Section II formulates the traffic model as an event-driven
stochastic process, and illustrates how vehicle behaviors
are encoded in the traffic model. Section III illustrates the
effectiveness of the model through case studies. Section IV
discusses applications of the method and concludes the paper.
Lane 1Lane 2Lane 3Lane 4T3T2T1T4Incoming flowOutgoing flowTimeLane1234t⇤1t⇤2t⇤3t⇤4t⇤5t⇤6t⇤7t⇤8t⇤9TimeLane1234¯t1¯t2¯t3¯t4¯t5¯t6¯t7¯t8¯t9TimeLane1234¯t1¯t2¯t3¯t4¯t5¯t6¯t7¯t8¯t9 II. TRAFFIC MODEL
This section introduces the traffic model. The traffic delay
at an intersection is modeled as an event-driven stochastic
process. An event is defined as the introduction of a new
vehicle. This section then describes how the microscopic
vehicle interactions affect the macroscopic dynamics.
A. Traffic Model at Intersections
Consider an intersection with K incoming lanes. A conflict
is identified if two incoming lanes intersect. These relation-
ships can be described in a conflict graph G with the nodes
being the incoming lanes and the links representing conflicts.
For example, Fig. 1a shows one road configuration with four
incoming lanes, and Fig. 1b shows the conflict graph. T k is
the delay at lane k, which will be introduced in (1).
When there are conflicts, vehicles from the corresponding
lanes cannot occupy the intersection at the same time. Let
t∗i be the desired time for vehicle i to pass the center of the
intersection. The vehicles are numbered such that t∗i < t∗i+1.
Fig. 2a shows the desired time of occupancy (centered at t∗i )
for vehicles coming in the four lanes in Fig. 1a. According
to the graph in Fig. 1b, the scenario in Fig. 2a is infeasible
as vehicles 1 through 4 cannot occupy the intersection at
the same time. Based on the FIFO policy, vehicles 2 and 3
yield to vehicle 1. Vehicles 4 and 5 yield to vehicles 2 and 3,
and so on. Let ¯ti be the actual time for vehicle i to pass the
center of the intersection. Fig. 2b shows the actual time of
occupancy when all vehicles adopt FIFO. The actual time of
occupancy may change when the policy changes, resulting
in different traffic delay. For example, Fig. 2c corresponds to
another policy that will be introduced in Section III.
Poisson process {t∗1, t∗2, . . .} with parameter λ =(cid:80)
The traffic at the intersection is modeled as an event-driven
stochastic system with the state being the traffic delay and the
input being the incoming traffic. It is assumed that the desired
passing time for incoming vehicles from lane k follows a
Poisson distribution with parameter λk. The traffic flows
from different lanes are independent. Since the combination
of multiple Poisson processes is a Poisson distribution [5],
the incoming traffic from all lanes can be described as one
k λk. The
input to the model is chosen to be the random arrival interval
between vehicle i + 1 and i, i.e., xi = t∗i+1 − t∗i , and the
lane number si+1 for vehicle i + 1. For all i, the probability
density for xi = x is px(x) = λe−λx. The probability of
λ . The delay for lane k considering i
si+1 = k is Ps(k) = λk
vehicles is denoted T k
i , which captures the difference between
the actual passing time and the traffic-free passing time of
those vehicles, i.e.,
T k
i = max
sj =k,j≤i
¯t(i)
j − t∗i ,
(1)
where ¯t(i)
j denotes the actual passing time for vehicle j only
considering the interactions among the first i vehicles, e.g.,
vehicle i + 1 has not approached the intersection yet. Here,
¯t(i)
corresponds to an equilibrium in microscopic vehicle
j
interactions which will be introduced in (8). It may differ
from ¯t(k)
j
for k (cid:54)= i.
Define Ti := [T 1
i , . . . , T K
i
]T . The dynamics of the traffic
delay at the intersection is determined by
Ti+1 = F(Ti, xi, si+1),
(2)
where the function F depends on the policies adopted by
the vehicles and the road topology defined by the conflict
graph G. Given (2), the conditional probability density of
Ti+1 given Ti, xi and si+1 is
x
τ
k
k
=
=
(cid:90)
Ps(k)
(cid:90)
(cid:90)
pTi+1(t)
(3)
where δ(·) is the delta function. The probability density is
pTi+1(t Ti, xi, si+1) = δ(t = F(Ti, xi, si+1)),
(cid:88)
(cid:88)
. . .(cid:82) (tk)+
which involves integration over a manifold. The cu-
is denoted as PTi(t) =
mulative probability of Ti
pTi(τ 1, . . . , τ k)dτ 1 . . . dτ k where t =
pTi+1(t τ , x, k)pTi(τ )dτ px(x)dx
−∞
−∞
[t1, . . . , tk]. The problems of interest are:
• Does the sequence {pTi}i converge in L1-norm?
Divergence corresponds to the formation of congestion,
i.e., the case that the expected delay keeps growing.
δ(0)pTi (τ )px(x)dτ dx,
(cid:82) (t1)+
F (τ ,x,k)=t
Ps(k)
(4)
• If converged, what is the steady state distribution of
pT := limi→∞ pTi?
From the steady state distribution, we may compute the
expected delay.
These two problems will be considered in the case studies
in Section III. Moreover, from the distribution of the lane
delays, we can compute the scalar delay introduced by the
(i + 1)th vehicle as
di+1 =
¯t(i+1)
j
+ ¯t(i+1)
i+1 − t∗i+1.
(5)
(cid:88)
(cid:16)
j≤i
(cid:17)
j
− ¯t(i)
(cid:88)
In the case that the introduction of a new vehicle only affects
the last vehicle in other lanes (which is usually the case),
di+1 = T si+1
i+1 +
(T k
i+1 − T k
i + xi).
(6)
B. Microscopic Interactions
k(cid:54)=si+1
It is assumed that the vehicles at intersections have fixed
paths. When interacting with other vehicles, they only change
their speed profiles to adjust the time to pass the intersection.
Such simplification is widely adopted [6], [7]. In this paper,
we further reduce the high dimensional speed profile for
vehicle i to a single state ti which denotes the time for
vehicle i to pass the center of the intersection. Since the
mapping from ti to a speed profile is surjective, interactions
can be analyzed using ti's.
The policy of vehicle i is denoted
ti(k) = f (t∗i , t−i(k − 1)),
(7)
where k denotes time step. The subscript −i denotes all
other indices except i. The one step delay is due to reaction
TABLE I: The mapping (2) under FIFO for si+1 = 1.
Region
1
2
3
4
Domain
i < xi − ∆s
T 1
i < xi − ∆d
T 2
i ≥ xi − ∆s
T 1
i < xi − ∆d
T 2
T 2
i < T 1
i
i ≥ xi − ∆d
T 2
T 2
i < T 1
i
i ≥ xi − ∆d
T 2
T 2
i > T 1
i
(a) Domain
(b) Value
Fig. 3: Illustration of the mapping (2) under FIFO for si+1 = 1.
Value
T 1
i+1 = 0
i+1 = −∆d
T 2
T 1
i+1 = T 1
i + ∆s − xi
i+1 = −∆d
T 2
T 1
i+1 = T 1
T 2
i+1 = T 2
T 1
i+1 = T 2
T 2
i+1 = T 2
i + ∆s − xi
i − xi
i + ∆d − xi
i − xi
time. An equilibrium is achieved if the vehicles do not have
incentives to adjust the passing time. Such equilibrium may
be broken with a new vehicle. It is assumed that the time for
the vehicles to achieve a new equilibrium is negligible. The
assumption is true when the flow rate is low. Every event
then leads to one equilibrium. The actual passing time ¯t(i)
j
when i vehicles are considered lies at the ith equilibrium
such that
¯t(i)
j = f (t∗j , ¯t(i)
−j),∀j ≤ i.
(cid:88)
The average delay of the vehicles satisfies
1
N
(¯t(N )
i − t∗i ) = lim
N→∞
¯d = lim
N→∞
1
N
i
(8)
di,
(9)
(cid:88)
i
where the second equality is due to (5). According to the
central limit theorem, the system is ergodic such that the
average delay of all vehicles equals the expected delay
introduced by any event in the steady state,
E(di).
(10)
E( ¯d) = lim
i→∞
III. CASE STUDIES
To illustrate the effectiveness of the model, this section
derives traffic properties under two frequently used policies
by analysis, event-driven simulation (EDS) as well as con-
ventional time-driven traffic simulation. The two policies
are first-in-first-out (FIFO) policy [8] and flexible order
(FO) policy [9]. For simplicity, we only consider a two-
lane intersection (which is equivalent to lane merging). More
detailed analyses are discussed in the extended version [10].
A policy specifies 1) the passing order, and 2) the temporal
gap between two consecutive vehicles. The temporal gap
refers to the time distance or headway maintained between
vehicles. Denote ∆d and ∆s to be the temporal gap between
vehicles from different lanes and the temporal gap between
vehicles from the same lane respectively. The gap may be
affected by vehicle speed, uncertainties in perception, and
etc. When the traffic flow rate is low, we assume ∆s = 0.
A. Case 1: Lane Merging with FIFO
Under FIFO, the passing order is determined by the desired
arrival times {t∗i }i such that the actual passing time for vehicle
i should be after the actual passing times for all conflicting
vehicles j such that j < i.As the passing order is fixed, the
actual passing time will not be affected by later vehicles, i.e.,
¯t(i)
j = ¯t(j)
for all j < i. For vehicle i,
j
¯t(i)
i
:= max{t∗i ,Di,Si},
(11)
where
Di = max
Si = max
j
j
(¯t(i)
j + ∆d) s.t. j < i, (sj, si) ∈ G,
(¯t(i)
j + ∆s) s.t. j < i, sj = si.
(12)
(13)
i = max{T j
The effect of FIFO is illustrated in Fig. 2b. Following from
(1) and (11), the equation (2) for FIFO can be computed,
which is listed in Table I and shown in Fig. 3. Only the case
for si+1 = 1 is shown. The case for si+1 = 2 can be obtained
by switching superscripts 1 and 2. To bound the domain from
below, let T j
i ,−∆d} for all i and j ∈ {1, 2}. The
mapping is piece-wise smooth with four smooth components.
Region 1 corresponds to where there is a sufficient gap in
both lanes for the (i + 1)th vehicle to pass without delay.
Regions 2 and 3 correspond to where the last vehicle is from
the ego lane and it causes delay for the (i + 1)th vehicle.
Region 4 corresponds to where the last vehicle is from the
other lane and delays the (i + 1)th vehicle.
Given the dynamic equation, the distribution of traffic
delay in (4) can be computed. The propagation of pTi for
λ1 = 0.1 s−1, λ2 = 0.5 s−1, ∆d = 2 s, ∆s = 1 s is shown
in Fig. 4a by an event-driven simulation of (2) with 10, 000
particles. At iteration 1, Ps(1) percent of particles are at
(0,−∆d), while the others are at (−∆d, 0). Every particle
corresponds to a traffic scenario. For conventional time-driven
traffic simulation, it is computationally expensive to obtain
distributions with 10, 000 traffic scenarios. However, with
the event-driven simulation under the proposed model, the
distributions can be obtained in real time. The distribution
approached steady state at iteration 8 with a unique pattern.
Theoretical analysis [10] also verifies this pattern.
B. Case 2: Lane Merging with FO
FO allows high priority vehicles to yield to low priority
vehicles if low priority vehicles can arrive earlier. The
passing order may change over time. At step i, let ¯t(i−1)
:=
max{t∗i , maxj<i,sj =si(¯t(i−1)
+ ∆s)} be the earliest desired
time for vehicle i to pass considering its front vehicles in
the same lane. Sort the list {¯t(i−1)
} in
i−1 , ¯t(i−1)
, . . . , ¯t(i−1)
1
j
i
i
xixi s d00T1iT2i1234 d132 d0T1i+1T2i+14 2
0
6
Delay in lane 1 [s]
4
8
2
0
6
Delay in lane 1 [s]
4
(a) Case 1: Lane Merging with FIFO. Iterations 2, 3, 4, 8.
0
−2
−2
8
6
4
2
8
6
4
2
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
0
−2
−2
8
6
4
2
8
6
4
2
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
0
−2
−2
8
8
0.2
0.1
0
0.2
0.1
0
8
2
0
6
Delay in lane 1 [s]
4
2
0
6
Delay in lane 1 [s]
4
0
−2
−2
8
6
4
2
8
6
4
2
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
0
−2
−2
2
0
6
Delay in lane 1 [s]
4
2
0
6
Delay in lane 1 [s]
4
8
8
0
−2
−2
8
6
4
2
8
6
4
2
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
0
−2
−2
0
−2
−2
2
0
6
Delay in lane 1 [s]
4
8
2
0
6
Delay in lane 1 [s]
4
Fig. 4: Event-driven simulation with pTi for λ1 = 0.1 s−1, λ2 = 0.5 s−1, ∆d = 2 s, and ∆s = 1 s with 10000 particles.
(b) Case 2: Lane Merging with FO. Iterations 2, 3, 4, 8.
ascending order and record the ranking in an injection Q.
Ties are broken by index. For the first vehicle in Q, i.e.,
vehicle k = Q−1(1), ¯t(i)
. By induction, assuming
that ¯t(i)
j
for Q(j) < Q(k) has been computed, then
k := ¯t(i−1)
k
k := max{¯t(i−1)
¯t(i)
k
,Di
k},
k,S i
(14)
where
Di
k = max
S i
k = max
j
j
j + ∆d) s.t. Q(j) < Q(k), (sj, sk) ∈ G,(15)
(¯t(i)
(¯t(i)
(16)
j + ∆s) s.t. Q(j) < Q(k), sj = sk.
(a) Domain
(b) Value
Fig. 5: Illustration of the mapping (2) under FO for si+1 = 1.
Under FO, the actual passing time may change at every
step. There is a distributed algorithm [11] for this policy
where the vehicles do not need to compute the global passing
order. Fig. 2c shows the effect of FO. Vehicles in the same
direction tend to form groups and pass together.
Following from (1) and (14), the dynamic equation (2)
for FO can be computed, which is listed in Table II and
illustrated in Fig. 5 for si+1 = 1. There are eight smooth
components in the mapping. Regions 1 to 4 are the same as in
the FIFO case such that vehicle i + 1 passes the intersection
after all other vehicles. Regions 5 to 8 correspond to where
vehicle i + 1 passes the intersection before the last vehicle
in the other lane. In regions 5 and 7, vehicle i + 1 does not
experience delay due to sufficient gap in the ego lane. The
last vehicle in the other lane is delayed in region 5, and not
delayed in region 7. Regions 6 and 8 correspond to where
the (i + 1)th vehicle is delayed by the last vehicle in the ego
lane but can still go before the last vehicle in the other lane.
Delay is caused in the other lane in region 6.
Given the dynamic equation, the distribution of delay in (4)
can be computed. Fig. 4b shows the event-driven simulation
with the same conditions as the FIFO case. FO generates
less delay than FIFO. However, the distribution under FO no
longer has the "zebra" pattern shown in FIFO. We investigate
the steady state distribution of delay for ∆s = 0 and leave the
case of ∆s > 0 for future work. When ∆s = 0, the mapping
pTi (cid:55)→ pTi+1 is a contraction as shown in Fig. 6. Proposition 1
provides a solution of pT when ∆s = 0 and λ1 = λ2. As the
problem is symmetric, define g(t) := pT(t, t−∆d) = pT(t−
∆d, t). The function g(t) represents half of the probability
part of the function be g(t) and the delta component be(cid:98)g(t),
density that t equals the maximum lane delay. Let the finite
which is nonzero only at 0 and ∆d.
Proposition 1 (Steady State Distribution for ∆s = 0 under
xixi s d00T1iT2i12345678 d86571432 d0T1i+1T2i+1 8
6
4
2
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
0
−2
−2
2
0
6
Delay in lane 1 [s]
4
8
8
6
4
2
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
0
−2
−2
2
0
6
Delay in lane 1 [s]
4
8
8
6
4
2
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
0
−2
−2
8
6
4
2
]
s
[
2
e
n
a
l
n
i
y
a
l
e
D
0
−2
−2
0.2
0.1
0
8
2
0
6
Delay in lane 1 [s]
4
2
0
6
Delay in lane 1 [s]
4
8
(a) Iteration 1.
(b) Iteration 3.
(c) Iteration 10.
(d) Iteration 20.
Fig. 6: Illustration of the convergence of (4) under FO for λ1 = 1.1 s−1, λ2 = 0.5 s−1, ∆d = 2 s, and ∆s = 0 s with 10000 particles.
TABLE II: The mapping (2) under FO for si+1 = 1.
Domain
i < xi − ∆s
T 1
i < xi − ∆d
T 2
i ≥ xi − ∆s
T 1
i < xi − ∆d
T 2
T 2
i < T 1
i
i ≥ xi − ∆d
T 2
T 2
i < T 1
i
i ∈ [xi − ∆d, xi)
T 2
T 2
i > T 1
i
i ∈ [xi, xi + ∆d)
T 2
i < xi − ∆s
T 1
i ∈ [∆d , ∆d + ∆s]
i − T 1
T 2
i ≥ xi − ∆s
T 1
i < xi − ∆s
T 1
i ≥ xi + ∆d
T 2
i − T 1
T 2
i > xi + ∆d + ∆s
i ≥ xi − ∆s
T 1
1
2
3
4
5
6
7
8
Value
T 1
i+1 = 0
i+1 = −∆d
T 2
T 1
i+1 = T 1
i + ∆s − xi
i+1 = −∆d
T 2
T 1
i+1 = T 1
T 2
i+1 = T 2
T 1
i+1 = T 2
T 2
i+1 = T 2
i + ∆s − xi
i − xi
i + ∆d − xi
i − xi
T 1
i+1 = 0
T 2
i+1 = ∆d
T 1
i+1 = T 1
i − xi + ∆s
i − xi + ∆s + ∆d
i+1 = T 1
T 2
T 1
i+1 = 0
i+1 = T 2
T 2
i − xi
T 1
i+1 = T 1
i − xi + ∆s
i − xi
T 2
i+1 = T 2
g(t) = Ce
where C =
FO). If ∆s = 0 and λ1 = λ2, we have
λ
2 ∆d
2e
λ
C, (17)
λ
1
2 −
λ(1+e
λ
∆d +e
2
2
λ
−λ∆d )
− λ
2
C, (cid:98)g(∆d) =
(cid:21) .
2 t, (cid:98)g(0) =
(cid:20)
(cid:104)(cid:82) ∞
∆d−t g(t + x − ∆d)pxdx +(cid:82) ∞
∆d−1
8
e
(cid:105)
Proof. For t ∈ (0, ∆d), using the mapping in regions 3,
4, and 8, we get the following steady state relationship
.
g(t) = 1
0 g(t + x)pxdx
2
Multiply both sides by e−λt, and then differentiate with
respect to t to get
λ
2
g.
Hence, g = Ce
g(cid:48) − λg = −
the constant C. Due to symmetry, (cid:82) ∆d
the fact that g = g +(cid:98)g and g = Ce
λ
2 t for some constant C. Now we solve for
2. Using
g(t)dt = 1
(18)
.
1
2
(19)
(cid:98)g(0) +
2C(e
λ
2 ∆d − 1)
λ
λ
0
2 t, we get
+(cid:98)g(∆d) =
0
0
0
2
0
1
2
at
2
0
of
By
the
τ +∆d
region
Consider
(cid:82) ∞
∆d
changing
point mass
integration, we
pxdxg(τ )dτ + 1
2
2
e−λ∆d + 1
(cid:98)g(0) =
0 e−λτ g(τ )dτ. Hence,
0 e−λ(τ +∆d)g(τ )dτ + 1
1. The
g(τ )dτ pxdx + 1
2
order
(cid:82) ∞
(cid:82) x
(cid:82) ∞
(cid:82) ∞
is
0 g(τ )dτ pxdx.
get
τ pxdxg(τ )dτ =
(cid:82) x−∆d
(cid:82) ∞
(cid:98)g(0) = 1
(cid:82) ∞
(cid:82) ∞
(cid:98)g(0) = 1
(cid:82) ∞
where I =(cid:82) ∞
I =(cid:98)g(0) + e−λ∆d(cid:98)g(∆d) +
Consider region 5. The point mass at ∆d is (cid:98)g(∆d) =
(cid:82) ∞
(cid:82) ∆d
(cid:98)g(∆d) = 1
g(τ )dτ pxdx. By changing the order of integration,
(cid:20) 1
(1 − e−λτ )g(τ )dτ.
(22)
2C(1 − e−
λ
(cid:82) ∆d
(cid:21)
0 e−λτ g(τ )dτ. Plugging in the expression of
0 pxdxg(τ )dτ = 1
g(τ ), we have
(cid:82) ∆d
(cid:98)g(∆d) =
(cid:82) τ
Then,
(20)
(21)
I,
2 )
λ∆d
2
0
2
0
1
2
2
x
.
.
0
1
2
2 − I
We combine (19) to (22) to verify Proposition 1.
According to (17), the probability of zero-delay (2(cid:98)g(0))
and the probability of ∆d-delay (2(cid:98)g(∆d)) only depend on
λ∆d, i.e., the ratio between the temporal gap and the arrival
interval. Fig. 7 illustrates those relationships. When the ratio
between the temporal gap and the arrival interval increases,
the probability of zero-delay decreases while the probability
of ∆d-delay increases. Fig. 7 also illustrates the result from
EDS, which verifies Proposition 1.
We validate the event-driven model against the time-driven
traffic simulation. By ergodicity (10), the mean delay of all
vehicles in the time-driven traffic simulation should equal
the expectation of the delay induced by any event in the
steady state. The statistical mean delay under each scenario
is obtained through the time-driven traffic simulation for
10 min. The details of the simulation is provided in an earlier
journal [11]. The expected steady state delay is computed
using the result from Proposition 1. Given (6), the steady
state probability density of delay for t ∈ [0, ∆d) satisfies
pd(t) =
[g(t + x) + g(t + x− ∆d) + g(x− t + ∆d)]pxdx.
(23)
(cid:90) ∞
0
y
t
i
l
i
b
a
b
o
r
P
1
8
.
0
6
.
0
4
.
0
2
.
0
0
0
Analytical 0-delay (2(cid:98)g(0))
Analytical ∆d-delay (2(cid:98)g(∆d))
Probability of 0-delay from EDS
Probability of ∆d-delay from EDS
]
s
[
d
y
a
l
e
d
d
e
t
c
e
p
x
E
1
5
.
0
0
0
Analytical
Approximated
EDS
Traffic Sim
0.2
0.4
0.6
0.8
1
1.2
Traffic flow rate λ [s−1]
1
2
3
λ∆d
4
5
6
Fig. 9: The expected delay for ∆d = 1.5 s and different λ.
Fig. 7: The probability of delay with respect to parameter λ∆d.
∆d = 1 s Analytical
∆d = 3 s Analytical
∆d = 4 s EDS
∆d = 2 s Analytical
∆d = 4 s Analytical
y
t
i
l
b
a
b
o
r
P
1
5
.
0
0
0
0.5
1
1.5
2
2.5
3
3.5
4
Delay d [s]
Fig. 8: Illustration of Pd in (24) for λ = 1 s−1 with different ∆d.
2 ∆d−λt(cid:105)
λ
1 − e−λt
.
+
2
(cid:104)
(cid:90) ∆d
Applying (17), the distribution of steady state delay becomes
Pd(t) =
4C
λ
e
λ
2 t − e
λ
2 (∆d−t) + e
(24)
Fig. 8 shows the distributions. The shaded area is obtained
though EDS with 104 particles, which validates (24).
The expected delay can be computed as:
∆d
2
e−λ∆d − 1
2 ∆d + e− λ
λ
2
d
0
2λ
∆d
+
2λ(e
(λ∆d )2
E(d) =
tdPd(t) =
= λ∆2
4 .
.
2 ∆d − 1)
(25)
When λ → 0, i.e., the traffic flow rate is low, E(d) →
2 + −λ∆d+
When ∆d = 1.5 s, Fig. 9 shows the analytical expected
delay (25), the approximation function λ∆2
4 , the expected
delay obtained through EDS, and the statistical mean delay
in the simulation. When the flow rate is low, the delays
obtained through the four methods align well. As the flow
increases, the model underestimates the traffic delay because
the assumption ∆s = 0 is only valid for small λ.
IV. DISCUSSION AND CONCLUSION
d
This paper introduced an analytical
traffic model for
unmanaged intersections. The macroscopic property, i.e.,
delay at the intersection, was modeled as an event-driven
stochastic dynamic process. The macroscopic dynamics
encoded the equilibrium resulted from microscopic vehicle
interactions. Both the vehicle policies and the road topology
could affect the macroscopic dynamics. With the model, the
distribution of delay can be obtained through either direct
analysis or event-driven simulation, which are more efficient
than conventional time-driven traffic simulation and capture
more microscopic details than conventional macroscopic flow
models. The steady state traffic properties were studied, and
the accuracy was verified in simulation.
The potential applications of the analytical model include
1) efficient verification or comparison of policies through
analysis or event-driven simulation; 2) policy optimization
(e.g., choosing optimal ∆d) with respect to macroscopic
objectives; 3) real-time traffic prediction for intersections;
and 4) infrastructure optimization (e.g., designing better road
structure and network) to improve traffic efficiency.
REFERENCES
[1] S. P. Hoogendoorn and P. H. L. Bovy, "State-of-the-art of vehicular
traffic flow modelling," Proceedings of the Institution of Mechanical
Engineers, Part I: Journal of Systems and Control Engineering, vol.
215, no. 4, pp. 283 -- 303, 2001.
[2] P. Gora and I. Rub, "Traffic models for self-driving connected cars,"
Transportation Research Procedia, vol. 14, pp. 2207 -- 2216, 2016,
transport Research Arena TRA2016.
[3] R. Corthout, G. Flotterod, F. Viti, and C. M. Tamp`ere, "Non-unique
flows in macroscopic first-order intersection models," Transportation
Research Part B: Methodological, vol. 46, no. 3, pp. 343 -- 359, 2012.
[4] G. Flotterod and J. Rohde, "Operational macroscopic modeling of
complex urban road intersections," Transportation Research Part B:
Methodological, vol. 45, no. 6, pp. 903 -- 922, 2011.
[5] C. Gardiner, Stochastic Methods. Springer, 2009, vol. 4.
[6] F. Altch´e, X. Qian, and A. de La Fortelle, "Time-optimal coordination
of mobile robots along specified paths," in IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), Oct 2016.
[7] X. Qian, F. Altch´e, J. Gr´egoire, and A. de La Fortelle, "Autonomous
intersection management systems: criteria, implementation and evalua-
tion," IET Intelligent Transport Systems, vol. 11, no. 3, pp. 182 -- 189,
April 2017.
[8] K. Dresner and P. Stone, "Multiagent traffic management: A reservation-
based intersection control mechanism," in International Joint Confer-
ence on Autonomous Agents and Multiagent Systems, 2004.
[9] M. Ahmane, A. Abbas-Turki, F. Perronnet, J. Wu, A. El Moudni,
J. Buisson, and R. Zeo, "Modeling and controlling an isolated urban
intersection based on cooperative vehicles," Transportation Research
Part C: Emerging Technologies, vol. 28, pp. 44 -- 62, 2013.
[10] C. Liu and M. J. Kochenderfer, "Analyzing traffic delay at unmanaged
intersections," arXiv:1806.02660, 2018.
[11] C. Liu, C. W. Lin, S. Shiraishi, and M. Tomizuka, "Distributed conflict
resolution for connected autonomous vehicles," IEEE Transactions on
Intelligent Vehicles, vol. 3, no. 1, pp. 18 -- 29, March 2018.
|
1809.07392 | 1 | 1809 | 2018-09-19T20:05:55 | Improved Bounds on Information Dissemination by Manhattan Random Waypoint Model | [
"cs.MA"
] | With the popularity of portable wireless devices it is important to model and predict how information or contagions spread by natural human mobility -- for understanding the spreading of deadly infectious diseases and for improving delay tolerant communication schemes. Formally, we model this problem by considering $M$ moving agents, where each agent initially carries a \emph{distinct} bit of information. When two agents are at the same location or in close proximity to one another, they share all their information with each other. We would like to know the time it takes until all bits of information reach all agents, called the \textit{flood time}, and how it depends on the way agents move, the size and shape of the network and the number of agents moving in the network.
We provide rigorous analysis for the \MRWP model (which takes paths with minimum number of turns), a convenient model used previously to analyze mobile agents, and find that with high probability the flood time is bounded by $O\big(N\log M\lceil(N/M) \log(NM)\rceil\big)$, where $M$ agents move on an $N\times N$ grid. In addition to extensive simulations, we use a data set of taxi trajectories to show that our method can successfully predict flood times in both experimental settings and the real world. | cs.MA | cs |
Improved Bounds on Information Dissemination by Manha(cid:130)an
Random Waypoint Model
Aria Rezaei
Stony Brook University
[email protected]
Je(cid:130) M. Phillips
University of Utah
je(cid:130)[email protected]
Jie Gao
Stony Brook University
[email protected]
Csaba D. T´oth
Cal State Univ. Northridge
[email protected]
ABSTRACT
With the popularity of portable wireless devices it is important to
model and predict how information or contagions spread by nat-
ural human mobility -- for understanding the spreading of deadly
infectious diseases and for improving delay tolerant communica-
tion schemes. Formally, we model this problem by considering M
moving agents, where each agent initially carries a distinct bit of
information. When two agents are at the same location or in close
proximity to one another, they share all their information with
each other. We would like to know the time it takes until all bits
of information reach all agents, called the (cid:131)ood time, and how it
depends on the way agents move, the size and shape of the network
and the number of agents moving in the network.
We provide rigorous analysis for the Manha(cid:138)an Random Way-
point model (which takes paths with minimum number of turns),
a convenient model used previously to analyze mobile agents,
and (cid:128)nd that with high probability the (cid:131)ood time is bounded by
O(cid:0)N log M(cid:100)(N/M) log(N M)(cid:101)(cid:1), where M agents move on an N × N
grid. In addition to extensive simulations, we use a data set of taxi
trajectories to show that our method can successfully predict (cid:131)ood
times in both experimental se(cid:138)ings and the real world.
KEYWORDS
Mobile Information Systems, Mobile Agents, Information Dissemi-
nation
1 INTRODUCTION
It has always been an interesting research topic to understand
human mobility and how contagions spread via such motion. One of
the motivations is to understand how infectious diseases spread by
moving agents. (cid:140)ink of a strain of an infectious virus such as SARS
or Ebola. When two individuals are at the same location at the same
time, there is a possibility for one to spread that strain of virus to the
other. (cid:140)erefore the spread of contagions in a population is highly
dependent on the density of the population and how individuals
move around. In another example, human motion can be used
for our bene(cid:128)t. It has become common that people carry wireless
devices around. Short range low cost wireless communication
can be established at the contact events to allow energy e(cid:129)cient
information exchange. In this case the mobility model in(cid:131)uences
how long it takes for a piece of information from one node to reach
all nodes in the network.
(cid:140)ere have been two main approaches to study human mobility
in the literature: data-driven methods versus theoretical analysis.
In recent years wireless technology has made it possible to collect
a large amount of mobility data through wireless devices. (cid:140)ere
has been a lot of work on (cid:128)nding exciting pa(cid:138)erns in human move-
ments and information spreading in real-world data [1, 19, 23]. It
has been shown that human mobility is immensely complex. Accu-
rate models can be built using historic tra(cid:129)c data to predict agents'
locations and social ties [20, 42]. But these models each work only
for a speci(cid:128)c scenario. It is unclear how these models can help us
analyze asymptotic behavior of moving agents or whether a model
generalizes to a di(cid:130)erent geographical location, a di(cid:130)erent travel
modality, or a di(cid:130)erent group of people. Furthermore, long-term
mobility data can be identity revealing. Even with great e(cid:130)orts
to anonymize the data and with removing big fractions of it, in-
dividuals are identi(cid:128)able by their movement pa(cid:138)erns. A seminal
work revealed that 4 randomly selected points in an hourly location
sequence of a person recorded over 15 months via cellphone anten-
nas is enough to make that person identi(cid:128)able among 1.5 million
individuals [17]. As a result, mobility data sets are usually not
published by companies due to concerns over user privacy, except
for a few special cases of shared vehicles (taxis or shared bikes).
On the other hand, an extensive amount of work has been ded-
icated to mathematical models of mobility and their asymptotic
behaviors. Although these models cannot compete with the ac-
curacy of data-driven models in the presence of enough historic
mobility data, they have been used for their rigorous analysis and
their ability to predict future events with provable certainty. Over
the years theoretical models have evolved from simplistic models,
inspired by known physical phenomena in real world, to more
sophisticated ones, taking into account the complexity of human
mobility. We brie(cid:131)y review these models and their analytical results
below. For a comprehensive survey on these models refer to [8].
• Random Walk: Perhaps the most studied movement model.
Inspired by the movement of (cid:131)oating particles in a liquid
or gas, called Brownian Motion [22], in its simplest version
an agent starts its movement in an arbitrary node in a
given network. At each time step, the agent chooses one
of the neighbors of its current node uniformly at random
and moves to that neighbor. (cid:140)ere are variations in which
agents can rest in their current position for a period of time
or use a non-uniform transition probability when choosing
a neighbor.
• Random Direction Model: In this model, an agent chooses
a random direction and possibly a random velocity, then
moves in that direction until it collides with the boundary
of the network. (cid:140)e agent then chooses a new direction
and velocity and continues as before.
• Random Way-point: In this model, an agent, starting
from an initial position in the network, chooses the next
destination from all nodes in the network uniformly at ran-
dom. (cid:140)en, using one of the shortest paths, moves towards
the destination and a(cid:137)er reaching it (cid:128)nds a new destina-
tion with the same method. (cid:140)is has been widely used in
modeling human motion and in many prior simulations
for mobile networks [6, 7, 26, 29, 37].
• Manhattan Random Way-point : (cid:140)is is a special case
of the Random Way-point model [6]. In a grid (or torus)
networks, agents move to the destination with as few turns
as possible. (cid:140)us they travel (cid:128)rst horizontally and then
vertically (or vice versa) to the destination. (cid:140)is model is
inspired by the fact that in urban streets, turning can be
time-consuming [13, 16].
• L´evy Walk: Studies on intelligent moving agents, espe-
cially humans, have revealed that the distance to the next
destination chosen by such agents seems to follow a fat-
tailed distribution [2, 25, 43]. A popular movement model
in this category is called L´evy Walk [41]. (cid:140)is model is sim-
ilar to Random Way-point, but instead of choosing a new
destination uniformly at random, an agent chooses a node
as its next destination with probability proportional to the
inverse of its distance from the agent's current position, to
some power α > 0.
1.1 Our Contributions
In this paper we provide improved upper bounds on the rate of
information dissemination when agents move according to the Man-
ha(cid:138)an Random Way-point model. (cid:140)rough a series of simulations
we show that, combined with bounds on Random Walk, our bounds
lead to a new conjecture on the time it takes for information to
disseminate through a network when agents' movements follow
the L´evy Walk model, a challenging question that remains widely
open. Finally, we report the result of a series of experiments we
have designed, which show that our model is capable of predicting
trends on experimental se(cid:138)ings, as well as real-world data.
We formally de(cid:128)ne our problem as follows. Consider a set of
M autonomous agents, each starting at time 0 with a unique bit
of information, bi, at a node selected uniformly at random in an
N × N torus1 denoted by G. Agents all follow the same movement
model and they share information with each other when they meet
during their move. Meeting is de(cid:128)ned as being at the same location
at the same time, where the location can be inside a node or on
an edge between two nodes. Agents start their movement at the
beginning of each time step in a synchronized fashion. We consider
uniform speed for all agents2 and transmission radius is practically
0 as agents have to be collocated in order to pass along information,
for simplicity. However our (cid:128)ndings can be extended for arbitrary
constant transmission radius.
In the above se(cid:138)ing, we are interested in (cid:128)nding bounds on
the time it takes until every agent (cid:128)nds out about every piece of
information. (cid:140)is value is called the (cid:131)ood time (TF ). An equally
important statistic is the time it takes for all agents to learn a
speci(cid:128)c bit of information, called broadcast time and denoted by
TB. Clearly TB ≤ TF . Using the union bound, any upper bound on
TB extends to TF , too, if the probability of the bound occurring is
su(cid:129)ciently high3. Both TF and TB are important statistics in various
applications. In the case of disease spreading, TB corresponds to the
time when an infection of any agent would have been passed to the
entire population. In a delay-tolerant wireless mobile network, TF
corresponds to the time in the past, from which we can assume all
information has been shared across the network; we can predicate
the start of a new protocol based on assuming all agents are up to
date a(cid:137)er this delay. In mobile social networks, TB corresponds to
the time it takes a new piece of information to permeate society.
Overall, both TF and TB are important statistics which capture
information (cid:131)ow in a network, and will be the focus of our study.
Our contributions are as follows:
M
• We (cid:128)nd a new upper bound for TF and TB that is tight
for a wide range of se(cid:138)ings. Speci(cid:128)cally, when M agents
move on an N × N grid with torus topology, we show
TB ≤ TF = O(N log M(cid:100) N
log(N M)(cid:101)). (cid:140)is bound im-
proves upon recent upper bounds for topologies with more
complex boundary conditions.
• We analyze the relation between Random Walk, Manha(cid:138)an
Random Way-point and L´evy Walk. (cid:140)rough simulations,
we show empirically that the L´evy Walk model can be
understood by carefully interpolating between the heavily
studied Random Walk model and our new results on the
Manha(cid:138)an Random Way-point model.
• We validate the theoretical bounds in a number of empirical
studies using simulated scenarios, bike and taxi trajectory
data sets.
2 BACKGROUND AND RELATED WORKS
Since the advent of social networks, researchers studied how in-
formation (e.g., rumors or viral videos) spreads through a network
[21, 24, 30]. (cid:140)e rate of change in these networks is slow enough
that they can be considered static throughout the course of infor-
mation dissemination. (cid:140)is assumption simpli(cid:128)es the mathematical
models tremendously. (cid:140)ese studies do not (cid:128)t dynamic mobile
networks, due to the high rate of topology changes. (cid:140)e spreading
behavior heavily depends on how agents move [32].
For analyzing the mobility models mentioned earlier, arguably
the simplest model for a geographically spread network is a 2D grid.
In some metropolitan se(cid:138)ings the downtown area is a reasonable
grid. Wrapping the grid around to a torus has been commonly
adopted in prior papers that analyze information spreading in a
mobile network. (cid:140)e bene(cid:128)t of the torus model is to get rid of
1A grid where nodes in the boundaries are each connected to their corresponding
node in the opposite boundary.
2In some previous de(cid:128)nitions of Random Way-point the agents choose their speed
uniformly at random from a range. But this choice will lead to the average moving
speed to be decreasing over time [45]. Also in reality vehicles/pedestrians o(cid:137)en move
with a (cid:128)xed speed.
3If Pr{TB > T } < 1/N d , for some d > 0, we know that Pr{TF > T } < M/N d <
1/N d(cid:48), when d is su(cid:129)ciently large.
2
the boundary e(cid:130)ect, thus simplifying the analysis. (cid:140)e torus also
removes the boundary e(cid:130)ect in the Random Way-point model,
which o(cid:137)en caused unwanted artifacts in simulations [3, 11, 39].
For direct comparison, we review prior results in the same format
as ours, where N and M are decoupled and there are no assump-
tions on their ratio, near-zero transmission radius (R ≈ 0), and
where each move consists of walking along a path from source
node to destination node, rather than jumping to the destination
instantaneously.
(cid:140)e Random Walk movement model has been extensively stud-
ied and tight bounds on many characteristics have been fully re-
solved [18, 31, 33]. One of the tightest bounds for Random Walk
is given by Pa(cid:138)erin et al. [38]. (cid:140)ey found that even with a very
small transmission radius, broadcast time does not depend on the
relation between the mobility speed and the transmission radius.
(cid:140)ey prove that with high probability (w.h.p):
TB = O
N
.
(1)
(cid:16)
(cid:109)(cid:17)
(cid:108) N√
M
Keeping track of a bit of information b, they (cid:128)rst divide the initial
grid into smaller cells and (cid:128)nd the time by which an arbitrary cell
is in(cid:128)ltrated by an agent carrying b. (cid:140)en they show that this
in(cid:128)ltrating agent will inform the majority of agents near this cell
and spread the information locally. A(cid:137)er the local spread is done
within the cell, information leaks into adjacent cells and this whole
process is repeated. Ultimately, every cell is in(cid:128)ltrated and every
agent in the network (cid:128)nds out about b.
Clementi et al. [14] have proved bounds for TF in Manha(cid:138)an-
like grids where agents move according to the Manha(cid:138)an Random
Way-point model. (cid:140)eir se(cid:138)ing slightly di(cid:130)ers from ours as they
do not employ boundary loops, which results in two zones with
very di(cid:130)erent tra(cid:129)c of agents. On the one hand, nodes in the
central zone are visited by agents across all nodes in the grid with
high probability. On the other hand, the periphery, called suburb
areas, are starved of agents: the probability that an agent passes
through these areas is signi(cid:128)cantly lower than that of the central
zone. (cid:140)ey prove that w.h.p. TF = O(cid:0)N/R + N
(cid:17)
the transmission radius and v is the agents' speed. We can rewrite
this in our se(cid:138)ing as:
M(cid:1), where R is
3/vR
(cid:16)
3
2
(2)
TF = O
N + N
M
.
(cid:140)ey found that the bulk of the Flooding Time is devoted to carrying
the information to the "suburbs," as it requires a (cid:131)ow of informed
agents traveling from the central zone to the suburbs.
A di(cid:130)erent line of work focused on solving the problem in a
general platform, oblivious to geometric considerations. Clementi
et al. [15] have derived an upper bound for TF , and the mixing time4
of the Markov chain corresponding to agent movements, as well as
how independent the collisions between di(cid:130)erent pairs are, play
a major role in the analysis of TF . Applying Manha(cid:138)an Random
Way-point model to their general bound yields the following:
(cid:16) N
(cid:0) N
2 + 1(cid:1) log2
2
(cid:17)
M
,
TF = O
vmax
MR
4Mixing time is the time needed for a Markov chain to reach its stationary distribution,
starting from an arbitrary distribution.
3
(cid:16)
(cid:108) N
2
(cid:109)(cid:17)
where vmax is the maximum speed of any agent. Rewri(cid:138)en in our
se(cid:138)ing, the bound is:
TF = O
N
(3)
Since their method is very general, their bound is not competi-
tive with the bounds on speci(cid:128)c movement models and networks.
Table 1 shows that our bound is a signi(cid:128)cant improvement over
these bounds on di(cid:130)erent movement models.
M
.
Table 1: Recent bounds on Manhattan Random Way-point
(MRWP), Random Way-point (RWP) and Random Walk
(RW).
Ours
Authors
Model
MRWP
Clementi et al. 2010 [14] MRWP
RWP
Clementi et al. 2015 [15]
Pe(cid:138)arin et al. 2011 [38]
RW
Bound
O(N (cid:100)N/M(cid:101))
3/M)
O(N + N
O(N (cid:100)N
2/M(cid:101))
O(N (cid:100)N/√
M(cid:101))
Rigorous analysis of the L´evy Walk model is much more chal-
lenging due to the strong spatiotemporal correlation [4, 34, 40]. (cid:140)e
most relevant work is done by Wang et al. [44] where the authors
have analyzed the distribution of the minimum time needed until a
piece of information reaches a certain region. However, to the best
of our knowledge, there are no bounds for TF when the agents are
moving according to a L´evy Walk.
Besides theoretical work, there has been a lot of empirical anal-
ysis of how information can spread through opportunistic peer
communication among mobile agents using simulations or empiri-
cal evaluations. Protocols for reducing communication cost (e.g., to
avoid a message be delivered to a node multiple times) have been
studied extensively [10, 28, 36]. Last, there has also been work on
a model assuming that a supporting static wireless network is in
place, which helps to cache and propagate these events [46]. (cid:140)is
model is di(cid:130)erent from ours.
3 BOUNDS ON FLOOD TIME
In this section, we present and prove our main theoretical results
on the (cid:131)ooding time on torus networks. We start with a trivial
lower bound.
Theorem 3.1. For M agents initially positioned at uniformly ran-
dom nodes and moving with constant speed in an N × N torus, with
constant probability, we have:
TF ≥ TB ≥ Ω(N).
(4)
Proof. Let bi be the bit of information initially carried by agent
Ai, and Aj the farthest agent to Ai at time 0. (cid:140)e time it takes until
bi reaches Aj, denoted by T(Aj , bi), is a lower bound for both TF
and TB. Since we are assuming uniformly random initial positions,
with 1/2 probability the distance between Ai and Aj is at least N/2.
Also note that information can only travel as fast as agents can.
(cid:140)is means that at each time step the distance between Aj and
the closest copy of bi is reduced by at most 2 units. As a result,
with probability 1/2, T(Aj , bi) is at least N/4, which completes the
proof.
(cid:3)
Now we move forward to our main theorem on the upper bound
for Manha(cid:138)an Random Way-point model.
Theorem 3.2. For M agents moving according to the Manha(cid:136)an
Random Way-point model with constant speed in an N × N torus,
with high probability5, we have:
TB ≤ TF = O
N log M
.
(5)
(cid:16)
log(N M)(cid:109)(cid:17)
(cid:108) N
M
W present the proof in two parts. (cid:140)e (cid:128)rst part is to analyze
when and with what probability two agents have a collocation
event so they can share information. (cid:140)is involves a few necessary
conditions: time-wise the moves made by two agents with a collo-
cation event need to overlap; geometrically their trajectories need
to have an intersection; and thirdly, they arrive at the intersection
at the same time. Next we need to analyze the global property: how
information sharing enabled by collocation events leads to global
dissemination. (cid:140)e next two subsections focus on these two parts
of the proof respectively.
3.1 Bounding Collocation Probability
We consider agents moving non-stop following Manha(cid:138)an Random
Way-point model, and partition the mobility trace of each agent into
moves between randomly chosen destinations. First, observe that
any move by an agent following Manha(cid:138)an Random Way-point
model takes Ω(N) time with constant probability. We call the two
straight parts of a movement, one horizontal and the other vertical,
segments. Note that any move in this model can have at most two
segments of di(cid:130)erent directions appearing in an arbitrary order.
We say two agents have a connection during their moves if they
are at the same location at the same time. For that to happen, the
moves of the two agents must at least overlap over time.
De(cid:128)nition 3.3. Strongly overlapping moves are moves made
by two agents where the time interval of a segment of an agent's
move is completely contained within the time interval of the other
agent's whole move.
Note that this overlap only needs to happen in the time interval
of two moves and no condition is imposed on their geometric
locations.
Lemma 3.4. Every move Mi of an agent Ai strongly overlaps with
at least one of the moves of another agent Aj , say Mj . (cid:138)e starting
moment of the moves can be at most N time-steps apart and with con-
stant probability, a segment of Mi will have a time duration overlap
of Ω(N) with a segment of Mj .
j
Proof. First consider the move M−
j
of Aj that is still active
when Mi starts. If M−
ends a(cid:137)er the (cid:128)rst segment of Mi, then M−
j
covers the (cid:128)rst segment of Mi and the two moves strongly overlap
(Figure 1 case (i)). If not, then consider the next move of Aj, M +
. If
j
ends a(cid:137)er Mi ends, M +
completely covers Mi's second segment
M +
j
j
and the two moves strongly overlap (Figure 1 case (ii)). If not, then
must be completely covered by Mi and the two moves strongly
M +
j
overlap (Figure 1 case (iii)). We have shown that a strong overlap
occurs. Since the time duration of each move can be at most N , the
5High probability in our work means at least 1 − 1/N d for some constant d > 0.
starting point of Mi is at most N time-steps apart from the starting
points of both M−
j
and M +
j
.
Figure 1: Any move of one agent strongly overlaps with at
least one of the moves of another agent.
Without loss of generality, assume a segment Sj of Mj is covered
completely by Mi. With probability 1/2, the time duration of Sj
(equal to its length due to the constant speed assumption) is at least
N/4. Now take the segment of Mi with the longest time duration
overlap with Sj, and denote it by Si. (cid:140)is overlap should be at least
Sj/2. As a result, with probability 1/2, the time duration of this
overlap between Si and Sj is at least N/8, as required.
(cid:3)
Figure 2: (cid:135)e segments are trimmed to duration of their
overlap, (cid:96). Le(cid:133): (cid:135)e collocation event of two agents hap-
pens if the distance between the intersection point (P) and
the starting point of both segments is equal ((cid:96)1 = (cid:96)2). Right:
Connection happens if there is a non-zero overlap between
two segments and there are 2(cid:96) such placements for the blue
segment if the red segment is (cid:128)xed.
Lemma 3.5. If two agents Ai and Aj have two segments Si and Sj
2).
with a time interval overlap of (cid:96), they meet with probability Θ((cid:96)/N
Proof. We trim both segments to the duration of their overlap,
making them of equal length (cid:96). All cases of a connection between
the two agents can be reduced to three main cases below by rotating
the torus or swapping the agents:
(1) Si is horizontal and Sj vertical. Here, the two agents con-
nect, if the two segments intersect geometrically at a point
P and P is at equal distance from the starting points of
Si and Sj (see Figure 2, le(cid:137)). Since we assume that G is a
2
torus, we can (cid:128)x Si's position in our analysis. Out of all N
possible placements of Sj on G, there are (cid:96) placements that
result in an intersection that meets the above condition.
Hence the probability of a connection between Ai and Aj
is (cid:96)/N
2 in this scenario.
4
MiM−j(i)M+j(ii)(iii)M−jM+jM−j(cid:96)1(cid:96)2P(cid:96) (2) Both Si and Sj are horizontal and in opposite directions.
In this case, any geometric intersection is enough for a
connection to happen (see Figure 2, right). If both agents
move in the same row, there would be 2(cid:96) placements of Sj
once Si's position is (cid:128)xed that results in an intersection.
(cid:140)e probability of both agents moving in the same row is
1/N , which makes the overall probability 2(cid:96)/N
(3) Both Si and Sj are horizontal and in the same direction. In
this case, the starting points of Si and Sj have to be in the
same exact node, which happens with probability 1/N
2.
Based on the 4 possible directions of each segment, there are
16 possible cases for Si and Sj, all of which happens with equal
probability and can be reduced to one of the 3 cases above. (cid:140)e
overall probability of a connection between two agents can be
bounded as:
2.
(cid:96)
N
2 ≤ Pr{Connect(Ai , Aj , (cid:96))} ≤ 32 (cid:96)
2 .
N
(6)
(cid:3)
An immediate consequence of Lemmas 3.5 and 3.4 is:
Lemma 3.6. Two agents with strongly overlapping moves have a
connection with probability Θ(1/N).
Proof. For the probability of two strongly overlapping agents
connecting we have:
N/2
P((cid:96) = x) Pr{Connect(Ai , Aj , x)}, (7)
Pr{Connect(Ai , Aj)} =
where P((cid:96) = x) is the probability that the time duration of the
overlap between segments of Ai and Aj is x. From Lemma 3.4 we
know that P((cid:96) ≥ N/8) ≥ 1/2. As a result we can rewrite (7) as:
x =1
Pr{Connect(Ai , Aj)} ≥ 1
2
N/8
2 ≥ 1
16N
N
.
(cid:140)e upper bound in (6) yields:
Pr{Connect(Ai , Aj)} ≤ max
x
Pr{Connect(Ai , Aj , x)} ≤ 16
N
.
(cid:3)
3.2 Bounding the Flood Time
Given the probability of two agents sharing information during
their moves, we are now ready to argue for how information propa-
gates to the entire network. For simplicity, we (cid:128)nd an upper bound
for the broadcast time TB (one speci(cid:128)c message reaching everyone)
and extend it to the (cid:131)ood time TF (all messages reaching everyone).
(cid:140)ere are two issues that we need to address. First, the positions of
an agent are temporally correlated -- but fortunately, if su(cid:129)ciently
far apart in time, the positions of agents are independent which
will help to simplify our analysis. Second, we need to track agents
who have been informed and who have not, and analyze how in-
formation spreads from the informed ones to the uninformed ones.
Lemma 3.7. (cid:138)e locations of two agents at 2N steps apart are
independent of each other.
Xi +2. Also, nodes visited between two consecutive destinations are
only dependent on those two destinations. Let Pt and Pt +2N be the
position of A at time t and t + 2N . Since the agents move along the
shortest path towards their destination, a move is at most N steps
long. As a result, at least two destinations will be visited between
time t and t + 2N by A. Take the destination immediately a(cid:137)er Pt ,
Xi, and immediately before Pt +2N , Xj. As noted above, Pt is only
dependent on Xi−1 and Xi, while Pt +2N is only dependent on Xj
and Xj+1, where Xi and Xj are distinct positions. (cid:140)is makes Pt
and Pt +2N completely independent.
(cid:3)
As mentioned earlier, we track the spread of a single bit of in-
formation b among the agents. We (cid:128)rst divide the time span of the
whole process into windows of size 6N time steps, called cycles.
th cycle starts at time 6kN and ends at time 6(k + 1)N . Each
(cid:140)e k
agent would visit at least 4 destinations between time (6k + 1)N
and (6k +5)N , which yield at least 3 moves independent of other cy-
cles in an agent's trajectory (note the 2N margin between selected
moves in each cycle, which guarantees independence). According
to Lemma 3.4, we know that a move by an agent will strongly over-
lap with a move by another agent and the starting point of the two
moves are at most N time steps apart. As a result, for each pair of
agents and each cycle, we can (cid:128)nd two strongly overlapping moves
independent of other cycles, which according to Lemma 3.6 have a
c/N chance of connection, where c is a constant. (cid:140)is essentially
makes collocation events between a (cid:128)xed pair of agents in di(cid:130)erent
cycles i.i.d.
We now divide the whole process into consecutive epochs. Each
epoch consists of si cycles, and starts when we have a set Ii of
agents who know about b (referred to as informed agents) and a
set Ui of agents who do not (referred to as uninformed agents). An
epoch ends when the number of informed agents doubles, or the
number of uninformed agents drops to zero. For the broadcast time
of b, TB, we can write:
TB ≤ 6N
si .
(8)
log M
i =1
We now (cid:128)nd the number of cycles needed for that w.h.p. each
th epoch.
agent in Ii is paired with a distinct agent in Ui during the i
By arti(cid:128)cially forcing informed agents to (cid:128)nd distinct partners, we
only slow down the process of information spread, and an upper
bound found in this manner is valid as an upper bound for the
main problem. (cid:140)e reason we are require distinct partners for each
informed agent is to ensure that connections between di(cid:130)erent
pairs are independent.
(cid:140)e probability of an informed agent A ∈ Ii connecting to any
A(cid:48) ∈ Ui is as follows (arrow shows the direction of information
exchange):
Pr{(cid:154)A
Pr{∃A
(cid:48) ∈ Ui : A → A
(cid:48) ∈ Ui : A → A
(cid:16)1 − c
(cid:17)si Ui
(cid:17)si Ui
(cid:48)} = 1 −(cid:16)1 − c
(cid:48)} =
N
,
N
.
(9)
Proof. Let the sequence of destinations chosen by an agent A
be (cid:104)X0, X1, . . . , Xk(cid:105). Observe that regardless of what Xi is, every
node in the torus (including Xi) has the same probability of being
Let there be an arbitrary order for agents in Ii and one for agents
in Ui. (cid:140)e (cid:128)rst informed agent can match to any of the Ui unin-
formed agents. A(cid:137)er the (cid:128)rst matching is done, there will be Ui−1
5
potential matches for the second informed agent and so on. Assum-
ing that there are q pairs at the end of this epoch (q ≤ min(Ii, Ui))
and using Equation (9), the probability of this happening (i.e., hav-
ing q pairs of matched informed/uninformed agents) is:
Pr{q-matching} =
N
N
i =0
(cid:1)si(M−q−i)(cid:17)
(cid:16)1 −(cid:0)1 − c
q
≥(cid:16)1 −(cid:0)1 − c
(cid:1)si(M/2)(cid:17)q/2
≥(cid:16)1 − exp(cid:16)−sic(M/2)
(cid:17)(cid:17)q/2
≥ 1 − q exp(cid:16)−sic(M/2)
(cid:17)
(cid:17) ≥ 1 − 1
(cid:17) ≥ qN d
N d
N
N
.
si ≥ 2N
cM
2dN
cM
si >
(log q + d log N)
log(N M).
For the above to happen w.h.p., for d > 1, we need:
1 − q exp(cid:16)−sic(M/2)
exp(cid:16) sic(M/2)
N
N
(cid:140)e last step is due to the fact that q < M. To make sure that si is a
non-zero integer and we have at least one cycle, we set:
Substituting (10) into (8), we have:
si =
.
cM
(cid:108)2dN
log(N M)(cid:109)
log(N M)(cid:109)
(cid:108)2dN
log M
log(N M)(cid:109)(cid:17)
(cid:108) N
i =1
N log M
cM
(cid:16)
TB ≤ 6N
.
M
= O
(12)
Since our bound for TB works for arbitrarily high probability (1 −
1/N d, for a constant d > 0), it extends to TF using the Union Bound.
(cid:140)is completes the proof of (cid:140)eorem 3.2.
(cid:140)is is a pre(cid:138)y tight bound. Consider a semi-dense scenario
where M ≈ N , our bound becomes O(N), which nearly meets the
trivial lower bound of Equation (4).
Compared to previously mentioned bounds for the Random Way-
point model in (2) and (3), our bound is stronger. For the same
movement model, although under slightly di(cid:130)erent network as-
sumptions, Clementi et al. found the bound of O(N + N
3/M) [14],
which is mainly due to the choice of not having a torus as they
intended to study the impact of rarely visited areas on the total
information spread time. Our bound also improves that of [15] by
a huge margin. (cid:140)is can be due to the fact that their method is
a general framework to (cid:128)nd an upper bound for TF . Our version
of the Manha(cid:138)an Random Way-point model assumes that agents
complete their move in one coordinate then start moving in another.
As suggested in [15], this assumption increases the probability of
connection between two agents, which in turn leads to a be(cid:138)er
upper bound for TF and TB. Further, the Manha(cid:138)an Random Way-
point model is a be(cid:138)er (cid:128)t for mobility in urban areas it implicitly
incorporates the cost of turning during movement.
(10)
(11)
Depending on the application, one can think of various exten-
sions to our model, such as an arbitrary transmission radius or
random waiting time between two consecutive moves of an agent.
In this case, an approach similar to ours can be adopted to (cid:128)nd a
bound for TF if these three components are available: (1) suitably
sized independent time windows (called cycle here), (2) guarantee
of a long enough time interval overlap between segments of moves
by two agents and (3) probability of connection between those two
segments.
4 EXPERIMENTS
In this section, through a series of experiments we test the accuracy
of our discovered bound. First, we test our model where agents are
moving in a torus-like grid, following Manha(cid:138)an Random Way-
point model. Next, using bike sharing records in 3 major cities,
we create synthetic trajectories in real-world road networks, and
compare simulated behavior of (cid:131)ood time against our model. Finally,
using GPS traces of taxis in a major city, we verify our model against
a real-world case of information dissemination via mobile agents.
4.1 Simulated Movements in a Grid
We simulate the movement of agents in a torus-like grid following
a Manha(cid:138)an Random Way-point model, and compare the aver-
age (cid:131)ood times, TF , against our bound. First, we (cid:128)x N to 100
and (cid:128)nd TF for M = {25, 50, · · · , 1000} by averaging TF over 100
realizations. To (cid:128)t the resulting values to our bound, we use func-
tion f (M) = c1
log(c2M) log(c3M), where each ci is a positive
M
constant, accounting for (cid:128)xed N and constants in our asymptotic
analysis. (cid:140)e results of the simulation along with (cid:128)(cid:138)ed values are
shown in Figure 3, middle. Our bound has accurately captured the
2, is equal to
changes in TF as the coe(cid:129)cient of determination, R
0.9894. Next, using a similar procedure, we (cid:128)x M to 100 and (cid:128)nd
TF for N = {25, 50, · · · , 500}. We (cid:128)t the values to the function
f (N) = c1N
2(log N + c2), where again each ci is a positive con-
stant. (cid:140)e simulation results and the (cid:128)(cid:138)ed values are depicted in
Figure 3. As expected, we showed good performance here too by
2 values is
yielding an R
coincidental).
2 of 0.9894 (equality between the two R
4.2 Simulated Movements in Real Networks
To test our model against non-grid networks, we use the bike rental
records of 3 major US cities [9, 12, 27]. Each city has a unique road
network and a set of (cid:128)xed stations, {S1, S2, · · · , Sk}, which are used
as the set of possible destinations each agent can choose from. (cid:140)e
goal here is to test our model against a se(cid:138)ing beyond the grid
network and uniformly random selection of destinations. (cid:140)e data
sets include an origin and a destination station for each trip made.
Using these records, we can estimate the probability of choosing a
destination Sj given that an agent is currently positioned in station
Si, called the transition probability between Si and Sj and denoted
by P(Si , Sj). We can also calculate the probability of initiating a
trajectory from any given station, from here on called initiation
probability and denoted by P(Si). To gradually move our tests
away from the theoretical se(cid:138)ings, we use the following movement
models throughout our simulations:
6
Figure 3: (le(cid:133)) (cid:135)e sigmoid-like behavior of L´evy Walk, sandwiched by Random Walk and Manhattan Random Way-point.
(middle, right) Our bound accurately captures the changes in TF as M and N are tweaked while other parameters are (cid:128)xed.
Figure 4: Actual (empty circles) and (cid:128)tted (dotted lines) TF for two movement policies in 3 cities.
(1) Similar to Section 4.1, we select each station in the sequence
of stations visited by an agent uniformly at random. (cid:140)is is
equivalent to the Random Way-point model and denoted
by RWP in this experiment.
(2) Next, we use the calculated P(Si) and P(Si , Sj) values to
build synthetic trajectories. We call this model DATA
throughout this experiment.
In each experiment, a(cid:137)er selecting a sequence of stations visited
by each agent, we (cid:128)nd the shortest paths between consecutive
stations using Routino [5] and OpenStreetMap [35] extracts. Two
agents will connect, if at any time t they are closer than 100 meters
from each other. Finally, for each of these experiments, we iterate
over 10 values of M between 5 and 2000, and report the average
(cid:131)ood time (TF ) by aggregating over 25 realizations.
7
0246810α1000200030004000TFvs.αLWMRWPRW102103#Agents(M)103TFvs.#AgentsActualFitted100200300400500GridSize(N)0100002000030000TFvs.GridSizeActualFitted101102103Agents(M)103104FloodTime(TF)Boston(RWP)R2=0.998101102103Agents(M)103104NYC(RWP)R2=0.999101102103Agents(M)103104D.C.(RWP)R2=0.995101102103Agents(M)103104FloodTime(TF)Boston(DATA)R2=1.000101102103Agents(M)103104NYC(DATA)R2=0.998101102103Agents(M)103104D.C.(DATA)R2=0.934 Table 2: Fitting score, R
cities.
2, for all movement policies in all
City
Boston
New York City
Washington, D.C.
RWP DATA
0.998
1.000
0.998
0.999
0.995
0.934
Using a function similar to Section 4.1, we can (cid:128)t the simulation
results to our bound. Figure 4 shows the actual and (cid:128)(cid:138)ed values
2 value of the (cid:128)(cid:138)ing. In these
for the 3 cities, along with the R
simulations, our bound closely approximates the (cid:131)ood time in the
simulations, even when the movement policy used is data-speci(cid:128)c
2 values care
rather than the Random Way-point model. (cid:140)e R
compared in Table 2. Across all se(cid:138)ings, we achieve > 0.93 (5 of
them > 0.99), which shows the (cid:131)exibility of our model to variations
of network and movement policy. Note that here the network was
a real-world road network, and far from a torus.
(cid:140)ere can be many di(cid:130)erent factors contributing to the (cid:131)ood
time in networks as complicated as urban maps, which are beyond
the scope of this study. Here, we tried to explore the limits of our
model's prediction capabilities by tweaking the se(cid:138)ings of experi-
ment in a controlled manner. Further investigation in the e(cid:130)ects of
structural properties of road networks, and di(cid:130)erent distributions
of frequent origins and destinations on the (cid:131)ood time is needed
to fully understand the process of information dissemination by
human mobility in real road networks.
4.3 Real-World Data
traces of taxi cabs in the city of Shenzhen in China [19]. Over the
course of 24 hours, the location of 9386 taxis are sampled every
1.01 minutes. We set the transmission radius to 100 meters and, for
simplicity, assume that connections can happen only on sampled
points in time. To generate di(cid:130)erent numbers of moving agents
(M), we have to subsample from the set of all taxis. Since these
trajectories are (cid:128)xed, we cannot extend them in the event of having
no information (cid:131)ood. Hence, we (cid:128)lter out those taxis that meet less
than 100 distinct taxis during the whole 24 hours, and 3905 taxis
will remain. We iterate over 10 di(cid:130)erent values of M between 500
and 3000, each time (cid:128)nding the (cid:131)ood time in hours. We average the
results of 25 realizations for each M and report it. (cid:140)e results are
shown in Figure 5. We have followed the same procedure to (cid:128)t the
simulation values to our bound. (cid:140)e resulting (cid:128)(cid:136)ed line is drawn
2 value of 0.996. (cid:140)is shows that our
in Figure 5, achieving an R
model is capable of predicting (cid:131)ood times for real-world scenarios
to some degree. It is worth noting that the real-world experiments
did not have signi(cid:128)cant (cid:131)uctuations in the (cid:131)ood time value and,
similar to controlled experiments in the two sections before, shows
a smooth behavior, even with only hundreds of moving agents in
some cases.
5 TOWARDS BOUNDING THE L´EVY WALK
Compared to other mobility models, the L´evy Walk if far less stud-
ied. Formally, in a L´evy Walk, given a constant α > 0, an agent
th destination, X = Xi, chooses node Y as its next
positioned at its i
destination, Xi +1, with the following probability:
Pr{Xi +1 = Y Xi = X} =
1
ZY − Xα ,
(13)
where Z is the normalizing factor and X − Y is the distance
between nodes X and Y, such as Manha(cid:136)an Distance in a grid,
Euclidean Distance in the 2D plane or Graph Shortest Path Distance
in any given network. Figure 6 compares a Random, a L´evy and a
Random Way-point walker, simulated for 250 steps. Notice that a
Random Way-point walker tends to take big steps and cover a vast
area in the grid, while a Random walker is concentrated to a small
area around its initial position. A L´evy walker shows a mixture
of the two behaviors. It roams around in a small area most of the
time, but occasionally makes a long move to a di(cid:130)erent region in
the grid.
Figure 5: Actual (empty circles) and (cid:128)tted (dotted lines) TF
for Shenzhen.
Next, we try to (cid:128)t our model to real-world GPS traces. Ideally, one
may want to experiment on personal trajectories, as the behavior of
a single moving agent is best understood by looking at individuals'
mobility traces. However, due to the sensitivity of such data, large
and high-quality data sets containing personal mobility traces are
extremely rare. As a substitute, we can study the mobility of shared
vehicles, as we did in the previous section. Here, we study GPS
Figure 6: A Random Way-point walker (le(cid:133)), a random
walker (middle), and a L´evy walker (α = 2) a(cid:133)er 250 steps.
8
100020003000Agents(M)05101520TF(hours)R2=0.996050100020406080100RWP4045505055606570RandomWalk02040406080100120140160L'evyWalk To compare how information propagates in the three movement
models, we have to (cid:128)rst observe that Random Walk (or Brownian
motion) and Random Way-point can be thought of as two extreme
ends of the spectrum of all possible L´evy Walks. In (13), se(cid:138)ing
α to 0 (and applying the corresponding Z value) yields a constant
probability regardless of the distance between X and Y, similar
to Random Way-point. On the other hand, given any time limit
T , we can make α high enough so that w.h.p. no agent selects a
destination more than one unit distance away at any time t ≤ T ,
e(cid:130)ectively forcing them to follow Random Walk.
Figure 3, le(cid:137), shows the simulated results for the average (cid:131)ood
time, denoted by TF , of a system where agents are moving by
Random Walk, Manha(cid:138)an Random Way-point or L´evy Walk model
with di(cid:130)erent values of α. Both N and M are set to 100, α goes from 0
to 10 in 0.1 increments, and each point is created by aggregating the
results of 100 realizations. Additionally, with our newly discovered
bound for the Manha(cid:138)an Random Way-point model, the bounds
for TF in Random Walk and Manha(cid:138)an Random Way-point have
go(cid:138)en very close. We now have a reason to believe that any future
bound for L´evy Walk should be close to either of the bounds for
these two movement models. And since their bounds are close, it
is worth investigating whether or not a careful interpolation of the
bounds for Random Walk and Manha(cid:138)an Random Way-point is a
good predictor of how a L´evy Walker moves in a network.
6 CONCLUSION
(cid:140)anks to ever-present portable devices, there has been a grow-
ing interest in a be(cid:138)er understanding of mobile networks (also
called vehicular networks), where autonomous agents move inde-
pendently and are capable of carrying and transmi(cid:138)ing information.
We studied the case of M agents moving in an N × N torus.
ha(cid:138)an Random Way-point model, TF = O(N(cid:6)N/M(cid:7)), that is tight
We made a new improvement to the (cid:131)ood time bound for Man-
for a wide range of problem se(cid:138)ings. To the best of our knowl-
edge, this bound is stronger than all previous bounds found for this
movement model. (cid:140)rough extensive experiments, we showed that
our bound can accurately predict (cid:131)ood time for a wide variety of
simulated and real-world se(cid:138)ings.
Lastly, given the shrinking di(cid:130)erence between the bounds for
Random Walk and Random Way-point, and the fact that L´evy Walk
behaves in between the former two movement models, it is now
worth investigating whether a careful interpolation of Random
Walk and Random Way-point can describe L´evy Walk accurately
enough. Finding theoretical bounds for L´evy Walk can be a valuable
future work that further expands our knowledge of the relation
between these three movement models and ultimately of human
mobility.
ACKNOWLEDGEMENTS
A. Rezaei and J. Gao acknowledge support through NSF CCF-1535900,
CNS-1618391, and DMS-1737812. J. Phillips acknowledges sup-
port by NSF CCF-1350888, ACI-1443046, CNS-1514520, and CNS-
1564287. Research by C.D. T´oth was supported in part by NSF
CCF-1422311 and CCF-1423615. (cid:140)e experiments were conducted
with equipment purchased through NSF CISE Research Infrastruc-
ture Grant No. 1405641. (cid:140)e authors thank Dagstuhl Seminar 15111
9
on Computational Geometry during which some of the ideas were
developed.
[3] Christian Be(cid:138)ste(cid:138)er. 2001.
for Simulation of Wireless Networks.
[2] Albert-l´aszl´o Barab´asi. 2005. (cid:140)e origin of bursts and heavy tails in human
REFERENCES
[1] Paolo Bajardi, Chiara Pole(cid:138)o, Jose J Ramasco, Michele Tizzoni, Vi(cid:138)oria Colizza,
and Alessandro Vespignani. 2011. Human mobility networks, travel restrictions,
and the global spread of 2009 H1N1 pandemic. PloS one 6, 1 (2011), e16591.
dynamics. Nature 435, 7039 (2005), 207 -- 211.
Smooth is Be(cid:138)er than Sharp: A Ran-
dom Mobility Model
cite-
seerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.23.3460 (2001).
[4] B Birand, M Zafer, G Zussman, and K W Lee. 2011. Dynamic Graph Properties
of Mobile Networks under Levy Walk Mobility. In 2011 IEEE Eighth International
Conference on Mobile Ad-Hoc and Sensor Systems. 292 -- 301.
[5] Andrew M. Bishop. 2017. Router for OpenStreetMap Data. h(cid:138)p://www.routino.
org/. (2017).
J Y Le Boudec and M Vojnovic. 2006. (cid:140)e Random Trip Model: Stability, Stationary
Regime, and Perfect Simulation. IEEE/ACM Trans. Netw. 14, 6 (Dec. 2006), 1153 --
1166.
Josh Broch, David A Maltz, David B Johnson, Yih-Chun Hu, and Jorjeta Jetcheva.
1998. A Performance Comparison of Multi-hop Wireless Ad Hoc Network
Routing Protocols. In Proceedings of the 4th Annual ACM/IEEE International
Conference on Mobile Computing and Networking (MobiCom '98). ACM, New
York, NY, USA, 85 -- 97.
[8] Tracy Camp, Je(cid:130) Boleng, and Vanessa Davies. 2002. A survey of mobility models
for ad hoc network research. Wireless communications and mobile computing 2,
5 (2002), 483 -- 502.
h(cid:138)ps://www.
[9] Capital Bikeshare. 2017. Washington DC Trip Histories.
[6]
[7]
capitalbikeshare.com/system-data. (2017).
[10] Maurice Chu, Horst Haussecker, and Feng Zhao. 2002. Scalable information-
driven sensor querying and routing for ad hoc heterogeneous sensor networks.
(cid:138)e International Journal of High Performance Computing Applications 16, 3 (2002),
293 -- 313.
[11] Tommy Chu and Ioanis Nikolaidis. 2002. On the Artifacts of Random Waypoint
[12] Citi Bike. 2017. New York City Trip Histories. h(cid:138)ps://www.citibikenyc.com/
Simulations. In International Conference on Internet Computing.
system-data. (2017).
Information
[13] Andrea Clementi, Angelo Monti, and Riccardo Silvestri. 2011. Modelling mobility:
[15] Andrea Clementi, Riccardo Silvestri, and Luca Trevisan. 2015.
[18] Tassos Dimitriou, Sotiris Nikoletseas, and Paul Spirakis. 2006. (cid:140)e infection time
[14] Andrea Clementi, Angelo Monti, and Riccardo Silvestri. 2013. Fast (cid:131)ooding over
A discrete revolution. Ad Hoc Networks 9, 6 (Aug. 2011), 998 -- 1014.
Manha(cid:138)an. Distributed computing 26, 1 (2013), 25 -- 38.
spreading in dynamic graphs. Distributed Computing 28, 1 (2015), 55 -- 73.
[16] Pilu Crescenzi, Miriam Di Ianni, Andrea Marino, Gianluca Rossi, and Paola Vocca.
2009. Spatial Node Distribution of Manha(cid:138)an Path Based Random Waypoint
Mobility Models with Applications. In Structural Information and Communication
Complexity (Lecture Notes in Computer Science). Springer, Berlin, Heidelberg, 154 --
166.
[17] Yves-Alexandre De Montjoye, C´esar A Hidalgo, Michel Verleysen, and Vincent D
Blondel. 2013. Unique in the crowd: (cid:140)e privacy bounds of human mobility.
Scienti(cid:128)c reports 3 (2013), 1376.
of graphs. Discrete Applied Mathematics 154, 18 (2006), 2577 -- 2589.
Jiaxin Ding, Jie Gao, and Hui Xiong. 2015. Understanding and modelling informa-
tion dissemination pa(cid:138)erns in vehicle-to-vehicle networks. In Proceedings of the
23rd SIGSPATIAL International Conference on Advances in Geographic Information
Systems. ACM, 41.
[20] Nathan Eagle, Alex (Sandy) Pentland, and David Lazer. 2009. Inferring friendship
network structure by using mobile phone data. Proceedings of the National
Academy of Sciences 106, 36 (2009), 15274 -- 15278. h(cid:138)ps://doi.org/10.1073/pnas.
0900282106 arXiv:h(cid:138)p://www.pnas.org/content/106/36/15274.full.pdf+html
[21] David Easley and Jon Kleinberg. 2010. Networks, crowds, and markets: Reasoning
about a highly connected world. Cambridge University Press.
[22] Richard P Feynman, Robert B Leighton, and Ma(cid:138)hew Sands. 2013. (cid:138)e Feynman
Lectures on Physics, Desktop Edition Volume I. Vol. 1. Basic books.
[23] Enrique Frias-Martinez, Graham Williamson, and Vanessa Frias-Martinez. 2011.
An agent-based model of epidemic spread using human mobility and social
network information. In Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE
(cid:138)ird Inernational Conference on Social Computing (SocialCom), 2011 IEEE (cid:138)ird
International Conference on. IEEE, 57 -- 64.
[24] Manuel Gomez Rodriguez, Jure Leskovec, and Andreas Krause. 2010. Inferring
networks of di(cid:130)usion and in(cid:131)uence. In Proceedings of the 16th ACM SIGKDD
international conference on Knowledge discovery and data mining. ACM, 1019 --
1028.
[19]
[25] Marta C Gonz´alez, C´esar A Hidalgo, and Albert-L´aszl´o Barab´asi. 2008. Under-
standing individual human mobility pa(cid:138)erns. Nature 453, 7196 (2008), 779 -- 782.
[26] Yih-Chun Hu and David B Johnson. 2000. Caching Strategies in On-demand
Routing Protocols for Wireless Ad Hoc Networks. In Proceedings of the 6th Annual
International Conference on Mobile Computing and Networking (MobiCom '00).
ACM, New York, NY, USA, 231 -- 242.
[27] Hubway. 2017. Boston Trip Histories. h(cid:138)ps://www.thehubway.com/system-data.
(2017).
[32]
Media.
[28] Chalermek Intanagonwiwat, Ramesh Govindan, and Deborah Estrin. 2000. Di-
rected di(cid:130)usion: A scalable and robust communication paradigm for sensor
networks. In Proceedings of the 6th annual international conference on Mobile
computing and networking. ACM, 56 -- 67.
[29] David B Johnson and David A Maltz. 1996. Dynamic Source Routing in Ad Hoc
Wireless Networks. In Mobile Computing. Springer, Boston, MA, 153 -- 181.
[30] David Kempe, Jon Kleinberg, and ´Eva Tardos. 2003. Maximizing the spread of
in(cid:131)uence through a social network. In Proceedings of the ninth ACM SIGKDD
international conference on Knowledge discovery and data mining. ACM, 137 -- 146.
[31] Harry Kesten and Vladas Sidoravicius. 2005. (cid:140)e spread of a rumor or infection
in a moving population. Annals of Probability (2005), 2402 -- 2462.
Jon Kleinberg. 2007. Computing: (cid:140)e wireless epidemic. Nature 449, 7160 (2007),
287 -- 288.
[33] Gregoyr Lawler. 2013. Intersections of random walks. Springer Science & Business
[34] K Lee, Y Kim, S Chong, I Rhee, and Y Yi. 2011. Delay-capacity tradeo(cid:130)s for
mobile networks with L´evy walks and L´evy (cid:131)ights. In 2011 Proceedings IEEE
INFOCOM. 3128 -- 3136.
Planet dump retrieved from
h(cid:138)ps://planet.osm.org . h(cid:138)ps://www.openstreetmap.org. (2017).
[35] OpenStreetMap contributors. 2017.
[36] Wei Peng and Xi-Cheng Lu. 2000. On the reduction of broadcast redundancy in
mobile ad hoc networks. In Proceedings of the 1st ACM international symposium
on Mobile ad hoc networking & computing. IEEE Press, 129 -- 130.
[37] C E Perkins, E M Royer, S R Das, and M K Marina. 2001. Performance comparison
of two on-demand routing protocols for ad hoc networks. IEEE Pers. Commun. 8,
1 (Feb. 2001), 16 -- 28.
[38] Alberto Pe(cid:138)arin, Andrea Pietracaprina, Geppino Pucci, and Eli Upfal. 2011. Tight
bounds on information dissemination in sparse mobile networks. In Proceedings
of the 30th annual ACM SIGACT-SIGOPS symposium on Principles of distributed
computing. ACM, 355 -- 362.
[39] E M Royer, P M Melliar-Smith, and L E Moser. 2001. An analysis of the optimum
node density for ad hoc mobile networks. In ICC 2001. IEEE International Confer-
ence on Communications. Conference Record (Cat. No.01CH37240), Vol. 3. 857 -- 861
vol.3.
[40] K Shinki, M Nishida, and N Hayashibara. 2017. Message Dissemination Using
L´evy Flight on Unit Disk Graphs. In 2017 IEEE 31st International Conference on
Advanced Information Networking and Applications (AINA). 355 -- 362.
[41] Michael F Shlesinger, Joseph Kla(cid:137)er, and Gert Zumofen. 1999. Above, below and
beyond Brownian motion. American Journal of Physics 67, 12 (1999), 1253 -- 1259.
[42] Chaoming Song, Zehui (cid:139), Nicholas Blumm, and Albert-Lszl Barabsi. 2010.
Limits of Predictability in Human Mobility. Science 327, 5968 (2010), 1018 -- 1021.
h(cid:138)ps://doi.org/10.1126/science.1177170
[43] G M Viswanathan, F Bartumeus, Sergey V. Buldyrev, J Catalan, U L Fulco, Shlomo
Havlin, M G E da Luz, M L Lyra, E P Raposo, and H Eugene Stanley. 2002. L´evy
(cid:131)ight random searches in biological phenomena. Physica A: Statistical Mechanics
and its Applications 314, 1 (Nov. 2002), 208 -- 213.
[44] S Wang, X Wang, X Cheng, J Huang, and R Bie. 2014. (cid:140)e Tempo-Spatial Infor-
mation Dissemination Properties of Mobile Opportunistic Networks with Levy
Mobility. In 2014 IEEE 34th International Conference on Distributed Computing
Systems. 124 -- 133.
J Yoon, M Liu, and B Noble. 2003. Random waypoint considered harmful. In IEEE
INFOCOM 2003. Twenty-second Annual Joint Conference of the IEEE Computer
and Communications Societies (IEEE Cat. No.03CH37428), Vol. 2. 1312 -- 1321 vol.2.
[46] Dengpan Zhou and Jie Gao. 2009. Opportunistic Processing and (cid:139)ery of Motion
Trajectories in Wireless Sensor Networks. In Proceedings of the 28th Annual IEEE
Conference on Computer Communications (INFOCOM'09). 1197 -- 1205.
[45]
10
|
1701.05108 | 1 | 1701 | 2017-01-18T15:36:46 | On the Computational Complexity of Variants of Combinatorial Voter Control in Elections | [
"cs.MA"
] | Voter control problems model situations in which an external agent tries toaffect the result of an election by adding or deleting the fewest number of voters. The goal of the agent is to make a specific candidate either win (\emph{constructive} control) or lose (\emph{destructive} control) the election. We study the constructive and destructive voter control problems whenadding and deleting voters have a \emph{combinatorial flavor}: If we add (resp.\ delete) a voter~$v$, we also add (resp.\ delete) a bundle~$\kappa(v) $ of voters that are associated with~$v$. While the bundle~$\kappa(v)$ may have more than one voter, a voter may also be associated with more than one voter. We analyze the computational complexity of the four voter control problems for the Plurality rule. We obtain that, in general, making a candidate lose is computationally easier than making her win. In particular, if the bundling relation is symmetric (i.e.\ $\forall w\colon w \in \kappa(v) \Leftrightarrow v \in \kappa(w) $), and if each voter has at most two voters associated with him, then destructive control is polynomial-time solvable while the constructive variant remains $\NP$-hard. Even if the bundles are disjoint (i.e.\ $\forall w\colon w \in \kappa(v) \Leftrightarrow \kappa(v) = \kappa(w) $), the constructive problem variants remain intractable. Finally, the minimization variant of constructive control by adding voters does not admit an efficient approximation algorithm, unless P=NP. | cs.MA | cs | On the Computational Complexity of Variants of
Combinatorial Voter Control in Elections
Leon Kellerhals, Viatcheslav Korenwein, Philipp Zschoche⋆,
Robert Bredereck⋆⋆, and Jiehua Chen
{leon.kellerhals, viatcheslav.korenwein, zschoche}@campus.tu-berlin.de
{robert.bredereck, jiehua.chen}@tu-berlin.de
TU Berlin, Germany
Abstract. Voter control problems model situations in which an external
agent tries to affect the result of an election by adding or deleting the
fewest number of voters. The goal of the agent is to make a specific
candidate either win (constructive control) or lose (destructive control)
the election. We study the constructive and destructive voter control
problems when adding and deleting voters have a combinatorial flavor:
If we add (resp. delete) a voter v, we also add (resp. delete) a bundle κ(v)
of voters that are associated with v. While the bundle κ(v) may have
more than one voter, a voter may also be associated with more than one
voter. We analyze the computational complexity of the four voter control
problems for the Plurality rule.
We obtain that, in general, making a candidate lose is computationally
easier than making her win. In particular, if the bundling relation is sym-
metric (i.e. ∀w : w ∈ κ(v) ⇔ v ∈ κ(w)), and if each voter has at most
two voters associated with him, then destructive control is polynomial-
time solvable while the constructive variant remains NP-hard. Even if the
bundles are disjoint (i.e. ∀w : w ∈ κ(v) ⇔ κ(v) = κ(w)), the constructive
problem variants remain intractable. Finally, the minimization variant
of constructive control by adding voters does not admit an efficient ap-
proximation algorithm, unless P = NP.
1 Introduction
Since the seminal paper by Bartholdi III et al. [3] on controlling an elec-
tion by adding or deleting the fewest number of voters or candidates with
the goal of making a specific candidate to win (constructive control), a lot
of research has been devoted to the study of control for different voting
rules [16, 14, 24, 23, 4, 21], on different control modes [17, 18], or even
on other controlling goals (e.g. aiming at several candidates' victory or a
⋆ PZ was supported by the Stiftung Begabtenförderung berufliche Bildung (SBB).
⋆⋆ RB was from September 2016 to September 2017 on postdoctoral leave at the Uni-
versity of Oxford (GB), supported by the DFG fellowship BR 5207/2.
7
1
0
2
n
a
J
8
1
]
A
M
.
s
c
[
1
v
8
0
1
5
0
.
1
0
7
1
:
v
i
X
r
a
specific candidate's defeat) [20, 26]. Recently, Bulteau et al. [8] introduced
combinatorial structures to constructive control by adding voters: When a
voter is added, a bundle of other voters is added as well. A combinatorial
structure of the voter set allows us to model situations where an external
agent hires speakers to convince whole groups of people to participate in
(or abstain from) an election. In such a scenario, convincing a whole group
of voters comes at the fixed cost of paying a speaker. Bulteau et al. [8]
model this by defining a bundle of associated voters for each voter which
will be convinced to vote "for free" when this voter is added or deleted.
Moreover, the bundles of different voters could overlap. For instance, con-
vincing two bundles of two voters each to participate in the election could
result in adding a total of four, three or even just two voters.
We extend the work of Bulteau et al. [8] and investigate the cases
where the agent wants to make a specific candidate win or lose by adding
(resp. deleting) the fewest number of bundles. We study one of the simplest
voting rules, the Plurality rule, where each voter gives one point to his
favorite candidate, and the candidate with most points becomes a winner.
Accordingly, an election consists of a set C of candidates and a set V of
voters who each have a favorite candidate. Since real world elections typ-
ically contain only a small number of candidates, and a bundle of voters
may correspond to a family with just a few members, we are especially
interested in situations where the election has only few candidates and the
bundle of each voter is small. Our goal is to ensure that a specific candidate
p becomes a winner (or a loser) of a given election, by convincing as few
voters from an unregistered voter set W as possible (or as few voters from
V as possible), together with the voters in their bundles, to participate
(or not to participate) in the election. We study the combinatorial voter
control problems from both the classical and the parameterized complex-
ity point of view. We confirm Bulteau et al.'s conjecture [8] that for the
Plurality rule, the three problem variants: combinatorial constructive con-
trol by deleting voters and combinatorial destructive control by adding
(resp. deleting) voters, behave similarly in complexity to the results of
combinatorial constructive control by adding voters: They are NP-hard
and intractable even for very restricted cases. We can also identify several
special cases, where the complexity of the four problems behave differently.
For instance, we find that constructive control tends to be computationally
harder than destructive control. We summarize our results in Table 1.
Related Work. Bartholdi III et al. [3] introduced the complexity study
of election control problems and showed that for the Plurality rule, the
non-combinatorial variant of the voter control problems can be solved in
2
linear time by a using simple greedy strategy. We refer the readers to the
work by Rothe and Schend [27], Faliszewski and Rothe [15] for general
expositions on election control problems.
In the original election control setting, a unit modification of the elec-
tion concerns usually a single voter or candidate. The idea of adding com-
binatorial structure to election voter control was initiated by Bulteau et al.
[8]: Instead of adding a voter at each time, one adds a "bundle" of vot-
ers to the election, and the bundles added to the election could intersect
with each other. They showed that combinatorial constructive control by
adding the fewest number of bundles becomes notorious hard, even for the
Plurality rule and for only two candidates. Chen [9] mentioned that even
if each bundle has only two voters and the underlying bundling graph is
acyclic, the problem still remains NP-hard. Bulteau et al. [8] and Chen [9]
conjectured that
"the combinatorial addition of voters for destructive control, and
combinatorial deletion of voters for either constructive or destruc-
tive control behave similarly to combinatorial addition of voters
for constructive control."
The combinatorial structure notion for voter control has also been
extended to candidate control [10] and electoral shift bribery [7].
Paper Outline. In Section 2, we introduce the notation used through-
out the paper. In Section 3 we formally define the four problem variants,
summarize our contributions, present results in which the four problem
variants (constructive or destructive, adding voters or deleting voters) be-
have similarly, and provide reductions between the problem variants. Sec-
tions 4 to 6 present our main results on three special cases (1) when the
bundles and the number of candidates are small, (2) when the bundles are
disjoint, and (3) when the solution size could be unlimited. We conclude
in Section 7 with several future research directions.
2 Preliminaries
The notation we use in this paper is based on Bulteau et al. [8]. We assume
familiarity with standard notions regarding algorithms and complexity
theory. For each z ∈ N we denote by [z] the set {1, . . . , z}.
Elections. An election E = (C, V ) consists of a set C of m candidates
and a set V of voters. Each voter v ∈ V has a favorite candidate c and we
call voter v a c-voter. Note that since we focus on the Plurality rule, we
simplify the notion of the preferences of voters in an election to the favorite
3
candidate of each voter. For each candidate c ∈ C and each subset V ′ ⊆ V
of voters, her (Plurality) score sc(V ′) is defined as the number of voters
from V ′ that have her as favorite candidate. We say that a candidate c is
a winner of election (C, V ) if c has the highest score sc(V ). For the sake
of convenience, for each C ′ ⊆ C, a C ′-voter denotes a voter whose favorite
candidate belongs to C ′.
Combinatorial Bundling Functions. Given a voter set X, a combina-
torial bundling function κ : X → 2X (abbreviated as bundling function)
is a function that assigns a set of voters to each voter; we require that
x ∈ κ(x). For the sake of convenience, for each subset X ′ ⊆ X, we define
κ(X ′) = Sx∈X ′ κ(x). For a voter x ∈ X, κ(x) is named x's bundle; x is
called the leader of the bundle. We let b denote the maximum bundle size
of a given κ. Formally, b = maxx∈X κ(x). One can think of the bundling
function as subsets of voters that can be added at a unit cost (e.g. κ(x)
is a group of voters influenced by x).
Bundling graphs. The bundling graph of an election is a model of how
the voter's bundle functions interact with each other.
Let κ : X → 2X be a bundling function. The bundling graph Gκ =
(V (Gκ), E(Gκ)) is a simple, directed graph: For each voter x ∈ X there is
a vertex ux ∈ V (Gκ). For each two distinct voters y, z ∈ X with z ∈ κ(y),
there is an arc (uy, uz) ∈ E(Gκ).
We consider three special cases of bundling functions/graphs which
we think are natural in real world. We say that a bundling function κ is
symmetric if for each two distinct voters x, y ∈ X, it holds that y ∈ κ(x)
if and only if x ∈ κ(y). The bundling graph for a symmetric bundling
function always has two directed arcs connecting each two vertices. Thus,
we can assume the graph to be undirected.
We say that κ is disjoint if for each two distinct voters x, y ∈ X, it
holds that either κ(x) = κ(y) or κ(x) ∩ κ(y) = ∅. It is an easy exercise
to verify that disjoint bundling functions are symmetric and the corre-
sponding undirected bundling graphs consist only of disjoint complete
subgraphs.
We say that κ is anonymous if for each two distinct voters x and y
with the same favorite candidate, it holds that κ(x) = κ(y), and that for
all other voters z we have x ∈ κ(z) if and only if y ∈ κ(z).
Parameterized Complexity. An instance (I, r) of a parameterized prob-
lem consists of the actual instance I and of an integer r referred to as the
parameter [13, 19, 25]. A parameterized problem is called fixed-parameter
tractable (in FPT) if there is an algorithm that solves each instance (I, r)
4
in f (r) · IO(1) time, where f is a computable function depending only on
the parameter r.
There is also a hierarchy of hardness classes for parameterized prob-
lems, of which the most important ones are W[1] and W[2]. One can show
that a parameterized problem L is (presumably) not in FPT by devis-
ing a parameterized reduction from a W[1]-hard or a W[2]-hard problem
to L. A parameterized reduction from a parameterized problem L to an-
other parameterized problem L′ is a function that acts as follows: For two
computable functions f and g, given an instance (I, r) of problem L, it
computes in f (r) · IO(1) time an instance (I ′, r′) of problem L′ so that
r′ ≤ g(r) and that (I, r) ∈ L if and only if (I ′, r′) ∈ L′. For a survey of
research on parameterized complexity in computational social choice, we
refer to Betzler et al. [5] and Bredereck et al. [6].
3 Central Problem
We consider the problem of combinatorial voter control in four variants.
The variants differ in whether they are constructive or destructive, mean-
ing that the goal is to make one selected candidate win or lose the election.
This goal can be achieved by either adding voters to or deleting voters from
the given election. Due to space constraints, we only provide the definition
of constructive control. Destructive control is defined analogously.
Combinatorial Constructive Control by Adding
(resp. Deleting) Voters [C-Cons-Add (resp. C-Cons-Del)]
Input: An election E = (C, V ), a set W of unregistered voters with
V ∩ W = ∅, a bundling function κ : W → 2W (resp. κ : V → 2V ),
a preferred winner p ∈ C, and an integer k ∈ N.
Quest.: Is there a size-at-most k subset W ′ ⊆ W (resp. V ′ ⊆ V ) of
voters such that p wins the election (C, V ∪ κ(W ′)) (resp. (C, V \
κ(V ′)))?
Throughout this work, when speaking of the "adding" or "deleting"
variants, we mean those variants in which voters are added or, respectively,
deleted. In similar fashion, we speak of the constructive and destructive
(abbr. by "Cons" and by "Des", respectively) problem variants. Further,
we refer to the set W ′ of voters as the solution for the "adding" variants
(the set V ′ of voters for the "deleting" variants, respectively) and denote
k as the solution size.
Our Contributions. We study both the classical and the parameterized
complexity of the four voter control variants. We are particularly inter-
ested in the real-world setting where the given election has a small number
5
Table 1: Computational complexity of the four combinatorial voter con-
trol variants with the Plurality rule. The parameters are "the solution size
k", "the number m of candidates" and "the maximum bundle size b". We
refer to I as the instance size. The rows distinguish between different
maximum bundle sizes b and the number m of candidates. All parameter-
ized intractability results are for the parameter "solution size k". ILP-FPT
means FPT based on a formulation as an integer linear program and the
result is for the parameter "number m of candidates".
C-Cons-Add C-Cons-Del C-Des-Add C-Des-Del References
κ symmetric
b = 2
b = 3
m = 2
O(I)
P
O(mI)
O(mI)
O(I5)
O(I5)
O(I5)
O(I5)
m unbounded
NP-h
NP-h
O(mI5)
O(mI5)
Obs 2, Thm 3
Thm 5
Thm 2,
Cor 1+2
Obs 1, Prop 2,
Cor 2
b unbounded
m = 2
m unbounded and
W[2]-h
W[2]-h
W[2]-h
W[2]-h
[8], Thm 1
κ disjoint
W[1]-h
W[2]-h
O(mI)
O(mI)
Thm 4+5
κ anonymous
ILP-FPT
ILP-FPT
ILP-FPT
ILP-FPT
Thm 1
κ arbitrary
b = 3, m = 2
W[1]-h
W[1]-h
W[1]-h
W[1]-h
Thm 1
of candidates and where only a few voters are associated to a voter. On the
one hand, we were able to confirm the conjecture given by Bulteau et al. [8]
and Chen [9] that when parameterized by the solution size, C-Cons-Del,
C-Des-Add, and C-Des-Del are all intractable even for just two candi-
dates or for bundle sizes of at most three, and that when parameterized by
the number of candidates, they are fixed-parameter tractable for anony-
mous bundling functions. On the other hand, we identify interesting spe-
cial cases where the four problems differ in their computational complexity.
We conclude that in general, destructive control tends to be easier than
constructive control: For symmetric bundles with at most three voters,
C-Cons-Add is known to be NP-hard, while both destructive problem
variants are polynomial-time solvable. For disjoint bundles, constructive
control is parameterized intractable (for the parameter "solution size k"),
while destructive control is polynomial-time solvable. Unlike for C-Cons-
Del, a polynomial-time approximation algorithm for C-Cons-Add does
not exist, unless P = NP. Our results are gathered in Table 1.
6
The following theorem summarizes the conjecture given by Bulteau
et al. [8] and Chen [9]. The proofs are deferred to Appendices A.1 to A.3.
Theorem 1. All four combinatorial voter control variants are
(i) W[2]-hard with respect to the solution size k, even for only two can-
didates and for symmetric bundling functions κ
(ii) W[1]-hard with respect to the solution size k, even for only two can-
didates and for bundle sizes of at most three.
(iii) fixed-parameter tractable with respect to the number m of candidates
if the bundling function κ is anonymous.
Relations between the four problem variants. We provide some re-
ductions between the problem variants. They are used in several sections
of this paper. The key idea for the reduction from destructive control to
constructive control is to guess the candidate that will defeat the distin-
guished candidate and ask whether one can make this candidate win the
election. The key idea for the reduction from the "deleting" to the "adding"
problem variants is to build the "complement" of the registered voter set.
Proposition 1. For each X ∈ {Add, Del}, C-Des-X with m candi-
dates is Turing reducible to C-Cons-X with two candidates. For each
Y ∈ {Cons, Des}, C-Y -Del with two candidates is many-one reducible
to C-Y -Add with two candidates. All these reductions preserve the prop-
erty of symmetry of the bundling functions.
4 Controlling Voters with Symmetric and Small Bundles
In this section, we study combinatorial voter control when the voter bun-
dles are symmetric and small. This could be the case when a voter's bun-
dle models his close friends (including himself), close relatives, or office
mates. Typically, this kind of relations is symmetric, and the number of
friends, relatives, or office mates is small. We show that for symmetric
bundles and for bundles size at most three, both destructive problem vari-
ants become polynomial-time solvable, while both constructive variants
remain NP-hard. However, if there are only two candidates, then we can
use dynamic programming to also solve the constructive control variants
in polynomial time. If we restrict the bundle size to be at most two, then
all four problem variants can be solved in polynomial time via simple
greedy algorithms.
7
As already observed in Section 2, we only need to consider the undi-
rected version of the bundling graph for symmetric bundles. Moreover, if
the bundle size is at most two, then the resulting bundling graph consists
of only cycles and trees. However, Bulteau et al. [8] already observed that
C-Cons-Add is NP-hard even if the resulting bundling graph solely con-
sists of cycles, and Chen [9] observed that C-Cons-Add remains NP-hard
even if the resulting bundling graph consists of only directed trees of depth
at most three.
Observation 1. C-Cons-Add is NP-hard even for symmetric bundling
functions with maximum bundle size b = 3.
It turns out that the reduction used by Bulteau et al. [8] to show
Observation 1 can be adapted to show NP-hardness for the deleting case.
Proposition 2. C-Cons-Del is NP-hard even for symmetric bundling
functions with maximum bundle size b = 3.
If, in addition to the bundles being symmetric and of size at most
three, we have only two candidates, then we can solve C-Cons-Add in
polynomial time. First of all, due to these constraints, we can assume
that the bundling graph Gκ is undirected and consists of only cycles and
paths. Then, it is easy to verify that we can consider each cycle and each
path separately. Finally, we devise a dynamic program for the case when
the bundling graph is a path or a cycle, maximizing the score difference
between our preferred candidate p and the other candidate. The crucial
idea behind the dynamic program is that the bundles of a minimum-size
solution induce a subgraph where each connected component is small.
Lemma 1. Let (E = (C, V ), W, κ, p, k) be a C-Cons-Add instance such
that C = {p, g}, and κ is symmetric with Gκ being a path. Then, finding
a size-at-most-k subset W ′ ⊆ W of voters such that the score difference
between p and q in κ(W ′) is maximum can be solved in O(W 5) time,
where W is the size of the unregistered voter set W .
Proof. Since Gκ is a path, each bundle has at most three voters. We denote
the path in Gκ by (w1, w2, . . . , wW ) and introduce some definitions for
this proof. The set W (s, t) := {wi ∈ W s ≤ i ≤ t} contains all voters on
a sequence from ws to wt. For every subset W ′ ⊆ W we define gap(W ′) :=
sp(κ(W ′)) − sg(κ(W ′)) as the score difference between p and g. One can
observe that if W ′ is a solution for (E = (C, V ), W, κ, p, k) then gap(W ′) ≥
sg(V ) − sp(V ); note that we only have two candidates. An (s, t)-proper-
subset W ′ is a subset of W (s, t) such that κ(W ′) ⊆ W (s, t). A maximum
8
(s, t)-proper-subset W ′ additionally requires that each (s, t)-proper-subset
W ′′ ⊆ W with W ′′ = W ′ has gap(W ′′) ≤ gap(W ′).
We provide a dynamic program in which a table entry T [r, s, t] contains
a maximum (s, t)-proper-subset W ′ of size r. We first initialize the table
entries for the case where t − s + 1 ≤ 9 and r ≤ 9 in linear time.
For t − s + 1 > 9, we compute the table entry T [r, s, t] by considering
every possible partition of W (s, t) into two disjoint parts.
T [r, s, t] := T [r − i, s, s + j] ∪ T [i, s + j + 1, t],
where i, j = arg max
gap(T [r − i, s, s + j]) + gap(T [i, s + j + 1, t]).
0≤i≤r
0≤j≤t−s−2
Note that a maximum (1, W )-proper-subset W ′ of size r − 1 could have
a higher gap than a maximum (1, W )-proper-subset W ′′ of size r.
To show the correctness of our program, we define the maximization
and minimization function on a set of voters W ′, which return the largest
and smallest index of all voters on the path induced by W ′, respectively:
max(W ′) := arg max
i∈W ′
{wi ∈ (W ′)} and min(W ′) := arg min
i∈W ′
{wi ∈ (W ′)}.
First, we use the following claim to show that each maximum (s, t)-proper-
subset W ′ can be partitioned into two (s, t)-proper-subsets W1, W2 such
that the two sets κ(W1) and κ(W2) are disjoint. (The formal proof of the
following claim can be found in the Appendix.)
Claim 1. Let W ′ be a maximum (s, t)-proper-subset and (max κ(W ′) −
min κ(W ′)+ 1) > 9. Then, there is a j with s < j < t such that there is an
(s, j)-proper-subset W1 and a (j+1, t)-proper-subset W2 with W1+W2 ≤
W ′ and κ(W1 ∪ W2) = κ(W ′).
Now, we show that the two subsets W1 and W2 from Claim 1 are indeed
optimal: There is a j such that W1 is a maximum (s, j)-proper-subset and
W2 is a maximum (j + 1, t)-proper-subset.
Assume towards a contradiction that W2 is a (j + 1, t)-proper-subset
but not a maximum (j + 1, t)-proper-subset. Therefore, there exists a
maximum (j+1, t)-proper-subset W ′
2. This implies that
gap(W1 ∪ W ′
2) > gap(W1 ∪ W2). This is a contradiction to W ′ = W1 ∪ W2
being a maximum (s, t)-proper-subset. The case in which W1 is not a
maximum (s, j)-proper-subset is analogous.
2 where W2 = W ′
Altogether, we have shown that we can compute T [k, s, t] in constant
time if t − s + 1 ≤ 9, and that otherwise there exist an i and a j such
9
that T [k, s, t] = T [k − i, s, t − j] ∪ T [i, t − j + 1, t]. The dynamic program
considers all possible i and j. The table entry T [i, 1, W ] contains a subset
W ′ ⊆ W of size i with maximum gap such that κ(W ′) ⊆ W (1, W ), which
is identical to κ(W ′) ⊆ W .
This completes the correctness proof of our dynamic program. The ta-
ble has O(k ·W 2) entries. To compute one entry the dynamic program ac-
cesses O(k · W ) other table entries. Note that the value gap(T [i, s, t]) can
be computed and stored after the entry T [i, s, t] is computed. This takes
at most O(W ) steps. Thus, the dynamic program runs in O(W 5) time.
⊓⊔
The dynamic program can be used to solve the same problem on cycles.
Altogether, we obtain the following.
Theorem 2. C-Cons-Add with a symmetric bundling function, maxi-
mum bundle size of three, and for two candidates can be solved in O(W 5)
time, where W is the size of the unregistered voter set.
Proof. Let (E = (C, V ), W, κ, p, k) be a C-Cons-Add instance, where
the maximum bundle size b is three, κ is symmetric, and C = {p, g}. This
means that all connected components C1, . . . , Cℓ of Gκ are path or cycles.
Furthermore, all bundles only contain voters from one connected compo-
nent. We define a dynamic program in which each table entry A[i, s, t]
contains a solution W ′ ⊆ W of size i, where κ(W ′) ⊆ V (Cs) ∪ · · · ∪ V (Ct)
and 1 ≤ s ≤ t ≤ ℓ:
(i) If s = t = j, then A[i, s, t] = T [i, 1, V (Cj)], where T is the dynamic
program of Cj, depending on whether Cj is a path or cycle.
(ii) Otherwise, we build the table as follows:
A[d, s, t] = A[d − i, s, s + j] ∪ A[i, s + j + 1, t], where
i, j = arg max
gap(A[d − i, s, s + j]) + gap(A[i, s + j + 1, t]).
0≤i≤d
1≤j≤t−s−1
Each of the table entries A[i, j, j] can be computed in O(i2 ·V (Cj)3) time
(see Lemmas 1 and B.1) and each of the table entries A[i, s, t] for s < t
can be computed in O(k · ℓ) time. Since we have k · ℓ2 entries, the total
running time is
Pℓ
i=1 O(k2 · V (Cj)3) = O(k2)Pℓ
i=1 O(V (Ci)) = O(k2 · W 3).
⊓⊔
From the polynomial-time solvability of Theorem 2 and by Proposi-
tion 1, we obtain the following results:
Corollary 1. C-Cons-Del with a symmetric bundling function, a max-
imum bundle size of three and two candidates can be solved in O(V 5)
time, where V is the size of the voters.
10
Corollary 2. C-Des-Add and C-Des-Del with a symmetric bundling
function and maximum bundle size three can be solved in time O(m · W 5)
and O(m · V 5), respectively, where m is the number of candidates, and
W and V are the sizes of the unregistered and registered voter set, re-
spectively.
5 Controlling Voters with Disjoint Bundles
We have seen in Section 4 that the interaction between the bundles in-
fluences the computational complexity of our combinatorial voter control
problems. For instance, adding a voter v to the election may lead to adding
another voter v′ with v ∈ κ(v). This is crucial for the reductions used to
prove Theorem 1 and Observation 1. Thus, it would be interesting to know
whether the problem becomes tractable if it is not necessary to add two
bundles that share some voter(s). More specifically, we are interested in
the case where the bundles are disjoint, meaning that we do not need to
consider every single voter, but only the bundles as a whole, as it does
not matter which voters of a bundle we select.
First, we consider disjoint bundles of size at most two. This is the case
for voters who have a partner. If a voter is convinced to participate in or
leaves the election, then the partner is convinced to do the same. Note
that this is equivalent to having symmetric bundles of size at most two.
Bulteau et al. [8, Theorem 6] constructed a linear-time algorithm for C-
Cons-Add if the maximum bundle size is two and κ is a full-d bundling
function (which implies symmetry). We can verify that their algorithm
actually works for disjoint bundles of size at most two. Thus, we obtain
the following.
Observation 2. C-Cons-Add with a symmetric bundling function and
with bundles of size at most two can be solved in O(I) time, where I is
the input size.
If we want to delete instead of add voter bundles, the problem reduces
to finding a special variant of the f -Factor problem, which is a gener-
alization of the well-known matching problem and can still be solved in
polynomial time [1, 2].
Theorem 3. C-Cons-Del with a symmetric bundling function and with
bundles of size at most two can be solved in polynomial time.
If we drop the restriction on the bundle sizes but still require the
bundles to be disjoint, then C-Cons-Add and C-Cons-Del become pa-
rameterized intractable with respect to the solution size.
11
Theorem 4. Parameterized by the solution size k, C-Cons-Add and C-
Cons-Del are W[1]-hard and W[2]-hard respectively, even for disjoint
bundles.
Proof (with only the construction for the W[1]-hardness proof of C-Cons-
Add). We provide a parameterized reduction from the W[1]-complete
problem Independent Set (parameterized by the "solution size") which,
given an undirected graph G = (V (G), E(G)) and a natural number h ∈ N,
asks whether G admits a size-h independent set U ⊆ V (G), that is, all ver-
tices in U are pairwise non-adjacent. Let (G, h) be an Independent Set
instance with E(G) = {e1, . . . , em−1} and V (G) = {u1, . . . , un}. Without
loss of generality, we assume that G is connected and h ≥ 3. We construct
an election E = (C, V ) with candidate set C := {p} ∪ {gj ej ∈ E(G)}.
For each edge ej ∈ E, we construct h − 1 registered voters that all have gj
as their favorite candidate. In total, V consists of (h − 1) · (m − 1) voters.
The unregistered voter set W is constructed as follows: For each ver-
tex ui ∈ V (G), add a p-voter pi, and for each edge ej incident with ui,
add a gj-voter a(i)
j . The voters constructed for each vertex ui are bundled
by the bundling function κ. More formally, for each ui ∈ V (G) and each
ej ∈ E(G) with ui ∈ ej, it holds that
κ(pi) = κ(a(i)
j ) := {pi} ∪ {a(i)
j ′
ej ′ ∈ E(G) ∧ ui ∈ ej ′}.
To finalize the construction, we set k := h. The construction is both
a polynomial-time and a parameterized reduction, and all bundles are
disjoint. To show the correctness, we note that p can only win if only if
her score can be increased to at least h without giving any other candidate
more than one more point. The solution corresponds to exactly to a subset
of h vertices that are pairwise non-adjacent. The detailed correctness proof
and the remaining proof for the W[2]-hardness result can be found in the
⊓⊔
Appendix.
For destructive control, it is sufficient to guess a potential defeater d
out of m − 1 possible candidates that will have a higher score than p in
the final election and use a greedy strategy similar to the one used for
Observation 2 to obtain the following result.
Theorem 5. C-Des-Add and C-Des-Del with a symmetric bundling
function and disjoint bundles can be solved in O(m · I) time, where I
is the input size and m the number of candidates.
12
6 Controlling Voters with Unlimited Budget
To analyze election control, it is interesting to know whether a solution
exist at all, without bounding its size. Indeed, Bartholdi III et al. [3] al-
ready considered the case of unlimited solution size for the constructive
candidate control problem. They showed that the problem is already NP-
hard, even if the solution size is not bounded. (The non-combinatorial
destructive control by adding unlimited amount of candidates is shown
to be also NP-hard by Hemaspaandra et al. [20].) In contrast, the non-
combinatorial voter control variants are linear-time solvable via simple
greedy algorithms [3]. This leads to the question whether the combi-
natorial structure increases the complexity. To this end, we relax the
four problem variants so that the solution can be of arbitrary size and
call these problems C-Cons-Add-Unlim, C-Des-Add-Unlim, C-Cons-
Del-Unlim and C-Des-Del-Unlim.
First of all, we observe that C-Cons-Del-Unlim becomes trivial if
no unique winner is required.
Lemma 2. Let I = (E = (C, V ), κ, p) be a C-Cons-Del-Unlim in-
stance. Then I is a yes-instance.
If we consider a voting rule R that only returns unique winners, then
C-Des-Del-Unlim also becomes tractable since we only need to delete
all voters.
For the constructive adding voters case, we obtain NP-hardness. The
idea for the reduction derives from the W[1]-hardness proof of C-Cons-
Add shown by Bulteau et al. [8].
Lemma 3. C-Cons-Add-Unlim is NP-hard.
Lemma 3 immediately implies the following inapproximability result
for the optimization variant of C-Cons-Add (denoted as Min-C-Cons-
Add), aiming at minimizing the solution size.
Theorem 6. There is no polynomial-time approximation algorithm for
Min-C-Cons-Add, unless P = NP.
7 Conclusion
We extend the study of combinatorial voter control problems introduced
by Bulteau et al. [8] and obtain that the destructive control variants tend
to be computationally easier than their constructive cousins.
13
Our research leads to several open questions and further research op-
portunities. First, we have shown hardness results for the adding candidate
case: if the bundling function consists of disjoint cliques, then parameter-
ized by the solution size, C-Cons-Add is W[1]-hard and C-Des-Add is
W[2]-hard. If one could also determine the complexity upper bound, that
is, under the given restrictions, if C-Cons-Add would be contained in
W[1], then this would yield another difference in complexity between the
destructive and the constructive variants. This also leads to the question
whether the problem variants in their general setting are not only W[2]-
hard, but W[2]-complete.
Second, we have only shown that Min-C-Cons-Add is inapproximable
and Min-C-Des-Del is trivially polynomial-time solvable. For the other
two problem variants, we do not know whether they can be approximated
efficiently or not.
Another open question is whether there are FPT-results for any natural
combined parameters. As a starting point, we conjecture that all problem
variants can be formulated as a monadic second-order logic formula with
length of at most f (k, b, m) (where k is the solution size, b is the maxi-
mum bundle size, m is the number of candidates, and f is a computable
function). Courcelle and Engelfriet [11] showed that every graph problem
expressible as a monadic second-order logic formula ρ can be solved in
g(ρ, ω) · I time, where ω is the treewidth of the input graph and I is
the input size. Our conjecture would provide us with a fixed-parameter
tractability result with respect to the solution size, the maximum bun-
dle size, the number of candidates, and the treewidth of our bundling
graph Gκ.
We have studied the Plurality rule exclusively. Thus it is still open
which of our results also hold for other voting rules, especially for the
Condorcet rule. Since with two candidates, the Condorcet rule is equiva-
lent to the strict majority rule, we can easily adapt some of our results
to work for the Condorcet rule as well. Other results (i.e., the Turing
reductions) cannot be easily adapted to work for the Condorcet rule.
References
1. R. P. Anstee. An algorithmic proof of Tutte's f -factor theorem. Journal of Algo-
rithms, 6(1):112–131, 1985. 11
2. R. P. Anstee. Minimum vertex weighted deficiency of (g, f )-factors: A greedy
algorithm. Discrete Applied Mathematics, 44(1–3):247–260, 1993. 11, 33
3. J. J. Bartholdi III, C. A. Tovey, and M. A. Trick. How hard is it to control an
election? Mathematical and Computer Modelling, 16(8-9):27–40, 1992. 1, 2, 13
14
4. N. Betzler and J. Uhlmann. Parameterized complexity of candidate control in
elections and related digraph problems. Theoretical Computer Science, 410(52):
43–53, 2009. 1
5. N. Betzler, R. Bredereck, J. Chen, and R. Niedermeier. Studies in computational
aspects of voting. In The Multivariate Algorithmic Revolution and Beyond, pages
318–363. Springer, 2012. 5
6. R. Bredereck, J. Chen, P. Faliszewski, J. Guo, R. Niedermeier, and G. J. Woeg-
inger. Parameterized algorithmics for computational social choice: Nine research
challenges. Tsinghua Science and Technology, 19(4):358–373, 2014. 5
7. R. Bredereck, P. Faliszewski, R. Niedermeier, and N. Talmon. Large-scale election
campaigns: Combinatorial shift bribery. Journal of Artificial Intelligence Research,
55:603–652, 2016. 3
8. L. Bulteau, J. Chen, P. Faliszewski, R. Niedermeier, and N. Talmon. Combinatorial
voter control in elections. Theoretical Computer Science, 589:99–120, 2015. 2, 3,
6, 7, 8, 11, 13, 17, 19, 21, 26, 27, 36, 37
9. J. Chen. Exploiting Structure in Computationally Hard Voting Problems. PhD
thesis, Technische Universität Berlin, 2015. 3, 6, 7, 8
10. J. Chen, P. Faliszewski, R. Niedermeier, and N. Talmon. Elections with few voters:
Candidate control can be easy. In AAAI '15, pages 2045–2051, 2015. 3
11. B. Courcelle and J. Engelfriet. Graph structure and monadic second-order logic: A
language-theoretic approach, volume 138. Cambridge University Press, 2012. 14
12. M. J. A. N. C. de Condorcet. Essai sur l'application de l'analyse a la probilite des
decisions rendues a la pluralite dex voix, 1785. 17
13. R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity.
Springer, 2013. 4
14. G. Erdélyi, M. R. Fellows, J. Rothe, and L. Schend. Control complexity in Buck-
lin and fallback voting: A theoretical analysis. Journal of Computer and System
Sciences, 81(4):632–660, 2015. 1
15. P. Faliszewski and J. Rothe. Control and bribery in voting.
In F. Brandt,
V. Conitzer, U. Endriss, J. Lang, and A. D. Procaccia, editors, Handbook of Com-
putational Social Choice, chapter 7. Cambridge University Press, 2016. 3
16. P. Faliszewski, E. Hemaspaandra, L. Hemaspaandra, and J. Rothe. Llull and
Copeland voting computationally resist bribery and constructive control. Journal
of Artificial Intelligence Research, 35:275–341, 2009. 1
17. P. Faliszewski, E. Hemaspaandra, and L. Hemaspaandra. Multimode control at-
tacks on elections. Journal of Artificial Intelligence Research, 40:305–351, 2011.
1
18. P. Faliszewski, E. Hemaspaandra, and L. A. Hemaspaandra. Weighted electoral
control. Journal of Artificial Intelligence Research, 52:507–542, 2015. 1
19. J. Flum and M. Grohe. Parameterized Complexity Theory. Springer, 2006. 4
20. E. Hemaspaandra, L. A. Hemaspaandra, and J. Rothe. Anyone but him: The
complexity of precluding an alternative. Artificial Intelligence, 171(5):255–285,
2007. 2, 13
21. L. A. Hemaspaandra, R. Lavaee, and C. Menton. Schulze and ranked-pairs vot-
ing are fixed-parameter tractable to bribe, manipulate, and control. Annals of
Mathematics and Artificial Intelligence, 77(3-4):191–223, 2016. 1
22. H. W. Lenstra Jr. Integer programming with a fixed number of variables. Mathe-
matics of operations research, 8(4):538–548, 1983. 21, 23
23. H. Liu and D. Zhu. Parameterized complexity of control problems in Maximin
election. Information Processing Letters, 110(10):383–388, 2010. 1
15
24. H. Liu, H. Feng, D. Zhu, and J. Luan. Parameterized computational complexity
of control problems in voting systems. Theoretical Computer Science, 410(27–29):
2746–2753, 2009. 1
25. R. Niedermeier.
Invitation to Fixed-Parameter Algorithms. Oxford University
Press, 2006. 4
26. A. D. Procaccia, J. S. Rosenschein, and A. Zohar. Multi-winner elections: Com-
plexity of manipulation, control and winner-determination. In IJCAI '07, pages
1476–1481, 2007. 2
27. J. Rothe and L. Schend. Challenges to complexity shields that are supposed to pro-
tect elections against manipulation and control: A survey. Annals of Mathematics
and Artificial Intelligence, 68(1–3):161–193, 2013. 3
16
Appendix
A Similarities in Complexity between the Problem
Variants
In this section we provide the theorems and proofs for results in which
the four problem variants behave similarly in complexity which are sum-
marized in the main text as Theorem 1.
First, we provide hardness results with different constraints on the pa-
rameters of the problem variants. Then, we show that the problem variants
are fixed-parameter tractable with respect to the number of candidates.
For this appendix, we introduce the Condorcet voting rule.
A candidate c is a Condorcet winner if it wins against every other
candidate in a head-to-head contest [12]. Formally, c is a Condorcet winner
if ∀c′ ∈ C \{c} :
{ v ∈ V c ≻v c′ } > { v ∈ V c′ ≻v c }. Condorcet's
voting rule returns a set consisting of the unique Condorcet winner if it
exists. Otherwise, it returns the empty set.
Note that, for the Condorcet rule, the problem definitions stated in
Section 3 need to be modified as the preferred winner (loser) p needs to
win (lose) the election evaluated by the Condorcet rule.
A.1 W[2]-Hard for the Solution Size
Bulteau et al. [8] originally stated that C-Cons-Add is W[2]-hard with
respect to the solution size for the Plurality and for the Condorcet voting
rule. Theorem A.1 uses their proof concept and extends their result for
the other three variants of the combinatorial voter control problem.
Theorem A.1. For both Plurality and Condorcet, C-Cons-Del, C-Des-
Add and C-Des-Del are all W[2]-hard with respect to the solution size
k, even if there are only two candidates and even if the bundling function
κ is symmetric.
Proof. We first consider the Plurality rule and provide a parameterized
reduction from the W[2]-complete problem Dominating Set parameter-
ized by the solution size h.
Dominating Set
Input: An undirected graph G = (V (G), E(G)) and a natural
number h ∈ N.
Question: Is there a dominating set of size at most h, that is, a
vertex subset U ⊆ V (G) with U ≤ h such that each vertex from
V (G) \ U is adjacent to at least one vertex from U ?
17
Let (G, h) be a Dominating Set instance. We construct an election
E = (C, V ) with C = {p, g}, where p is our preferred candidate. We define
the voter set for our three problem variants differently.
For C-Des-Add:
- The registered voter set V consists of V (G) − 1 p-voters (and no
g-voters).
- The unregistered voter set W consists of one g-voter wi for each
vertex ui ∈ V (G).
For C-Cons-Del, we define the voter set V such that
- V consists of one g-voter wi for each vertex ui ∈ V (G) and
- no p-voters.1
For C-Des-Del, we define the voter set V such that
- V consist of one p-voter wi for each vertex ui ∈ V (G) and
- one g-voter w′ with bundle κ(w′) = {w′}.
We define the bundle κ(wi) as the closed neighborhood of ui, formally
κ(wi) = wi ∪ {wj {ui, uj} ∈ E(G)}. Finalizing our construction, we set
k := h.
It is clear that our construction is both a polynomial reduction and
a parameterized reduction with respect to k. Also, it is obvious that the
bundling function used in the construction is symmetric.
We exemplary show for C-Cons-Del that there is a dominating set
of size h if and only if there is a subset V ′ of size at most k. The other
two variants can be proven analogously.
For the "only if" part, given a dominating set U of size at most h, we
define V ′ to be the corresponding voter set, that is, V ′ := {wi ui ∈ U }.
It is clear that V ′ ≤ h = k and p as well as g becomes winner, because
g loses V (G) points and has the same score as p.
For the "if" part, given a subset of the voters V ′ ⊆ V of size at most k
such that p is a winner of E′ = (C, V \ κ(V ′)), we define U to be the set of
vertices corresponding to the voters from V ′, that is, U := {ui wi ∈ V ′}.
It follows that U ≤ k = h and for each vertex ui ∈ V (G) \ U there must
be a vertex uj ∈ U which is a neighbor of ui, since otherwise we will still
have some g-voters and p will not become a winner.
Finally, we need to consider the Condorcet rule. For two candidates,
the Condorcet rule is equivalent to the strict majority rule and, hence,
the proof is analogous to the proof for the Plurality rule. We only need
1 For simplicity, we assume that every candidate is a winner in an election without
voters. However, if one requires a non-empty voter set, then one can easily adjust the
construction by adding one p-voter wp and one g-voter wg with κ(wp) = {wp} and
κ(wd) = {wd}. This makes the argumentation in the proof slightly more complicated.
18
to adapt the C-Des-Add such that the candidate p is the only winner.
We can accomplish that by adding to the V one p-voter w′ with bundle
κ(w′) = {w′}.
⊓⊔
A.2 W[1]-Hard for the Solution Size
From Bulteau et al. [8] we know that C-Cons-Add is W[1]-hard with
respect to the solution size for the Plurality voting rule, even when the
maximum bundle size is three. We show this for three other variants of the
combinatorial voter control and additional for Condorcet's voting rule.
Theorem A.2. For both Plurality and Condorcet, C-Cons-Add, C-Cons-
Del, C-Des-Add and C-Des-Del are all W[1]-hard with respect to the
solution size k, even if the maximum bundle size b is three and there are
only two candidates.
Proof. We first consider the Plurality rule and provide a parameterized
reduction from the W[1]-complete problem Clique parameterized by the
solution size h.
Clique
Input: An undirected graph G = (V (G), E(G)) and a natural
number h ∈ N.
Question: Is there a clique of size at least h, that is, a complete
subgraph with h vertices?
Let (G, h) be a Clique instance. Without loss of generality, we assume
h > 3. (If not, it can be solved in polynomial time.) We construct an
election E = (C, V ) with C = {p, g}, where p is our preferred candidate.
For each vertex u ∈ V (G) we define one vertex voter wu with the bundle
κ(wu) = {wu}. For each edge e = {u, u′} ∈ E(G) we define one edge voter
we with the bundle κ(we) = {we, wu, wu′}. Now, we finalize the voter set
definition which slightly differs for our three problem variants.
For C-Cons-Del, we define the voter set V such that:
- Vertex voters are p-voters.
- Edge voters are g-voters.
- We add a set Dp of dummy p-voters and a set Dg of dummy g-voters
with κ(w) = {w} for each w ∈ Dp ∪ Dg. We set set the cardinalities
of Dp and Dg such that sg(V ) − sp(V ) = (cid:0)h
2(cid:1) − h.
For C-Des-Del, we define the voter set V such that
- Vertex voters are g-voters.
- Edge voters are p-voters.
19
- We add a set Dp of dummy p-voters and a set Dg of dummy g-voters
with κ(w) = {w} for each w ∈ Dp ∪ Dg. We set set the cardinalities
of Dp and Dg such that sp(V ) − sg(V ) = (cid:0)h
2(cid:1) − h − 1.
For C-Des-Add:
- Vertex voters are unregistered g-voters.
- Edge voters are unregistered p-voters.
- There are no other voters in the unregistered voter set W .
- The registered voter set V consist of (cid:0)h
2(cid:1) − h − 1 p-voters.
Finally, we set k := (cid:0)h
2(cid:1).
It is clear that our construction is both a polynomial reduction and a
parameterized reduction with respect to k.
We exemplary show for C-Cons-Del that there is a clique U of size
at least h if and only if there is a subset V ′ of size at most k such that p
becomes a winner of E. The other two variants can be proven analogously.
For the "only if" part, given a clique U of size h, we construct V ′
by adding to it any edge voter we with e ∈ E(G[U ]). It is clear that
V ′ ≤ (cid:0)h
2(cid:1) = k. Observe that candidate p as well as candidate g become
winners. Candidate g loses (cid:0)h
2(cid:1) points and candidate p loses h points. Thus,
g and p have the same score.
For the "if" part, given a subset of the voters V ′ ⊆ V of size at most k
such that p is a winner of E′ = (C, V \ κ(V ′)), we define U to be the set
of vertices corresponding to the voters from V ′, that is, U := {u ∈ V (G)
we ∈ V ′ and u ∈ e}. We observe that deletion of vertex voters doesn't
reduce the score of g and removing a vertex voter from V ′ would lead to
a smaller solution. Hence, we can assume that V ′ does not contain any
vertex voters. In order to reduce the score of g, enough edge voters must
be removed, but a certain amount of vertex voters will be removed as well
since they are in the bundles of the edge voters. We denote the number of
indirectly removed vertex voters be x. Clearly x ≤ h, because otherwise
p loses more than h points, g loses at most (cid:0)h
2(cid:1) points, and g remains
the only winner. Assume towards a contradiction that x ≤ h − 2. The
score of g decreases by at least (cid:0)h
2(cid:1) − h in E′ (compared to E) so that
2 (cid:1) < (cid:0)h
V ′ contains at least (cid:0)h
2(cid:1) − h
for any h > 3. Hence, x ≥ h − 2 implying that there are at least (cid:0)h
2(cid:1) − 2
edge voters. Now, assume towards a contradiction that x = h − 1. Then,
2(cid:1) ≤ (cid:0)h−1
(cid:0)x
2(cid:1) − 2 for any h > 3. Thus, x = h implying that there are
exactly k = (cid:0)h
2(cid:1) edge voters in V ′ with altogether h different vertex voters
in their bundles. In this case, U is a clique of size h, since otherwise we
cannot have (cid:0)h
2(cid:1) edges incident to h vertices.
2(cid:1) − h edge voters. However, (cid:0)x
2(cid:1) ≤ (cid:0)h−2
2 (cid:1) < (cid:0)h
20
Finally, we need to consider the Condorcet rule. For two candidates,
the Condorcet rule is equivalent to the strict majority rule and, hence,
the proof is analogous to the proof for the Plurality rule. We only need to
adapt the C-Des-Add such that candidate p is the only winner. We can
accomplish that by adding to the V one p-voter d with bundle κ(d) = {d}.
⊓⊔
Condorcet-C-Cons-Add can be proved analogously.
A.3 Fixed-Parameter Tractability for the Number of
Candidates
Bulteau et al. [8] provide an integer linear program (ILP) that solves
Plurality- and Condorcet-C-Cons-Add for the case when the bundling
function is anonymous (see Section 2 for the corresponding definition)
and exploit Lenstra's theorem to show fixed-parameter tractability with
respect to the number of candidates. Their idea is to utilize the fact that
for anonymous bundling functions, voters with the same preference order
"lead" the same bundle of voters and "follow" the same voter as well. Thus,
with m candidates, we will have at most m! different bundles. By this ob-
servation, they introduce O(m!) variables, one variable for a bundle, to
encode whether to select a bundle to the solution. Indeed, as long as the
bundling function is anonymous, the same idea applies to the remaining
three combinatorial voter control variants. Although the technique is anal-
ogous, we provide the corresponding ILPs for Theorems A.3 to A.5 and
show the correctness for the sake of completeness.
Theorem A.3. For both Plurality and Condorcet, C-Cons-Del is fixed-
parameter tractable with respect to the number m of candidates, if the
bundling function κ is anonymous.
Proof. Given a C-Cons-Del instance ((C, V ), κ, p, k) with m candidates
and anonymous bundling function κ, we construct an integer linear pro-
gram (ILP) with at most O(m!) variables and at most O(V + m) con-
straints for C-Cons-Del similar to Bulteau et al. [8] for C-Cons-Add.
Fixed-parameter tractability follows because every ILP with ρ variables
and L input bits is solvable in O(ρ2.5ρ+o(ρ)L) time [22].
Since m candidates are given, there are at most m! voters with pair-
wise different preference orders. We denote these as ≻1, ≻2, . . . , ≻m! and
note that there are at most m! different bundles, because κ is leader-
anonymous.
We will use the following notation for the construction of the ILP:
21
1. Define κ(≻i) as the set of preference orders of the voters included in
the bundle of the voters with preference order ≻i; note that by the
anonymity, if κ(≻i) contains a preference order ≻, then every voter
with preference order ≻ is in the bundle of the voter that has preference
order ≻i.
2. Define κ−1(≻j) := {≻i≻j∈ κ(≻i)} as the set of preference orders that
include ≻j in their bundles.
3. Define Ni as the number of voters with preference order ≻i in V .
4. For each candidate a ∈ C, define F (a) as the set of preference orders
in which a ∈ C is ranked first and
5. let s(a) be the initial score in election (C, V ).
To encode a solution W ′, for each preference order ≻i, i ∈ [m!], we intro-
duce two Boolean variables, xi and yi. The intended meaning of xi = 1 is
that the sought solution contains a voter with preference order ≻i. The
intended meaning of yi = 1 is that κ(V ′) contains a voter with preference
order ≻i.
Now, we are ready to state the integer linear problem. Note that it
suffices to find a feasible solution. Thus, we do not need to specify any
objective function.
xi ≤ k,
X
i∈[m!]
xi ≤ Ni,
xi ≤ m! · yj,
xi ≥ yj,
X
≻i∈κ−1(≻j )
X
≻i∈κ−1(≻j )
∀i ∈ [m!]
∀j ∈ [m!]
∀j ∈ [m!]
s(p) − X
≻j ∈F (p)
Nj · yj ≥ s(a) − X
≻j∈F (a)
Nj · yj,
∀a ∈ C \ {p}
xi, yi ∈ {0, 1},
∀i ∈ [m!]
(1)
(2)
(3)
(4)
(5)
(6)
Constraint (1) ensures that at most k voters are added to the solution.
Constraint (2) ensures that the voters added to the solution are indeed
present in V . Constraints (3) and (4) ensure that variables yj, 1 ≤ j ≤ m!,
have correct values. Indeed, if for some preference order ≻i we have xi = 1
and ≻j∈ κ(≻i), then constraint (3) ensures that yj = 1. On the other
hand, if for some preference order ≻j we have that for each preference
order ≻i with ≻j∈ κ(≻i) it holds that xi = 0, then constraint (4) ensures
that yj = 0. Constraint (5) ensures that p has a (plurality) score which is
22
at least as high as the score of every candidate (which makes p a winner).
Clearly, there is a solution for this integer linear program if and only if
there is a solution for C-Cons-Del with the Plurality rule.
For the case of the Condorcet rule, we need to define the following
additional parameters: Let s(a, b) := {v ∈ V a ≻v b} denote the
number of voters that prefer candidate a over candidate b and P (a, b)
denote the set of preference orders in which a is preferred to b. We modify
only constraint (5) as follows:
s(p, a) − X
Nj · yj ≥ s(a, p) − X
≻j∈P (p,a)
≻j∈P (a,p)
Nj · yj + 1
∀a ∈ C \ {p}
(7)
This ensures that p can beat every other given candidate a in a head-to-
head contest if and only if there is a solution to the ILP.
As for the running time, it is clear that both ILPs have O(m) variables
and O(V + m) constraints. By the famous result of Lenstra Jr [22], fixed-
⊓⊔
parameter tractability follows.
Theorem A.4. For both Plurality and Condorcet, C-Des-Add is fixed-
parameter tractable with respect to the number m of candidates, if the
bundling function κ is anonymous.
Proof. As in the proof for Theorem A.3, we construct an integer linear
program (ILP) with at most O(m!) variables and constraints, and use the
same notation, except the following two:
– Ni denotes the number of voters with preference order ≻i in W .
– The intended meaning of yi = 1 is that κ(W ′) contains a voter with
preference order ≻i.
The constraints for the ILP are as follows:
(8)
∀i ∈ [m!]
(9)
∀j ∈ [m!] (10)
∀j ∈ [m!] (11)
Nj · yj(cid:17),
∀a ∈ C \ {p} (12)
αa(cid:16)s(p) + X
Nj · yj + 1(cid:17) ≤ αa(cid:16)s(a) + X
≻j ∈F (p)
≻j∈F (a)
xi ≤ k,
X
i∈[m!]
xi ≤ Ni,
xi ≤ m! · yi,
xi ≥ yj,
X
≻i∈κ−1(≻j )
X
≻i∈κ−1(≻j )
23
X
a∈C\{p}
αa ≥ 1,
xi, yi, αa ∈ {0, 1},
(13)
∀i ∈ [m!] (14)
If αa = 0, constraint (12) is valid. If αa = 1, constraint (12) ensures that
p loses against candidate a. Constraint (13) ensures that at least one of
the Boolean variables αa has value 1 and, therefore, there exists at least
one candidate a ∈ C such that a wins against p.
For the Condorcet rule, alter constraint (12) as follows:
αa(cid:16)s(p, a) + X
Nj · yj + 1(cid:17) ≤ αa(cid:16)s(a, p) + X
Nj · yj(cid:17), ∀a ∈ C \ {p}
≻j∈O(p,a)
≻j∈O(a,p)
(15)
Here s(a, b) = {w ∈ W a ≻w b} is the number of unregistered voters
that prefer candidate a over candidate b.
Constraint (13) ensures that at least one of the Boolean variables αa
has value 1 while constraint (15) ensures that p loses against at least one
of the candidates a in a head-to-head contest.
We omit the reasoning for the running time as it is the same as the
⊓⊔
one shown for Theorem A.3.
Theorem A.5. For both Plurality and Condorcet, C-Des-Del is fixed-
parameter-tractable with respect to the number m of candidates, if the
bundling function κ is anonymous.
Proof. This ILP is almost the same as the one for C-Des-Add. The only
difference is that we use "−" instead of "+" in constraint (12):
αa(cid:16)s(p) − X
Nj · yj + 1(cid:17) ≤ αa(cid:16)s(a) − X
Nj · yj(cid:17), ∀a ∈ C \ {p} (16)
≻j∈F (p)
≻j∈F (a)
Note that, in this problem variant, we do not try to find a subset of
unregistered voters W ′ ⊆ W to add to the election, but a subset of the
registered voters V ′ ⊆ V to remove from the election. The definitions for
Ni and yi change accordingly.
For the Condorcet rule, alter constraint (16) as follows:
αa(cid:16)s(p, a) − X
Nj · yj + 1(cid:17) ≤ αa(cid:16)s(a, p) − X
Nj · yj(cid:17), ∀a ∈ C \ {p}
≻j∈O(p,a)
≻j∈O(a,p)
(17)
24
Here s(a, b) = {w ∈ W a ≻w b} is the number of unregistered voters
that prefer candidate a over candidate b. We omit the reasoning for the
⊓⊔
running time as it is the same as the one shown for Theorem A.3.
B Missing Proofs
B.1 Proof of Proposition 1
Proposition B.1. For each X ∈ {Add, Del}, C-Des-X with m candi-
dates is Turing reducible to C-Cons-X with two candidates.
Proof. First, we provide a Turing reduction from C-Des-Add to C-Cons-
Add. Let I = (E = (C, V ), W, κ, p, k) be a C-Des-Add instance with
candidate set C = {p, g1, . . . , gm−1}. We compose m − 1 instances of the
problem C-Cons-Add Ji = (Ei = (Ci, Vi), W ′, κi, gi, k), where
– Ci := {p, gi},
– Vi := {v ∈ V v's more preferred candidate is either p or gi} ∪ {vd},
where vd is an additional p-voter,
– Wi := {w ∈ W w's more preferred candidate is either p or gi}, and
– κi : Wi → 2Wi with κi(w) := κ(w) ∩ Wi.
We show that I has a solution of size at most k if and only if at least one
of the instances Ji has a solution of size at most k.
For the "only if" part, let W ′
only two candidates p and gj, this implies sp(Vi∪κi(W ′
in Ji. Since we added an additional p-voter to Ji, sp(V ∪ κ(W ′
κ(W ′
for I.
i ⊆ Wi be a solution for Ji. Since Ji has
i ))
i )) < sgi(V ∪
i ⊆ W is a solution
i )) in I. Thus p loses election (C, V ∪κ(W ′
i )) ≤ sgi(Vi∪κi(W ′
i )) and W ′
For the "if" part, let W ′ ⊆ W be a solution for I, meaning that p loses
election (V ∪ κ(W ′)). Thus, there exists a gi ∈ C with sgi(V ∪ κ(W ′)) >
i := W ′∩Wi. Since gi and p are the only candidates in
sp(V ∪κ(W ′)). Let W ′
Ci, and since Vi has one additional p-voter compared to V , it follows that
sgi(Vi ∪κi(W ′
i )),
and W ′
i )). Thus, gi wins election (Ci, Vi ∪κi(W ′
i )) ≥ sp(Vi ∪κi(W ′
i is a solution for Ji.
For the Turing reduction from C-Des-Del to C-Cons-Del, the con-
struction of the reduction is similar to the one from C-Des-Add to
C-Cons-Add. Set the bundling functions κi : Vi → 2Vi with κi(v) :=
κ(v) ∩ Vi and κi(vd) = {vd}.
⊓⊔
Proposition B.2. For each Y ∈ {Cons, Des}, C-Y -Del with two can-
didates is many-one reducible to C-Y -Add with two candidates.
25
Proof. First, we provide a polynomial-time reduction from C-Cons-Del
to C-Cons-Add. Let I = (E = (C, V ), κ, p, k) be a C-Cons-Del in-
stance with C = {p, g}. We define V to be the complement voter set,
that is, it contains the same voters as V but all p-voters become g-voters
and all g-voters become p-voters. We construct a C-Cons-Add instance
J = (E = (C, V ), W, κ, p, k), where W := V . Clearly, the construction
of J can be implemented in polynomial time. We now claim that V ′ is a
size-k solution for I if and only if W ′ := V ′ is a size-k solution for J.
First, we observe that for V ′ to be a solution for I, it must hold that
sp(V ) − sp(κ(V ′)) ≥ sg(V ) − sg(κ(V ′)). Similarly for W ′ to be a solution
for I, it must hold that sp(V ) + sp(κ(W ′)) ≥ sg(V ) + sg(κ(W ′)). As per
definition of the complement voter set and because of V ′ = W ′, we know
that
sp(κ(V ′)) = sg(κ(W ′)) and sg(κ(V ′)) = sp(κ(W ′)).
Thus, it holds that
sp(V ) − sp(κ(V ′)) ≥ sg(V ) − sg(κ(V ′))
if and only if sp(V ) + sg(κ(V ′)) ≥ sg(V ) + sp(κ(V ′))
if and only if sp(V ) + sp(κ(W ′)) ≥ sg(V ) + sg(κ(W ′)),
implying that V ′ is a solution for I if and only if W ′ is a solution for J.
The reduction from C-Des-Del to C-Des-Add works analogously.
⊓⊔
B.2 Proof of Observation 1
Observation 1. C-Cons-Add is NP-hard even for symmetric bundling
functions with maximum bundle size b = 3.
Proof. Bulteau et al. [8, Theorem 7] have shown NP-hardness of C-Cons-
Add for full-d bundling functions2 and maximum bundle size b = 3. Since
a full-d bundling function is symmetric [8, Observation 1]. Thus, NP-
hardness of C-Cons-Add for symmetric bundling functions and b ≤ 3
⊓⊔
follows.
B.3 Proof of Proposition 2
Proposition 2. C-Cons-Del is NP-hard even for symmetric bundling
functions with maximum bundle size b = 3.
2 Full-d bundling functions are defined by Bulteau et al. [8, Section 2].
26
Proof. Bulteau et al. [8, Theorem 7] have shown NP-hardness of C-Cons-
Add for full-d bundling functions3 and maximum bundle size b ≤ 3. Their
idea was to construct a cycle for each variable x: The cycle contains ver-
tices that correspond to the clauses containing either x or x and are con-
nected through some p-voters in such a way that one must take the vertices
corresponding to the clauses with the same literal, that is, either x or x.
We utilize this construction to show the hardness for the deleting case
by reducing from the following NP-complete 3SAT problem [8].
(2-2)-3SAT
Input: A collection F of size-two-or-three clauses over the variable
set X = {x1, . . . , xn}, such that each clause has either two or three
literals, and each variable appears exactly four times, twice as a
positive literal and twice as a negative literal.
Question: Is there a truth assignment that satisfies all the clauses
in F?
Let I = (F, X ) be a (2-2)-3SAT instance. Now, we construct an
instance for C-Cons-Del as follows: Let p be the candidate whose victory
we want to ensure and let d be the winner of the original election. For
each clause fj = (ℓ1
j ) ∈ F, we introduce a candidate cj. Thus,
the candidate set is C = {p, d} ∪ {cj fj ∈ F}.
The voter set is composed of three groups:
j ∨ ℓ2
j ∨ ℓ3
1. For each variable xi, for each clause fj that contains xi as a literal
(that is, either positive or negated variable), we introduce a d-voter,
denoted as vj
i (we call him a variable voter ), and a cj-voter, denoted as
uj
i (we call him a clause voter ). Now, let fj, fr, j < r, be the clauses
containing xi, and let fs, ft, s < t be the two clauses containing
xi. We construct the bundles of these the clause voters corresponding
to fj, fr, fs, ft and the variable voters such that the bundling graph
forms a cycle with the d-voters between each two clause voters. More
precisely, let
κ(uj
κ(us
κ(ur
κ(ut
i ) := {uj
i ) := {us
i ) := {ur
i) := {ut
i , vj
i , vs
i , vr
i, vt
i , vs
i , vr
i , vt
i, vj
i }, κ(vs
i }, κ(vr
κ(vt
i},
κ(vj
i },
i ) := {vs
i ) := {vr
i) := {vt
i ):= {vj
i , uj
i , us
i, ur
i , us
i , us
i }
i , ur
i }
i , ut
i}
i , uj
i }.
3 See footnote 2.
27
2. For each clause Cj that contains only two literals, we introduce a cj-
voter uj
c and set his bundle to be the singleton κ(uj
c) = {uj
c}.
d , v−
d with the
3. We introduce two p-voters w+
p , w−
p and two d-voters v+
following bundles:
κ(w+
κ(v+
p ) := {w1
d ) := {v+
p, v+
d },
d , w+
p },
κ(w−
κ(v−
p ):= {w−
d ):= {v−
p , v−
d , w−
d },
p }.
The bundling graph for the first group of voters is depicted in the left
figure and the bundling graph for the second group of voters is depicted
in the middle figure, and the bundling graph for the last group of voters
is depicted in the right figure.
ur
i
vr
i
vt
i
us
i
ut
i
vs
i
vj
i
uj
i
uj
c
w+
p
w−
p
v+
d
v−
d
This completes the construction of the voters and their bundles. It is
straight forward that the construction runs in polynomial time.
By simple calculation, we have that the score difference between can-
didate d and p is 4 · n and the score difference between candidate cj and
p is one.
Now, we show that the given (F, X ) is a yes instance for (2-2)-3SAT
if and only if the constructed instance (C, V, κ, p, k = 2n) is a yes instance
for C-Cons-Del.
For the "only if" part, let σ : X → {0, 1} be a satisfying assignment
for I. Consider some variable xi. Let fj, fr, j < r, be the clauses con-
taining xi and let fs, ft, s < t be the two clauses containing xi. Now, if
σ(xi) = 1, then we add to V ′ the two clause voters uj
i ; otherwise we
add to V ′ the two clause voters us
i. We do this for each variable xi.
Since each of the bundles of uj
i contains two p-voters and σ is a
satisfying assignment, one can verify that V ′ = 2 and, after deleting the
bundles corresponding to V ′, only two d-voters and for each clause Cj at
most two cj-voters remain. Since there are in total two p-voters, p will
co-win with d (and with some clause candidates).
i and ur
i and ur
i and ut
28
For the "if" part, let V ′ be set of at most 2 · n voters, such that p wins
election (C, V \ κ(V ′)). First of all, we observe that the score difference
between d and p is 4 · n and deleting voters will not increase the score of
any candidate. Thus, in order to make p win, d has to lose at least 4 · n
points. Since each bundle contains at most two d-voters, it follows that
V ′ = 2n and for each voter v ∈ V ′, it must be that κ(v) contains two d-
voters. This precludes including any voter from the last two groups to V ′.
Moreover, for each two voters u, v ∈ V ′ that are on the same cycle (of the
bundling graph), by the construction of the bundles, we must have that
κ(u) ∩ κ(v) = ∅ since otherwise we will not delete enough d-voters. This
implies that, for each variable x, V ′ contains the two clause voters that
correspond to the clauses with either x or x. Now, if it is the case for x,
then we set σ(x) = 1; otherwise σ(x) = 0. The assignment σ is obviously
valid. It remains to show that σ satisfies F. Now, observe that each clause
candidate cj has to lose at least one point because of the original score
difference. This implies that there must be some literal ℓi from Cj such
that uj
i ∈ V ′, by our assignment we also set ℓi to 1, satisfying Cj. Thus,
our constructed σ satisfies all clauses.
B.4 Proof of Claim 1
Let W ′ be a maximum (s, t)-proper-subset and (max κ(W ′)−
Claim 1.
min κ(W ′) + 1) > 9. Then, there is a j with s < j < t, such that there
is a (s, j)-proper-subset W1 and a (j + 1, t)-proper-subset W2 with W1 +
W2 ≤ W ′ and κ(W1 ∪ W2) = κ(W ′).
Proof. Let s′ := min κ(W ′) and t′ := max κ(W ′). We split this in two
cases.
First, we consider the case in which k ≤ 3. Note that t′ − s′ + 1 > 9
and s ≤ s′ ≤ t′ ≤ t. Therefore, W (s′, t′) > 9. Since the maximum
bundle size is 3, we know that κ(W ′) is at most 9. This implies that
there is a wj ∈ W (s′, t′) \ κ(W ′). We set W1 := { wi ∈ W ′ i ≤ j } and
W2 := { wi ∈ W ′ j < i }. Thus, W1+W2 = W ′, κ(W1∪W2) = κ(W ′)
and W1 is a (s, j)-proper-subset and W2 is a (j + 1, t)-proper-subset.
Now, let us consider the case in which k > 3. If the set W ′ can be
partitioned into two disjoint subsets W1 and W2 where W1 ∪ W2 = W ′,
W1 ∩ W2 = W ′, κ(W1) ∩ κ(W2) = ∅ and κ(W ′) = κ(W1 ∪ W2), we
are done. So assume there is no such partition. That means there is no
wj ∈ W (s′, t′) \ κ(W ′), otherwise we could split κ(W ′) at position i as
we did in case k ≤ 3. Therefore, W (s′, t′) = κ(W ′). Furthermore, we
can conclude that there are two bundles of size 2 (the endpoints of the
29
path) and that all other bundles are of size 3, because of the maximum
bundle size and Gκ being a path. At least every second voter on the path
is in W ′, otherwise we could split κ(W ′) at position i or i + 1, where
wi, wi+1 6∈ W ′. Therefore, W ′ \ {ws′, ws′+1, . . . , ws′+8} ≤ W ′ − 4. One
can observe that κ({ws′+1, ws′+4, ws′+7}) = {ws′, ws′+1, . . . , ws′+8}. For
the set W ′′ := (W ′\{ws′, ws′+1, . . . , ws′+8})∪{ws′+1, ws′+4, ws′+7} it holds
that W ′′ ≤ W ′ and κ(W ′) = κ(W ′′). Furthermore, one can observe that
ws′+4 is the only element which has ws′+5 in its bundle and ws′+7 is the
only element which has ws′+6 in its bundle. That means we can split
W ′′ at this point. We set j := s′ + 5, W1 := { wi ∈ W ′
i ≤ j } and
W2 := { wi ∈ W ′ j < i }. Thus, W1+W2 ≤ W ′, κ(W1∪W2) = κ(W ′)
⊓⊔
and W1 is a (s, j)-proper-subset and W2 is a (j, t)-proper-subset.
The idea for cycles is similar.
Lemma B.1. If C-Cons-Add has only two candidates p, q and symmet-
ric bundling function κ with bundles of size at most three such that the
bundling graph Gκ is a cycle, then finding a size-at-most-k subset W ′ ⊆ W
of voters such that the score difference between p and q in κ(W ′) is max-
imum can be solved in O(W 5) time, where W is the size of the set W
of unregistered voters.
Proof. Let I = (E = (C, V ), W, κ, p, k) be a C-Cons-Add instance, where
b = 3, κ is symmetric and C = {p, g}. Let w1 . . . wW w1 be the cycle in
Gκ. Suppose that W ≤ 9. We compute sp(V ) and sg(V ) in linear time.
Every W ′ ⊆ W of size at most k ≤ W ≤ 9 is a solution if κ(W ′) contains
sg(V ) − sp(V ) more p-voter than g-voter. Since W is upper bounded by
a constant, the size of the power set of W is also upper bounded by a
constant. Therefore, we can check in constant time whether one of the
subsets of W is a solution.
Now let us consider the interesting case. Suppose that W > 9. The
idea is to break the cycle into a path and solve it with Lemma 1. We create
nine C-Cons-Add instances I1, . . . , I9 where Ii := (E = (C, V ), W ∪
{wb, we}, κi, p, k), and wb and we are g-voters and
κi : W → 2W , w 7→
{wb, wi},
{we, wi+1},
(κ(w) \ {wi+1}) ∪ {we},
(κ(w) \ {wi}) ∪ {wb},
κ(w),
if w = wb
if w = we
if w = wi
if w = wi+1
otherwise,
where i ∈ [9]. The bundling function κi is modified in such a way that the
30
w1 w2 w3 w4 w5 wb
we w6 w7 w8 w9
. . .
Fig. 1: Part of the construction used in Lemma B.1. Specifically, it shows
how the cycle is reduced to a path in I5.
instance Ii becomes a path and has g-voter on its endpoints, depicted in
Figure 1. Therefore, we can solve all 9 instances Ii in O(W 5), where W
is the size of the unregistered voters (see Lemma 1).
It remains to be shown that I is a yes-instance if and only there it
exists an i ∈ [9] such that Ii is a yes-instance.
For the "only if" part, let Ii, for some i ∈ [9], be a yes-instance. Thus,
there is a set W ′ ⊆ W ∪ {wb, we} of size at most k such that p ∈ Plu-
rality(C, V ∪ κi(W ′)). One can observe that the bundles of wb and we
are useless, because they each have only two voters and at least one of
them is a g-voter. For this reason we assume without loss of generality
that W ′ ⊆ W . Since we replaced an edge between wi and wi+1 with
two g-voters wb and we that are not connected, we can conclude that
wb ∈ κi(W ′) ⇔ wi ∈ W ′ and we ∈ κi(W ′) ⇔ wi+1 ∈ W ′. We now check
what happens when wi or wi+1 is in W ′.
(i) Assume wi, wi+1 ∈ W ′. Then κi(W ′) \ {wb, we} = κ(W ′). The set
κ(W ′) has 2 g-voter less than κi(W ′). Thus, p ∈ Plurality(C, V ∪
κ(W ′)).
(ii) Assume wi ∈ W ′ and wi+1 6∈ W ′. Then κ(W ′) = (κi(W ′) \ {wb}) ∪
{wi+1}. In contrast to κi(W ′), κ(W ′) loses the g-voter wb and may
win the voter wi+1. It does not matter whether wi+1 is a p-voter
or g-voter, as sg(κ(W ′)) ≤ sg(κi(W ′)) and sp(κ(W ′)) ≥ sp(κi(W ′)).
Thus, p ∈ Plurality(C, V ∪ κ(W ′)).
(iii) Assume wi+1 ∈ W ′ and wi 6∈ W ′. This case is analogous to (ii).
In the case where neither wi nor wi+1 is in W ′, we immediately get
κi(W ′) = κ(W ′) from the definition of κi. Therefore, I is a yes-instance.
For the "if" part, let I be a yes-instance. Thus, there is a set W ′ ⊆ W
of size at most k such that p ∈ Plurality(C, V ∪ κ(W ′)). It is sufficient
to show that for some i, 1 ≤ i ≤ 9 it holds that wi, wi+1 6∈ W ′, because
these are the only cases in which κi differs from κ. Thus, κ(W ′) = κi(W ′),
which means that Ii is a yes-instance.
Now, assume that such an i does not exist. Then, on the path w1 . . . w9
at least every second voter is in W ′. Therefore, W ′ \ {w1, . . . , w9} ≤
W ′ − 4. One can observe that κ({w2, w5, w8}) = {w1, . . . , w9}. For the
31
set W ′′ := (W ′ \ {w1, . . . , w9}) ∪ {w2, w5, w8} it holds that W ′′ ≤ W ′
and κ(W ′) = κ(W ′′), where W ′′ ≤ W ′. Furthermore, one can observe
that w6, w7 6∈ W ′′. That means κ(W ′′) = κ5(W ′′) and hence p ∈ Plural-
ity(C, V ∪ κ5(W ′′)). This implies that I5 is a yes-instance.
⊓⊔
B.5 Proof of Corollary 2
Corollary 2. C-Des-Add and C-Des-Del with a symmetric bundling
function and maximum bundle size three can be solved in time O(m·W 5)
and O(m · V 5), respectively, where m is the number of candidates, and
W and V are the sizes of the unregistered and registered voter set,
respectively.
Proof. To obtain the result for C-Des-Add, use the Turing reduction
from C-Des-Add to C-Cons-Add (Proposition 1). Note that the in-
stances produced by the Turing reduction only have two candidates. Thus,
we can use Theorem 2.
Analogously, we can use Corollary 1 to obtain the result for C-Des-
Del.
As in both Turing reductions we create O(m) instances, we need to
solve O(m) instances, resulting in a running time of O(m · n5) in both
cases, where n is the number of either the unregistered (for C-Des-Add)
⊓⊔
or the registered (for C-Des-Del) voters .
B.6 Proof of Theorem 3
Theorem 3. C-Cons-Del with a symmetric bundling function and with
bundles of size at most two can be solved in polynomial time.
Proof. Let I = ((C, V ), κ, p, k) be a C-Cons-Del instance, where the
bundling function κ is symmetric and each bundle has at most two voters.
We first consider the case where each bundle has size exactly two. Since
deleting bundles never increases any candidate's score, we can assume
that we do not delete any bundle that contains a p-voter. Under this
assumption, the original score of p equals her score in the final election.
Furthermore, if we know the score of p in the final election, then we know
for each remaining candidate ci 6= p how many ci-voters we need to delete
to make her have no more score than p; let di be the score difference
between ci and p. Due to this observation, we can construct a multigraph
(with loops), which contains a vertex ui for each candidate ci except p,
and where for each bundle with two voters supporting ci and cj there is
an edge incident to the vertices vi and vj (note that if ci = cj, then the
32
edge is a loop). Now, deleting minimum bundles to make each candidate ci
lose at least di points is equivalent to finding a subgraph with minimum
number of edges where each vertex has degree di. Anstee [2] showed that
the latter problem can be solved in polynomial time.
To handle the case where some bundle has only one voter, we first
observe that if our preferred candidate p has zero points, then we need to
delete all bundles. Note that the underlying bundling graph consist of only
disjoint edges and isolated vertices. The minimum number of bundles to
be deleted equals the number of edges plus the number of isolated vertices.
Now, we consider the case where our preferred candidate p has at least
one point. For each bundle with only one voter vi, we introduce a dummy
candidate di and a dummy di-voter, and bundle the voter with vi together.
Since each of the dummy candidates has exactly one point, the original
instance (with p having at least one point) is a yes instance if and only
if the modified instance is a yes instance. In the modified instance, every
bundle has size exactly two. The first case applies.
B.7 Proof of Theorem 4
In order to prove Theorem 4, we first show W[1]-hardness of C-Cons-Add
(Lemma B.2) and then W[2]-hardness of C-Cons-Del (Lemma B.3).
Lemma B.2. C-Cons-Add is W[1]-hard, even for disjoint bundling func-
tion.
Proof. We provide a parameterized reduction from the W[1]-complete
problem Independent Set (parameterized by the "solution size"), de-
fined as follows:
Independent Set
Input: An undirected Graph G = (V (G), E(G)) and a natural
number h ∈ N.
Question: Does G admit a size-h independent set, that is, all ver-
tices in U are pairwise non-adjacent.?
Let (G, h) be an Independent Set instance with E(G) = {e1, . . . ,
em−1} and V (G) = {u1, . . . , un}. Without loss of generality, we assume
that G is connected and h ≥ 3. We construct an election E = (C, V )
with candidate set C := {p} ∪ { gj ej ∈ E(G) }. For each edge ej ∈ E,
we construct h − 1 registered voters that all prefer gj most. In total, V
consists of (h − 1)(m − 1) voters.
The unregistered voter set W is constructed as follows: For each vertex
ui ∈ V (G), add a p-voter pi, and for each edge ej incident with ui, add
33
a gj-voter a(i)
j . The voters constructed for each vertex ui are bundled by
κ. More formally, for each ui ∈ V (G) and each ej ∈ E(G) with ui ∈ ej, it
holds that
κ(pi) = κ(a(i)
j ) := {pi} ∪ {a(i)
j
ej ∈ E(G) ∧ ui ∈ ej}.
To finalize the construction, we set k := h.
The construction is both a polynomial-time and a parameterized re-
duction, and all bundles are disjoint. To show the correctness, we note
that p can only win if only if her score can be increased to at least h
without giving any other candidate more than one more point. The so-
lution corresponds to exactly to a subset of h vertices that are pairwise
non-adjacent.
Now, it remains to show that G has a size-h independent set if and
only if ((C, V ), W, κ, p, k) is a yes-instance for C-Cons-Add.
For the "only if" part, suppose that U ⊆ V (G) is a size-h independent
set for G. Define the subset W ′ as the voters pi ∈ W corresponding to the
vertices ui ∈ U . Obviously, W ′ = h = k. As per definition, U does not
contain adjacent vertices, each candidate gj may achieve a score increase of
at most one, while p achieves a score increase of k. As the initial difference
in scores between p and every candidate gj is k − 1, p co-wins the final
election with each candidate gj such that ui ∈ ej for some voter pi ∈ W ′.
For the "if" part, suppose that there is a subset W ′ of size at most k
such that p is a winner of the Plurality election (C, V ∪ κ(W ′)). First,
we claim that W ′ = k and κ(W ′) contains exactly k p-voters: Since the
original score difference between p and any other candidate gj is k − 1 and
since each unregistered voter's bundle contains exactly one p-voter and
at least one non-p-voter, it follows that in the final election at least one
non-p-candidate has score of at least k, but p can have a score increase of
at most k. This means
(1) that p's final score must increase to k, that is, κ(W ′) must have exactly
k p-voters, and
(2) that no other candidate can have a score increase of more than one.
Now, define U := {ui pi ∈ κ(W ′)}. Due Observation (1) we have
that U = k = h. To show that U is also an independent set, we con-
sider two arbitrary vertices ui, uℓ ∈ U . Suppose for the sake of contradic-
tion that {ui, uℓ} ∈ E(G), denote this edge by ej. By the construction
of our bundling function, it must hold that a(i)
j ∈ κ(W ′). Then, gj
would have a score increase of at least two-a contradiction to our obser-
⊓⊔
vation (2).
j , a(ℓ)
34
Lemma B.3. C-Cons-Del is W[2]-hard, even for disjoint bundling func-
tion.
Proof. We provide a parameterized reduction from the W[2]-complete
problem Dominating Set (see Appendix A.1 for the corresponding def-
inition).
Let (G = (V (G), E(G)), h) be a Dominating Set instance with
E(G) = {e1, . . . , em′} and V (G) = {u1, . . . , un′}. Let ∆max denote the
maximum degree of G. We construct an election E = (C, V ) with candi-
date set C = {p, g1, . . . , gn′}.
The voter set V consists of two groups:
1. For each vertex ui ∈ V (G), add a gi-voter, denote as vertex voter v(i)
,
and for each of its neighbors uj ∈ N (ui) add agj-voter, denote as
neighbor voter v(i)
. Note that, for each vertex ui, we have added ex-
j
actly N [ui] voters that all most prefer gi. The vertex voters and
neighbor voters constructed due to vertex ui are contained in one bun-
dle. Formally, the bundling function for these voters looks as follows.
i
∀ui ∈ V (G) : κ(v(i)
i ) := {v(i)
j
uj ∈ N [ui]}, and
∀uj ∈ N (ui) : κ(v(i)
j ) := κ(v(i)
i )
2. For each vertex ui ∈ V (G), add ∆max + 1 − N [ui] gi-voters. Further-
more, add ∆max p-voters to V . All these voters are bundled together.
Note that in this bundle, each candidate gi has exactly N [ui] points
less than p.
We use D to denote the set of all voters constructed in the second group.
Finalizing the construction, set k := h.
The construction is both a polynomial-time and a parameterized re-
duction, and the bundling function is disjoint. Note that for all candidates
gi have score sgi(V ) = ∆max + 1, and the difference in score between every
gi and the candidate p is sgi(V ) − sp(V ) = ∆max + 1 − ∆max = 1.
It remains to show that there is a dominating set of size at most h
if and only if there is a subset V ′ of voters of size at most k, such that
if their respective bundles are deleted from the election, p becomes a
Plurality winner of the election.
For the "only if" part, given a dominating set U of size at most h, we
define V ′ to be the corresponding voter set, that is, V ′ := {v(i)
ui ∈ U }.
i
It is easy to verify that V ′ ≤ h = k and p has a score of ∆max, while
all other candidates have a score of at most ∆max. Thus, p is a Plurality-
winner of the election (C, V \ κ(V ′)).
35
For the "if" part, suppose that there is a subset V ′ ⊆ V of size at most
k such that p is a winner of the Plurality election (C, V \κ(V ′)). First of all,
since all voters in D are bundled together such that p has more supporters
than any other candidate in this bundle, by the disjoint property of the
bundles, we know that p will also win the election (C, V \ κ(V ′ \ D).
Thus, we can assume that V ′ does not contain any voter from the second
group D This means that in the final election, p will have score ∆max,
implying that each gi has to lose at least one point. In other words,
κ(V ′) contains at least one gi-voter for each candidate gi.
(∗)
Now, we define the vertex subset U := {ui u(i)
i ∈ κ(V ′)} and show
that it is a dominating set of size at most h. Since all bundles are disjoint
and V ′ ≤ k, it is clear that U ≤ k = h. To show that U is a dominating
set, we consider an arbitrary vertex uj /∈ U . By (∗), we know that κ(V ′)
contains a gj-voter; note that he is from the first group. Let this voter be
v(i)
. Due to the construction of our bundling function, it follows that κ(V ′)
j
also contains v(i)
i
ui ∈ U .
, implying that U contains ui. Thus uj is dominated by
⊓⊔
B.8 Proof of Theorem 5
Theorem 5. C-Des-Add and C-Des-Del with a symmetric bundling
function and disjoint bundles can be solved in O(m · I) time, where I
is the input size and m the number of candidates.
Proof. We consider C-Des-Add first. Let I = (E = (C, V ), W, κ, p, k)
be a C-Des-Add instance with k being symmetric. To make p lose, it
is enough to add voters such that there is some other candidate c that
has a higher score than p in the final election. Due to this observation,
we "guess" the defeater c ∈ C \ {p} and we use an even simpler greedy
strategy than the one used for Observation 2 (see Theorem 6 by Bulteau
et al. [8]) to add voters to maximize the score difference between c and
p: Since the bundles are disjoint, we greedily add bundles that maximally
improve the score difference between c and p (in favor of c). If using up
k bundles, we can make c have at least the same score as p, then return
yes; otherwise we continue with the next defeater. It is easy to verify that
this approach is correct and the running time is O(m · I).
The strategy for C-Des-Del works analogously.
36
B.9 Proof of Lemma 2
Lemma 2. Let I = (E = (C, V ), κ, p) be a C-Cons-Del-Unlim in-
stance. Then I is a yes-instance.
Proof. Let I = (E = (C, V ), κ, p, k) be a C-Cons-Del-Unlim instance.
Set V ′ := V . Since there is no voter in the election (C, V \ κ(V ′)), all
candidates have the same score. Thus p ∈ Plurality(C, V \ V ′).
⊓⊔
B.10 Proof of Lemma 3
Lemma 3. C-Cons-Add-Unlim is NP-hard.
Proof. To prove NP-hardness, we extend the proof idea for W[1]-hardness
from Bulteau et al. [8]. Let (G = (V (G), E(G)), h) be a Clique instance.
(The definition of Clique can be found in Appendix A.2.) Without loss of
generality, we assume h ≥ 4. We construct an election E := (C, V ) where
C := {p, g, x}. V consists of (cid:0)h
2(cid:1) p-voters and (2 · (cid:0)h
2(cid:1) − h) g-voters. The
set of unregistered voters W is composed as follows:
- For each vertex u ∈ V (G), add a g-voter wu ∈ W with κ(wu) = {wu}.
We call wu a vertex-voter.
- For each edge e = {u, u′} ∈ E(G) we add a p-voter we ∈ W and two
e) =
e ∈ W such that κ(we) = {we, wu, wu′, w′
e ) = {w′′
e }. We call we an edge-voter.
x-voters w′
{w′
e, w′′
e} and κ(w′′
e, w′′
e }, κ(w′
Obviously, our construction is a polynomial reduction. It remains to show
that there is a clique of size at least h if and only if there is a subset W ′
such that p is a winner of the Plurality election (C, V ∪ κ(W ′)).
For the "if" part, suppose that there is a subset W ′ ⊆ W such that
p ∈ Plurality(C, V ∪ κ(W ′)). We show that the vertex set U := { u ∈
V (G) we ∈ W ′ ∧ u ∈ e } is a clique of size h in G. First, we observe that
a solution W ′ ⊆ W cannot have more than (cid:0)h
2(cid:1) edge-voters, otherwise x
achieves a higher score than p. Second, a solution W ′ must have (cid:0)h
2(cid:1) edge-
voters which have only h different vertex-voters in its bundles (a clique in
G). For a precise argumentation see the proof of Theorem A.2.
For the "only if" part, suppose that U ⊆ V (G) is a size-h clique for
G. We construct the subset W ′ by adding any edge voter we with e ∈
E(G[U ′]). Now it is easy to check that p ∈ Plurality(C, V ∪ κ(W ′)).
⊓⊔
(Compare with the proof of Theorem A.2.)
B.11 Proof of Theorem 6
Theorem 6. There is no polynomial-time approximation algorithm for
Min-C-Cons-Add, unless P = NP.
37
Proof. Assume towards a contradiction that there is an approximation
algorithm Aα for Min-C-Cons-Add which runs in polynomial time and
provides a solution Wα such that α · OP T ≥ Wα, where OP T is a
solution of minimum size. Let (E, W, κ, p) be an C-Cons-Add-Unlim
instance. One can create a Min-C-Cons-Add instance (E, W, κ, p) and
compute a solution Wα of size α · OP T in polynomial time. Wα is a
solution for the instance (E, W, κ, p) of the NP-hard problem C-Cons-
⊓⊔
Add-Unlim. This is a contradiction unless P = NP.
38
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.